Spaces:
Configuration error
Configuration error
add demo
Browse files- .gitattributes +2 -0
- .gitmodules +3 -0
- LICENSE +21 -0
- NOTICE +263 -0
- README.md +8 -6
- app.py +376 -0
- assets/NOTICE +8 -0
- assets/pexels-heyho-5998120_19mm.jpg +3 -0
- assets/pexels-itsterrymag-12639296_24mm.jpg +3 -0
- extern/ZoeDepth +1 -0
- extern/splatting-0.0.1-py3-none-any.whl +3 -0
- genwarp/GenWarp.py +546 -0
- genwarp/__init__.py +1 -0
- genwarp/models/__init__.py +4 -0
- genwarp/models/attention.py +499 -0
- genwarp/models/motion_module.py +399 -0
- genwarp/models/mutual_self_attention.py +420 -0
- genwarp/models/pose_guider.py +63 -0
- genwarp/models/resnet.py +265 -0
- genwarp/models/transformer_2d.py +409 -0
- genwarp/models/transformer_3d.py +179 -0
- genwarp/models/unet_2d_blocks.py +1087 -0
- genwarp/models/unet_2d_condition.py +1324 -0
- genwarp/models/unet_3d.py +645 -0
- genwarp/models/unet_3d_blocks.py +885 -0
- genwarp/ops.py +130 -0
- requirements.txt +14 -0
.gitattributes
CHANGED
|
@@ -33,3 +33,5 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
|
|
| 33 |
*.zip filter=lfs diff=lfs merge=lfs -text
|
| 34 |
*.zst filter=lfs diff=lfs merge=lfs -text
|
| 35 |
*tfevents* filter=lfs diff=lfs merge=lfs -text
|
|
|
|
|
|
|
|
|
| 33 |
*.zip filter=lfs diff=lfs merge=lfs -text
|
| 34 |
*.zst filter=lfs diff=lfs merge=lfs -text
|
| 35 |
*tfevents* filter=lfs diff=lfs merge=lfs -text
|
| 36 |
+
assets/*.jpg filter=lfs diff=lfs merge=lfs -text
|
| 37 |
+
extern/splatting-0.0.1-py3-none-any.whl filter=lfs diff=lfs merge=lfs -text
|
.gitmodules
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[submodule "extern/ZoeDepth"]
|
| 2 |
+
path = extern/ZoeDepth
|
| 3 |
+
url = [email protected]:isl-org/ZoeDepth.git
|
LICENSE
ADDED
|
@@ -0,0 +1,21 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
MIT License
|
| 2 |
+
|
| 3 |
+
Copyright (c) 2024 Sony Research Inc.
|
| 4 |
+
|
| 5 |
+
Permission is hereby granted, free of charge, to any person obtaining a copy
|
| 6 |
+
of this software and associated documentation files (the "Software"), to deal
|
| 7 |
+
in the Software without restriction, including without limitation the rights
|
| 8 |
+
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
|
| 9 |
+
copies of the Software, and to permit persons to whom the Software is
|
| 10 |
+
furnished to do so, subject to the following conditions:
|
| 11 |
+
|
| 12 |
+
The above copyright notice and this permission notice shall be included in all
|
| 13 |
+
copies or substantial portions of the Software.
|
| 14 |
+
|
| 15 |
+
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
|
| 16 |
+
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
|
| 17 |
+
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
|
| 18 |
+
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
|
| 19 |
+
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
|
| 20 |
+
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
|
| 21 |
+
SOFTWARE.
|
NOTICE
ADDED
|
@@ -0,0 +1,263 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
This repository contains files and parts of codes adapted or modified from thrid-party repositories under other licenses. Below are list of the reporitories. Adapted files are specified in top lines of each file.
|
| 2 |
+
|
| 3 |
+
-----------------------------
|
| 4 |
+
Moore-AnimateAnyone
|
| 5 |
+
Apache License, Version 2.0
|
| 6 |
+
Copyright @2023-2024 Moore Threads Technology Co., Ltd.
|
| 7 |
+
https://github.com/MooreThreads/Moore-AnimateAnyone
|
| 8 |
+
|
| 9 |
+
-----------------------------
|
| 10 |
+
magic-animate
|
| 11 |
+
BSD 3-Clause License
|
| 12 |
+
Copyright (c) Bytedance Inc.
|
| 13 |
+
https://github.com/magic-research/magic-animate
|
| 14 |
+
|
| 15 |
+
-----------------------------
|
| 16 |
+
AnimateDiff
|
| 17 |
+
Apache License, Version 2.0
|
| 18 |
+
https://github.com/guoyww/AnimateDiff
|
| 19 |
+
|
| 20 |
+
-----------------------------
|
| 21 |
+
Diffusers
|
| 22 |
+
Apache License, Version 2.0
|
| 23 |
+
Copyright (c) Hugging Face Inc.
|
| 24 |
+
https://github.com/huggingface/diffusers
|
| 25 |
+
|
| 26 |
+
|
| 27 |
+
================================================================================
|
| 28 |
+
|
| 29 |
+
|
| 30 |
+
Apache License
|
| 31 |
+
Version 2.0, January 2004
|
| 32 |
+
http://www.apache.org/licenses/
|
| 33 |
+
|
| 34 |
+
TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
|
| 35 |
+
|
| 36 |
+
1. Definitions.
|
| 37 |
+
|
| 38 |
+
"License" shall mean the terms and conditions for use, reproduction,
|
| 39 |
+
and distribution as defined by Sections 1 through 9 of this document.
|
| 40 |
+
|
| 41 |
+
"Licensor" shall mean the copyright owner or entity authorized by
|
| 42 |
+
the copyright owner that is granting the License.
|
| 43 |
+
|
| 44 |
+
"Legal Entity" shall mean the union of the acting entity and all
|
| 45 |
+
other entities that control, are controlled by, or are under common
|
| 46 |
+
control with that entity. For the purposes of this definition,
|
| 47 |
+
"control" means (i) the power, direct or indirect, to cause the
|
| 48 |
+
direction or management of such entity, whether by contract or
|
| 49 |
+
otherwise, or (ii) ownership of fifty percent (50%) or more of the
|
| 50 |
+
outstanding shares, or (iii) beneficial ownership of such entity.
|
| 51 |
+
|
| 52 |
+
"You" (or "Your") shall mean an individual or Legal Entity
|
| 53 |
+
exercising permissions granted by this License.
|
| 54 |
+
|
| 55 |
+
"Source" form shall mean the preferred form for making modifications,
|
| 56 |
+
including but not limited to software source code, documentation
|
| 57 |
+
source, and configuration files.
|
| 58 |
+
|
| 59 |
+
"Object" form shall mean any form resulting from mechanical
|
| 60 |
+
transformation or translation of a Source form, including but
|
| 61 |
+
not limited to compiled object code, generated documentation,
|
| 62 |
+
and conversions to other media types.
|
| 63 |
+
|
| 64 |
+
"Work" shall mean the work of authorship, whether in Source or
|
| 65 |
+
Object form, made available under the License, as indicated by a
|
| 66 |
+
copyright notice that is included in or attached to the work
|
| 67 |
+
(an example is provided in the Appendix below).
|
| 68 |
+
|
| 69 |
+
"Derivative Works" shall mean any work, whether in Source or Object
|
| 70 |
+
form, that is based on (or derived from) the Work and for which the
|
| 71 |
+
editorial revisions, annotations, elaborations, or other modifications
|
| 72 |
+
represent, as a whole, an original work of authorship. For the purposes
|
| 73 |
+
of this License, Derivative Works shall not include works that remain
|
| 74 |
+
separable from, or merely link (or bind by name) to the interfaces of,
|
| 75 |
+
the Work and Derivative Works thereof.
|
| 76 |
+
|
| 77 |
+
"Contribution" shall mean any work of authorship, including
|
| 78 |
+
the original version of the Work and any modifications or additions
|
| 79 |
+
to that Work or Derivative Works thereof, that is intentionally
|
| 80 |
+
submitted to Licensor for inclusion in the Work by the copyright owner
|
| 81 |
+
or by an individual or Legal Entity authorized to submit on behalf of
|
| 82 |
+
the copyright owner. For the purposes of this definition, "submitted"
|
| 83 |
+
means any form of electronic, verbal, or written communication sent
|
| 84 |
+
to the Licensor or its representatives, including but not limited to
|
| 85 |
+
communication on electronic mailing lists, source code control systems,
|
| 86 |
+
and issue tracking systems that are managed by, or on behalf of, the
|
| 87 |
+
Licensor for the purpose of discussing and improving the Work, but
|
| 88 |
+
excluding communication that is conspicuously marked or otherwise
|
| 89 |
+
designated in writing by the copyright owner as "Not a Contribution."
|
| 90 |
+
|
| 91 |
+
"Contributor" shall mean Licensor and any individual or Legal Entity
|
| 92 |
+
on behalf of whom a Contribution has been received by Licensor and
|
| 93 |
+
subsequently incorporated within the Work.
|
| 94 |
+
|
| 95 |
+
2. Grant of Copyright License. Subject to the terms and conditions of
|
| 96 |
+
this License, each Contributor hereby grants to You a perpetual,
|
| 97 |
+
worldwide, non-exclusive, no-charge, royalty-free, irrevocable
|
| 98 |
+
copyright license to reproduce, prepare Derivative Works of,
|
| 99 |
+
publicly display, publicly perform, sublicense, and distribute the
|
| 100 |
+
Work and such Derivative Works in Source or Object form.
|
| 101 |
+
|
| 102 |
+
3. Grant of Patent License. Subject to the terms and conditions of
|
| 103 |
+
this License, each Contributor hereby grants to You a perpetual,
|
| 104 |
+
worldwide, non-exclusive, no-charge, royalty-free, irrevocable
|
| 105 |
+
(except as stated in this section) patent license to make, have made,
|
| 106 |
+
use, offer to sell, sell, import, and otherwise transfer the Work,
|
| 107 |
+
where such license applies only to those patent claims licensable
|
| 108 |
+
by such Contributor that are necessarily infringed by their
|
| 109 |
+
Contribution(s) alone or by combination of their Contribution(s)
|
| 110 |
+
with the Work to which such Contribution(s) was submitted. If You
|
| 111 |
+
institute patent litigation against any entity (including a
|
| 112 |
+
cross-claim or counterclaim in a lawsuit) alleging that the Work
|
| 113 |
+
or a Contribution incorporated within the Work constitutes direct
|
| 114 |
+
or contributory patent infringement, then any patent licenses
|
| 115 |
+
granted to You under this License for that Work shall terminate
|
| 116 |
+
as of the date such litigation is filed.
|
| 117 |
+
|
| 118 |
+
4. Redistribution. You may reproduce and distribute copies of the
|
| 119 |
+
Work or Derivative Works thereof in any medium, with or without
|
| 120 |
+
modifications, and in Source or Object form, provided that You
|
| 121 |
+
meet the following conditions:
|
| 122 |
+
|
| 123 |
+
(a) You must give any other recipients of the Work or
|
| 124 |
+
Derivative Works a copy of this License; and
|
| 125 |
+
|
| 126 |
+
(b) You must cause any modified files to carry prominent notices
|
| 127 |
+
stating that You changed the files; and
|
| 128 |
+
|
| 129 |
+
(c) You must retain, in the Source form of any Derivative Works
|
| 130 |
+
that You distribute, all copyright, patent, trademark, and
|
| 131 |
+
attribution notices from the Source form of the Work,
|
| 132 |
+
excluding those notices that do not pertain to any part of
|
| 133 |
+
the Derivative Works; and
|
| 134 |
+
|
| 135 |
+
(d) If the Work includes a "NOTICE" text file as part of its
|
| 136 |
+
distribution, then any Derivative Works that You distribute must
|
| 137 |
+
include a readable copy of the attribution notices contained
|
| 138 |
+
within such NOTICE file, excluding those notices that do not
|
| 139 |
+
pertain to any part of the Derivative Works, in at least one
|
| 140 |
+
of the following places: within a NOTICE text file distributed
|
| 141 |
+
as part of the Derivative Works; within the Source form or
|
| 142 |
+
documentation, if provided along with the Derivative Works; or,
|
| 143 |
+
within a display generated by the Derivative Works, if and
|
| 144 |
+
wherever such third-party notices normally appear. The contents
|
| 145 |
+
of the NOTICE file are for informational purposes only and
|
| 146 |
+
do not modify the License. You may add Your own attribution
|
| 147 |
+
notices within Derivative Works that You distribute, alongside
|
| 148 |
+
or as an addendum to the NOTICE text from the Work, provided
|
| 149 |
+
that such additional attribution notices cannot be construed
|
| 150 |
+
as modifying the License.
|
| 151 |
+
|
| 152 |
+
You may add Your own copyright statement to Your modifications and
|
| 153 |
+
may provide additional or different license terms and conditions
|
| 154 |
+
for use, reproduction, or distribution of Your modifications, or
|
| 155 |
+
for any such Derivative Works as a whole, provided Your use,
|
| 156 |
+
reproduction, and distribution of the Work otherwise complies with
|
| 157 |
+
the conditions stated in this License.
|
| 158 |
+
|
| 159 |
+
5. Submission of Contributions. Unless You explicitly state otherwise,
|
| 160 |
+
any Contribution intentionally submitted for inclusion in the Work
|
| 161 |
+
by You to the Licensor shall be under the terms and conditions of
|
| 162 |
+
this License, without any additional terms or conditions.
|
| 163 |
+
Notwithstanding the above, nothing herein shall supersede or modify
|
| 164 |
+
the terms of any separate license agreement you may have executed
|
| 165 |
+
with Licensor regarding such Contributions.
|
| 166 |
+
|
| 167 |
+
6. Trademarks. This License does not grant permission to use the trade
|
| 168 |
+
names, trademarks, service marks, or product names of the Licensor,
|
| 169 |
+
except as required for reasonable and customary use in describing the
|
| 170 |
+
origin of the Work and reproducing the content of the NOTICE file.
|
| 171 |
+
|
| 172 |
+
7. Disclaimer of Warranty. Unless required by applicable law or
|
| 173 |
+
agreed to in writing, Licensor provides the Work (and each
|
| 174 |
+
Contributor provides its Contributions) on an "AS IS" BASIS,
|
| 175 |
+
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
|
| 176 |
+
implied, including, without limitation, any warranties or conditions
|
| 177 |
+
of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
|
| 178 |
+
PARTICULAR PURPOSE. You are solely responsible for determining the
|
| 179 |
+
appropriateness of using or redistributing the Work and assume any
|
| 180 |
+
risks associated with Your exercise of permissions under this License.
|
| 181 |
+
|
| 182 |
+
8. Limitation of Liability. In no event and under no legal theory,
|
| 183 |
+
whether in tort (including negligence), contract, or otherwise,
|
| 184 |
+
unless required by applicable law (such as deliberate and grossly
|
| 185 |
+
negligent acts) or agreed to in writing, shall any Contributor be
|
| 186 |
+
liable to You for damages, including any direct, indirect, special,
|
| 187 |
+
incidental, or consequential damages of any character arising as a
|
| 188 |
+
result of this License or out of the use or inability to use the
|
| 189 |
+
Work (including but not limited to damages for loss of goodwill,
|
| 190 |
+
work stoppage, computer failure or malfunction, or any and all
|
| 191 |
+
other commercial damages or losses), even if such Contributor
|
| 192 |
+
has been advised of the possibility of such damages.
|
| 193 |
+
|
| 194 |
+
9. Accepting Warranty or Additional Liability. While redistributing
|
| 195 |
+
the Work or Derivative Works thereof, You may choose to offer,
|
| 196 |
+
and charge a fee for, acceptance of support, warranty, indemnity,
|
| 197 |
+
or other liability obligations and/or rights consistent with this
|
| 198 |
+
License. However, in accepting such obligations, You may act only
|
| 199 |
+
on Your own behalf and on Your sole responsibility, not on behalf
|
| 200 |
+
of any other Contributor, and only if You agree to indemnify,
|
| 201 |
+
defend, and hold each Contributor harmless for any liability
|
| 202 |
+
incurred by, or claims asserted against, such Contributor by reason
|
| 203 |
+
of your accepting any such warranty or additional liability.
|
| 204 |
+
|
| 205 |
+
END OF TERMS AND CONDITIONS
|
| 206 |
+
|
| 207 |
+
APPENDIX: How to apply the Apache License to your work.
|
| 208 |
+
|
| 209 |
+
To apply the Apache License to your work, attach the following
|
| 210 |
+
boilerplate notice, with the fields enclosed by brackets "[]"
|
| 211 |
+
replaced with your own identifying information. (Don't include
|
| 212 |
+
the brackets!) The text should be enclosed in the appropriate
|
| 213 |
+
comment syntax for the file format. We also recommend that a
|
| 214 |
+
file or class name and description of purpose be included on the
|
| 215 |
+
same "printed page" as the copyright notice for easier
|
| 216 |
+
identification within third-party archives.
|
| 217 |
+
|
| 218 |
+
Copyright [yyyy] [name of copyright owner]
|
| 219 |
+
|
| 220 |
+
Licensed under the Apache License, Version 2.0 (the "License");
|
| 221 |
+
you may not use this file except in compliance with the License.
|
| 222 |
+
You may obtain a copy of the License at
|
| 223 |
+
|
| 224 |
+
http://www.apache.org/licenses/LICENSE-2.0
|
| 225 |
+
|
| 226 |
+
Unless required by applicable law or agreed to in writing, software
|
| 227 |
+
distributed under the License is distributed on an "AS IS" BASIS,
|
| 228 |
+
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
| 229 |
+
See the License for the specific language governing permissions and
|
| 230 |
+
limitations under the License.
|
| 231 |
+
|
| 232 |
+
|
| 233 |
+
================================================================================
|
| 234 |
+
|
| 235 |
+
|
| 236 |
+
BSD 3-Clause License
|
| 237 |
+
|
| 238 |
+
Copyright 2023 MagicAnimate Team All rights reserved.
|
| 239 |
+
|
| 240 |
+
Redistribution and use in source and binary forms, with or without
|
| 241 |
+
modification, are permitted provided that the following conditions are met:
|
| 242 |
+
|
| 243 |
+
1. Redistributions of source code must retain the above copyright notice, this
|
| 244 |
+
list of conditions and the following disclaimer.
|
| 245 |
+
|
| 246 |
+
2. Redistributions in binary form must reproduce the above copyright notice,
|
| 247 |
+
this list of conditions and the following disclaimer in the documentation
|
| 248 |
+
and/or other materials provided with the distribution.
|
| 249 |
+
|
| 250 |
+
3. Neither the name of the copyright holder nor the names of its
|
| 251 |
+
contributors may be used to endorse or promote products derived from
|
| 252 |
+
this software without specific prior written permission.
|
| 253 |
+
|
| 254 |
+
THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
|
| 255 |
+
AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
|
| 256 |
+
IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
|
| 257 |
+
DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
|
| 258 |
+
FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
|
| 259 |
+
DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
|
| 260 |
+
SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
|
| 261 |
+
CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
|
| 262 |
+
OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
|
| 263 |
+
OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
README.md
CHANGED
|
@@ -1,13 +1,15 @@
|
|
| 1 |
---
|
| 2 |
-
title:
|
| 3 |
-
emoji:
|
| 4 |
-
colorFrom:
|
| 5 |
-
colorTo:
|
| 6 |
sdk: gradio
|
| 7 |
sdk_version: 4.42.0
|
| 8 |
app_file: app.py
|
| 9 |
pinned: false
|
| 10 |
license: mit
|
|
|
|
|
|
|
|
|
|
|
|
|
| 11 |
---
|
| 12 |
-
|
| 13 |
-
Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
|
|
|
|
| 1 |
---
|
| 2 |
+
title: GenWarp
|
| 3 |
+
emoji: 🌃
|
| 4 |
+
colorFrom: purple
|
| 5 |
+
colorTo: blue
|
| 6 |
sdk: gradio
|
| 7 |
sdk_version: 4.42.0
|
| 8 |
app_file: app.py
|
| 9 |
pinned: false
|
| 10 |
license: mit
|
| 11 |
+
preload_from_hub:
|
| 12 |
+
- Sony/genwarp
|
| 13 |
+
- stabilityai/sd-vae-ft-mse diffusion_pytorch_model.safetensors
|
| 14 |
+
- lambdalabs/sd-image-variations-diffusers image_encoder/pytorch_model.bin
|
| 15 |
---
|
|
|
|
|
|
app.py
ADDED
|
@@ -0,0 +1,376 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import sys
|
| 2 |
+
import os
|
| 3 |
+
from subprocess import check_call
|
| 4 |
+
import tempfile
|
| 5 |
+
|
| 6 |
+
from os.path import basename, splitext, join
|
| 7 |
+
from io import BytesIO
|
| 8 |
+
|
| 9 |
+
import numpy as np
|
| 10 |
+
from scipy.spatial import KDTree
|
| 11 |
+
from PIL import Image
|
| 12 |
+
|
| 13 |
+
import torch
|
| 14 |
+
import torch.nn.functional as F
|
| 15 |
+
from torchvision.transforms.functional import to_tensor, to_pil_image
|
| 16 |
+
from einops import rearrange
|
| 17 |
+
import gradio as gr
|
| 18 |
+
from huggingface_hub import hf_hub_download
|
| 19 |
+
|
| 20 |
+
from extern.ZoeDepth.zoedepth.utils.misc import colorize
|
| 21 |
+
|
| 22 |
+
from gradio_model3dgscamera import Model3DGSCamera
|
| 23 |
+
|
| 24 |
+
IMAGE_SIZE = 512
|
| 25 |
+
NEAR, FAR = 0.01, 100
|
| 26 |
+
FOVY = np.deg2rad(55)
|
| 27 |
+
|
| 28 |
+
def download_models():
|
| 29 |
+
models = [
|
| 30 |
+
{
|
| 31 |
+
'repo': 'stabilityai/sd-vae-ft-mse',
|
| 32 |
+
'sub': None,
|
| 33 |
+
'dst': 'checkpoints/sd-vae-ft-mse',
|
| 34 |
+
'files': ['config.json', 'diffusion_pytorch_model.safetensors'],
|
| 35 |
+
'token': None
|
| 36 |
+
},
|
| 37 |
+
{
|
| 38 |
+
'repo': 'lambdalabs/sd-image-variations-diffusers',
|
| 39 |
+
'sub': 'image_encoder',
|
| 40 |
+
'dst': 'checkpoints',
|
| 41 |
+
'files': ['config.json', 'pytorch_model.bin'],
|
| 42 |
+
'token': None
|
| 43 |
+
},
|
| 44 |
+
{
|
| 45 |
+
'repo': 'Sony/genwarp',
|
| 46 |
+
'sub': 'multi1',
|
| 47 |
+
'dst': 'checkpoints',
|
| 48 |
+
'files': ['config.json', 'denoising_unet.pth', 'pose_guider.pth', 'reference_unet.pth'],
|
| 49 |
+
'token': None
|
| 50 |
+
}
|
| 51 |
+
]
|
| 52 |
+
|
| 53 |
+
for model in models:
|
| 54 |
+
for file in model['files']:
|
| 55 |
+
hf_hub_download(
|
| 56 |
+
repo_id=model['repo'],
|
| 57 |
+
subfolder=model['sub'],
|
| 58 |
+
filename=file,
|
| 59 |
+
local_dir=model['dst'],
|
| 60 |
+
token=model['token']
|
| 61 |
+
)
|
| 62 |
+
|
| 63 |
+
# Crop the image to the shorter side.
|
| 64 |
+
def crop(img: Image) -> Image:
|
| 65 |
+
W, H = img.size
|
| 66 |
+
if W < H:
|
| 67 |
+
left, right = 0, W
|
| 68 |
+
top, bottom = np.ceil((H - W) / 2.), np.floor((H - W) / 2.) + W
|
| 69 |
+
else:
|
| 70 |
+
left, right = np.ceil((W - H) / 2.), np.floor((W - H) / 2.) + H
|
| 71 |
+
top, bottom = 0, H
|
| 72 |
+
return img.crop((left, top, right, bottom))
|
| 73 |
+
|
| 74 |
+
def unproject(depth):
|
| 75 |
+
fovy_deg = 55
|
| 76 |
+
H, W = depth.shape[2:4]
|
| 77 |
+
|
| 78 |
+
mean_depth = depth.mean(dim=(2, 3)).squeeze().item()
|
| 79 |
+
|
| 80 |
+
viewport_mtx = get_viewport_matrix(
|
| 81 |
+
IMAGE_SIZE, IMAGE_SIZE,
|
| 82 |
+
batch_size=1
|
| 83 |
+
).to(depth)
|
| 84 |
+
|
| 85 |
+
# Projection matrix.
|
| 86 |
+
fovy = torch.ones(1) * FOVY
|
| 87 |
+
proj_mtx = get_projection_matrix(
|
| 88 |
+
fovy=fovy,
|
| 89 |
+
aspect_wh=1.,
|
| 90 |
+
near=NEAR,
|
| 91 |
+
far=FAR
|
| 92 |
+
).to(depth)
|
| 93 |
+
|
| 94 |
+
view_mtx = camera_lookat(
|
| 95 |
+
torch.tensor([[0., 0., 0.]]),
|
| 96 |
+
torch.tensor([[0., 0., 1.]]),
|
| 97 |
+
torch.tensor([[0., -1., 0.]])
|
| 98 |
+
).to(depth)
|
| 99 |
+
|
| 100 |
+
scr_mtx = (viewport_mtx @ proj_mtx).to(depth)
|
| 101 |
+
|
| 102 |
+
grid = torch.stack(torch.meshgrid(
|
| 103 |
+
torch.arange(W), torch.arange(H), indexing='xy'), dim=-1
|
| 104 |
+
).to(depth)[None] # BHW2
|
| 105 |
+
|
| 106 |
+
screen = F.pad(grid, (0, 1), 'constant', 0)
|
| 107 |
+
screen = F.pad(screen, (0, 1), 'constant', 1)
|
| 108 |
+
screen_flat = rearrange(screen, 'b h w c -> b (h w) c')
|
| 109 |
+
|
| 110 |
+
eye = screen_flat @ torch.linalg.inv_ex(
|
| 111 |
+
scr_mtx.float()
|
| 112 |
+
)[0].mT.to(depth)
|
| 113 |
+
eye = eye * rearrange(depth, 'b c h w -> b (h w) c')
|
| 114 |
+
eye[..., 3] = 1
|
| 115 |
+
|
| 116 |
+
points = eye @ torch.linalg.inv_ex(view_mtx.float())[0].mT.to(depth)
|
| 117 |
+
points = points[0, :, :3]
|
| 118 |
+
|
| 119 |
+
# Translate to the origin.
|
| 120 |
+
points[..., 2] -= mean_depth
|
| 121 |
+
camera_pos = (0, 0, -mean_depth)
|
| 122 |
+
view_mtx = camera_lookat(
|
| 123 |
+
torch.tensor([[0., 0., -mean_depth]]),
|
| 124 |
+
torch.tensor([[0., 0., 0.]]),
|
| 125 |
+
torch.tensor([[0., -1., 0.]])
|
| 126 |
+
).to(depth)
|
| 127 |
+
|
| 128 |
+
return points, camera_pos, view_mtx, proj_mtx
|
| 129 |
+
|
| 130 |
+
def calc_dist2(points: np.ndarray):
|
| 131 |
+
dists, _ = KDTree(points).query(points, k=4)
|
| 132 |
+
mean_dists = (dists[:, 1:] ** 2).mean(1)
|
| 133 |
+
return mean_dists
|
| 134 |
+
|
| 135 |
+
def save_as_splat(
|
| 136 |
+
filepath: str,
|
| 137 |
+
xyz: np.ndarray,
|
| 138 |
+
rgb: np.ndarray
|
| 139 |
+
):
|
| 140 |
+
# To gaussian splat
|
| 141 |
+
inv_sigmoid = lambda x: np.log(x / (1 - x))
|
| 142 |
+
dist2 = np.clip(calc_dist2(xyz), a_min=0.0000001, a_max=None)
|
| 143 |
+
scales = np.repeat(np.log(np.sqrt(dist2))[..., np.newaxis], 3, axis=1)
|
| 144 |
+
rots = np.zeros((xyz.shape[0], 4))
|
| 145 |
+
rots[:, 0] = 1
|
| 146 |
+
opacities = inv_sigmoid(0.1 * np.ones((xyz.shape[0], 1)))
|
| 147 |
+
|
| 148 |
+
sorted_indices = np.argsort((
|
| 149 |
+
-np.exp(np.sum(scales, axis=-1, keepdims=True))
|
| 150 |
+
/ (1 + np.exp(-opacities))
|
| 151 |
+
).squeeze())
|
| 152 |
+
|
| 153 |
+
buffer = BytesIO()
|
| 154 |
+
for idx in sorted_indices:
|
| 155 |
+
position = xyz[idx]
|
| 156 |
+
scale = np.exp(scales[idx]).astype(np.float32)
|
| 157 |
+
rot = rots[idx].astype(np.float32)
|
| 158 |
+
color = np.concatenate(
|
| 159 |
+
(rgb[idx], 1 / (1 + np.exp(-opacities[idx]))),
|
| 160 |
+
axis=-1
|
| 161 |
+
)
|
| 162 |
+
buffer.write(position.tobytes())
|
| 163 |
+
buffer.write(scale.tobytes())
|
| 164 |
+
buffer.write((color * 255).clip(0, 255).astype(np.uint8).tobytes())
|
| 165 |
+
buffer.write(
|
| 166 |
+
((rot / np.linalg.norm(rot)) * 128 + 128)
|
| 167 |
+
.clip(0, 255)
|
| 168 |
+
.astype(np.uint8)
|
| 169 |
+
.tobytes()
|
| 170 |
+
)
|
| 171 |
+
|
| 172 |
+
with open(filepath, "wb") as f:
|
| 173 |
+
f.write(buffer.getvalue())
|
| 174 |
+
|
| 175 |
+
def view_from_rt(position, rotation):
|
| 176 |
+
t = np.array(position)
|
| 177 |
+
euler = np.array(rotation)
|
| 178 |
+
|
| 179 |
+
cx = np.cos(euler[0])
|
| 180 |
+
sx = np.sin(euler[0])
|
| 181 |
+
cy = np.cos(euler[1])
|
| 182 |
+
sy = np.sin(euler[1])
|
| 183 |
+
cz = np.cos(euler[2])
|
| 184 |
+
sz = np.sin(euler[2])
|
| 185 |
+
R = np.array([
|
| 186 |
+
cy * cz + sy * sx * sz,
|
| 187 |
+
-cy * sz + sy * sx * cz,
|
| 188 |
+
sy * cx,
|
| 189 |
+
cx * sz,
|
| 190 |
+
cx * cz,
|
| 191 |
+
-sx,
|
| 192 |
+
-sy * cz + cy * sx * sz,
|
| 193 |
+
sy * sz + cy * sx * cz,
|
| 194 |
+
cy * cx
|
| 195 |
+
])
|
| 196 |
+
view_mtx = np.array([
|
| 197 |
+
[R[0], R[1], R[2], 0],
|
| 198 |
+
[R[3], R[4], R[5], 0],
|
| 199 |
+
[R[6], R[7], R[8], 0],
|
| 200 |
+
[
|
| 201 |
+
-t[0] * R[0] - t[1] * R[3] - t[2] * R[6],
|
| 202 |
+
-t[0] * R[1] - t[1] * R[4] - t[2] * R[7],
|
| 203 |
+
-t[0] * R[2] - t[1] * R[5] - t[2] * R[8],
|
| 204 |
+
1
|
| 205 |
+
]
|
| 206 |
+
]).T
|
| 207 |
+
|
| 208 |
+
B = np.array([
|
| 209 |
+
[1, 0, 0, 0],
|
| 210 |
+
[0, -1, 0, 0],
|
| 211 |
+
[0, 0, -1, 0],
|
| 212 |
+
[0, 0, 0, 1]
|
| 213 |
+
])
|
| 214 |
+
return B @ view_mtx
|
| 215 |
+
|
| 216 |
+
|
| 217 |
+
# Setup.
|
| 218 |
+
download_models()
|
| 219 |
+
|
| 220 |
+
mde = torch.hub.load(
|
| 221 |
+
'./extern/ZoeDepth',
|
| 222 |
+
'ZoeD_N',
|
| 223 |
+
source='local',
|
| 224 |
+
pretrained=True,
|
| 225 |
+
trust_repo=True
|
| 226 |
+
)
|
| 227 |
+
|
| 228 |
+
import spaces
|
| 229 |
+
|
| 230 |
+
check_call([
|
| 231 |
+
sys.executable, '-m', 'pip', 'install',
|
| 232 |
+
'extern/splatting-0.0.1-py3-none-any.whl'
|
| 233 |
+
])
|
| 234 |
+
|
| 235 |
+
from genwarp import GenWarp
|
| 236 |
+
from genwarp.ops import (
|
| 237 |
+
camera_lookat, get_projection_matrix, get_viewport_matrix
|
| 238 |
+
)
|
| 239 |
+
|
| 240 |
+
# GenWarp
|
| 241 |
+
genwarp_cfg = dict(
|
| 242 |
+
pretrained_model_path='checkpoints',
|
| 243 |
+
checkpoint_name='multi1',
|
| 244 |
+
half_precision_weights=True
|
| 245 |
+
)
|
| 246 |
+
genwarp_nvs = GenWarp(cfg=genwarp_cfg, device='cpu')
|
| 247 |
+
|
| 248 |
+
|
| 249 |
+
with tempfile.TemporaryDirectory() as tmpdir:
|
| 250 |
+
with gr.Blocks(
|
| 251 |
+
title='GenWarp Demo',
|
| 252 |
+
css='img {display: inline;}'
|
| 253 |
+
) as demo:
|
| 254 |
+
# Internal states.
|
| 255 |
+
src_image = gr.State()
|
| 256 |
+
src_depth = gr.State()
|
| 257 |
+
proj_mtx = gr.State()
|
| 258 |
+
src_view_mtx = gr.State()
|
| 259 |
+
|
| 260 |
+
# Blocks.
|
| 261 |
+
gr.Markdown(
|
| 262 |
+
"""
|
| 263 |
+
# GenWarp: Single Image to Novel Views with Semantic-Preserving Generative Warping
|
| 264 |
+
[](https://genwarp-nvs.github.io/)
|
| 265 |
+
[](https://huggingface.co/spaces/Sony/GenWarp)
|
| 266 |
+
[](https://github.com/sony/genwarp/)
|
| 267 |
+
[](https://huggingface.co/Sony/genwarp)
|
| 268 |
+
[](https://arxiv.org/abs/2405.17251)
|
| 269 |
+
|
| 270 |
+
## Introduction
|
| 271 |
+
This is an official demo for the paper "[GenWarp: Single Image to Novel Views with Semantic-Preserving Generative Warping](https://genwarp-nvs.github.io/)". Genwarp can generate novel view images from a single input conditioned on camera poses. In this demo, we offer a basic use of inference of the model. For detailed information, please refer the [paper](https://arxiv.org/abs/2405.17251).
|
| 272 |
+
|
| 273 |
+
## How to Use
|
| 274 |
+
1. Upload a reference image to "Reference Input"
|
| 275 |
+
- You can also select a image from "Examples"
|
| 276 |
+
2. Move the camera to your desired view in "Unprojected 3DGS" 3D viewer
|
| 277 |
+
3. Hit "Generate a novel view" button and check the result
|
| 278 |
+
|
| 279 |
+
"""
|
| 280 |
+
)
|
| 281 |
+
file = gr.File(label='Reference Input', file_types=['image'])
|
| 282 |
+
examples = gr.Examples(
|
| 283 |
+
examples=['./assets/pexels-heyho-5998120_19mm.jpg',
|
| 284 |
+
'./assets/pexels-itsterrymag-12639296_24mm.jpg'],
|
| 285 |
+
inputs=file
|
| 286 |
+
)
|
| 287 |
+
with gr.Row():
|
| 288 |
+
image_widget = gr.Image(
|
| 289 |
+
label='Reference View', type='filepath',
|
| 290 |
+
interactive=False
|
| 291 |
+
)
|
| 292 |
+
depth_widget = gr.Image(label='Estimated Depth', type='pil')
|
| 293 |
+
viewer = Model3DGSCamera(
|
| 294 |
+
label = 'Unprojected 3DGS',
|
| 295 |
+
width=IMAGE_SIZE,
|
| 296 |
+
height=IMAGE_SIZE,
|
| 297 |
+
camera_width=IMAGE_SIZE,
|
| 298 |
+
camera_height=IMAGE_SIZE,
|
| 299 |
+
camera_fx=IMAGE_SIZE / (np.tan(FOVY / 2.)) / 2.,
|
| 300 |
+
camera_fy=IMAGE_SIZE / (np.tan(FOVY / 2.)) / 2.,
|
| 301 |
+
camera_near=NEAR,
|
| 302 |
+
camera_far=FAR
|
| 303 |
+
)
|
| 304 |
+
button = gr.Button('Generate a novel view', size='lg', variant='primary')
|
| 305 |
+
with gr.Row():
|
| 306 |
+
warped_widget = gr.Image(
|
| 307 |
+
label='Warped Image', type='pil', interactive=False
|
| 308 |
+
)
|
| 309 |
+
gen_widget = gr.Image(
|
| 310 |
+
label='Generated View', type='pil', interactive=False
|
| 311 |
+
)
|
| 312 |
+
|
| 313 |
+
# Callbacks
|
| 314 |
+
@spaces.GPU
|
| 315 |
+
def cb_mde(image_file: str):
|
| 316 |
+
image = to_tensor(crop(Image.open(
|
| 317 |
+
image_file
|
| 318 |
+
).convert('RGB')).resize((IMAGE_SIZE, IMAGE_SIZE)))[None].cuda()
|
| 319 |
+
depth = mde.cuda().infer(image)
|
| 320 |
+
depth_image = to_pil_image(colorize(depth[0]))
|
| 321 |
+
return to_pil_image(image[0]), depth_image, image.cpu().detach(), depth.cpu().detach()
|
| 322 |
+
|
| 323 |
+
@spaces.GPU
|
| 324 |
+
def cb_3d(image, depth, image_file):
|
| 325 |
+
xyz, camera_pos, view_mtx, proj_mtx = unproject(depth.cuda())
|
| 326 |
+
rgb = rearrange(image, 'b c h w -> b (h w) c')[0]
|
| 327 |
+
splat_file = join(tmpdir, f'./{splitext(basename(image_file))[0]}.splat')
|
| 328 |
+
save_as_splat(splat_file, xyz.cpu().detach().numpy(), rgb.cpu().detach().numpy())
|
| 329 |
+
return (splat_file, camera_pos, None), view_mtx.cpu().detach(), proj_mtx.cpu().detach()
|
| 330 |
+
|
| 331 |
+
@spaces.GPU
|
| 332 |
+
def cb_generate(viewer, image, depth, src_view_mtx, proj_mtx):
|
| 333 |
+
image = image.cuda()
|
| 334 |
+
depth = depth.cuda()
|
| 335 |
+
src_view_mtx = src_view_mtx.cuda()
|
| 336 |
+
proj_mtx = proj_mtx.cuda()
|
| 337 |
+
src_camera_pos = viewer[1]
|
| 338 |
+
src_camera_rot = viewer[2]
|
| 339 |
+
tar_view_mtx = view_from_rt(src_camera_pos, src_camera_rot)
|
| 340 |
+
tar_view_mtx = torch.from_numpy(tar_view_mtx).to(image)
|
| 341 |
+
rel_view_mtx = (
|
| 342 |
+
tar_view_mtx @ torch.linalg.inv(src_view_mtx.to(image))
|
| 343 |
+
).to(image)
|
| 344 |
+
|
| 345 |
+
# GenWarp.
|
| 346 |
+
renders = genwarp_nvs.to('cuda')(
|
| 347 |
+
src_image=image.half(),
|
| 348 |
+
src_depth=depth.half(),
|
| 349 |
+
rel_view_mtx=rel_view_mtx.half(),
|
| 350 |
+
src_proj_mtx=proj_mtx.half(),
|
| 351 |
+
tar_proj_mtx=proj_mtx.half()
|
| 352 |
+
)
|
| 353 |
+
|
| 354 |
+
warped = renders['warped']
|
| 355 |
+
synthesized = renders['synthesized']
|
| 356 |
+
warped_pil = to_pil_image(warped[0])
|
| 357 |
+
synthesized_pil = to_pil_image(synthesized[0])
|
| 358 |
+
|
| 359 |
+
return warped_pil, synthesized_pil
|
| 360 |
+
|
| 361 |
+
# Events
|
| 362 |
+
file.change(
|
| 363 |
+
fn=cb_mde,
|
| 364 |
+
inputs=file,
|
| 365 |
+
outputs=[image_widget, depth_widget, src_image, src_depth]
|
| 366 |
+
).then(
|
| 367 |
+
fn=cb_3d,
|
| 368 |
+
inputs=[src_image, src_depth, image_widget],
|
| 369 |
+
outputs=[viewer, src_view_mtx, proj_mtx])
|
| 370 |
+
button.click(
|
| 371 |
+
fn=cb_generate,
|
| 372 |
+
inputs=[viewer, src_image, src_depth, src_view_mtx, proj_mtx],
|
| 373 |
+
outputs=[warped_widget, gen_widget])
|
| 374 |
+
|
| 375 |
+
if __name__ == '__main__':
|
| 376 |
+
demo.launch()
|
assets/NOTICE
ADDED
|
@@ -0,0 +1,8 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
Images are taken from Pexels
|
| 2 |
+
https://www.pexels.com/
|
| 3 |
+
|
| 4 |
+
pexels-itsterrymag-12639296_24mm.jpg
|
| 5 |
+
https://www.pexels.com/ja-jp/photo/12639296/
|
| 6 |
+
|
| 7 |
+
pexels-heyho-5998120_19mm.jpg
|
| 8 |
+
https://www.pexels.com/ja-jp/photo/5998120/
|
assets/pexels-heyho-5998120_19mm.jpg
ADDED
|
Git LFS Details
|
assets/pexels-itsterrymag-12639296_24mm.jpg
ADDED
|
Git LFS Details
|
extern/ZoeDepth
ADDED
|
@@ -0,0 +1 @@
|
|
|
|
|
|
|
| 1 |
+
Subproject commit edb6daf45458569e24f50250ef1ed08c015f17a7
|
extern/splatting-0.0.1-py3-none-any.whl
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:26d488928a774f4677a0f6cdd9f2a2a63ee73502d90676f507444cc21ecd069d
|
| 3 |
+
size 5189840
|
genwarp/GenWarp.py
ADDED
|
@@ -0,0 +1,546 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
from os.path import join
|
| 2 |
+
from typing import Union, Optional, List, Dict, Tuple, Any
|
| 3 |
+
from dataclasses import dataclass
|
| 4 |
+
import inspect
|
| 5 |
+
|
| 6 |
+
from omegaconf import OmegaConf, DictConfig
|
| 7 |
+
from jaxtyping import Float
|
| 8 |
+
|
| 9 |
+
import torch
|
| 10 |
+
from torch import Tensor
|
| 11 |
+
import torch.nn.functional as F
|
| 12 |
+
from einops import rearrange, repeat
|
| 13 |
+
|
| 14 |
+
from diffusers import AutoencoderKL, DDIMScheduler
|
| 15 |
+
from diffusers.image_processor import VaeImageProcessor
|
| 16 |
+
from transformers import CLIPVisionModelWithProjection, CLIPImageProcessor
|
| 17 |
+
|
| 18 |
+
from .models import (
|
| 19 |
+
PoseGuider,
|
| 20 |
+
UNet2DConditionModel,
|
| 21 |
+
UNet3DConditionModel,
|
| 22 |
+
ReferenceAttentionControl
|
| 23 |
+
)
|
| 24 |
+
from .ops import get_viewport_matrix, forward_warper
|
| 25 |
+
|
| 26 |
+
class GenWarp():
|
| 27 |
+
@dataclass
|
| 28 |
+
class Config():
|
| 29 |
+
pretrained_model_path: str = ''
|
| 30 |
+
checkpoint_name: str = ''
|
| 31 |
+
half_precision_weights: bool = False
|
| 32 |
+
height: int = 512
|
| 33 |
+
width: int = 512
|
| 34 |
+
num_inference_steps: int = 20
|
| 35 |
+
guidance_scale: float = 3.5
|
| 36 |
+
|
| 37 |
+
cfg: Config
|
| 38 |
+
|
| 39 |
+
class Embedder():
|
| 40 |
+
def __init__(self, **kwargs) -> None:
|
| 41 |
+
self.kwargs = kwargs
|
| 42 |
+
self.create_embedding_fn()
|
| 43 |
+
|
| 44 |
+
def create_embedding_fn(self) -> None:
|
| 45 |
+
embed_fns = []
|
| 46 |
+
d = self.kwargs['input_dims']
|
| 47 |
+
out_dim = 0
|
| 48 |
+
if self.kwargs['include_input']:
|
| 49 |
+
embed_fns.append(lambda x : x)
|
| 50 |
+
out_dim += d
|
| 51 |
+
|
| 52 |
+
max_freq = self.kwargs['max_freq_log2']
|
| 53 |
+
N_freqs = self.kwargs['num_freqs']
|
| 54 |
+
|
| 55 |
+
if self.kwargs['log_sampling']:
|
| 56 |
+
freq_bands = 2.**torch.linspace(0., max_freq, steps=N_freqs)
|
| 57 |
+
else:
|
| 58 |
+
freq_bands = torch.linspace(2.**0., 2.**max_freq, steps=N_freqs)
|
| 59 |
+
|
| 60 |
+
for freq in freq_bands:
|
| 61 |
+
for p_fn in self.kwargs['periodic_fns']:
|
| 62 |
+
embed_fns.append(lambda x, p_fn=p_fn, freq=freq : p_fn(x * freq))
|
| 63 |
+
out_dim += d
|
| 64 |
+
|
| 65 |
+
self.embed_fns = embed_fns
|
| 66 |
+
self.out_dim = out_dim
|
| 67 |
+
|
| 68 |
+
def embed(self, inputs) -> Tensor:
|
| 69 |
+
return torch.cat([fn(inputs) for fn in self.embed_fns], -1)
|
| 70 |
+
|
| 71 |
+
def __init__(
|
| 72 |
+
self,
|
| 73 |
+
cfg: Optional[Union[dict, DictConfig]] = None,
|
| 74 |
+
device: Optional[str] = 'cpu'
|
| 75 |
+
) -> None:
|
| 76 |
+
self.cfg = OmegaConf.structured(self.Config(**cfg))
|
| 77 |
+
self.model_path = join(
|
| 78 |
+
self.cfg.pretrained_model_path, self.cfg.checkpoint_name
|
| 79 |
+
)
|
| 80 |
+
self.device = device
|
| 81 |
+
self.configure()
|
| 82 |
+
|
| 83 |
+
def configure(self) -> None:
|
| 84 |
+
print(f"Loading GenWarp...")
|
| 85 |
+
|
| 86 |
+
# Configurations.
|
| 87 |
+
self.dtype = (
|
| 88 |
+
torch.float16 if self.cfg.half_precision_weights else torch.float32
|
| 89 |
+
)
|
| 90 |
+
self.viewport_mtx: Float[Tensor, 'B 4 4'] = get_viewport_matrix(
|
| 91 |
+
self.cfg.width, self.cfg.height,
|
| 92 |
+
batch_size=1, device=self.device
|
| 93 |
+
).to(self.dtype)
|
| 94 |
+
|
| 95 |
+
# Load models.
|
| 96 |
+
self.load_models()
|
| 97 |
+
|
| 98 |
+
# Timestep
|
| 99 |
+
self.scheduler.set_timesteps(
|
| 100 |
+
self.cfg.num_inference_steps, device=self.device)
|
| 101 |
+
self.num_train_timesteps = self.scheduler.config.num_train_timesteps
|
| 102 |
+
|
| 103 |
+
print(f"Loaded GenWarp.")
|
| 104 |
+
|
| 105 |
+
def load_models(self) -> None:
|
| 106 |
+
# VAE.
|
| 107 |
+
self.vae = AutoencoderKL.from_pretrained(
|
| 108 |
+
join(self.cfg.pretrained_model_path, 'sd-vae-ft-mse')
|
| 109 |
+
).to(self.device, dtype=self.dtype)
|
| 110 |
+
|
| 111 |
+
# Image processor.
|
| 112 |
+
self.vae_scale_factor = \
|
| 113 |
+
2 ** (len(self.vae.config.block_out_channels) - 1)
|
| 114 |
+
self.vae_image_processor = VaeImageProcessor(
|
| 115 |
+
vae_scale_factor=self.vae_scale_factor, do_convert_rgb=True
|
| 116 |
+
)
|
| 117 |
+
self.clip_image_processor = CLIPImageProcessor()
|
| 118 |
+
|
| 119 |
+
# Image encoder.
|
| 120 |
+
self.image_encoder = CLIPVisionModelWithProjection.from_pretrained(
|
| 121 |
+
join(self.cfg.pretrained_model_path, 'image_encoder')
|
| 122 |
+
).to(self.device, dtype=self.dtype)
|
| 123 |
+
|
| 124 |
+
# Reference Unet.
|
| 125 |
+
self.reference_unet = UNet2DConditionModel.from_config(
|
| 126 |
+
UNet2DConditionModel.load_config(
|
| 127 |
+
join(self.model_path, 'config.json')
|
| 128 |
+
)).to(self.device, dtype=self.dtype)
|
| 129 |
+
self.reference_unet.load_state_dict(torch.load(
|
| 130 |
+
join(self.model_path, 'reference_unet.pth'),
|
| 131 |
+
map_location='cpu'),
|
| 132 |
+
)
|
| 133 |
+
|
| 134 |
+
# Denoising Unet.
|
| 135 |
+
self.denoising_unet = UNet3DConditionModel.from_pretrained_2d(
|
| 136 |
+
join(self.model_path, 'config.json'),
|
| 137 |
+
join(self.model_path, 'denoising_unet.pth')
|
| 138 |
+
).to(self.device, dtype=self.dtype)
|
| 139 |
+
self.unet_in_channels = self.denoising_unet.config.in_channels
|
| 140 |
+
|
| 141 |
+
# Pose guider.
|
| 142 |
+
self.pose_guider = PoseGuider(
|
| 143 |
+
conditioning_embedding_channels=320,
|
| 144 |
+
conditioning_channels=11,
|
| 145 |
+
).to(self.device, dtype=self.dtype)
|
| 146 |
+
self.pose_guider.load_state_dict(torch.load(
|
| 147 |
+
join(self.model_path, 'pose_guider.pth'),
|
| 148 |
+
map_location='cpu'),
|
| 149 |
+
)
|
| 150 |
+
|
| 151 |
+
# Noise scheduler
|
| 152 |
+
sched_kwargs = OmegaConf.to_container(OmegaConf.create({
|
| 153 |
+
'num_train_timesteps': 1000,
|
| 154 |
+
'beta_start': 0.00085,
|
| 155 |
+
'beta_end': 0.012,
|
| 156 |
+
'beta_schedule': 'scaled_linear',
|
| 157 |
+
'steps_offset': 1,
|
| 158 |
+
'clip_sample': False
|
| 159 |
+
}))
|
| 160 |
+
sched_kwargs.update(
|
| 161 |
+
rescale_betas_zero_snr=True,
|
| 162 |
+
timestep_spacing='trailing',
|
| 163 |
+
prediction_type='v_prediction',
|
| 164 |
+
)
|
| 165 |
+
self.scheduler = DDIMScheduler(**sched_kwargs)
|
| 166 |
+
|
| 167 |
+
self.vae.requires_grad_(False)
|
| 168 |
+
self.image_encoder.requires_grad_(False)
|
| 169 |
+
self.reference_unet.requires_grad_(False)
|
| 170 |
+
self.denoising_unet.requires_grad_(False)
|
| 171 |
+
self.pose_guider.requires_grad_(False)
|
| 172 |
+
|
| 173 |
+
# Coordinates embedding.
|
| 174 |
+
self.embedder = self.get_embedder(2)
|
| 175 |
+
|
| 176 |
+
def to(self, device: str):
|
| 177 |
+
self.device = device
|
| 178 |
+
self.viewport_mtx = self.viewport_mtx.to(device)
|
| 179 |
+
self.vae = self.vae.to(device)
|
| 180 |
+
self.image_encoder = self.image_encoder.to(device)
|
| 181 |
+
self.reference_unet = self.reference_unet.to(device)
|
| 182 |
+
self.denoising_unet = self.denoising_unet.to(device)
|
| 183 |
+
self.pose_guider = self.pose_guider.to(device)
|
| 184 |
+
|
| 185 |
+
return self
|
| 186 |
+
|
| 187 |
+
def get_embedder(self, multires):
|
| 188 |
+
embed_kwargs = {
|
| 189 |
+
'include_input' : True,
|
| 190 |
+
'input_dims' : 2,
|
| 191 |
+
'max_freq_log2' : multires-1,
|
| 192 |
+
'num_freqs' : multires,
|
| 193 |
+
'log_sampling' : True,
|
| 194 |
+
'periodic_fns' : [torch.sin, torch.cos],
|
| 195 |
+
}
|
| 196 |
+
|
| 197 |
+
embedder_obj = self.Embedder(**embed_kwargs)
|
| 198 |
+
embed = lambda x, eo=embedder_obj : eo.embed(x)
|
| 199 |
+
return embed
|
| 200 |
+
|
| 201 |
+
def __call__(
|
| 202 |
+
self,
|
| 203 |
+
src_image: Float[Tensor, 'B C H W'],
|
| 204 |
+
src_depth: Float[Tensor, 'B C H W'],
|
| 205 |
+
rel_view_mtx: Float[Tensor, 'B 4 4'],
|
| 206 |
+
src_proj_mtx: Float[Tensor, 'B 4 4'],
|
| 207 |
+
tar_proj_mtx: Float[Tensor, 'B 4 4'],
|
| 208 |
+
) -> Dict[str, Tensor]:
|
| 209 |
+
""" Perform NVS.
|
| 210 |
+
"""
|
| 211 |
+
batch_size = src_image.shape[0]
|
| 212 |
+
|
| 213 |
+
# Rearrange and resize.
|
| 214 |
+
src_image = self.preprocess_image(src_image)
|
| 215 |
+
src_depth = self.preprocess_image(src_depth)
|
| 216 |
+
viewport_mtx = repeat(
|
| 217 |
+
self.viewport_mtx, 'b h w -> (repeat b) h w',
|
| 218 |
+
repeat=batch_size)
|
| 219 |
+
|
| 220 |
+
pipe_args = dict(
|
| 221 |
+
src_image=src_image,
|
| 222 |
+
src_depth=src_depth,
|
| 223 |
+
rel_view_mtx=rel_view_mtx,
|
| 224 |
+
src_proj_mtx=src_proj_mtx,
|
| 225 |
+
tar_proj_mtx=tar_proj_mtx,
|
| 226 |
+
viewport_mtx=viewport_mtx
|
| 227 |
+
)
|
| 228 |
+
|
| 229 |
+
# Prepare inputs.
|
| 230 |
+
conditions, renders = self.prepare_conditions(**pipe_args)
|
| 231 |
+
|
| 232 |
+
# NVS.
|
| 233 |
+
latents_clean = self.perform_nvs(
|
| 234 |
+
**pipe_args,
|
| 235 |
+
**conditions
|
| 236 |
+
)
|
| 237 |
+
|
| 238 |
+
# Decode to images.
|
| 239 |
+
synthesized = self.decode_latents(latents_clean)
|
| 240 |
+
|
| 241 |
+
inference_out = {
|
| 242 |
+
'synthesized': synthesized,
|
| 243 |
+
'warped': renders['warped'],
|
| 244 |
+
'mask': renders['mask'],
|
| 245 |
+
'correspondence': conditions['correspondence']
|
| 246 |
+
}
|
| 247 |
+
|
| 248 |
+
return inference_out
|
| 249 |
+
|
| 250 |
+
def preprocess_image(
|
| 251 |
+
self,
|
| 252 |
+
image: Float[Tensor, 'B C H W']
|
| 253 |
+
) -> Float[Tensor, 'B C H W']:
|
| 254 |
+
image = F.interpolate(
|
| 255 |
+
image, (self.cfg.height, self.cfg.width)
|
| 256 |
+
)
|
| 257 |
+
return image
|
| 258 |
+
|
| 259 |
+
def get_image_prompt(
|
| 260 |
+
self,
|
| 261 |
+
src_image: Float[Tensor, 'B C H W']
|
| 262 |
+
) -> Float[Tensor, '2 B L']:
|
| 263 |
+
ref_image_for_clip = self.vae_image_processor.preprocess(
|
| 264 |
+
src_image, height=224, width=224
|
| 265 |
+
)
|
| 266 |
+
ref_image_for_clip = ref_image_for_clip * 0.5 + 0.5
|
| 267 |
+
|
| 268 |
+
clip_image = self.clip_image_processor.preprocess(
|
| 269 |
+
ref_image_for_clip, return_tensors='pt'
|
| 270 |
+
).pixel_values
|
| 271 |
+
|
| 272 |
+
clip_image_embeds = self.image_encoder(
|
| 273 |
+
clip_image.to(self.device, dtype=self.image_encoder.dtype)
|
| 274 |
+
).image_embeds
|
| 275 |
+
|
| 276 |
+
image_prompt_embeds = clip_image_embeds.unsqueeze(1)
|
| 277 |
+
uncond_image_prompt_embeds = torch.zeros_like(image_prompt_embeds)
|
| 278 |
+
|
| 279 |
+
image_prompt_embeds = torch.cat(
|
| 280 |
+
[uncond_image_prompt_embeds, image_prompt_embeds], dim=0
|
| 281 |
+
)
|
| 282 |
+
|
| 283 |
+
return image_prompt_embeds
|
| 284 |
+
|
| 285 |
+
def encode_images(
|
| 286 |
+
self,
|
| 287 |
+
rgb: Float[Tensor, 'B C H W']
|
| 288 |
+
) -> Float[Tensor, 'B C H W']:
|
| 289 |
+
rgb = self.vae_image_processor.preprocess(rgb)
|
| 290 |
+
latents = self.vae.encode(rgb).latent_dist.mean
|
| 291 |
+
latents = latents * 0.18215
|
| 292 |
+
return latents
|
| 293 |
+
|
| 294 |
+
def decode_latents(
|
| 295 |
+
self,
|
| 296 |
+
latents: Float[Tensor, 'B C H W']
|
| 297 |
+
) -> Float[Tensor, 'B C H W']:
|
| 298 |
+
latents = 1 / 0.18215 * latents
|
| 299 |
+
rgb = []
|
| 300 |
+
for frame_idx in range(latents.shape[0]):
|
| 301 |
+
rgb.append(self.vae.decode(latents[frame_idx : frame_idx + 1]).sample)
|
| 302 |
+
rgb = torch.cat(rgb)
|
| 303 |
+
rgb = (rgb / 2 + 0.5).clamp(0, 1)
|
| 304 |
+
return rgb.squeeze(2)
|
| 305 |
+
|
| 306 |
+
def get_reference_controls(
|
| 307 |
+
self,
|
| 308 |
+
batch_size: int
|
| 309 |
+
) -> Tuple[ReferenceAttentionControl, ReferenceAttentionControl]:
|
| 310 |
+
reader = ReferenceAttentionControl(
|
| 311 |
+
self.denoising_unet,
|
| 312 |
+
do_classifier_free_guidance=True,
|
| 313 |
+
mode='read',
|
| 314 |
+
batch_size=batch_size,
|
| 315 |
+
fusion_blocks='full',
|
| 316 |
+
feature_fusion_type='attention_full_sharing'
|
| 317 |
+
)
|
| 318 |
+
writer = ReferenceAttentionControl(
|
| 319 |
+
self.reference_unet,
|
| 320 |
+
do_classifier_free_guidance=True,
|
| 321 |
+
mode='write',
|
| 322 |
+
batch_size=batch_size,
|
| 323 |
+
fusion_blocks='full',
|
| 324 |
+
feature_fusion_type='attention_full_sharing'
|
| 325 |
+
)
|
| 326 |
+
|
| 327 |
+
return reader, writer
|
| 328 |
+
|
| 329 |
+
def prepare_extra_step_kwargs(
|
| 330 |
+
self,
|
| 331 |
+
generator,
|
| 332 |
+
eta
|
| 333 |
+
) -> Dict[str, Any]:
|
| 334 |
+
accepts_eta = 'eta' in set(
|
| 335 |
+
inspect.signature(self.scheduler.step).parameters.keys()
|
| 336 |
+
)
|
| 337 |
+
extra_step_kwargs = {}
|
| 338 |
+
if accepts_eta:
|
| 339 |
+
extra_step_kwargs['eta'] = eta
|
| 340 |
+
|
| 341 |
+
# check if the scheduler accepts generator
|
| 342 |
+
accepts_generator = 'generator' in set(
|
| 343 |
+
inspect.signature(self.scheduler.step).parameters.keys()
|
| 344 |
+
)
|
| 345 |
+
if accepts_generator:
|
| 346 |
+
extra_step_kwargs['generator'] = generator
|
| 347 |
+
return extra_step_kwargs
|
| 348 |
+
|
| 349 |
+
def get_pose_features(
|
| 350 |
+
self,
|
| 351 |
+
src_embed: Float[Tensor, 'B C H W'],
|
| 352 |
+
trg_embed: Float[Tensor, 'B C H W'],
|
| 353 |
+
do_classifier_guidance: bool = True
|
| 354 |
+
) -> Tuple[Tensor, Tensor]:
|
| 355 |
+
pose_cond_tensor = src_embed.unsqueeze(2)
|
| 356 |
+
pose_cond_tensor = pose_cond_tensor.to(
|
| 357 |
+
device=self.device, dtype=self.pose_guider.dtype
|
| 358 |
+
)
|
| 359 |
+
pose_cond_tensor_2 = trg_embed.unsqueeze(2)
|
| 360 |
+
pose_cond_tensor_2 = pose_cond_tensor_2.to(
|
| 361 |
+
device=self.device, dtype=self.pose_guider.dtype
|
| 362 |
+
)
|
| 363 |
+
pose_fea = self.pose_guider(pose_cond_tensor)
|
| 364 |
+
pose_fea_2 = self.pose_guider(pose_cond_tensor_2)
|
| 365 |
+
|
| 366 |
+
if do_classifier_guidance:
|
| 367 |
+
pose_fea = torch.cat([pose_fea] * 2)
|
| 368 |
+
pose_fea_2 = torch.cat([pose_fea_2] * 2)
|
| 369 |
+
|
| 370 |
+
return pose_fea, pose_fea_2
|
| 371 |
+
|
| 372 |
+
@torch.no_grad()
|
| 373 |
+
def prepare_conditions(
|
| 374 |
+
self,
|
| 375 |
+
src_image: Float[Tensor, 'B C H W'],
|
| 376 |
+
src_depth: Float[Tensor, 'B C H W'],
|
| 377 |
+
rel_view_mtx: Float[Tensor, 'B 4 4'],
|
| 378 |
+
src_proj_mtx: Float[Tensor, 'B 4 4'],
|
| 379 |
+
tar_proj_mtx: Float[Tensor, 'B 4 4'],
|
| 380 |
+
viewport_mtx: Float[Tensor, 'B 4 4']
|
| 381 |
+
) -> Tuple[Dict[str, Tensor], Dict[str, Tensor]]:
|
| 382 |
+
# Prepare inputs.
|
| 383 |
+
B = src_image.shape[0]
|
| 384 |
+
H, W = src_image.shape[2:4]
|
| 385 |
+
src_scr_mtx = (viewport_mtx @ src_proj_mtx).to(src_proj_mtx)
|
| 386 |
+
mvp_mtx = (tar_proj_mtx @ rel_view_mtx).to(rel_view_mtx)
|
| 387 |
+
|
| 388 |
+
# Coordinate grids.
|
| 389 |
+
grid: Float[Tensor, 'H W C'] = torch.stack(torch.meshgrid(
|
| 390 |
+
torch.arange(W), torch.arange(H), indexing='xy'), dim=-1
|
| 391 |
+
).to(self.device, dtype=self.dtype)
|
| 392 |
+
|
| 393 |
+
# Unproject depth.
|
| 394 |
+
screen = F.pad(grid, (0, 1), 'constant', 0) # z=0 (z doesn't matter)
|
| 395 |
+
screen = F.pad(screen, (0, 1), 'constant', 1) # w=1
|
| 396 |
+
screen = repeat(screen, 'h w c -> b h w c', b=B)
|
| 397 |
+
screen_flat = rearrange(screen, 'b h w c -> b (h w) c')
|
| 398 |
+
# To eye coordinates.
|
| 399 |
+
eye = screen_flat @ torch.linalg.inv_ex(
|
| 400 |
+
src_scr_mtx.float()
|
| 401 |
+
)[0].mT.to(self.dtype)
|
| 402 |
+
# Overwrite depth.
|
| 403 |
+
eye = eye * rearrange(src_depth, 'b c h w -> b (h w) c')
|
| 404 |
+
eye[..., 3] = 1
|
| 405 |
+
|
| 406 |
+
# Coordinates embedding.
|
| 407 |
+
coords = torch.stack((grid[..., 0]/H, grid[..., 1]/W), dim=-1)
|
| 408 |
+
embed = repeat(self.embedder(coords), 'h w c -> b c h w', b=B)
|
| 409 |
+
|
| 410 |
+
# Warping.
|
| 411 |
+
input_image: Float[Tensor, 'B C H W'] = torch.cat(
|
| 412 |
+
[embed, src_image], dim=1
|
| 413 |
+
)
|
| 414 |
+
output_image = forward_warper(
|
| 415 |
+
input_image, screen_flat[..., :2], eye,
|
| 416 |
+
mvp_mtx=mvp_mtx, viewport_mtx=viewport_mtx
|
| 417 |
+
)
|
| 418 |
+
warped_embed = output_image['warped'][:, :embed.shape[1]]
|
| 419 |
+
warped_image = output_image['warped'][:, embed.shape[1]:]
|
| 420 |
+
# mask == 1 where there's no pixel
|
| 421 |
+
mask = output_image['mask']
|
| 422 |
+
correspondence = output_image['correspondence']
|
| 423 |
+
|
| 424 |
+
# Conditions.
|
| 425 |
+
src_coords_embed = torch.cat(
|
| 426 |
+
[embed, torch.zeros_like(mask, device=mask.device)], dim=1)
|
| 427 |
+
trg_coords_embed = torch.cat([warped_embed, mask], dim=1)
|
| 428 |
+
|
| 429 |
+
# Outputs.
|
| 430 |
+
conditions = dict(
|
| 431 |
+
src_coords_embed=src_coords_embed,
|
| 432 |
+
trg_coords_embed=trg_coords_embed,
|
| 433 |
+
correspondence=correspondence
|
| 434 |
+
)
|
| 435 |
+
|
| 436 |
+
renders = dict(
|
| 437 |
+
warped=warped_image,
|
| 438 |
+
mask=1 - mask # mask == 1 where there's a pixel
|
| 439 |
+
)
|
| 440 |
+
|
| 441 |
+
return conditions, renders
|
| 442 |
+
|
| 443 |
+
def perform_nvs(
|
| 444 |
+
self,
|
| 445 |
+
src_image,
|
| 446 |
+
src_coords_embed,
|
| 447 |
+
trg_coords_embed,
|
| 448 |
+
correspondence,
|
| 449 |
+
eta: float=0.0,
|
| 450 |
+
generator: Optional[Union[torch.Generator, List[torch.Generator]]]=None,
|
| 451 |
+
**kwargs,
|
| 452 |
+
) -> Float[Tensor, 'B C H W']:
|
| 453 |
+
batch_size = src_image.shape[0]
|
| 454 |
+
|
| 455 |
+
# For the cross attention.
|
| 456 |
+
reference_control_reader, reference_control_writer = \
|
| 457 |
+
self.get_reference_controls(batch_size)
|
| 458 |
+
|
| 459 |
+
# Prepare extra step kwargs.
|
| 460 |
+
extra_step_kwargs = self.prepare_extra_step_kwargs(
|
| 461 |
+
generator, eta
|
| 462 |
+
)
|
| 463 |
+
|
| 464 |
+
with torch.no_grad():
|
| 465 |
+
# Create fake inputs. It'll be replaced by pure noise.
|
| 466 |
+
latents = torch.randn(
|
| 467 |
+
batch_size,
|
| 468 |
+
self.unet_in_channels,
|
| 469 |
+
self.cfg.height // self.vae_scale_factor,
|
| 470 |
+
self.cfg.width // self.vae_scale_factor
|
| 471 |
+
).to(self.device, dtype=src_image.dtype)
|
| 472 |
+
initial_t = torch.tensor(
|
| 473 |
+
[self.num_train_timesteps - 1] * batch_size
|
| 474 |
+
).to(self.device, dtype=torch.long)
|
| 475 |
+
|
| 476 |
+
# Add noise.
|
| 477 |
+
noise = torch.randn_like(latents)
|
| 478 |
+
latents_noisy_start = self.scheduler.add_noise(
|
| 479 |
+
latents, noise, initial_t
|
| 480 |
+
)
|
| 481 |
+
latents_noisy_start = latents_noisy_start.unsqueeze(2)
|
| 482 |
+
|
| 483 |
+
# Prepare clip image embeds.
|
| 484 |
+
image_prompt_embeds = self.get_image_prompt(src_image)
|
| 485 |
+
|
| 486 |
+
# Prepare ref image latents.
|
| 487 |
+
ref_image_latents = self.encode_images(src_image)
|
| 488 |
+
|
| 489 |
+
# Prepare pose condition image.
|
| 490 |
+
pose_fea, pose_fea_2 = self.get_pose_features(
|
| 491 |
+
src_coords_embed, trg_coords_embed
|
| 492 |
+
)
|
| 493 |
+
|
| 494 |
+
pose_fea = pose_fea[:, :, 0, ...]
|
| 495 |
+
|
| 496 |
+
# Forward reference images.
|
| 497 |
+
self.reference_unet(
|
| 498 |
+
ref_image_latents.repeat(2, 1, 1, 1),
|
| 499 |
+
torch.zeros(batch_size * 2).to(ref_image_latents),
|
| 500 |
+
encoder_hidden_states=image_prompt_embeds,
|
| 501 |
+
pose_cond_fea=pose_fea,
|
| 502 |
+
return_dict=False,
|
| 503 |
+
)
|
| 504 |
+
# Update the denosing net with reference features.
|
| 505 |
+
reference_control_reader.update(
|
| 506 |
+
reference_control_writer,
|
| 507 |
+
correspondence=correspondence
|
| 508 |
+
)
|
| 509 |
+
|
| 510 |
+
timesteps = self.scheduler.timesteps
|
| 511 |
+
latents_noisy = latents_noisy_start
|
| 512 |
+
for t in timesteps:
|
| 513 |
+
# Prepare latents.
|
| 514 |
+
latent_model_input = torch.cat([latents_noisy] * 2)
|
| 515 |
+
latent_model_input = self.scheduler.scale_model_input(
|
| 516 |
+
latent_model_input, t
|
| 517 |
+
)
|
| 518 |
+
|
| 519 |
+
# Denoise.
|
| 520 |
+
noise_pred = self.denoising_unet(
|
| 521 |
+
latent_model_input,
|
| 522 |
+
t,
|
| 523 |
+
encoder_hidden_states=image_prompt_embeds,
|
| 524 |
+
pose_cond_fea=pose_fea_2,
|
| 525 |
+
return_dict=False,
|
| 526 |
+
)[0]
|
| 527 |
+
|
| 528 |
+
# CFG.
|
| 529 |
+
noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
|
| 530 |
+
noise_pred = noise_pred_uncond + self.cfg.guidance_scale * (
|
| 531 |
+
noise_pred_text - noise_pred_uncond
|
| 532 |
+
)
|
| 533 |
+
|
| 534 |
+
# t -> t-1
|
| 535 |
+
latents_noisy = self.scheduler.step(
|
| 536 |
+
noise_pred, t, latents_noisy, **extra_step_kwargs,
|
| 537 |
+
return_dict=False
|
| 538 |
+
)[0]
|
| 539 |
+
|
| 540 |
+
# Noise disappears eventually
|
| 541 |
+
latents_clean = latents_noisy
|
| 542 |
+
|
| 543 |
+
reference_control_reader.clear()
|
| 544 |
+
reference_control_writer.clear()
|
| 545 |
+
|
| 546 |
+
return latents_clean.squeeze(2)
|
genwarp/__init__.py
ADDED
|
@@ -0,0 +1 @@
|
|
|
|
|
|
|
| 1 |
+
from .GenWarp import GenWarp
|
genwarp/models/__init__.py
ADDED
|
@@ -0,0 +1,4 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
from .pose_guider import PoseGuider
|
| 2 |
+
from .unet_2d_condition import UNet2DConditionModel
|
| 3 |
+
from .unet_3d import UNet3DConditionModel
|
| 4 |
+
from .mutual_self_attention import ReferenceAttentionControl
|
genwarp/models/attention.py
ADDED
|
@@ -0,0 +1,499 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# This code is adapted from below and then modified.
|
| 2 |
+
# -----------------------------------------------------------------------------
|
| 3 |
+
# Moore-AnimateAnyone
|
| 4 |
+
# Apache License, Version 2.0
|
| 5 |
+
# Copyright @2023-2024 Moore Threads Technology Co., Ltd.
|
| 6 |
+
# https://github.com/MooreThreads/Moore-AnimateAnyone
|
| 7 |
+
# -----------------------------------------------------------------------------
|
| 8 |
+
# Diffusers
|
| 9 |
+
# Apache License, Version 2.0
|
| 10 |
+
# Copyright (c) Hugging Face Inc.
|
| 11 |
+
# https://github.com/huggingface/diffusers
|
| 12 |
+
# ==============================================================================
|
| 13 |
+
|
| 14 |
+
# Adapted from https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention.py
|
| 15 |
+
|
| 16 |
+
from typing import Any, Dict, Optional
|
| 17 |
+
|
| 18 |
+
import torch
|
| 19 |
+
from diffusers.models.attention import AdaLayerNorm, Attention, FeedForward
|
| 20 |
+
from diffusers.models.embeddings import SinusoidalPositionalEmbedding
|
| 21 |
+
from einops import rearrange
|
| 22 |
+
from torch import nn
|
| 23 |
+
|
| 24 |
+
|
| 25 |
+
class BasicTransformerBlock(nn.Module):
|
| 26 |
+
r"""
|
| 27 |
+
A basic Transformer block.
|
| 28 |
+
|
| 29 |
+
Parameters:
|
| 30 |
+
dim (`int`): The number of channels in the input and output.
|
| 31 |
+
num_attention_heads (`int`): The number of heads to use for multi-head attention.
|
| 32 |
+
attention_head_dim (`int`): The number of channels in each head.
|
| 33 |
+
dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
|
| 34 |
+
cross_attention_dim (`int`, *optional*): The size of the encoder_hidden_states vector for cross attention.
|
| 35 |
+
activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to be used in feed-forward.
|
| 36 |
+
num_embeds_ada_norm (:
|
| 37 |
+
obj: `int`, *optional*): The number of diffusion steps used during training. See `Transformer2DModel`.
|
| 38 |
+
attention_bias (:
|
| 39 |
+
obj: `bool`, *optional*, defaults to `False`): Configure if the attentions should contain a bias parameter.
|
| 40 |
+
only_cross_attention (`bool`, *optional*):
|
| 41 |
+
Whether to use only cross-attention layers. In this case two cross attention layers are used.
|
| 42 |
+
double_self_attention (`bool`, *optional*):
|
| 43 |
+
Whether to use two self-attention layers. In this case no cross attention layers are used.
|
| 44 |
+
upcast_attention (`bool`, *optional*):
|
| 45 |
+
Whether to upcast the attention computation to float32. This is useful for mixed precision training.
|
| 46 |
+
norm_elementwise_affine (`bool`, *optional*, defaults to `True`):
|
| 47 |
+
Whether to use learnable elementwise affine parameters for normalization.
|
| 48 |
+
norm_type (`str`, *optional*, defaults to `"layer_norm"`):
|
| 49 |
+
The normalization layer to use. Can be `"layer_norm"`, `"ada_norm"` or `"ada_norm_zero"`.
|
| 50 |
+
final_dropout (`bool` *optional*, defaults to False):
|
| 51 |
+
Whether to apply a final dropout after the last feed-forward layer.
|
| 52 |
+
attention_type (`str`, *optional*, defaults to `"default"`):
|
| 53 |
+
The type of attention to use. Can be `"default"` or `"gated"` or `"gated-text-image"`.
|
| 54 |
+
positional_embeddings (`str`, *optional*, defaults to `None`):
|
| 55 |
+
The type of positional embeddings to apply to.
|
| 56 |
+
num_positional_embeddings (`int`, *optional*, defaults to `None`):
|
| 57 |
+
The maximum number of positional embeddings to apply.
|
| 58 |
+
"""
|
| 59 |
+
|
| 60 |
+
def __init__(
|
| 61 |
+
self,
|
| 62 |
+
dim: int,
|
| 63 |
+
num_attention_heads: int,
|
| 64 |
+
attention_head_dim: int,
|
| 65 |
+
dropout=0.0,
|
| 66 |
+
cross_attention_dim: Optional[int] = None,
|
| 67 |
+
activation_fn: str = "geglu",
|
| 68 |
+
num_embeds_ada_norm: Optional[int] = None,
|
| 69 |
+
attention_bias: bool = False,
|
| 70 |
+
only_cross_attention: bool = False,
|
| 71 |
+
double_self_attention: bool = False,
|
| 72 |
+
upcast_attention: bool = False,
|
| 73 |
+
norm_elementwise_affine: bool = True,
|
| 74 |
+
norm_type: str = "layer_norm", # 'layer_norm', 'ada_norm', 'ada_norm_zero', 'ada_norm_single'
|
| 75 |
+
norm_eps: float = 1e-5,
|
| 76 |
+
final_dropout: bool = False,
|
| 77 |
+
attention_type: str = "default",
|
| 78 |
+
positional_embeddings: Optional[str] = None,
|
| 79 |
+
num_positional_embeddings: Optional[int] = None,
|
| 80 |
+
):
|
| 81 |
+
super().__init__()
|
| 82 |
+
self.only_cross_attention = only_cross_attention
|
| 83 |
+
|
| 84 |
+
self.use_ada_layer_norm_zero = (
|
| 85 |
+
num_embeds_ada_norm is not None
|
| 86 |
+
) and norm_type == "ada_norm_zero"
|
| 87 |
+
self.use_ada_layer_norm = (
|
| 88 |
+
num_embeds_ada_norm is not None
|
| 89 |
+
) and norm_type == "ada_norm"
|
| 90 |
+
self.use_ada_layer_norm_single = norm_type == "ada_norm_single"
|
| 91 |
+
self.use_layer_norm = norm_type == "layer_norm"
|
| 92 |
+
|
| 93 |
+
if norm_type in ("ada_norm", "ada_norm_zero") and num_embeds_ada_norm is None:
|
| 94 |
+
raise ValueError(
|
| 95 |
+
f"`norm_type` is set to {norm_type}, but `num_embeds_ada_norm` is not defined. Please make sure to"
|
| 96 |
+
f" define `num_embeds_ada_norm` if setting `norm_type` to {norm_type}."
|
| 97 |
+
)
|
| 98 |
+
|
| 99 |
+
if positional_embeddings and (num_positional_embeddings is None):
|
| 100 |
+
raise ValueError(
|
| 101 |
+
"If `positional_embedding` type is defined, `num_positition_embeddings` must also be defined."
|
| 102 |
+
)
|
| 103 |
+
|
| 104 |
+
if positional_embeddings == "sinusoidal":
|
| 105 |
+
self.pos_embed = SinusoidalPositionalEmbedding(
|
| 106 |
+
dim, max_seq_length=num_positional_embeddings
|
| 107 |
+
)
|
| 108 |
+
else:
|
| 109 |
+
self.pos_embed = None
|
| 110 |
+
|
| 111 |
+
# Define 3 blocks. Each block has its own normalization layer.
|
| 112 |
+
# 1. Self-Attn
|
| 113 |
+
if self.use_ada_layer_norm:
|
| 114 |
+
self.norm1 = AdaLayerNorm(dim, num_embeds_ada_norm)
|
| 115 |
+
elif self.use_ada_layer_norm_zero:
|
| 116 |
+
self.norm1 = AdaLayerNormZero(dim, num_embeds_ada_norm)
|
| 117 |
+
else:
|
| 118 |
+
self.norm1 = nn.LayerNorm(
|
| 119 |
+
dim, elementwise_affine=norm_elementwise_affine, eps=norm_eps
|
| 120 |
+
)
|
| 121 |
+
|
| 122 |
+
self.attn1 = Attention(
|
| 123 |
+
query_dim=dim,
|
| 124 |
+
heads=num_attention_heads,
|
| 125 |
+
dim_head=attention_head_dim,
|
| 126 |
+
dropout=dropout,
|
| 127 |
+
bias=attention_bias,
|
| 128 |
+
cross_attention_dim=cross_attention_dim if only_cross_attention else None,
|
| 129 |
+
upcast_attention=upcast_attention,
|
| 130 |
+
)
|
| 131 |
+
|
| 132 |
+
# 2. Cross-Attn
|
| 133 |
+
if cross_attention_dim is not None or double_self_attention:
|
| 134 |
+
# We currently only use AdaLayerNormZero for self attention where there will only be one attention block.
|
| 135 |
+
# I.e. the number of returned modulation chunks from AdaLayerZero would not make sense if returned during
|
| 136 |
+
# the second cross attention block.
|
| 137 |
+
self.norm2 = (
|
| 138 |
+
AdaLayerNorm(dim, num_embeds_ada_norm)
|
| 139 |
+
if self.use_ada_layer_norm
|
| 140 |
+
else nn.LayerNorm(
|
| 141 |
+
dim, elementwise_affine=norm_elementwise_affine, eps=norm_eps
|
| 142 |
+
)
|
| 143 |
+
)
|
| 144 |
+
self.attn2 = Attention(
|
| 145 |
+
query_dim=dim,
|
| 146 |
+
cross_attention_dim=cross_attention_dim
|
| 147 |
+
if not double_self_attention
|
| 148 |
+
else None,
|
| 149 |
+
heads=num_attention_heads,
|
| 150 |
+
dim_head=attention_head_dim,
|
| 151 |
+
dropout=dropout,
|
| 152 |
+
bias=attention_bias,
|
| 153 |
+
upcast_attention=upcast_attention,
|
| 154 |
+
) # is self-attn if encoder_hidden_states is none
|
| 155 |
+
else:
|
| 156 |
+
self.norm2 = None
|
| 157 |
+
self.attn2 = None
|
| 158 |
+
|
| 159 |
+
# 3. Feed-forward
|
| 160 |
+
if not self.use_ada_layer_norm_single:
|
| 161 |
+
self.norm3 = nn.LayerNorm(
|
| 162 |
+
dim, elementwise_affine=norm_elementwise_affine, eps=norm_eps
|
| 163 |
+
)
|
| 164 |
+
|
| 165 |
+
self.ff = FeedForward(
|
| 166 |
+
dim,
|
| 167 |
+
dropout=dropout,
|
| 168 |
+
activation_fn=activation_fn,
|
| 169 |
+
final_dropout=final_dropout,
|
| 170 |
+
)
|
| 171 |
+
|
| 172 |
+
# 4. Fuser
|
| 173 |
+
if attention_type == "gated" or attention_type == "gated-text-image":
|
| 174 |
+
self.fuser = GatedSelfAttentionDense(
|
| 175 |
+
dim, cross_attention_dim, num_attention_heads, attention_head_dim
|
| 176 |
+
)
|
| 177 |
+
|
| 178 |
+
# 5. Scale-shift for PixArt-Alpha.
|
| 179 |
+
if self.use_ada_layer_norm_single:
|
| 180 |
+
self.scale_shift_table = nn.Parameter(torch.randn(6, dim) / dim**0.5)
|
| 181 |
+
|
| 182 |
+
# let chunk size default to None
|
| 183 |
+
self._chunk_size = None
|
| 184 |
+
self._chunk_dim = 0
|
| 185 |
+
|
| 186 |
+
def set_chunk_feed_forward(self, chunk_size: Optional[int], dim: int = 0):
|
| 187 |
+
# Sets chunk feed-forward
|
| 188 |
+
self._chunk_size = chunk_size
|
| 189 |
+
self._chunk_dim = dim
|
| 190 |
+
|
| 191 |
+
def forward(
|
| 192 |
+
self,
|
| 193 |
+
hidden_states: torch.FloatTensor,
|
| 194 |
+
attention_mask: Optional[torch.FloatTensor] = None,
|
| 195 |
+
encoder_hidden_states: Optional[torch.FloatTensor] = None,
|
| 196 |
+
encoder_attention_mask: Optional[torch.FloatTensor] = None,
|
| 197 |
+
timestep: Optional[torch.LongTensor] = None,
|
| 198 |
+
cross_attention_kwargs: Dict[str, Any] = None,
|
| 199 |
+
class_labels: Optional[torch.LongTensor] = None,
|
| 200 |
+
) -> torch.FloatTensor:
|
| 201 |
+
# Notice that normalization is always applied before the real computation in the following blocks.
|
| 202 |
+
# 0. Self-Attention
|
| 203 |
+
batch_size = hidden_states.shape[0]
|
| 204 |
+
|
| 205 |
+
if self.use_ada_layer_norm:
|
| 206 |
+
norm_hidden_states = self.norm1(hidden_states, timestep)
|
| 207 |
+
elif self.use_ada_layer_norm_zero:
|
| 208 |
+
norm_hidden_states, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.norm1(
|
| 209 |
+
hidden_states, timestep, class_labels, hidden_dtype=hidden_states.dtype
|
| 210 |
+
)
|
| 211 |
+
elif self.use_layer_norm:
|
| 212 |
+
norm_hidden_states = self.norm1(hidden_states)
|
| 213 |
+
elif self.use_ada_layer_norm_single:
|
| 214 |
+
shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = (
|
| 215 |
+
self.scale_shift_table[None] + timestep.reshape(batch_size, 6, -1)
|
| 216 |
+
).chunk(6, dim=1)
|
| 217 |
+
norm_hidden_states = self.norm1(hidden_states)
|
| 218 |
+
norm_hidden_states = norm_hidden_states * (1 + scale_msa) + shift_msa
|
| 219 |
+
norm_hidden_states = norm_hidden_states.squeeze(1)
|
| 220 |
+
else:
|
| 221 |
+
raise ValueError("Incorrect norm used")
|
| 222 |
+
|
| 223 |
+
if self.pos_embed is not None:
|
| 224 |
+
norm_hidden_states = self.pos_embed(norm_hidden_states)
|
| 225 |
+
|
| 226 |
+
# 1. Retrieve lora scale.
|
| 227 |
+
lora_scale = (
|
| 228 |
+
cross_attention_kwargs.get("scale", 1.0)
|
| 229 |
+
if cross_attention_kwargs is not None
|
| 230 |
+
else 1.0
|
| 231 |
+
)
|
| 232 |
+
|
| 233 |
+
# 2. Prepare GLIGEN inputs
|
| 234 |
+
cross_attention_kwargs = (
|
| 235 |
+
cross_attention_kwargs.copy() if cross_attention_kwargs is not None else {}
|
| 236 |
+
)
|
| 237 |
+
gligen_kwargs = cross_attention_kwargs.pop("gligen", None)
|
| 238 |
+
|
| 239 |
+
attn_output = self.attn1(
|
| 240 |
+
norm_hidden_states,
|
| 241 |
+
encoder_hidden_states=encoder_hidden_states
|
| 242 |
+
if self.only_cross_attention
|
| 243 |
+
else None,
|
| 244 |
+
attention_mask=attention_mask,
|
| 245 |
+
**cross_attention_kwargs,
|
| 246 |
+
)
|
| 247 |
+
if self.use_ada_layer_norm_zero:
|
| 248 |
+
attn_output = gate_msa.unsqueeze(1) * attn_output
|
| 249 |
+
elif self.use_ada_layer_norm_single:
|
| 250 |
+
attn_output = gate_msa * attn_output
|
| 251 |
+
|
| 252 |
+
hidden_states = attn_output + hidden_states
|
| 253 |
+
if hidden_states.ndim == 4:
|
| 254 |
+
hidden_states = hidden_states.squeeze(1)
|
| 255 |
+
|
| 256 |
+
# 2.5 GLIGEN Control
|
| 257 |
+
if gligen_kwargs is not None:
|
| 258 |
+
hidden_states = self.fuser(hidden_states, gligen_kwargs["objs"])
|
| 259 |
+
|
| 260 |
+
# 3. Cross-Attention
|
| 261 |
+
if self.attn2 is not None:
|
| 262 |
+
if self.use_ada_layer_norm:
|
| 263 |
+
norm_hidden_states = self.norm2(hidden_states, timestep)
|
| 264 |
+
elif self.use_ada_layer_norm_zero or self.use_layer_norm:
|
| 265 |
+
norm_hidden_states = self.norm2(hidden_states)
|
| 266 |
+
elif self.use_ada_layer_norm_single:
|
| 267 |
+
# For PixArt norm2 isn't applied here:
|
| 268 |
+
# https://github.com/PixArt-alpha/PixArt-alpha/blob/0f55e922376d8b797edd44d25d0e7464b260dcab/diffusion/model/nets/PixArtMS.py#L70C1-L76C103
|
| 269 |
+
norm_hidden_states = hidden_states
|
| 270 |
+
else:
|
| 271 |
+
raise ValueError("Incorrect norm")
|
| 272 |
+
|
| 273 |
+
if self.pos_embed is not None and self.use_ada_layer_norm_single is False:
|
| 274 |
+
norm_hidden_states = self.pos_embed(norm_hidden_states)
|
| 275 |
+
|
| 276 |
+
attn_output = self.attn2(
|
| 277 |
+
norm_hidden_states,
|
| 278 |
+
encoder_hidden_states=encoder_hidden_states,
|
| 279 |
+
attention_mask=encoder_attention_mask,
|
| 280 |
+
**cross_attention_kwargs,
|
| 281 |
+
)
|
| 282 |
+
hidden_states = attn_output + hidden_states
|
| 283 |
+
|
| 284 |
+
# 4. Feed-forward
|
| 285 |
+
if not self.use_ada_layer_norm_single:
|
| 286 |
+
norm_hidden_states = self.norm3(hidden_states)
|
| 287 |
+
|
| 288 |
+
if self.use_ada_layer_norm_zero:
|
| 289 |
+
norm_hidden_states = (
|
| 290 |
+
norm_hidden_states * (1 + scale_mlp[:, None]) + shift_mlp[:, None]
|
| 291 |
+
)
|
| 292 |
+
|
| 293 |
+
if self.use_ada_layer_norm_single:
|
| 294 |
+
norm_hidden_states = self.norm2(hidden_states)
|
| 295 |
+
norm_hidden_states = norm_hidden_states * (1 + scale_mlp) + shift_mlp
|
| 296 |
+
|
| 297 |
+
ff_output = self.ff(norm_hidden_states, scale=lora_scale)
|
| 298 |
+
|
| 299 |
+
if self.use_ada_layer_norm_zero:
|
| 300 |
+
ff_output = gate_mlp.unsqueeze(1) * ff_output
|
| 301 |
+
elif self.use_ada_layer_norm_single:
|
| 302 |
+
ff_output = gate_mlp * ff_output
|
| 303 |
+
|
| 304 |
+
hidden_states = ff_output + hidden_states
|
| 305 |
+
if hidden_states.ndim == 4:
|
| 306 |
+
hidden_states = hidden_states.squeeze(1)
|
| 307 |
+
|
| 308 |
+
return hidden_states
|
| 309 |
+
|
| 310 |
+
|
| 311 |
+
class WarpedFeatureInjector(nn.Module):
|
| 312 |
+
def __init__(self, dim: int):
|
| 313 |
+
super().__init__()
|
| 314 |
+
|
| 315 |
+
self.dim = dim
|
| 316 |
+
# Additional convolutional layers
|
| 317 |
+
self.conv1 = nn.Conv2d(dim, dim, kernel_size=3, stride=1, padding=1, bias=False)
|
| 318 |
+
self.conv2 = nn.Conv2d(dim, dim, kernel_size=3, stride=1, padding=1, bias=False)
|
| 319 |
+
self.conv3 = nn.Conv2d(dim, dim, kernel_size=3, stride=1, padding=1, bias=False)
|
| 320 |
+
# Initialize convolutional layers
|
| 321 |
+
nn.init.kaiming_normal_(self.conv1.weight, mode='fan_out', nonlinearity='relu')
|
| 322 |
+
nn.init.kaiming_normal_(self.conv2.weight, mode='fan_out', nonlinearity='relu')
|
| 323 |
+
nn.init.kaiming_normal_(self.conv3.weight, mode='fan_out', nonlinearity='relu')
|
| 324 |
+
|
| 325 |
+
# Zero convolution
|
| 326 |
+
self.out_conv = nn.Conv2d(
|
| 327 |
+
dim, dim, kernel_size=(1, 1), stride=(1, 1), padding=(0, 0), bias=False
|
| 328 |
+
)
|
| 329 |
+
nn.init.zeros_(self.out_conv.weight.data)
|
| 330 |
+
def forward(self, x):
|
| 331 |
+
# Apply convolutional layers
|
| 332 |
+
x = self.conv1(x)
|
| 333 |
+
x = self.conv2(x)
|
| 334 |
+
x = self.conv3(x)
|
| 335 |
+
|
| 336 |
+
# Apply zero convolution
|
| 337 |
+
x = self.out_conv(x)
|
| 338 |
+
|
| 339 |
+
return x
|
| 340 |
+
|
| 341 |
+
|
| 342 |
+
|
| 343 |
+
class TemporalBasicTransformerBlock(nn.Module):
|
| 344 |
+
def __init__(
|
| 345 |
+
self,
|
| 346 |
+
dim: int,
|
| 347 |
+
num_attention_heads: int,
|
| 348 |
+
attention_head_dim: int,
|
| 349 |
+
dropout=0.0,
|
| 350 |
+
cross_attention_dim: Optional[int] = None,
|
| 351 |
+
activation_fn: str = "geglu",
|
| 352 |
+
num_embeds_ada_norm: Optional[int] = None,
|
| 353 |
+
attention_bias: bool = False,
|
| 354 |
+
only_cross_attention: bool = False,
|
| 355 |
+
upcast_attention: bool = False,
|
| 356 |
+
unet_use_cross_frame_attention=None,
|
| 357 |
+
unet_use_temporal_attention=None,
|
| 358 |
+
use_zero_convs=False,
|
| 359 |
+
):
|
| 360 |
+
super().__init__()
|
| 361 |
+
self.only_cross_attention = only_cross_attention
|
| 362 |
+
self.use_ada_layer_norm = num_embeds_ada_norm is not None
|
| 363 |
+
self.unet_use_cross_frame_attention = unet_use_cross_frame_attention
|
| 364 |
+
self.unet_use_temporal_attention = unet_use_temporal_attention
|
| 365 |
+
|
| 366 |
+
if use_zero_convs:
|
| 367 |
+
# self.zero_conv = nn.Conv2d(
|
| 368 |
+
# dim, dim, kernel_size=(1, 1), stride=(1, 1), padding=(0, 0), bias=False
|
| 369 |
+
# )
|
| 370 |
+
# nn.init.zeros_(self.zero_conv.weight.data)
|
| 371 |
+
self.zero_conv = WarpedFeatureInjector(dim)
|
| 372 |
+
|
| 373 |
+
else:
|
| 374 |
+
self.zero_conv = None
|
| 375 |
+
|
| 376 |
+
# SC-Attn
|
| 377 |
+
self.attn1 = Attention(
|
| 378 |
+
query_dim=dim,
|
| 379 |
+
heads=num_attention_heads,
|
| 380 |
+
dim_head=attention_head_dim,
|
| 381 |
+
dropout=dropout,
|
| 382 |
+
bias=attention_bias,
|
| 383 |
+
upcast_attention=upcast_attention,
|
| 384 |
+
)
|
| 385 |
+
self.norm1 = (
|
| 386 |
+
AdaLayerNorm(dim, num_embeds_ada_norm)
|
| 387 |
+
if self.use_ada_layer_norm
|
| 388 |
+
else nn.LayerNorm(dim)
|
| 389 |
+
)
|
| 390 |
+
|
| 391 |
+
# Cross-Attn
|
| 392 |
+
if cross_attention_dim is not None:
|
| 393 |
+
self.attn2 = Attention(
|
| 394 |
+
query_dim=dim,
|
| 395 |
+
cross_attention_dim=cross_attention_dim,
|
| 396 |
+
heads=num_attention_heads,
|
| 397 |
+
dim_head=attention_head_dim,
|
| 398 |
+
dropout=dropout,
|
| 399 |
+
bias=attention_bias,
|
| 400 |
+
upcast_attention=upcast_attention,
|
| 401 |
+
)
|
| 402 |
+
else:
|
| 403 |
+
self.attn2 = None
|
| 404 |
+
|
| 405 |
+
if cross_attention_dim is not None:
|
| 406 |
+
self.norm2 = (
|
| 407 |
+
AdaLayerNorm(dim, num_embeds_ada_norm)
|
| 408 |
+
if self.use_ada_layer_norm
|
| 409 |
+
else nn.LayerNorm(dim)
|
| 410 |
+
)
|
| 411 |
+
else:
|
| 412 |
+
self.norm2 = None
|
| 413 |
+
|
| 414 |
+
# Feed-forward
|
| 415 |
+
self.ff = FeedForward(dim, dropout=dropout, activation_fn=activation_fn)
|
| 416 |
+
self.norm3 = nn.LayerNorm(dim)
|
| 417 |
+
self.use_ada_layer_norm_zero = False
|
| 418 |
+
|
| 419 |
+
# Temp-Attn
|
| 420 |
+
assert unet_use_temporal_attention is not None
|
| 421 |
+
if unet_use_temporal_attention:
|
| 422 |
+
self.attn_temp = Attention(
|
| 423 |
+
query_dim=dim,
|
| 424 |
+
heads=num_attention_heads,
|
| 425 |
+
dim_head=attention_head_dim,
|
| 426 |
+
dropout=dropout,
|
| 427 |
+
bias=attention_bias,
|
| 428 |
+
upcast_attention=upcast_attention,
|
| 429 |
+
)
|
| 430 |
+
nn.init.zeros_(self.attn_temp.to_out[0].weight.data)
|
| 431 |
+
self.norm_temp = (
|
| 432 |
+
AdaLayerNorm(dim, num_embeds_ada_norm)
|
| 433 |
+
if self.use_ada_layer_norm
|
| 434 |
+
else nn.LayerNorm(dim)
|
| 435 |
+
)
|
| 436 |
+
|
| 437 |
+
def forward(
|
| 438 |
+
self,
|
| 439 |
+
hidden_states,
|
| 440 |
+
encoder_hidden_states=None,
|
| 441 |
+
timestep=None,
|
| 442 |
+
attention_mask=None,
|
| 443 |
+
video_length=None,
|
| 444 |
+
):
|
| 445 |
+
norm_hidden_states = (
|
| 446 |
+
self.norm1(hidden_states, timestep)
|
| 447 |
+
if self.use_ada_layer_norm
|
| 448 |
+
else self.norm1(hidden_states)
|
| 449 |
+
)
|
| 450 |
+
|
| 451 |
+
if self.unet_use_cross_frame_attention:
|
| 452 |
+
hidden_states = (
|
| 453 |
+
self.attn1(
|
| 454 |
+
norm_hidden_states,
|
| 455 |
+
attention_mask=attention_mask,
|
| 456 |
+
video_length=video_length,
|
| 457 |
+
)
|
| 458 |
+
+ hidden_states
|
| 459 |
+
)
|
| 460 |
+
else:
|
| 461 |
+
hidden_states = (
|
| 462 |
+
self.attn1(norm_hidden_states, attention_mask=attention_mask)
|
| 463 |
+
+ hidden_states
|
| 464 |
+
)
|
| 465 |
+
|
| 466 |
+
if self.attn2 is not None:
|
| 467 |
+
# Cross-Attention
|
| 468 |
+
norm_hidden_states = (
|
| 469 |
+
self.norm2(hidden_states, timestep)
|
| 470 |
+
if self.use_ada_layer_norm
|
| 471 |
+
else self.norm2(hidden_states)
|
| 472 |
+
)
|
| 473 |
+
hidden_states = (
|
| 474 |
+
self.attn2(
|
| 475 |
+
norm_hidden_states,
|
| 476 |
+
encoder_hidden_states=encoder_hidden_states,
|
| 477 |
+
attention_mask=attention_mask,
|
| 478 |
+
)
|
| 479 |
+
+ hidden_states
|
| 480 |
+
)
|
| 481 |
+
|
| 482 |
+
# Feed-forward
|
| 483 |
+
hidden_states = self.ff(self.norm3(hidden_states)) + hidden_states
|
| 484 |
+
|
| 485 |
+
# Temporal-Attention
|
| 486 |
+
if self.unet_use_temporal_attention:
|
| 487 |
+
d = hidden_states.shape[1]
|
| 488 |
+
hidden_states = rearrange(
|
| 489 |
+
hidden_states, "(b f) d c -> (b d) f c", f=video_length
|
| 490 |
+
)
|
| 491 |
+
norm_hidden_states = (
|
| 492 |
+
self.norm_temp(hidden_states, timestep)
|
| 493 |
+
if self.use_ada_layer_norm
|
| 494 |
+
else self.norm_temp(hidden_states)
|
| 495 |
+
)
|
| 496 |
+
hidden_states = self.attn_temp(norm_hidden_states) + hidden_states
|
| 497 |
+
hidden_states = rearrange(hidden_states, "(b d) f c -> (b f) d c", d=d)
|
| 498 |
+
|
| 499 |
+
return hidden_states
|
genwarp/models/motion_module.py
ADDED
|
@@ -0,0 +1,399 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# This code is adapted from below.
|
| 2 |
+
# -----------------------------------------------------------------------------
|
| 3 |
+
# Moore-AnimateAnyone
|
| 4 |
+
# Apache License, Version 2.0
|
| 5 |
+
# Copyright @2023-2024 Moore Threads Technology Co., Ltd.
|
| 6 |
+
# https://github.com/MooreThreads/Moore-AnimateAnyone
|
| 7 |
+
# -----------------------------------------------------------------------------
|
| 8 |
+
# AnimateDiff
|
| 9 |
+
# Apache License, Version 2.0G
|
| 10 |
+
# https://github.com/guoyww/AnimateDiff
|
| 11 |
+
# ==============================================================================
|
| 12 |
+
|
| 13 |
+
# Adapt from https://github.com/guoyww/AnimateDiff/blob/main/animatediff/models/motion_module.py
|
| 14 |
+
import math
|
| 15 |
+
from dataclasses import dataclass
|
| 16 |
+
from typing import Callable, Optional
|
| 17 |
+
|
| 18 |
+
import torch
|
| 19 |
+
from torch import nn
|
| 20 |
+
from diffusers.models.attention import FeedForward
|
| 21 |
+
from diffusers.models.attention_processor import Attention, AttnProcessor
|
| 22 |
+
from diffusers.utils import BaseOutput
|
| 23 |
+
from diffusers.utils.import_utils import is_xformers_available
|
| 24 |
+
from einops import rearrange, repeat
|
| 25 |
+
|
| 26 |
+
def zero_module(module):
|
| 27 |
+
# Zero out the parameters of a module and return it.
|
| 28 |
+
for p in module.parameters():
|
| 29 |
+
p.detach().zero_()
|
| 30 |
+
return module
|
| 31 |
+
|
| 32 |
+
|
| 33 |
+
@dataclass
|
| 34 |
+
class TemporalTransformer3DModelOutput(BaseOutput):
|
| 35 |
+
sample: torch.FloatTensor
|
| 36 |
+
|
| 37 |
+
|
| 38 |
+
if is_xformers_available():
|
| 39 |
+
import xformers
|
| 40 |
+
import xformers.ops
|
| 41 |
+
else:
|
| 42 |
+
xformers = None
|
| 43 |
+
|
| 44 |
+
|
| 45 |
+
def get_motion_module(in_channels, motion_module_type: str, motion_module_kwargs: dict):
|
| 46 |
+
if motion_module_type == "Vanilla":
|
| 47 |
+
return VanillaTemporalModule(
|
| 48 |
+
in_channels=in_channels,
|
| 49 |
+
**motion_module_kwargs,
|
| 50 |
+
)
|
| 51 |
+
else:
|
| 52 |
+
raise ValueError
|
| 53 |
+
|
| 54 |
+
|
| 55 |
+
class VanillaTemporalModule(nn.Module):
|
| 56 |
+
def __init__(
|
| 57 |
+
self,
|
| 58 |
+
in_channels,
|
| 59 |
+
num_attention_heads=8,
|
| 60 |
+
num_transformer_block=2,
|
| 61 |
+
attention_block_types=("Temporal_Self", "Temporal_Self"),
|
| 62 |
+
cross_frame_attention_mode=None,
|
| 63 |
+
temporal_position_encoding=False,
|
| 64 |
+
temporal_position_encoding_max_len=24,
|
| 65 |
+
temporal_attention_dim_div=1,
|
| 66 |
+
zero_initialize=True,
|
| 67 |
+
):
|
| 68 |
+
super().__init__()
|
| 69 |
+
|
| 70 |
+
self.temporal_transformer = TemporalTransformer3DModel(
|
| 71 |
+
in_channels=in_channels,
|
| 72 |
+
num_attention_heads=num_attention_heads,
|
| 73 |
+
attention_head_dim=in_channels
|
| 74 |
+
// num_attention_heads
|
| 75 |
+
// temporal_attention_dim_div,
|
| 76 |
+
num_layers=num_transformer_block,
|
| 77 |
+
attention_block_types=attention_block_types,
|
| 78 |
+
cross_frame_attention_mode=cross_frame_attention_mode,
|
| 79 |
+
temporal_position_encoding=temporal_position_encoding,
|
| 80 |
+
temporal_position_encoding_max_len=temporal_position_encoding_max_len,
|
| 81 |
+
)
|
| 82 |
+
|
| 83 |
+
if zero_initialize:
|
| 84 |
+
self.temporal_transformer.proj_out = zero_module(
|
| 85 |
+
self.temporal_transformer.proj_out
|
| 86 |
+
)
|
| 87 |
+
|
| 88 |
+
def forward(
|
| 89 |
+
self,
|
| 90 |
+
input_tensor,
|
| 91 |
+
temb,
|
| 92 |
+
encoder_hidden_states,
|
| 93 |
+
attention_mask=None,
|
| 94 |
+
anchor_frame_idx=None,
|
| 95 |
+
):
|
| 96 |
+
hidden_states = input_tensor
|
| 97 |
+
hidden_states = self.temporal_transformer(
|
| 98 |
+
hidden_states, encoder_hidden_states, attention_mask
|
| 99 |
+
)
|
| 100 |
+
|
| 101 |
+
output = hidden_states
|
| 102 |
+
return output
|
| 103 |
+
|
| 104 |
+
|
| 105 |
+
class TemporalTransformer3DModel(nn.Module):
|
| 106 |
+
def __init__(
|
| 107 |
+
self,
|
| 108 |
+
in_channels,
|
| 109 |
+
num_attention_heads,
|
| 110 |
+
attention_head_dim,
|
| 111 |
+
num_layers,
|
| 112 |
+
attention_block_types=(
|
| 113 |
+
"Temporal_Self",
|
| 114 |
+
"Temporal_Self",
|
| 115 |
+
),
|
| 116 |
+
dropout=0.0,
|
| 117 |
+
norm_num_groups=32,
|
| 118 |
+
cross_attention_dim=768,
|
| 119 |
+
activation_fn="geglu",
|
| 120 |
+
attention_bias=False,
|
| 121 |
+
upcast_attention=False,
|
| 122 |
+
cross_frame_attention_mode=None,
|
| 123 |
+
temporal_position_encoding=False,
|
| 124 |
+
temporal_position_encoding_max_len=24,
|
| 125 |
+
):
|
| 126 |
+
super().__init__()
|
| 127 |
+
|
| 128 |
+
inner_dim = num_attention_heads * attention_head_dim
|
| 129 |
+
|
| 130 |
+
self.norm = torch.nn.GroupNorm(
|
| 131 |
+
num_groups=norm_num_groups, num_channels=in_channels, eps=1e-6, affine=True
|
| 132 |
+
)
|
| 133 |
+
self.proj_in = nn.Linear(in_channels, inner_dim)
|
| 134 |
+
|
| 135 |
+
self.transformer_blocks = nn.ModuleList(
|
| 136 |
+
[
|
| 137 |
+
TemporalTransformerBlock(
|
| 138 |
+
dim=inner_dim,
|
| 139 |
+
num_attention_heads=num_attention_heads,
|
| 140 |
+
attention_head_dim=attention_head_dim,
|
| 141 |
+
attention_block_types=attention_block_types,
|
| 142 |
+
dropout=dropout,
|
| 143 |
+
norm_num_groups=norm_num_groups,
|
| 144 |
+
cross_attention_dim=cross_attention_dim,
|
| 145 |
+
activation_fn=activation_fn,
|
| 146 |
+
attention_bias=attention_bias,
|
| 147 |
+
upcast_attention=upcast_attention,
|
| 148 |
+
cross_frame_attention_mode=cross_frame_attention_mode,
|
| 149 |
+
temporal_position_encoding=temporal_position_encoding,
|
| 150 |
+
temporal_position_encoding_max_len=temporal_position_encoding_max_len,
|
| 151 |
+
)
|
| 152 |
+
for d in range(num_layers)
|
| 153 |
+
]
|
| 154 |
+
)
|
| 155 |
+
self.proj_out = nn.Linear(inner_dim, in_channels)
|
| 156 |
+
|
| 157 |
+
def forward(self, hidden_states, encoder_hidden_states=None, attention_mask=None):
|
| 158 |
+
assert (
|
| 159 |
+
hidden_states.dim() == 5
|
| 160 |
+
), f"Expected hidden_states to have ndim=5, but got ndim={hidden_states.dim()}."
|
| 161 |
+
video_length = hidden_states.shape[2]
|
| 162 |
+
hidden_states = rearrange(hidden_states, "b c f h w -> (b f) c h w")
|
| 163 |
+
|
| 164 |
+
batch, channel, height, weight = hidden_states.shape
|
| 165 |
+
residual = hidden_states
|
| 166 |
+
|
| 167 |
+
hidden_states = self.norm(hidden_states)
|
| 168 |
+
inner_dim = hidden_states.shape[1]
|
| 169 |
+
hidden_states = hidden_states.permute(0, 2, 3, 1).reshape(
|
| 170 |
+
batch, height * weight, inner_dim
|
| 171 |
+
)
|
| 172 |
+
hidden_states = self.proj_in(hidden_states)
|
| 173 |
+
|
| 174 |
+
# Transformer Blocks
|
| 175 |
+
for block in self.transformer_blocks:
|
| 176 |
+
hidden_states = block(
|
| 177 |
+
hidden_states,
|
| 178 |
+
encoder_hidden_states=encoder_hidden_states,
|
| 179 |
+
video_length=video_length,
|
| 180 |
+
)
|
| 181 |
+
|
| 182 |
+
# output
|
| 183 |
+
hidden_states = self.proj_out(hidden_states)
|
| 184 |
+
hidden_states = (
|
| 185 |
+
hidden_states.reshape(batch, height, weight, inner_dim)
|
| 186 |
+
.permute(0, 3, 1, 2)
|
| 187 |
+
.contiguous()
|
| 188 |
+
)
|
| 189 |
+
|
| 190 |
+
output = hidden_states + residual
|
| 191 |
+
output = rearrange(output, "(b f) c h w -> b c f h w", f=video_length)
|
| 192 |
+
|
| 193 |
+
return output
|
| 194 |
+
|
| 195 |
+
|
| 196 |
+
class TemporalTransformerBlock(nn.Module):
|
| 197 |
+
def __init__(
|
| 198 |
+
self,
|
| 199 |
+
dim,
|
| 200 |
+
num_attention_heads,
|
| 201 |
+
attention_head_dim,
|
| 202 |
+
attention_block_types=(
|
| 203 |
+
"Temporal_Self",
|
| 204 |
+
"Temporal_Self",
|
| 205 |
+
),
|
| 206 |
+
dropout=0.0,
|
| 207 |
+
norm_num_groups=32,
|
| 208 |
+
cross_attention_dim=768,
|
| 209 |
+
activation_fn="geglu",
|
| 210 |
+
attention_bias=False,
|
| 211 |
+
upcast_attention=False,
|
| 212 |
+
cross_frame_attention_mode=None,
|
| 213 |
+
temporal_position_encoding=False,
|
| 214 |
+
temporal_position_encoding_max_len=24,
|
| 215 |
+
):
|
| 216 |
+
super().__init__()
|
| 217 |
+
|
| 218 |
+
attention_blocks = []
|
| 219 |
+
norms = []
|
| 220 |
+
|
| 221 |
+
for block_name in attention_block_types:
|
| 222 |
+
attention_blocks.append(
|
| 223 |
+
VersatileAttention(
|
| 224 |
+
attention_mode=block_name.split("_")[0],
|
| 225 |
+
cross_attention_dim=cross_attention_dim
|
| 226 |
+
if block_name.endswith("_Cross")
|
| 227 |
+
else None,
|
| 228 |
+
query_dim=dim,
|
| 229 |
+
heads=num_attention_heads,
|
| 230 |
+
dim_head=attention_head_dim,
|
| 231 |
+
dropout=dropout,
|
| 232 |
+
bias=attention_bias,
|
| 233 |
+
upcast_attention=upcast_attention,
|
| 234 |
+
cross_frame_attention_mode=cross_frame_attention_mode,
|
| 235 |
+
temporal_position_encoding=temporal_position_encoding,
|
| 236 |
+
temporal_position_encoding_max_len=temporal_position_encoding_max_len,
|
| 237 |
+
)
|
| 238 |
+
)
|
| 239 |
+
norms.append(nn.LayerNorm(dim))
|
| 240 |
+
|
| 241 |
+
self.attention_blocks = nn.ModuleList(attention_blocks)
|
| 242 |
+
self.norms = nn.ModuleList(norms)
|
| 243 |
+
|
| 244 |
+
self.ff = FeedForward(dim, dropout=dropout, activation_fn=activation_fn)
|
| 245 |
+
self.ff_norm = nn.LayerNorm(dim)
|
| 246 |
+
|
| 247 |
+
def forward(
|
| 248 |
+
self,
|
| 249 |
+
hidden_states,
|
| 250 |
+
encoder_hidden_states=None,
|
| 251 |
+
attention_mask=None,
|
| 252 |
+
video_length=None,
|
| 253 |
+
):
|
| 254 |
+
for attention_block, norm in zip(self.attention_blocks, self.norms):
|
| 255 |
+
norm_hidden_states = norm(hidden_states)
|
| 256 |
+
hidden_states = (
|
| 257 |
+
attention_block(
|
| 258 |
+
norm_hidden_states,
|
| 259 |
+
encoder_hidden_states=encoder_hidden_states
|
| 260 |
+
if attention_block.is_cross_attention
|
| 261 |
+
else None,
|
| 262 |
+
video_length=video_length,
|
| 263 |
+
)
|
| 264 |
+
+ hidden_states
|
| 265 |
+
)
|
| 266 |
+
|
| 267 |
+
hidden_states = self.ff(self.ff_norm(hidden_states)) + hidden_states
|
| 268 |
+
|
| 269 |
+
output = hidden_states
|
| 270 |
+
return output
|
| 271 |
+
|
| 272 |
+
|
| 273 |
+
class PositionalEncoding(nn.Module):
|
| 274 |
+
def __init__(self, d_model, dropout=0.0, max_len=24):
|
| 275 |
+
super().__init__()
|
| 276 |
+
self.dropout = nn.Dropout(p=dropout)
|
| 277 |
+
position = torch.arange(max_len).unsqueeze(1)
|
| 278 |
+
div_term = torch.exp(
|
| 279 |
+
torch.arange(0, d_model, 2) * (-math.log(10000.0) / d_model)
|
| 280 |
+
)
|
| 281 |
+
pe = torch.zeros(1, max_len, d_model)
|
| 282 |
+
pe[0, :, 0::2] = torch.sin(position * div_term)
|
| 283 |
+
pe[0, :, 1::2] = torch.cos(position * div_term)
|
| 284 |
+
self.register_buffer("pe", pe)
|
| 285 |
+
|
| 286 |
+
def forward(self, x):
|
| 287 |
+
x = x + self.pe[:, : x.size(1)]
|
| 288 |
+
return self.dropout(x)
|
| 289 |
+
|
| 290 |
+
|
| 291 |
+
class VersatileAttention(Attention):
|
| 292 |
+
def __init__(
|
| 293 |
+
self,
|
| 294 |
+
attention_mode=None,
|
| 295 |
+
cross_frame_attention_mode=None,
|
| 296 |
+
temporal_position_encoding=False,
|
| 297 |
+
temporal_position_encoding_max_len=24,
|
| 298 |
+
*args,
|
| 299 |
+
**kwargs,
|
| 300 |
+
):
|
| 301 |
+
super().__init__(*args, **kwargs)
|
| 302 |
+
assert attention_mode == "Temporal"
|
| 303 |
+
|
| 304 |
+
self.attention_mode = attention_mode
|
| 305 |
+
self.is_cross_attention = kwargs["cross_attention_dim"] is not None
|
| 306 |
+
|
| 307 |
+
self.pos_encoder = (
|
| 308 |
+
PositionalEncoding(
|
| 309 |
+
kwargs["query_dim"],
|
| 310 |
+
dropout=0.0,
|
| 311 |
+
max_len=temporal_position_encoding_max_len,
|
| 312 |
+
)
|
| 313 |
+
if (temporal_position_encoding and attention_mode == "Temporal")
|
| 314 |
+
else None
|
| 315 |
+
)
|
| 316 |
+
|
| 317 |
+
def extra_repr(self):
|
| 318 |
+
return f"(Module Info) Attention_Mode: {self.attention_mode}, Is_Cross_Attention: {self.is_cross_attention}"
|
| 319 |
+
|
| 320 |
+
def set_use_memory_efficient_attention_xformers(
|
| 321 |
+
self,
|
| 322 |
+
use_memory_efficient_attention_xformers: bool,
|
| 323 |
+
attention_op: Optional[Callable] = None,
|
| 324 |
+
):
|
| 325 |
+
if use_memory_efficient_attention_xformers:
|
| 326 |
+
if not is_xformers_available():
|
| 327 |
+
raise ModuleNotFoundError(
|
| 328 |
+
(
|
| 329 |
+
"Refer to https://github.com/facebookresearch/xformers for more information on how to install"
|
| 330 |
+
" xformers"
|
| 331 |
+
),
|
| 332 |
+
name="xformers",
|
| 333 |
+
)
|
| 334 |
+
elif not torch.cuda.is_available():
|
| 335 |
+
raise ValueError(
|
| 336 |
+
"torch.cuda.is_available() should be True but is False. xformers' memory efficient attention is"
|
| 337 |
+
" only available for GPU "
|
| 338 |
+
)
|
| 339 |
+
else:
|
| 340 |
+
try:
|
| 341 |
+
# Make sure we can run the memory efficient attention
|
| 342 |
+
_ = xformers.ops.memory_efficient_attention(
|
| 343 |
+
torch.randn((1, 2, 40), device="cuda"),
|
| 344 |
+
torch.randn((1, 2, 40), device="cuda"),
|
| 345 |
+
torch.randn((1, 2, 40), device="cuda"),
|
| 346 |
+
)
|
| 347 |
+
except Exception as e:
|
| 348 |
+
raise e
|
| 349 |
+
|
| 350 |
+
# XFormersAttnProcessor corrupts video generation and work with Pytorch 1.13.
|
| 351 |
+
# Pytorch 2.0.1 AttnProcessor works the same as XFormersAttnProcessor in Pytorch 1.13.
|
| 352 |
+
# You don't need XFormersAttnProcessor here.
|
| 353 |
+
# processor = XFormersAttnProcessor(
|
| 354 |
+
# attention_op=attention_op,
|
| 355 |
+
# )
|
| 356 |
+
processor = AttnProcessor()
|
| 357 |
+
else:
|
| 358 |
+
processor = AttnProcessor()
|
| 359 |
+
|
| 360 |
+
self.set_processor(processor)
|
| 361 |
+
|
| 362 |
+
def forward(
|
| 363 |
+
self,
|
| 364 |
+
hidden_states,
|
| 365 |
+
encoder_hidden_states=None,
|
| 366 |
+
attention_mask=None,
|
| 367 |
+
video_length=None,
|
| 368 |
+
**cross_attention_kwargs,
|
| 369 |
+
):
|
| 370 |
+
if self.attention_mode == "Temporal":
|
| 371 |
+
d = hidden_states.shape[1] # d means HxW
|
| 372 |
+
hidden_states = rearrange(
|
| 373 |
+
hidden_states, "(b f) d c -> (b d) f c", f=video_length
|
| 374 |
+
)
|
| 375 |
+
|
| 376 |
+
if self.pos_encoder is not None:
|
| 377 |
+
hidden_states = self.pos_encoder(hidden_states)
|
| 378 |
+
|
| 379 |
+
encoder_hidden_states = (
|
| 380 |
+
repeat(encoder_hidden_states, "b n c -> (b d) n c", d=d)
|
| 381 |
+
if encoder_hidden_states is not None
|
| 382 |
+
else encoder_hidden_states
|
| 383 |
+
)
|
| 384 |
+
|
| 385 |
+
else:
|
| 386 |
+
raise NotImplementedError
|
| 387 |
+
|
| 388 |
+
hidden_states = self.processor(
|
| 389 |
+
self,
|
| 390 |
+
hidden_states,
|
| 391 |
+
encoder_hidden_states=encoder_hidden_states,
|
| 392 |
+
attention_mask=attention_mask,
|
| 393 |
+
**cross_attention_kwargs,
|
| 394 |
+
)
|
| 395 |
+
|
| 396 |
+
if self.attention_mode == "Temporal":
|
| 397 |
+
hidden_states = rearrange(hidden_states, "(b d) f c -> (b f) d c", d=d)
|
| 398 |
+
|
| 399 |
+
return hidden_states
|
genwarp/models/mutual_self_attention.py
ADDED
|
@@ -0,0 +1,420 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# This code is adapted from below and then modified.
|
| 2 |
+
# -----------------------------------------------------------------------------
|
| 3 |
+
# Moore-AnimateAnyone
|
| 4 |
+
# Apache License, Version 2.0
|
| 5 |
+
# Copyright @2023-2024 Moore Threads Technology Co., Ltd.
|
| 6 |
+
# https://github.com/MooreThreads/Moore-AnimateAnyone
|
| 7 |
+
# -----------------------------------------------------------------------------
|
| 8 |
+
# magic-animate
|
| 9 |
+
# BSD 3-Clause License
|
| 10 |
+
# Copyright (c) Bytedance Inc.
|
| 11 |
+
# https://github.com/magic-research/magic-animate
|
| 12 |
+
# ==============================================================================
|
| 13 |
+
|
| 14 |
+
# Adapted from https://github.com/magic-research/magic-animate/blob/main/magicanimate/models/mutual_self_attention.py
|
| 15 |
+
from typing import Any, Dict, Optional
|
| 16 |
+
import math
|
| 17 |
+
|
| 18 |
+
import torch
|
| 19 |
+
from einops import rearrange
|
| 20 |
+
|
| 21 |
+
from .attention import TemporalBasicTransformerBlock
|
| 22 |
+
from .attention import BasicTransformerBlock
|
| 23 |
+
|
| 24 |
+
def torch_dfs(model: torch.nn.Module):
|
| 25 |
+
result = [model]
|
| 26 |
+
for child in model.children():
|
| 27 |
+
result += torch_dfs(child)
|
| 28 |
+
return result
|
| 29 |
+
|
| 30 |
+
|
| 31 |
+
class ReferenceAttentionControl:
|
| 32 |
+
def __init__(
|
| 33 |
+
self,
|
| 34 |
+
unet,
|
| 35 |
+
mode="write",
|
| 36 |
+
do_classifier_free_guidance=False,
|
| 37 |
+
attention_auto_machine_weight=float("inf"),
|
| 38 |
+
gn_auto_machine_weight=1.0,
|
| 39 |
+
style_fidelity=1.0,
|
| 40 |
+
reference_attn=True,
|
| 41 |
+
reference_adain=False,
|
| 42 |
+
fusion_blocks="midup",
|
| 43 |
+
batch_size=1,
|
| 44 |
+
feature_fusion_type=None,
|
| 45 |
+
) -> None:
|
| 46 |
+
self.unet = unet
|
| 47 |
+
assert mode in ["read", "write"]
|
| 48 |
+
assert fusion_blocks in ["midup", "full"]
|
| 49 |
+
self.reference_attn = reference_attn
|
| 50 |
+
self.reference_adain = reference_adain
|
| 51 |
+
self.fusion_blocks = fusion_blocks
|
| 52 |
+
self.feature_fusion_type = feature_fusion_type
|
| 53 |
+
|
| 54 |
+
self.mode = mode
|
| 55 |
+
self.do_classifier_free_guidance = do_classifier_free_guidance
|
| 56 |
+
self.attention_auto_machine_weight = attention_auto_machine_weight
|
| 57 |
+
self.gn_auto_machine_weight = gn_auto_machine_weight
|
| 58 |
+
self.style_fidelity = style_fidelity
|
| 59 |
+
self.batch_size = batch_size
|
| 60 |
+
|
| 61 |
+
self.register_reference_hooks(
|
| 62 |
+
mode,
|
| 63 |
+
do_classifier_free_guidance,
|
| 64 |
+
attention_auto_machine_weight,
|
| 65 |
+
gn_auto_machine_weight,
|
| 66 |
+
style_fidelity,
|
| 67 |
+
reference_attn,
|
| 68 |
+
reference_adain,
|
| 69 |
+
fusion_blocks,
|
| 70 |
+
batch_size=batch_size,
|
| 71 |
+
)
|
| 72 |
+
|
| 73 |
+
def rehook(self):
|
| 74 |
+
self.register_reference_hooks(
|
| 75 |
+
self.mode,
|
| 76 |
+
self.do_classifier_free_guidance,
|
| 77 |
+
self.attention_auto_machine_weight,
|
| 78 |
+
self.gn_auto_machine_weight,
|
| 79 |
+
self.style_fidelity,
|
| 80 |
+
self.reference_attn,
|
| 81 |
+
self.reference_adain,
|
| 82 |
+
self.fusion_blocks,
|
| 83 |
+
self.batch_size,
|
| 84 |
+
)
|
| 85 |
+
|
| 86 |
+
def register_reference_hooks(
|
| 87 |
+
self,
|
| 88 |
+
mode,
|
| 89 |
+
do_classifier_free_guidance,
|
| 90 |
+
attention_auto_machine_weight,
|
| 91 |
+
gn_auto_machine_weight,
|
| 92 |
+
style_fidelity,
|
| 93 |
+
reference_attn,
|
| 94 |
+
reference_adain,
|
| 95 |
+
dtype=torch.float16,
|
| 96 |
+
batch_size=1,
|
| 97 |
+
num_images_per_prompt=1,
|
| 98 |
+
device=torch.device("cpu"),
|
| 99 |
+
fusion_blocks="midup",
|
| 100 |
+
):
|
| 101 |
+
do_classifier_free_guidance = do_classifier_free_guidance
|
| 102 |
+
attention_auto_machine_weight = attention_auto_machine_weight
|
| 103 |
+
gn_auto_machine_weight = gn_auto_machine_weight
|
| 104 |
+
style_fidelity = style_fidelity
|
| 105 |
+
reference_attn = reference_attn
|
| 106 |
+
reference_adain = reference_adain
|
| 107 |
+
fusion_blocks = fusion_blocks
|
| 108 |
+
num_images_per_prompt = num_images_per_prompt
|
| 109 |
+
dtype = dtype
|
| 110 |
+
feature_fusion_type = self.feature_fusion_type
|
| 111 |
+
|
| 112 |
+
if do_classifier_free_guidance:
|
| 113 |
+
uc_mask = (
|
| 114 |
+
torch.Tensor(
|
| 115 |
+
[1] * batch_size * num_images_per_prompt * 16
|
| 116 |
+
+ [0] * batch_size * num_images_per_prompt * 16
|
| 117 |
+
)
|
| 118 |
+
.to(device)
|
| 119 |
+
.bool()
|
| 120 |
+
)
|
| 121 |
+
else:
|
| 122 |
+
uc_mask = (
|
| 123 |
+
torch.Tensor([0] * batch_size * num_images_per_prompt * 2)
|
| 124 |
+
.to(device)
|
| 125 |
+
.bool()
|
| 126 |
+
)
|
| 127 |
+
|
| 128 |
+
def hacked_basic_transformer_inner_forward(
|
| 129 |
+
self,
|
| 130 |
+
hidden_states: torch.FloatTensor,
|
| 131 |
+
attention_mask: Optional[torch.FloatTensor] = None,
|
| 132 |
+
encoder_hidden_states: Optional[torch.FloatTensor] = None,
|
| 133 |
+
encoder_attention_mask: Optional[torch.FloatTensor] = None,
|
| 134 |
+
timestep: Optional[torch.LongTensor] = None,
|
| 135 |
+
cross_attention_kwargs: Dict[str, Any] = None,
|
| 136 |
+
class_labels: Optional[torch.LongTensor] = None,
|
| 137 |
+
video_length=None,
|
| 138 |
+
):
|
| 139 |
+
if self.use_ada_layer_norm: # False
|
| 140 |
+
norm_hidden_states = self.norm1(hidden_states, timestep)
|
| 141 |
+
elif self.use_ada_layer_norm_zero:
|
| 142 |
+
(
|
| 143 |
+
norm_hidden_states,
|
| 144 |
+
gate_msa,
|
| 145 |
+
shift_mlp,
|
| 146 |
+
scale_mlp,
|
| 147 |
+
gate_mlp,
|
| 148 |
+
) = self.norm1(
|
| 149 |
+
hidden_states,
|
| 150 |
+
timestep,
|
| 151 |
+
class_labels,
|
| 152 |
+
hidden_dtype=hidden_states.dtype,
|
| 153 |
+
)
|
| 154 |
+
else:
|
| 155 |
+
norm_hidden_states = self.norm1(hidden_states)
|
| 156 |
+
|
| 157 |
+
cross_attention_kwargs = (
|
| 158 |
+
cross_attention_kwargs if cross_attention_kwargs is not None else {}
|
| 159 |
+
)
|
| 160 |
+
if self.only_cross_attention:
|
| 161 |
+
attn_output = self.attn1(
|
| 162 |
+
norm_hidden_states,
|
| 163 |
+
encoder_hidden_states=encoder_hidden_states
|
| 164 |
+
if self.only_cross_attention
|
| 165 |
+
else None,
|
| 166 |
+
attention_mask=attention_mask,
|
| 167 |
+
**cross_attention_kwargs,
|
| 168 |
+
)
|
| 169 |
+
else:
|
| 170 |
+
if mode == "write":
|
| 171 |
+
self.bank.append(norm_hidden_states.clone())
|
| 172 |
+
self.bank_unnorm.append(hidden_states.clone())
|
| 173 |
+
attn_output = self.attn1(
|
| 174 |
+
norm_hidden_states,
|
| 175 |
+
encoder_hidden_states=encoder_hidden_states
|
| 176 |
+
if self.only_cross_attention
|
| 177 |
+
else None,
|
| 178 |
+
attention_mask=attention_mask,
|
| 179 |
+
**cross_attention_kwargs,
|
| 180 |
+
)
|
| 181 |
+
if mode == "read":
|
| 182 |
+
bank_fea = [
|
| 183 |
+
rearrange(
|
| 184 |
+
d.unsqueeze(1).repeat(1, video_length, 1, 1),
|
| 185 |
+
"b t l c -> (b t) l c",
|
| 186 |
+
)
|
| 187 |
+
for d in self.bank
|
| 188 |
+
]
|
| 189 |
+
|
| 190 |
+
bank_fea_unnorm = [
|
| 191 |
+
rearrange(
|
| 192 |
+
d.unsqueeze(1).repeat(1, video_length, 1, 1),
|
| 193 |
+
"b t l c -> (b t) l c",
|
| 194 |
+
)
|
| 195 |
+
for d in self.bank_unnorm
|
| 196 |
+
]
|
| 197 |
+
|
| 198 |
+
|
| 199 |
+
modify_norm_hidden_states = torch.cat(
|
| 200 |
+
[norm_hidden_states] + bank_fea, dim=1
|
| 201 |
+
)
|
| 202 |
+
|
| 203 |
+
if feature_fusion_type == 'attention_full_sharing':
|
| 204 |
+
# Full sharing for ablation exp.
|
| 205 |
+
hidden_states_uc = (
|
| 206 |
+
self.attn1(
|
| 207 |
+
norm_hidden_states,
|
| 208 |
+
encoder_hidden_states=modify_norm_hidden_states,
|
| 209 |
+
attention_mask=None,
|
| 210 |
+
)
|
| 211 |
+
+ hidden_states
|
| 212 |
+
)
|
| 213 |
+
else:
|
| 214 |
+
raise ValueError("feature_fusion_type is not valid")
|
| 215 |
+
|
| 216 |
+
if do_classifier_free_guidance:
|
| 217 |
+
hidden_states_c = hidden_states_uc.clone()
|
| 218 |
+
_uc_mask = uc_mask.clone()
|
| 219 |
+
if hidden_states.shape[0] != _uc_mask.shape[0]:
|
| 220 |
+
_uc_mask = (
|
| 221 |
+
torch.Tensor(
|
| 222 |
+
[1] * (hidden_states.shape[0] // 2)
|
| 223 |
+
+ [0] * (hidden_states.shape[0] // 2)
|
| 224 |
+
)
|
| 225 |
+
.to(device)
|
| 226 |
+
.bool()
|
| 227 |
+
)
|
| 228 |
+
hidden_states_c[_uc_mask] = (
|
| 229 |
+
self.attn1(
|
| 230 |
+
norm_hidden_states[_uc_mask],
|
| 231 |
+
encoder_hidden_states=norm_hidden_states[_uc_mask],
|
| 232 |
+
attention_mask=None,
|
| 233 |
+
)
|
| 234 |
+
+ hidden_states[_uc_mask]
|
| 235 |
+
)
|
| 236 |
+
hidden_states = hidden_states_c.clone()
|
| 237 |
+
else:
|
| 238 |
+
hidden_states = hidden_states_uc
|
| 239 |
+
|
| 240 |
+
if self.attn2 is not None:
|
| 241 |
+
# Cross-Attention
|
| 242 |
+
norm_hidden_states = (
|
| 243 |
+
self.norm2(hidden_states, timestep)
|
| 244 |
+
if self.use_ada_layer_norm
|
| 245 |
+
else self.norm2(hidden_states)
|
| 246 |
+
)
|
| 247 |
+
hidden_states = (
|
| 248 |
+
self.attn2(
|
| 249 |
+
norm_hidden_states,
|
| 250 |
+
encoder_hidden_states=encoder_hidden_states,
|
| 251 |
+
attention_mask=None,
|
| 252 |
+
)
|
| 253 |
+
+ hidden_states
|
| 254 |
+
)
|
| 255 |
+
|
| 256 |
+
# Feed-forward
|
| 257 |
+
hidden_states = self.ff(self.norm3(hidden_states)) + hidden_states
|
| 258 |
+
|
| 259 |
+
# Temporal-Attention
|
| 260 |
+
if self.unet_use_temporal_attention:
|
| 261 |
+
d = hidden_states.shape[1]
|
| 262 |
+
hidden_states = rearrange(
|
| 263 |
+
hidden_states, "(b f) d c -> (b d) f c", f=video_length
|
| 264 |
+
)
|
| 265 |
+
norm_hidden_states = (
|
| 266 |
+
self.norm_temp(hidden_states, timestep)
|
| 267 |
+
if self.use_ada_layer_norm
|
| 268 |
+
else self.norm_temp(hidden_states)
|
| 269 |
+
)
|
| 270 |
+
hidden_states = (
|
| 271 |
+
self.attn_temp(norm_hidden_states) + hidden_states
|
| 272 |
+
)
|
| 273 |
+
hidden_states = rearrange(
|
| 274 |
+
hidden_states, "(b d) f c -> (b f) d c", d=d
|
| 275 |
+
)
|
| 276 |
+
|
| 277 |
+
return hidden_states
|
| 278 |
+
|
| 279 |
+
if self.use_ada_layer_norm_zero:
|
| 280 |
+
attn_output = gate_msa.unsqueeze(1) * attn_output
|
| 281 |
+
hidden_states = attn_output + hidden_states
|
| 282 |
+
|
| 283 |
+
if self.attn2 is not None:
|
| 284 |
+
norm_hidden_states = (
|
| 285 |
+
self.norm2(hidden_states, timestep)
|
| 286 |
+
if self.use_ada_layer_norm
|
| 287 |
+
else self.norm2(hidden_states)
|
| 288 |
+
)
|
| 289 |
+
|
| 290 |
+
attn_output = self.attn2(
|
| 291 |
+
norm_hidden_states,
|
| 292 |
+
encoder_hidden_states=encoder_hidden_states,
|
| 293 |
+
attention_mask=encoder_attention_mask,
|
| 294 |
+
**cross_attention_kwargs,
|
| 295 |
+
)
|
| 296 |
+
hidden_states = attn_output + hidden_states
|
| 297 |
+
|
| 298 |
+
norm_hidden_states = self.norm3(hidden_states)
|
| 299 |
+
|
| 300 |
+
if self.use_ada_layer_norm_zero:
|
| 301 |
+
norm_hidden_states = (
|
| 302 |
+
norm_hidden_states * (1 + scale_mlp[:, None]) + shift_mlp[:, None]
|
| 303 |
+
)
|
| 304 |
+
|
| 305 |
+
ff_output = self.ff(norm_hidden_states)
|
| 306 |
+
|
| 307 |
+
if self.use_ada_layer_norm_zero:
|
| 308 |
+
ff_output = gate_mlp.unsqueeze(1) * ff_output
|
| 309 |
+
|
| 310 |
+
hidden_states = ff_output + hidden_states
|
| 311 |
+
|
| 312 |
+
return hidden_states
|
| 313 |
+
|
| 314 |
+
if self.reference_attn:
|
| 315 |
+
if self.fusion_blocks == "midup":
|
| 316 |
+
attn_modules = [
|
| 317 |
+
module
|
| 318 |
+
for module in (
|
| 319 |
+
torch_dfs(self.unet.mid_block) + torch_dfs(self.unet.up_blocks)
|
| 320 |
+
)
|
| 321 |
+
if isinstance(module, BasicTransformerBlock)
|
| 322 |
+
or isinstance(module, TemporalBasicTransformerBlock)
|
| 323 |
+
]
|
| 324 |
+
elif self.fusion_blocks == "full":
|
| 325 |
+
attn_modules = [
|
| 326 |
+
module
|
| 327 |
+
for module in torch_dfs(self.unet)
|
| 328 |
+
if isinstance(module, BasicTransformerBlock)
|
| 329 |
+
or isinstance(module, TemporalBasicTransformerBlock)
|
| 330 |
+
]
|
| 331 |
+
attn_modules = sorted(
|
| 332 |
+
attn_modules, key=lambda x: -x.norm1.normalized_shape[0]
|
| 333 |
+
)
|
| 334 |
+
|
| 335 |
+
for i, module in enumerate(attn_modules):
|
| 336 |
+
module._original_inner_forward = module.forward
|
| 337 |
+
if isinstance(module, BasicTransformerBlock):
|
| 338 |
+
module.forward = hacked_basic_transformer_inner_forward.__get__(
|
| 339 |
+
module, BasicTransformerBlock
|
| 340 |
+
)
|
| 341 |
+
if isinstance(module, TemporalBasicTransformerBlock):
|
| 342 |
+
module.forward = hacked_basic_transformer_inner_forward.__get__(
|
| 343 |
+
module, TemporalBasicTransformerBlock
|
| 344 |
+
)
|
| 345 |
+
|
| 346 |
+
module.bank = []
|
| 347 |
+
module.bank_unnorm = []
|
| 348 |
+
module.correspondence = None
|
| 349 |
+
module.attn_weight = float(i) / float(len(attn_modules))
|
| 350 |
+
|
| 351 |
+
def update(self, writer, correspondence=None, dtype=torch.float16):
|
| 352 |
+
if self.reference_attn:
|
| 353 |
+
if self.fusion_blocks == "midup":
|
| 354 |
+
reader_attn_modules = [
|
| 355 |
+
module
|
| 356 |
+
for module in (
|
| 357 |
+
torch_dfs(self.unet.mid_block) + torch_dfs(self.unet.up_blocks)
|
| 358 |
+
)
|
| 359 |
+
if isinstance(module, TemporalBasicTransformerBlock)
|
| 360 |
+
]
|
| 361 |
+
writer_attn_modules = [
|
| 362 |
+
module
|
| 363 |
+
for module in (
|
| 364 |
+
torch_dfs(writer.unet.mid_block)
|
| 365 |
+
+ torch_dfs(writer.unet.up_blocks)
|
| 366 |
+
)
|
| 367 |
+
if isinstance(module, BasicTransformerBlock)
|
| 368 |
+
]
|
| 369 |
+
elif self.fusion_blocks == "full":
|
| 370 |
+
reader_attn_modules = [
|
| 371 |
+
module
|
| 372 |
+
for module in torch_dfs(self.unet)
|
| 373 |
+
if isinstance(module, TemporalBasicTransformerBlock)
|
| 374 |
+
]
|
| 375 |
+
writer_attn_modules = [
|
| 376 |
+
module
|
| 377 |
+
for module in torch_dfs(writer.unet)
|
| 378 |
+
if isinstance(module, BasicTransformerBlock)
|
| 379 |
+
]
|
| 380 |
+
reader_attn_modules = sorted(
|
| 381 |
+
reader_attn_modules, key=lambda x: -x.norm1.normalized_shape[0]
|
| 382 |
+
)
|
| 383 |
+
writer_attn_modules = sorted(
|
| 384 |
+
writer_attn_modules, key=lambda x: -x.norm1.normalized_shape[0]
|
| 385 |
+
)
|
| 386 |
+
for r, w in zip(reader_attn_modules, writer_attn_modules):
|
| 387 |
+
r.bank = [v.clone().to(dtype) for v in w.bank]
|
| 388 |
+
r.bank_unnorm = [v.clone().to(dtype) for v in w.bank_unnorm]
|
| 389 |
+
if correspondence is not None:
|
| 390 |
+
r.correspondence = [correspondence]
|
| 391 |
+
else:
|
| 392 |
+
r.correspondence = None
|
| 393 |
+
|
| 394 |
+
def clear(self):
|
| 395 |
+
if self.reference_attn:
|
| 396 |
+
if self.fusion_blocks == "midup":
|
| 397 |
+
reader_attn_modules = [
|
| 398 |
+
module
|
| 399 |
+
for module in (
|
| 400 |
+
torch_dfs(self.unet.mid_block) + torch_dfs(self.unet.up_blocks)
|
| 401 |
+
)
|
| 402 |
+
if isinstance(module, BasicTransformerBlock)
|
| 403 |
+
or isinstance(module, TemporalBasicTransformerBlock)
|
| 404 |
+
]
|
| 405 |
+
elif self.fusion_blocks == "full":
|
| 406 |
+
reader_attn_modules = [
|
| 407 |
+
module
|
| 408 |
+
for module in torch_dfs(self.unet)
|
| 409 |
+
if isinstance(module, BasicTransformerBlock)
|
| 410 |
+
or isinstance(module, TemporalBasicTransformerBlock)
|
| 411 |
+
]
|
| 412 |
+
reader_attn_modules = sorted(
|
| 413 |
+
reader_attn_modules, key=lambda x: -x.norm1.normalized_shape[0]
|
| 414 |
+
)
|
| 415 |
+
for r in reader_attn_modules:
|
| 416 |
+
r.bank.clear()
|
| 417 |
+
r.bank_unnorm.clear()
|
| 418 |
+
if r.correspondence is not None:
|
| 419 |
+
r.correspondence.clear()
|
| 420 |
+
r.correspondence = None
|
genwarp/models/pose_guider.py
ADDED
|
@@ -0,0 +1,63 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# This code is adapted from below and then modified.
|
| 2 |
+
# -----------------------------------------------------------------------------
|
| 3 |
+
# Moore-AnimateAnyone
|
| 4 |
+
# Apache License, Version 2.0
|
| 5 |
+
# Copyright @2023-2024 Moore Threads Technology Co., Ltd.
|
| 6 |
+
# https://github.com/MooreThreads/Moore-AnimateAnyone
|
| 7 |
+
# ==============================================================================
|
| 8 |
+
|
| 9 |
+
from typing import Tuple
|
| 10 |
+
|
| 11 |
+
import torch.nn as nn
|
| 12 |
+
import torch.nn.functional as F
|
| 13 |
+
from diffusers.models.modeling_utils import ModelMixin
|
| 14 |
+
|
| 15 |
+
from .motion_module import zero_module
|
| 16 |
+
from .resnet import InflatedConv3d
|
| 17 |
+
|
| 18 |
+
class PoseGuider(ModelMixin):
|
| 19 |
+
def __init__(
|
| 20 |
+
self,
|
| 21 |
+
conditioning_embedding_channels: int,
|
| 22 |
+
conditioning_channels: int = 3,
|
| 23 |
+
block_out_channels: Tuple[int] = (16, 32, 64, 128),
|
| 24 |
+
):
|
| 25 |
+
super().__init__()
|
| 26 |
+
self.conv_in = InflatedConv3d(
|
| 27 |
+
conditioning_channels, block_out_channels[0], kernel_size=3, padding=1
|
| 28 |
+
)
|
| 29 |
+
|
| 30 |
+
self.blocks = nn.ModuleList([])
|
| 31 |
+
|
| 32 |
+
for i in range(len(block_out_channels) - 1):
|
| 33 |
+
channel_in = block_out_channels[i]
|
| 34 |
+
channel_out = block_out_channels[i + 1]
|
| 35 |
+
self.blocks.append(
|
| 36 |
+
InflatedConv3d(channel_in, channel_in, kernel_size=3, padding=1)
|
| 37 |
+
)
|
| 38 |
+
self.blocks.append(
|
| 39 |
+
InflatedConv3d(
|
| 40 |
+
channel_in, channel_out, kernel_size=3, padding=1, stride=2
|
| 41 |
+
)
|
| 42 |
+
)
|
| 43 |
+
|
| 44 |
+
self.conv_out = zero_module(
|
| 45 |
+
InflatedConv3d(
|
| 46 |
+
block_out_channels[-1],
|
| 47 |
+
conditioning_embedding_channels,
|
| 48 |
+
kernel_size=3,
|
| 49 |
+
padding=1,
|
| 50 |
+
)
|
| 51 |
+
)
|
| 52 |
+
|
| 53 |
+
def forward(self, conditioning):
|
| 54 |
+
embedding = self.conv_in(conditioning)
|
| 55 |
+
embedding = F.silu(embedding)
|
| 56 |
+
|
| 57 |
+
for block in self.blocks:
|
| 58 |
+
embedding = block(embedding)
|
| 59 |
+
embedding = F.silu(embedding)
|
| 60 |
+
|
| 61 |
+
embedding = self.conv_out(embedding)
|
| 62 |
+
|
| 63 |
+
return embedding
|
genwarp/models/resnet.py
ADDED
|
@@ -0,0 +1,265 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# This code is adapted from below.
|
| 2 |
+
# -----------------------------------------------------------------------------
|
| 3 |
+
# Moore-AnimateAnyone
|
| 4 |
+
# Apache License, Version 2.0
|
| 5 |
+
# Copyright @2023-2024 Moore Threads Technology Co., Ltd.
|
| 6 |
+
# https://github.com/MooreThreads/Moore-AnimateAnyone
|
| 7 |
+
# -----------------------------------------------------------------------------
|
| 8 |
+
# Diffusers
|
| 9 |
+
# Apache License, Version 2.0
|
| 10 |
+
# Copyright (c) Hugging Face Inc.
|
| 11 |
+
# https://github.com/huggingface/diffusers
|
| 12 |
+
# ==============================================================================
|
| 13 |
+
|
| 14 |
+
# Adapted from https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/resnet.py
|
| 15 |
+
|
| 16 |
+
import torch
|
| 17 |
+
import torch.nn as nn
|
| 18 |
+
import torch.nn.functional as F
|
| 19 |
+
from einops import rearrange
|
| 20 |
+
|
| 21 |
+
|
| 22 |
+
class InflatedConv3d(nn.Conv2d):
|
| 23 |
+
def forward(self, x):
|
| 24 |
+
video_length = x.shape[2]
|
| 25 |
+
|
| 26 |
+
x = rearrange(x, "b c f h w -> (b f) c h w")
|
| 27 |
+
x = super().forward(x)
|
| 28 |
+
x = rearrange(x, "(b f) c h w -> b c f h w", f=video_length)
|
| 29 |
+
|
| 30 |
+
return x
|
| 31 |
+
|
| 32 |
+
|
| 33 |
+
class InflatedGroupNorm(nn.GroupNorm):
|
| 34 |
+
def forward(self, x):
|
| 35 |
+
video_length = x.shape[2]
|
| 36 |
+
|
| 37 |
+
x = rearrange(x, "b c f h w -> (b f) c h w")
|
| 38 |
+
x = super().forward(x)
|
| 39 |
+
x = rearrange(x, "(b f) c h w -> b c f h w", f=video_length)
|
| 40 |
+
|
| 41 |
+
return x
|
| 42 |
+
|
| 43 |
+
|
| 44 |
+
class Upsample3D(nn.Module):
|
| 45 |
+
def __init__(
|
| 46 |
+
self,
|
| 47 |
+
channels,
|
| 48 |
+
use_conv=False,
|
| 49 |
+
use_conv_transpose=False,
|
| 50 |
+
out_channels=None,
|
| 51 |
+
name="conv",
|
| 52 |
+
):
|
| 53 |
+
super().__init__()
|
| 54 |
+
self.channels = channels
|
| 55 |
+
self.out_channels = out_channels or channels
|
| 56 |
+
self.use_conv = use_conv
|
| 57 |
+
self.use_conv_transpose = use_conv_transpose
|
| 58 |
+
self.name = name
|
| 59 |
+
|
| 60 |
+
conv = None
|
| 61 |
+
if use_conv_transpose:
|
| 62 |
+
raise NotImplementedError
|
| 63 |
+
elif use_conv:
|
| 64 |
+
self.conv = InflatedConv3d(self.channels, self.out_channels, 3, padding=1)
|
| 65 |
+
|
| 66 |
+
def forward(self, hidden_states, output_size=None):
|
| 67 |
+
assert hidden_states.shape[1] == self.channels
|
| 68 |
+
|
| 69 |
+
if self.use_conv_transpose:
|
| 70 |
+
raise NotImplementedError
|
| 71 |
+
|
| 72 |
+
# Cast to float32 to as 'upsample_nearest2d_out_frame' op does not support bfloat16
|
| 73 |
+
dtype = hidden_states.dtype
|
| 74 |
+
if dtype == torch.bfloat16:
|
| 75 |
+
hidden_states = hidden_states.to(torch.float32)
|
| 76 |
+
|
| 77 |
+
# upsample_nearest_nhwc fails with large batch sizes. see https://github.com/huggingface/diffusers/issues/984
|
| 78 |
+
if hidden_states.shape[0] >= 64:
|
| 79 |
+
hidden_states = hidden_states.contiguous()
|
| 80 |
+
|
| 81 |
+
# if `output_size` is passed we force the interpolation output
|
| 82 |
+
# size and do not make use of `scale_factor=2`
|
| 83 |
+
if output_size is None:
|
| 84 |
+
hidden_states = F.interpolate(
|
| 85 |
+
hidden_states, scale_factor=[1.0, 2.0, 2.0], mode="nearest"
|
| 86 |
+
)
|
| 87 |
+
else:
|
| 88 |
+
hidden_states = F.interpolate(
|
| 89 |
+
hidden_states, size=output_size, mode="nearest"
|
| 90 |
+
)
|
| 91 |
+
|
| 92 |
+
# If the input is bfloat16, we cast back to bfloat16
|
| 93 |
+
if dtype == torch.bfloat16:
|
| 94 |
+
hidden_states = hidden_states.to(dtype)
|
| 95 |
+
|
| 96 |
+
# if self.use_conv:
|
| 97 |
+
# if self.name == "conv":
|
| 98 |
+
# hidden_states = self.conv(hidden_states)
|
| 99 |
+
# else:
|
| 100 |
+
# hidden_states = self.Conv2d_0(hidden_states)
|
| 101 |
+
hidden_states = self.conv(hidden_states)
|
| 102 |
+
|
| 103 |
+
return hidden_states
|
| 104 |
+
|
| 105 |
+
|
| 106 |
+
class Downsample3D(nn.Module):
|
| 107 |
+
def __init__(
|
| 108 |
+
self, channels, use_conv=False, out_channels=None, padding=1, name="conv"
|
| 109 |
+
):
|
| 110 |
+
super().__init__()
|
| 111 |
+
self.channels = channels
|
| 112 |
+
self.out_channels = out_channels or channels
|
| 113 |
+
self.use_conv = use_conv
|
| 114 |
+
self.padding = padding
|
| 115 |
+
stride = 2
|
| 116 |
+
self.name = name
|
| 117 |
+
|
| 118 |
+
if use_conv:
|
| 119 |
+
self.conv = InflatedConv3d(
|
| 120 |
+
self.channels, self.out_channels, 3, stride=stride, padding=padding
|
| 121 |
+
)
|
| 122 |
+
else:
|
| 123 |
+
raise NotImplementedError
|
| 124 |
+
|
| 125 |
+
def forward(self, hidden_states):
|
| 126 |
+
assert hidden_states.shape[1] == self.channels
|
| 127 |
+
if self.use_conv and self.padding == 0:
|
| 128 |
+
raise NotImplementedError
|
| 129 |
+
|
| 130 |
+
assert hidden_states.shape[1] == self.channels
|
| 131 |
+
hidden_states = self.conv(hidden_states)
|
| 132 |
+
|
| 133 |
+
return hidden_states
|
| 134 |
+
|
| 135 |
+
|
| 136 |
+
class ResnetBlock3D(nn.Module):
|
| 137 |
+
def __init__(
|
| 138 |
+
self,
|
| 139 |
+
*,
|
| 140 |
+
in_channels,
|
| 141 |
+
out_channels=None,
|
| 142 |
+
conv_shortcut=False,
|
| 143 |
+
dropout=0.0,
|
| 144 |
+
temb_channels=512,
|
| 145 |
+
groups=32,
|
| 146 |
+
groups_out=None,
|
| 147 |
+
pre_norm=True,
|
| 148 |
+
eps=1e-6,
|
| 149 |
+
non_linearity="swish",
|
| 150 |
+
time_embedding_norm="default",
|
| 151 |
+
output_scale_factor=1.0,
|
| 152 |
+
use_in_shortcut=None,
|
| 153 |
+
use_inflated_groupnorm=None,
|
| 154 |
+
):
|
| 155 |
+
super().__init__()
|
| 156 |
+
self.pre_norm = pre_norm
|
| 157 |
+
self.pre_norm = True
|
| 158 |
+
self.in_channels = in_channels
|
| 159 |
+
out_channels = in_channels if out_channels is None else out_channels
|
| 160 |
+
self.out_channels = out_channels
|
| 161 |
+
self.use_conv_shortcut = conv_shortcut
|
| 162 |
+
self.time_embedding_norm = time_embedding_norm
|
| 163 |
+
self.output_scale_factor = output_scale_factor
|
| 164 |
+
|
| 165 |
+
if groups_out is None:
|
| 166 |
+
groups_out = groups
|
| 167 |
+
|
| 168 |
+
assert use_inflated_groupnorm != None
|
| 169 |
+
if use_inflated_groupnorm:
|
| 170 |
+
self.norm1 = InflatedGroupNorm(
|
| 171 |
+
num_groups=groups, num_channels=in_channels, eps=eps, affine=True
|
| 172 |
+
)
|
| 173 |
+
else:
|
| 174 |
+
self.norm1 = torch.nn.GroupNorm(
|
| 175 |
+
num_groups=groups, num_channels=in_channels, eps=eps, affine=True
|
| 176 |
+
)
|
| 177 |
+
|
| 178 |
+
self.conv1 = InflatedConv3d(
|
| 179 |
+
in_channels, out_channels, kernel_size=3, stride=1, padding=1
|
| 180 |
+
)
|
| 181 |
+
|
| 182 |
+
if temb_channels is not None:
|
| 183 |
+
if self.time_embedding_norm == "default":
|
| 184 |
+
time_emb_proj_out_channels = out_channels
|
| 185 |
+
elif self.time_embedding_norm == "scale_shift":
|
| 186 |
+
time_emb_proj_out_channels = out_channels * 2
|
| 187 |
+
else:
|
| 188 |
+
raise ValueError(
|
| 189 |
+
f"unknown time_embedding_norm : {self.time_embedding_norm} "
|
| 190 |
+
)
|
| 191 |
+
|
| 192 |
+
self.time_emb_proj = torch.nn.Linear(
|
| 193 |
+
temb_channels, time_emb_proj_out_channels
|
| 194 |
+
)
|
| 195 |
+
else:
|
| 196 |
+
self.time_emb_proj = None
|
| 197 |
+
|
| 198 |
+
if use_inflated_groupnorm:
|
| 199 |
+
self.norm2 = InflatedGroupNorm(
|
| 200 |
+
num_groups=groups_out, num_channels=out_channels, eps=eps, affine=True
|
| 201 |
+
)
|
| 202 |
+
else:
|
| 203 |
+
self.norm2 = torch.nn.GroupNorm(
|
| 204 |
+
num_groups=groups_out, num_channels=out_channels, eps=eps, affine=True
|
| 205 |
+
)
|
| 206 |
+
self.dropout = torch.nn.Dropout(dropout)
|
| 207 |
+
self.conv2 = InflatedConv3d(
|
| 208 |
+
out_channels, out_channels, kernel_size=3, stride=1, padding=1
|
| 209 |
+
)
|
| 210 |
+
|
| 211 |
+
if non_linearity == "swish":
|
| 212 |
+
self.nonlinearity = lambda x: F.silu(x)
|
| 213 |
+
elif non_linearity == "mish":
|
| 214 |
+
self.nonlinearity = Mish()
|
| 215 |
+
elif non_linearity == "silu":
|
| 216 |
+
self.nonlinearity = nn.SiLU()
|
| 217 |
+
|
| 218 |
+
self.use_in_shortcut = (
|
| 219 |
+
self.in_channels != self.out_channels
|
| 220 |
+
if use_in_shortcut is None
|
| 221 |
+
else use_in_shortcut
|
| 222 |
+
)
|
| 223 |
+
|
| 224 |
+
self.conv_shortcut = None
|
| 225 |
+
if self.use_in_shortcut:
|
| 226 |
+
self.conv_shortcut = InflatedConv3d(
|
| 227 |
+
in_channels, out_channels, kernel_size=1, stride=1, padding=0
|
| 228 |
+
)
|
| 229 |
+
|
| 230 |
+
def forward(self, input_tensor, temb):
|
| 231 |
+
hidden_states = input_tensor
|
| 232 |
+
|
| 233 |
+
hidden_states = self.norm1(hidden_states)
|
| 234 |
+
hidden_states = self.nonlinearity(hidden_states)
|
| 235 |
+
|
| 236 |
+
hidden_states = self.conv1(hidden_states)
|
| 237 |
+
|
| 238 |
+
if temb is not None:
|
| 239 |
+
temb = self.time_emb_proj(self.nonlinearity(temb))[:, :, None, None, None]
|
| 240 |
+
|
| 241 |
+
if temb is not None and self.time_embedding_norm == "default":
|
| 242 |
+
hidden_states = hidden_states + temb
|
| 243 |
+
|
| 244 |
+
hidden_states = self.norm2(hidden_states)
|
| 245 |
+
|
| 246 |
+
if temb is not None and self.time_embedding_norm == "scale_shift":
|
| 247 |
+
scale, shift = torch.chunk(temb, 2, dim=1)
|
| 248 |
+
hidden_states = hidden_states * (1 + scale) + shift
|
| 249 |
+
|
| 250 |
+
hidden_states = self.nonlinearity(hidden_states)
|
| 251 |
+
|
| 252 |
+
hidden_states = self.dropout(hidden_states)
|
| 253 |
+
hidden_states = self.conv2(hidden_states)
|
| 254 |
+
|
| 255 |
+
if self.conv_shortcut is not None:
|
| 256 |
+
input_tensor = self.conv_shortcut(input_tensor)
|
| 257 |
+
|
| 258 |
+
output_tensor = (input_tensor + hidden_states) / self.output_scale_factor
|
| 259 |
+
|
| 260 |
+
return output_tensor
|
| 261 |
+
|
| 262 |
+
|
| 263 |
+
class Mish(torch.nn.Module):
|
| 264 |
+
def forward(self, hidden_states):
|
| 265 |
+
return hidden_states * torch.tanh(torch.nn.functional.softplus(hidden_states))
|
genwarp/models/transformer_2d.py
ADDED
|
@@ -0,0 +1,409 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# This code is adapted from below and then modified.
|
| 2 |
+
# -----------------------------------------------------------------------------
|
| 3 |
+
# Moore-AnimateAnyone
|
| 4 |
+
# Apache License, Version 2.0
|
| 5 |
+
# Copyright @2023-2024 Moore Threads Technology Co., Ltd.
|
| 6 |
+
# https://github.com/MooreThreads/Moore-AnimateAnyone
|
| 7 |
+
# -----------------------------------------------------------------------------
|
| 8 |
+
# Diffusers
|
| 9 |
+
# Apache License, Version 2.0
|
| 10 |
+
# Copyright (c) Hugging Face Inc.
|
| 11 |
+
# https://github.com/huggingface/diffusers
|
| 12 |
+
# ==============================================================================
|
| 13 |
+
|
| 14 |
+
# Adapted from https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/transformer_2d.py
|
| 15 |
+
from dataclasses import dataclass
|
| 16 |
+
from typing import Any, Dict, Optional
|
| 17 |
+
|
| 18 |
+
import torch
|
| 19 |
+
from diffusers.configuration_utils import ConfigMixin, register_to_config
|
| 20 |
+
# from diffusers.models.embeddings import CaptionProjection
|
| 21 |
+
from diffusers.models.lora import LoRACompatibleConv, LoRACompatibleLinear
|
| 22 |
+
from diffusers.models.modeling_utils import ModelMixin
|
| 23 |
+
from diffusers.models.normalization import AdaLayerNormSingle
|
| 24 |
+
from diffusers.utils import USE_PEFT_BACKEND, BaseOutput, deprecate, is_torch_version
|
| 25 |
+
from torch import nn
|
| 26 |
+
|
| 27 |
+
from .attention import BasicTransformerBlock
|
| 28 |
+
|
| 29 |
+
|
| 30 |
+
@dataclass
|
| 31 |
+
class Transformer2DModelOutput(BaseOutput):
|
| 32 |
+
"""
|
| 33 |
+
The output of [`Transformer2DModel`].
|
| 34 |
+
|
| 35 |
+
Args:
|
| 36 |
+
sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)` or `(batch size, num_vector_embeds - 1, num_latent_pixels)` if [`Transformer2DModel`] is discrete):
|
| 37 |
+
The hidden states output conditioned on the `encoder_hidden_states` input. If discrete, returns probability
|
| 38 |
+
distributions for the unnoised latent pixels.
|
| 39 |
+
"""
|
| 40 |
+
|
| 41 |
+
sample: torch.FloatTensor
|
| 42 |
+
ref_feature: torch.FloatTensor
|
| 43 |
+
|
| 44 |
+
|
| 45 |
+
class Transformer2DModel(ModelMixin, ConfigMixin):
|
| 46 |
+
"""
|
| 47 |
+
A 2D Transformer model for image-like data.
|
| 48 |
+
|
| 49 |
+
Parameters:
|
| 50 |
+
num_attention_heads (`int`, *optional*, defaults to 16): The number of heads to use for multi-head attention.
|
| 51 |
+
attention_head_dim (`int`, *optional*, defaults to 88): The number of channels in each head.
|
| 52 |
+
in_channels (`int`, *optional*):
|
| 53 |
+
The number of channels in the input and output (specify if the input is **continuous**).
|
| 54 |
+
num_layers (`int`, *optional*, defaults to 1): The number of layers of Transformer blocks to use.
|
| 55 |
+
dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
|
| 56 |
+
cross_attention_dim (`int`, *optional*): The number of `encoder_hidden_states` dimensions to use.
|
| 57 |
+
sample_size (`int`, *optional*): The width of the latent images (specify if the input is **discrete**).
|
| 58 |
+
This is fixed during training since it is used to learn a number of position embeddings.
|
| 59 |
+
num_vector_embeds (`int`, *optional*):
|
| 60 |
+
The number of classes of the vector embeddings of the latent pixels (specify if the input is **discrete**).
|
| 61 |
+
Includes the class for the masked latent pixel.
|
| 62 |
+
activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to use in feed-forward.
|
| 63 |
+
num_embeds_ada_norm ( `int`, *optional*):
|
| 64 |
+
The number of diffusion steps used during training. Pass if at least one of the norm_layers is
|
| 65 |
+
`AdaLayerNorm`. This is fixed during training since it is used to learn a number of embeddings that are
|
| 66 |
+
added to the hidden states.
|
| 67 |
+
|
| 68 |
+
During inference, you can denoise for up to but not more steps than `num_embeds_ada_norm`.
|
| 69 |
+
attention_bias (`bool`, *optional*):
|
| 70 |
+
Configure if the `TransformerBlocks` attention should contain a bias parameter.
|
| 71 |
+
"""
|
| 72 |
+
|
| 73 |
+
_supports_gradient_checkpointing = True
|
| 74 |
+
|
| 75 |
+
@register_to_config
|
| 76 |
+
def __init__(
|
| 77 |
+
self,
|
| 78 |
+
num_attention_heads: int = 16,
|
| 79 |
+
attention_head_dim: int = 88,
|
| 80 |
+
in_channels: Optional[int] = None,
|
| 81 |
+
out_channels: Optional[int] = None,
|
| 82 |
+
num_layers: int = 1,
|
| 83 |
+
dropout: float = 0.0,
|
| 84 |
+
norm_num_groups: int = 32,
|
| 85 |
+
cross_attention_dim: Optional[int] = None,
|
| 86 |
+
attention_bias: bool = False,
|
| 87 |
+
sample_size: Optional[int] = None,
|
| 88 |
+
num_vector_embeds: Optional[int] = None,
|
| 89 |
+
patch_size: Optional[int] = None,
|
| 90 |
+
activation_fn: str = "geglu",
|
| 91 |
+
num_embeds_ada_norm: Optional[int] = None,
|
| 92 |
+
use_linear_projection: bool = False,
|
| 93 |
+
only_cross_attention: bool = False,
|
| 94 |
+
double_self_attention: bool = False,
|
| 95 |
+
upcast_attention: bool = False,
|
| 96 |
+
norm_type: str = "layer_norm",
|
| 97 |
+
norm_elementwise_affine: bool = True,
|
| 98 |
+
norm_eps: float = 1e-5,
|
| 99 |
+
attention_type: str = "default",
|
| 100 |
+
caption_channels: int = None,
|
| 101 |
+
):
|
| 102 |
+
super().__init__()
|
| 103 |
+
self.use_linear_projection = use_linear_projection
|
| 104 |
+
self.num_attention_heads = num_attention_heads
|
| 105 |
+
self.attention_head_dim = attention_head_dim
|
| 106 |
+
inner_dim = num_attention_heads * attention_head_dim
|
| 107 |
+
|
| 108 |
+
conv_cls = nn.Conv2d if USE_PEFT_BACKEND else LoRACompatibleConv
|
| 109 |
+
linear_cls = nn.Linear if USE_PEFT_BACKEND else LoRACompatibleLinear
|
| 110 |
+
|
| 111 |
+
# 1. Transformer2DModel can process both standard continuous images of shape `(batch_size, num_channels, width, height)` as well as quantized image embeddings of shape `(batch_size, num_image_vectors)`
|
| 112 |
+
# Define whether input is continuous or discrete depending on configuration
|
| 113 |
+
self.is_input_continuous = (in_channels is not None) and (patch_size is None)
|
| 114 |
+
self.is_input_vectorized = num_vector_embeds is not None
|
| 115 |
+
self.is_input_patches = in_channels is not None and patch_size is not None
|
| 116 |
+
|
| 117 |
+
if norm_type == "layer_norm" and num_embeds_ada_norm is not None:
|
| 118 |
+
deprecation_message = (
|
| 119 |
+
f"The configuration file of this model: {self.__class__} is outdated. `norm_type` is either not set or"
|
| 120 |
+
" incorrectly set to `'layer_norm'`.Make sure to set `norm_type` to `'ada_norm'` in the config."
|
| 121 |
+
" Please make sure to update the config accordingly as leaving `norm_type` might led to incorrect"
|
| 122 |
+
" results in future versions. If you have downloaded this checkpoint from the Hugging Face Hub, it"
|
| 123 |
+
" would be very nice if you could open a Pull request for the `transformer/config.json` file"
|
| 124 |
+
)
|
| 125 |
+
deprecate(
|
| 126 |
+
"norm_type!=num_embeds_ada_norm",
|
| 127 |
+
"1.0.0",
|
| 128 |
+
deprecation_message,
|
| 129 |
+
standard_warn=False,
|
| 130 |
+
)
|
| 131 |
+
norm_type = "ada_norm"
|
| 132 |
+
|
| 133 |
+
if self.is_input_continuous and self.is_input_vectorized:
|
| 134 |
+
raise ValueError(
|
| 135 |
+
f"Cannot define both `in_channels`: {in_channels} and `num_vector_embeds`: {num_vector_embeds}. Make"
|
| 136 |
+
" sure that either `in_channels` or `num_vector_embeds` is None."
|
| 137 |
+
)
|
| 138 |
+
elif self.is_input_vectorized and self.is_input_patches:
|
| 139 |
+
raise ValueError(
|
| 140 |
+
f"Cannot define both `num_vector_embeds`: {num_vector_embeds} and `patch_size`: {patch_size}. Make"
|
| 141 |
+
" sure that either `num_vector_embeds` or `num_patches` is None."
|
| 142 |
+
)
|
| 143 |
+
elif (
|
| 144 |
+
not self.is_input_continuous
|
| 145 |
+
and not self.is_input_vectorized
|
| 146 |
+
and not self.is_input_patches
|
| 147 |
+
):
|
| 148 |
+
raise ValueError(
|
| 149 |
+
f"Has to define `in_channels`: {in_channels}, `num_vector_embeds`: {num_vector_embeds}, or patch_size:"
|
| 150 |
+
f" {patch_size}. Make sure that `in_channels`, `num_vector_embeds` or `num_patches` is not None."
|
| 151 |
+
)
|
| 152 |
+
|
| 153 |
+
# 2. Define input layers
|
| 154 |
+
self.in_channels = in_channels
|
| 155 |
+
|
| 156 |
+
self.norm = torch.nn.GroupNorm(
|
| 157 |
+
num_groups=norm_num_groups,
|
| 158 |
+
num_channels=in_channels,
|
| 159 |
+
eps=1e-6,
|
| 160 |
+
affine=True,
|
| 161 |
+
)
|
| 162 |
+
if use_linear_projection:
|
| 163 |
+
self.proj_in = linear_cls(in_channels, inner_dim)
|
| 164 |
+
else:
|
| 165 |
+
self.proj_in = conv_cls(
|
| 166 |
+
in_channels, inner_dim, kernel_size=1, stride=1, padding=0
|
| 167 |
+
)
|
| 168 |
+
|
| 169 |
+
# 3. Define transformers blocks
|
| 170 |
+
self.transformer_blocks = nn.ModuleList(
|
| 171 |
+
[
|
| 172 |
+
BasicTransformerBlock(
|
| 173 |
+
inner_dim,
|
| 174 |
+
num_attention_heads,
|
| 175 |
+
attention_head_dim,
|
| 176 |
+
dropout=dropout,
|
| 177 |
+
cross_attention_dim=cross_attention_dim,
|
| 178 |
+
activation_fn=activation_fn,
|
| 179 |
+
num_embeds_ada_norm=num_embeds_ada_norm,
|
| 180 |
+
attention_bias=attention_bias,
|
| 181 |
+
only_cross_attention=only_cross_attention,
|
| 182 |
+
double_self_attention=double_self_attention,
|
| 183 |
+
upcast_attention=upcast_attention,
|
| 184 |
+
norm_type=norm_type,
|
| 185 |
+
norm_elementwise_affine=norm_elementwise_affine,
|
| 186 |
+
norm_eps=norm_eps,
|
| 187 |
+
attention_type=attention_type,
|
| 188 |
+
)
|
| 189 |
+
for d in range(num_layers)
|
| 190 |
+
]
|
| 191 |
+
)
|
| 192 |
+
|
| 193 |
+
# 4. Define output layers
|
| 194 |
+
self.out_channels = in_channels if out_channels is None else out_channels
|
| 195 |
+
# TODO: should use out_channels for continuous projections
|
| 196 |
+
if use_linear_projection:
|
| 197 |
+
self.proj_out = linear_cls(inner_dim, in_channels)
|
| 198 |
+
else:
|
| 199 |
+
self.proj_out = conv_cls(
|
| 200 |
+
inner_dim, in_channels, kernel_size=1, stride=1, padding=0
|
| 201 |
+
)
|
| 202 |
+
|
| 203 |
+
# 5. PixArt-Alpha blocks.
|
| 204 |
+
self.adaln_single = None
|
| 205 |
+
self.use_additional_conditions = False
|
| 206 |
+
if norm_type == "ada_norm_single":
|
| 207 |
+
self.use_additional_conditions = self.config.sample_size == 128
|
| 208 |
+
# TODO(Sayak, PVP) clean this, for now we use sample size to determine whether to use
|
| 209 |
+
# additional conditions until we find better name
|
| 210 |
+
self.adaln_single = AdaLayerNormSingle(
|
| 211 |
+
inner_dim, use_additional_conditions=self.use_additional_conditions
|
| 212 |
+
)
|
| 213 |
+
|
| 214 |
+
self.caption_projection = None
|
| 215 |
+
# if caption_channels is not None:
|
| 216 |
+
# self.caption_projection = CaptionProjection(
|
| 217 |
+
# in_features=caption_channels, hidden_size=inner_dim
|
| 218 |
+
# )
|
| 219 |
+
|
| 220 |
+
self.gradient_checkpointing = False
|
| 221 |
+
|
| 222 |
+
def _set_gradient_checkpointing(self, module, value=False):
|
| 223 |
+
if hasattr(module, "gradient_checkpointing"):
|
| 224 |
+
module.gradient_checkpointing = value
|
| 225 |
+
|
| 226 |
+
def forward(
|
| 227 |
+
self,
|
| 228 |
+
hidden_states: torch.Tensor,
|
| 229 |
+
encoder_hidden_states: Optional[torch.Tensor] = None,
|
| 230 |
+
timestep: Optional[torch.LongTensor] = None,
|
| 231 |
+
added_cond_kwargs: Dict[str, torch.Tensor] = None,
|
| 232 |
+
class_labels: Optional[torch.LongTensor] = None,
|
| 233 |
+
cross_attention_kwargs: Dict[str, Any] = None,
|
| 234 |
+
attention_mask: Optional[torch.Tensor] = None,
|
| 235 |
+
encoder_attention_mask: Optional[torch.Tensor] = None,
|
| 236 |
+
return_dict: bool = True,
|
| 237 |
+
):
|
| 238 |
+
"""
|
| 239 |
+
The [`Transformer2DModel`] forward method.
|
| 240 |
+
|
| 241 |
+
Args:
|
| 242 |
+
hidden_states (`torch.LongTensor` of shape `(batch size, num latent pixels)` if discrete, `torch.FloatTensor` of shape `(batch size, channel, height, width)` if continuous):
|
| 243 |
+
Input `hidden_states`.
|
| 244 |
+
encoder_hidden_states ( `torch.FloatTensor` of shape `(batch size, sequence len, embed dims)`, *optional*):
|
| 245 |
+
Conditional embeddings for cross attention layer. If not given, cross-attention defaults to
|
| 246 |
+
self-attention.
|
| 247 |
+
timestep ( `torch.LongTensor`, *optional*):
|
| 248 |
+
Used to indicate denoising step. Optional timestep to be applied as an embedding in `AdaLayerNorm`.
|
| 249 |
+
class_labels ( `torch.LongTensor` of shape `(batch size, num classes)`, *optional*):
|
| 250 |
+
Used to indicate class labels conditioning. Optional class labels to be applied as an embedding in
|
| 251 |
+
`AdaLayerZeroNorm`.
|
| 252 |
+
cross_attention_kwargs ( `Dict[str, Any]`, *optional*):
|
| 253 |
+
A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
|
| 254 |
+
`self.processor` in
|
| 255 |
+
[diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
|
| 256 |
+
attention_mask ( `torch.Tensor`, *optional*):
|
| 257 |
+
An attention mask of shape `(batch, key_tokens)` is applied to `encoder_hidden_states`. If `1` the mask
|
| 258 |
+
is kept, otherwise if `0` it is discarded. Mask will be converted into a bias, which adds large
|
| 259 |
+
negative values to the attention scores corresponding to "discard" tokens.
|
| 260 |
+
encoder_attention_mask ( `torch.Tensor`, *optional*):
|
| 261 |
+
Cross-attention mask applied to `encoder_hidden_states`. Two formats supported:
|
| 262 |
+
|
| 263 |
+
* Mask `(batch, sequence_length)` True = keep, False = discard.
|
| 264 |
+
* Bias `(batch, 1, sequence_length)` 0 = keep, -10000 = discard.
|
| 265 |
+
|
| 266 |
+
If `ndim == 2`: will be interpreted as a mask, then converted into a bias consistent with the format
|
| 267 |
+
above. This bias will be added to the cross-attention scores.
|
| 268 |
+
return_dict (`bool`, *optional*, defaults to `True`):
|
| 269 |
+
Whether or not to return a [`~models.unet_2d_condition.UNet2DConditionOutput`] instead of a plain
|
| 270 |
+
tuple.
|
| 271 |
+
|
| 272 |
+
Returns:
|
| 273 |
+
If `return_dict` is True, an [`~models.transformer_2d.Transformer2DModelOutput`] is returned, otherwise a
|
| 274 |
+
`tuple` where the first element is the sample tensor.
|
| 275 |
+
"""
|
| 276 |
+
# ensure attention_mask is a bias, and give it a singleton query_tokens dimension.
|
| 277 |
+
# we may have done this conversion already, e.g. if we came here via UNet2DConditionModel#forward.
|
| 278 |
+
# we can tell by counting dims; if ndim == 2: it's a mask rather than a bias.
|
| 279 |
+
# expects mask of shape:
|
| 280 |
+
# [batch, key_tokens]
|
| 281 |
+
# adds singleton query_tokens dimension:
|
| 282 |
+
# [batch, 1, key_tokens]
|
| 283 |
+
# this helps to broadcast it as a bias over attention scores, which will be in one of the following shapes:
|
| 284 |
+
# [batch, heads, query_tokens, key_tokens] (e.g. torch sdp attn)
|
| 285 |
+
# [batch * heads, query_tokens, key_tokens] (e.g. xformers or classic attn)
|
| 286 |
+
if attention_mask is not None and attention_mask.ndim == 2:
|
| 287 |
+
# assume that mask is expressed as:
|
| 288 |
+
# (1 = keep, 0 = discard)
|
| 289 |
+
# convert mask into a bias that can be added to attention scores:
|
| 290 |
+
# (keep = +0, discard = -10000.0)
|
| 291 |
+
attention_mask = (1 - attention_mask.to(hidden_states.dtype)) * -10000.0
|
| 292 |
+
attention_mask = attention_mask.unsqueeze(1)
|
| 293 |
+
|
| 294 |
+
# convert encoder_attention_mask to a bias the same way we do for attention_mask
|
| 295 |
+
if encoder_attention_mask is not None and encoder_attention_mask.ndim == 2:
|
| 296 |
+
encoder_attention_mask = (
|
| 297 |
+
1 - encoder_attention_mask.to(hidden_states.dtype)
|
| 298 |
+
) * -10000.0
|
| 299 |
+
encoder_attention_mask = encoder_attention_mask.unsqueeze(1)
|
| 300 |
+
|
| 301 |
+
# Retrieve lora scale.
|
| 302 |
+
lora_scale = (
|
| 303 |
+
cross_attention_kwargs.get("scale", 1.0)
|
| 304 |
+
if cross_attention_kwargs is not None
|
| 305 |
+
else 1.0
|
| 306 |
+
)
|
| 307 |
+
|
| 308 |
+
# 1. Input
|
| 309 |
+
batch, _, height, width = hidden_states.shape
|
| 310 |
+
residual = hidden_states
|
| 311 |
+
|
| 312 |
+
hidden_states = self.norm(hidden_states)
|
| 313 |
+
if not self.use_linear_projection:
|
| 314 |
+
hidden_states = (
|
| 315 |
+
self.proj_in(hidden_states, scale=lora_scale)
|
| 316 |
+
if not USE_PEFT_BACKEND
|
| 317 |
+
else self.proj_in(hidden_states)
|
| 318 |
+
)
|
| 319 |
+
inner_dim = hidden_states.shape[1]
|
| 320 |
+
hidden_states = hidden_states.permute(0, 2, 3, 1).reshape(
|
| 321 |
+
batch, height * width, inner_dim
|
| 322 |
+
)
|
| 323 |
+
else:
|
| 324 |
+
inner_dim = hidden_states.shape[1]
|
| 325 |
+
hidden_states = hidden_states.permute(0, 2, 3, 1).reshape(
|
| 326 |
+
batch, height * width, inner_dim
|
| 327 |
+
)
|
| 328 |
+
hidden_states = (
|
| 329 |
+
self.proj_in(hidden_states, scale=lora_scale)
|
| 330 |
+
if not USE_PEFT_BACKEND
|
| 331 |
+
else self.proj_in(hidden_states)
|
| 332 |
+
)
|
| 333 |
+
|
| 334 |
+
# 2. Blocks
|
| 335 |
+
if self.caption_projection is not None:
|
| 336 |
+
batch_size = hidden_states.shape[0]
|
| 337 |
+
encoder_hidden_states = self.caption_projection(encoder_hidden_states)
|
| 338 |
+
encoder_hidden_states = encoder_hidden_states.view(
|
| 339 |
+
batch_size, -1, hidden_states.shape[-1]
|
| 340 |
+
)
|
| 341 |
+
|
| 342 |
+
ref_feature = hidden_states.reshape(batch, height, width, inner_dim)
|
| 343 |
+
for block in self.transformer_blocks:
|
| 344 |
+
if self.training and self.gradient_checkpointing:
|
| 345 |
+
|
| 346 |
+
def create_custom_forward(module, return_dict=None):
|
| 347 |
+
def custom_forward(*inputs):
|
| 348 |
+
if return_dict is not None:
|
| 349 |
+
return module(*inputs, return_dict=return_dict)
|
| 350 |
+
else:
|
| 351 |
+
return module(*inputs)
|
| 352 |
+
|
| 353 |
+
return custom_forward
|
| 354 |
+
|
| 355 |
+
ckpt_kwargs: Dict[str, Any] = (
|
| 356 |
+
{"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
|
| 357 |
+
)
|
| 358 |
+
hidden_states = torch.utils.checkpoint.checkpoint(
|
| 359 |
+
create_custom_forward(block),
|
| 360 |
+
hidden_states,
|
| 361 |
+
attention_mask,
|
| 362 |
+
encoder_hidden_states,
|
| 363 |
+
encoder_attention_mask,
|
| 364 |
+
timestep,
|
| 365 |
+
cross_attention_kwargs,
|
| 366 |
+
class_labels,
|
| 367 |
+
**ckpt_kwargs,
|
| 368 |
+
)
|
| 369 |
+
else:
|
| 370 |
+
hidden_states = block(
|
| 371 |
+
hidden_states,
|
| 372 |
+
attention_mask=attention_mask,
|
| 373 |
+
encoder_hidden_states=encoder_hidden_states,
|
| 374 |
+
encoder_attention_mask=encoder_attention_mask,
|
| 375 |
+
timestep=timestep,
|
| 376 |
+
cross_attention_kwargs=cross_attention_kwargs,
|
| 377 |
+
class_labels=class_labels,
|
| 378 |
+
)
|
| 379 |
+
|
| 380 |
+
# 3. Output
|
| 381 |
+
if self.is_input_continuous:
|
| 382 |
+
if not self.use_linear_projection:
|
| 383 |
+
hidden_states = (
|
| 384 |
+
hidden_states.reshape(batch, height, width, inner_dim)
|
| 385 |
+
.permute(0, 3, 1, 2)
|
| 386 |
+
.contiguous()
|
| 387 |
+
)
|
| 388 |
+
hidden_states = (
|
| 389 |
+
self.proj_out(hidden_states, scale=lora_scale)
|
| 390 |
+
if not USE_PEFT_BACKEND
|
| 391 |
+
else self.proj_out(hidden_states)
|
| 392 |
+
)
|
| 393 |
+
else:
|
| 394 |
+
hidden_states = (
|
| 395 |
+
self.proj_out(hidden_states, scale=lora_scale)
|
| 396 |
+
if not USE_PEFT_BACKEND
|
| 397 |
+
else self.proj_out(hidden_states)
|
| 398 |
+
)
|
| 399 |
+
hidden_states = (
|
| 400 |
+
hidden_states.reshape(batch, height, width, inner_dim)
|
| 401 |
+
.permute(0, 3, 1, 2)
|
| 402 |
+
.contiguous()
|
| 403 |
+
)
|
| 404 |
+
|
| 405 |
+
output = hidden_states + residual
|
| 406 |
+
if not return_dict:
|
| 407 |
+
return (output, ref_feature)
|
| 408 |
+
|
| 409 |
+
return Transformer2DModelOutput(sample=output, ref_feature=ref_feature)
|
genwarp/models/transformer_3d.py
ADDED
|
@@ -0,0 +1,179 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# This code is adapted from below and then modified.
|
| 2 |
+
# -----------------------------------------------------------------------------
|
| 3 |
+
# Moore-AnimateAnyone
|
| 4 |
+
# Apache License, Version 2.0
|
| 5 |
+
# Copyright @2023-2024 Moore Threads Technology Co., Ltd.
|
| 6 |
+
# https://github.com/MooreThreads/Moore-AnimateAnyone
|
| 7 |
+
# ==============================================================================
|
| 8 |
+
|
| 9 |
+
from dataclasses import dataclass
|
| 10 |
+
from typing import Optional
|
| 11 |
+
|
| 12 |
+
import torch
|
| 13 |
+
from diffusers.configuration_utils import ConfigMixin, register_to_config
|
| 14 |
+
from diffusers.models import ModelMixin
|
| 15 |
+
from diffusers.utils import BaseOutput
|
| 16 |
+
from diffusers.utils.import_utils import is_xformers_available
|
| 17 |
+
from einops import rearrange, repeat
|
| 18 |
+
from torch import nn
|
| 19 |
+
|
| 20 |
+
from .attention import TemporalBasicTransformerBlock
|
| 21 |
+
|
| 22 |
+
|
| 23 |
+
@dataclass
|
| 24 |
+
class Transformer3DModelOutput(BaseOutput):
|
| 25 |
+
sample: torch.FloatTensor
|
| 26 |
+
|
| 27 |
+
|
| 28 |
+
if is_xformers_available():
|
| 29 |
+
import xformers
|
| 30 |
+
import xformers.ops
|
| 31 |
+
else:
|
| 32 |
+
xformers = None
|
| 33 |
+
|
| 34 |
+
|
| 35 |
+
class Transformer3DModel(ModelMixin, ConfigMixin):
|
| 36 |
+
_supports_gradient_checkpointing = True
|
| 37 |
+
|
| 38 |
+
@register_to_config
|
| 39 |
+
def __init__(
|
| 40 |
+
self,
|
| 41 |
+
num_attention_heads: int = 16,
|
| 42 |
+
attention_head_dim: int = 88,
|
| 43 |
+
in_channels: Optional[int] = None,
|
| 44 |
+
num_layers: int = 1,
|
| 45 |
+
dropout: float = 0.0,
|
| 46 |
+
norm_num_groups: int = 32,
|
| 47 |
+
cross_attention_dim: Optional[int] = None,
|
| 48 |
+
attention_bias: bool = False,
|
| 49 |
+
activation_fn: str = "geglu",
|
| 50 |
+
num_embeds_ada_norm: Optional[int] = None,
|
| 51 |
+
use_linear_projection: bool = False,
|
| 52 |
+
only_cross_attention: bool = False,
|
| 53 |
+
upcast_attention: bool = False,
|
| 54 |
+
unet_use_cross_frame_attention=None,
|
| 55 |
+
unet_use_temporal_attention=None,
|
| 56 |
+
use_zero_convs=False,
|
| 57 |
+
):
|
| 58 |
+
super().__init__()
|
| 59 |
+
self.use_linear_projection = use_linear_projection
|
| 60 |
+
self.num_attention_heads = num_attention_heads
|
| 61 |
+
self.attention_head_dim = attention_head_dim
|
| 62 |
+
inner_dim = num_attention_heads * attention_head_dim
|
| 63 |
+
|
| 64 |
+
# Define input layers
|
| 65 |
+
self.in_channels = in_channels
|
| 66 |
+
|
| 67 |
+
self.norm = torch.nn.GroupNorm(
|
| 68 |
+
num_groups=norm_num_groups, num_channels=in_channels, eps=1e-6, affine=True
|
| 69 |
+
)
|
| 70 |
+
if use_linear_projection:
|
| 71 |
+
self.proj_in = nn.Linear(in_channels, inner_dim)
|
| 72 |
+
else:
|
| 73 |
+
self.proj_in = nn.Conv2d(
|
| 74 |
+
in_channels, inner_dim, kernel_size=1, stride=1, padding=0
|
| 75 |
+
)
|
| 76 |
+
|
| 77 |
+
# Define transformers blocks
|
| 78 |
+
self.transformer_blocks = nn.ModuleList(
|
| 79 |
+
[
|
| 80 |
+
TemporalBasicTransformerBlock(
|
| 81 |
+
inner_dim,
|
| 82 |
+
num_attention_heads,
|
| 83 |
+
attention_head_dim,
|
| 84 |
+
dropout=dropout,
|
| 85 |
+
cross_attention_dim=cross_attention_dim,
|
| 86 |
+
activation_fn=activation_fn,
|
| 87 |
+
num_embeds_ada_norm=num_embeds_ada_norm,
|
| 88 |
+
attention_bias=attention_bias,
|
| 89 |
+
only_cross_attention=only_cross_attention,
|
| 90 |
+
upcast_attention=upcast_attention,
|
| 91 |
+
unet_use_cross_frame_attention=unet_use_cross_frame_attention,
|
| 92 |
+
unet_use_temporal_attention=unet_use_temporal_attention,
|
| 93 |
+
use_zero_convs=use_zero_convs,
|
| 94 |
+
)
|
| 95 |
+
for d in range(num_layers)
|
| 96 |
+
]
|
| 97 |
+
)
|
| 98 |
+
|
| 99 |
+
# 4. Define output layers
|
| 100 |
+
if use_linear_projection:
|
| 101 |
+
self.proj_out = nn.Linear(in_channels, inner_dim)
|
| 102 |
+
else:
|
| 103 |
+
self.proj_out = nn.Conv2d(
|
| 104 |
+
inner_dim, in_channels, kernel_size=1, stride=1, padding=0
|
| 105 |
+
)
|
| 106 |
+
|
| 107 |
+
self.gradient_checkpointing = False
|
| 108 |
+
|
| 109 |
+
def _set_gradient_checkpointing(self, module, value=False):
|
| 110 |
+
if hasattr(module, "gradient_checkpointing"):
|
| 111 |
+
module.gradient_checkpointing = value
|
| 112 |
+
|
| 113 |
+
def forward(
|
| 114 |
+
self,
|
| 115 |
+
hidden_states,
|
| 116 |
+
encoder_hidden_states=None,
|
| 117 |
+
timestep=None,
|
| 118 |
+
return_dict: bool = True,
|
| 119 |
+
):
|
| 120 |
+
# Input
|
| 121 |
+
assert (
|
| 122 |
+
hidden_states.dim() == 5
|
| 123 |
+
), f"Expected hidden_states to have ndim=5, but got ndim={hidden_states.dim()}."
|
| 124 |
+
video_length = hidden_states.shape[2]
|
| 125 |
+
hidden_states = rearrange(hidden_states, "b c f h w -> (b f) c h w")
|
| 126 |
+
if encoder_hidden_states.shape[0] != hidden_states.shape[0]:
|
| 127 |
+
encoder_hidden_states = repeat(
|
| 128 |
+
encoder_hidden_states, "b n c -> (b f) n c", f=video_length
|
| 129 |
+
)
|
| 130 |
+
|
| 131 |
+
batch, channel, height, weight = hidden_states.shape
|
| 132 |
+
residual = hidden_states
|
| 133 |
+
|
| 134 |
+
hidden_states = self.norm(hidden_states)
|
| 135 |
+
if not self.use_linear_projection:
|
| 136 |
+
hidden_states = self.proj_in(hidden_states)
|
| 137 |
+
inner_dim = hidden_states.shape[1]
|
| 138 |
+
hidden_states = hidden_states.permute(0, 2, 3, 1).reshape(
|
| 139 |
+
batch, height * weight, inner_dim
|
| 140 |
+
)
|
| 141 |
+
else:
|
| 142 |
+
inner_dim = hidden_states.shape[1]
|
| 143 |
+
hidden_states = hidden_states.permute(0, 2, 3, 1).reshape(
|
| 144 |
+
batch, height * weight, inner_dim
|
| 145 |
+
)
|
| 146 |
+
hidden_states = self.proj_in(hidden_states)
|
| 147 |
+
|
| 148 |
+
# Blocks
|
| 149 |
+
for i, block in enumerate(self.transformer_blocks):
|
| 150 |
+
hidden_states = block(
|
| 151 |
+
hidden_states,
|
| 152 |
+
encoder_hidden_states=encoder_hidden_states,
|
| 153 |
+
timestep=timestep,
|
| 154 |
+
video_length=video_length,
|
| 155 |
+
)
|
| 156 |
+
|
| 157 |
+
# Output
|
| 158 |
+
if not self.use_linear_projection:
|
| 159 |
+
hidden_states = (
|
| 160 |
+
hidden_states.reshape(batch, height, weight, inner_dim)
|
| 161 |
+
.permute(0, 3, 1, 2)
|
| 162 |
+
.contiguous()
|
| 163 |
+
)
|
| 164 |
+
hidden_states = self.proj_out(hidden_states)
|
| 165 |
+
else:
|
| 166 |
+
hidden_states = self.proj_out(hidden_states)
|
| 167 |
+
hidden_states = (
|
| 168 |
+
hidden_states.reshape(batch, height, weight, inner_dim)
|
| 169 |
+
.permute(0, 3, 1, 2)
|
| 170 |
+
.contiguous()
|
| 171 |
+
)
|
| 172 |
+
|
| 173 |
+
output = hidden_states + residual
|
| 174 |
+
|
| 175 |
+
output = rearrange(output, "(b f) c h w -> b c f h w", f=video_length)
|
| 176 |
+
if not return_dict:
|
| 177 |
+
return (output,)
|
| 178 |
+
|
| 179 |
+
return Transformer3DModelOutput(sample=output)
|
genwarp/models/unet_2d_blocks.py
ADDED
|
@@ -0,0 +1,1087 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# This code is adapted from below and then modified.
|
| 2 |
+
# -----------------------------------------------------------------------------
|
| 3 |
+
# Moore-AnimateAnyone
|
| 4 |
+
# Apache License, Version 2.0
|
| 5 |
+
# Copyright @2023-2024 Moore Threads Technology Co., Ltd.
|
| 6 |
+
# https://github.com/MooreThreads/Moore-AnimateAnyone
|
| 7 |
+
# -----------------------------------------------------------------------------
|
| 8 |
+
# Diffusers
|
| 9 |
+
# Apache License, Version 2.0
|
| 10 |
+
# Copyright (c) Hugging Face Inc.
|
| 11 |
+
# https://github.com/huggingface/diffusers
|
| 12 |
+
# ==============================================================================
|
| 13 |
+
|
| 14 |
+
# Adapted from https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/unet_2d_blocks.py
|
| 15 |
+
from typing import Any, Dict, Optional, Tuple, Union
|
| 16 |
+
|
| 17 |
+
import numpy as np
|
| 18 |
+
import torch
|
| 19 |
+
import torch.nn.functional as F
|
| 20 |
+
from diffusers.models.activations import get_activation
|
| 21 |
+
from diffusers.models.attention_processor import Attention
|
| 22 |
+
from diffusers.models import DualTransformer2DModel
|
| 23 |
+
from diffusers.models.resnet import Downsample2D, ResnetBlock2D, Upsample2D
|
| 24 |
+
from diffusers.utils import is_torch_version, logging
|
| 25 |
+
from diffusers.utils.torch_utils import apply_freeu
|
| 26 |
+
from torch import nn
|
| 27 |
+
|
| 28 |
+
from .transformer_2d import Transformer2DModel
|
| 29 |
+
|
| 30 |
+
logger = logging.get_logger(__name__) # pylint: disable=invalid-name
|
| 31 |
+
|
| 32 |
+
|
| 33 |
+
def get_down_block(
|
| 34 |
+
down_block_type: str,
|
| 35 |
+
num_layers: int,
|
| 36 |
+
in_channels: int,
|
| 37 |
+
out_channels: int,
|
| 38 |
+
temb_channels: int,
|
| 39 |
+
add_downsample: bool,
|
| 40 |
+
resnet_eps: float,
|
| 41 |
+
resnet_act_fn: str,
|
| 42 |
+
transformer_layers_per_block: int = 1,
|
| 43 |
+
num_attention_heads: Optional[int] = None,
|
| 44 |
+
resnet_groups: Optional[int] = None,
|
| 45 |
+
cross_attention_dim: Optional[int] = None,
|
| 46 |
+
downsample_padding: Optional[int] = None,
|
| 47 |
+
dual_cross_attention: bool = False,
|
| 48 |
+
use_linear_projection: bool = False,
|
| 49 |
+
only_cross_attention: bool = False,
|
| 50 |
+
upcast_attention: bool = False,
|
| 51 |
+
resnet_time_scale_shift: str = "default",
|
| 52 |
+
attention_type: str = "default",
|
| 53 |
+
resnet_skip_time_act: bool = False,
|
| 54 |
+
resnet_out_scale_factor: float = 1.0,
|
| 55 |
+
cross_attention_norm: Optional[str] = None,
|
| 56 |
+
attention_head_dim: Optional[int] = None,
|
| 57 |
+
downsample_type: Optional[str] = None,
|
| 58 |
+
dropout: float = 0.0,
|
| 59 |
+
):
|
| 60 |
+
# If attn head dim is not defined, we default it to the number of heads
|
| 61 |
+
if attention_head_dim is None:
|
| 62 |
+
logger.warn(
|
| 63 |
+
f"It is recommended to provide `attention_head_dim` when calling `get_down_block`. Defaulting `attention_head_dim` to {num_attention_heads}."
|
| 64 |
+
)
|
| 65 |
+
attention_head_dim = num_attention_heads
|
| 66 |
+
|
| 67 |
+
down_block_type = (
|
| 68 |
+
down_block_type[7:]
|
| 69 |
+
if down_block_type.startswith("UNetRes")
|
| 70 |
+
else down_block_type
|
| 71 |
+
)
|
| 72 |
+
if down_block_type == "DownBlock2D":
|
| 73 |
+
return DownBlock2D(
|
| 74 |
+
num_layers=num_layers,
|
| 75 |
+
in_channels=in_channels,
|
| 76 |
+
out_channels=out_channels,
|
| 77 |
+
temb_channels=temb_channels,
|
| 78 |
+
dropout=dropout,
|
| 79 |
+
add_downsample=add_downsample,
|
| 80 |
+
resnet_eps=resnet_eps,
|
| 81 |
+
resnet_act_fn=resnet_act_fn,
|
| 82 |
+
resnet_groups=resnet_groups,
|
| 83 |
+
downsample_padding=downsample_padding,
|
| 84 |
+
resnet_time_scale_shift=resnet_time_scale_shift,
|
| 85 |
+
)
|
| 86 |
+
elif down_block_type == "CrossAttnDownBlock2D":
|
| 87 |
+
if cross_attention_dim is None:
|
| 88 |
+
raise ValueError(
|
| 89 |
+
"cross_attention_dim must be specified for CrossAttnDownBlock2D"
|
| 90 |
+
)
|
| 91 |
+
return CrossAttnDownBlock2D(
|
| 92 |
+
num_layers=num_layers,
|
| 93 |
+
transformer_layers_per_block=transformer_layers_per_block,
|
| 94 |
+
in_channels=in_channels,
|
| 95 |
+
out_channels=out_channels,
|
| 96 |
+
temb_channels=temb_channels,
|
| 97 |
+
dropout=dropout,
|
| 98 |
+
add_downsample=add_downsample,
|
| 99 |
+
resnet_eps=resnet_eps,
|
| 100 |
+
resnet_act_fn=resnet_act_fn,
|
| 101 |
+
resnet_groups=resnet_groups,
|
| 102 |
+
downsample_padding=downsample_padding,
|
| 103 |
+
cross_attention_dim=cross_attention_dim,
|
| 104 |
+
num_attention_heads=num_attention_heads,
|
| 105 |
+
dual_cross_attention=dual_cross_attention,
|
| 106 |
+
use_linear_projection=use_linear_projection,
|
| 107 |
+
only_cross_attention=only_cross_attention,
|
| 108 |
+
upcast_attention=upcast_attention,
|
| 109 |
+
resnet_time_scale_shift=resnet_time_scale_shift,
|
| 110 |
+
attention_type=attention_type,
|
| 111 |
+
)
|
| 112 |
+
raise ValueError(f"{down_block_type} does not exist.")
|
| 113 |
+
|
| 114 |
+
|
| 115 |
+
def get_up_block(
|
| 116 |
+
up_block_type: str,
|
| 117 |
+
num_layers: int,
|
| 118 |
+
in_channels: int,
|
| 119 |
+
out_channels: int,
|
| 120 |
+
prev_output_channel: int,
|
| 121 |
+
temb_channels: int,
|
| 122 |
+
add_upsample: bool,
|
| 123 |
+
resnet_eps: float,
|
| 124 |
+
resnet_act_fn: str,
|
| 125 |
+
resolution_idx: Optional[int] = None,
|
| 126 |
+
transformer_layers_per_block: int = 1,
|
| 127 |
+
num_attention_heads: Optional[int] = None,
|
| 128 |
+
resnet_groups: Optional[int] = None,
|
| 129 |
+
cross_attention_dim: Optional[int] = None,
|
| 130 |
+
dual_cross_attention: bool = False,
|
| 131 |
+
use_linear_projection: bool = False,
|
| 132 |
+
only_cross_attention: bool = False,
|
| 133 |
+
upcast_attention: bool = False,
|
| 134 |
+
resnet_time_scale_shift: str = "default",
|
| 135 |
+
attention_type: str = "default",
|
| 136 |
+
resnet_skip_time_act: bool = False,
|
| 137 |
+
resnet_out_scale_factor: float = 1.0,
|
| 138 |
+
cross_attention_norm: Optional[str] = None,
|
| 139 |
+
attention_head_dim: Optional[int] = None,
|
| 140 |
+
upsample_type: Optional[str] = None,
|
| 141 |
+
dropout: float = 0.0,
|
| 142 |
+
) -> nn.Module:
|
| 143 |
+
# If attn head dim is not defined, we default it to the number of heads
|
| 144 |
+
if attention_head_dim is None:
|
| 145 |
+
logger.warn(
|
| 146 |
+
f"It is recommended to provide `attention_head_dim` when calling `get_up_block`. Defaulting `attention_head_dim` to {num_attention_heads}."
|
| 147 |
+
)
|
| 148 |
+
attention_head_dim = num_attention_heads
|
| 149 |
+
|
| 150 |
+
up_block_type = (
|
| 151 |
+
up_block_type[7:] if up_block_type.startswith("UNetRes") else up_block_type
|
| 152 |
+
)
|
| 153 |
+
if up_block_type == "UpBlock2D":
|
| 154 |
+
return UpBlock2D(
|
| 155 |
+
num_layers=num_layers,
|
| 156 |
+
in_channels=in_channels,
|
| 157 |
+
out_channels=out_channels,
|
| 158 |
+
prev_output_channel=prev_output_channel,
|
| 159 |
+
temb_channels=temb_channels,
|
| 160 |
+
resolution_idx=resolution_idx,
|
| 161 |
+
dropout=dropout,
|
| 162 |
+
add_upsample=add_upsample,
|
| 163 |
+
resnet_eps=resnet_eps,
|
| 164 |
+
resnet_act_fn=resnet_act_fn,
|
| 165 |
+
resnet_groups=resnet_groups,
|
| 166 |
+
resnet_time_scale_shift=resnet_time_scale_shift,
|
| 167 |
+
)
|
| 168 |
+
elif up_block_type == "CrossAttnUpBlock2D":
|
| 169 |
+
if cross_attention_dim is None:
|
| 170 |
+
raise ValueError(
|
| 171 |
+
"cross_attention_dim must be specified for CrossAttnUpBlock2D"
|
| 172 |
+
)
|
| 173 |
+
return CrossAttnUpBlock2D(
|
| 174 |
+
num_layers=num_layers,
|
| 175 |
+
transformer_layers_per_block=transformer_layers_per_block,
|
| 176 |
+
in_channels=in_channels,
|
| 177 |
+
out_channels=out_channels,
|
| 178 |
+
prev_output_channel=prev_output_channel,
|
| 179 |
+
temb_channels=temb_channels,
|
| 180 |
+
resolution_idx=resolution_idx,
|
| 181 |
+
dropout=dropout,
|
| 182 |
+
add_upsample=add_upsample,
|
| 183 |
+
resnet_eps=resnet_eps,
|
| 184 |
+
resnet_act_fn=resnet_act_fn,
|
| 185 |
+
resnet_groups=resnet_groups,
|
| 186 |
+
cross_attention_dim=cross_attention_dim,
|
| 187 |
+
num_attention_heads=num_attention_heads,
|
| 188 |
+
dual_cross_attention=dual_cross_attention,
|
| 189 |
+
use_linear_projection=use_linear_projection,
|
| 190 |
+
only_cross_attention=only_cross_attention,
|
| 191 |
+
upcast_attention=upcast_attention,
|
| 192 |
+
resnet_time_scale_shift=resnet_time_scale_shift,
|
| 193 |
+
attention_type=attention_type,
|
| 194 |
+
)
|
| 195 |
+
|
| 196 |
+
raise ValueError(f"{up_block_type} does not exist.")
|
| 197 |
+
|
| 198 |
+
|
| 199 |
+
class AutoencoderTinyBlock(nn.Module):
|
| 200 |
+
"""
|
| 201 |
+
Tiny Autoencoder block used in [`AutoencoderTiny`]. It is a mini residual module consisting of plain conv + ReLU
|
| 202 |
+
blocks.
|
| 203 |
+
|
| 204 |
+
Args:
|
| 205 |
+
in_channels (`int`): The number of input channels.
|
| 206 |
+
out_channels (`int`): The number of output channels.
|
| 207 |
+
act_fn (`str`):
|
| 208 |
+
` The activation function to use. Supported values are `"swish"`, `"mish"`, `"gelu"`, and `"relu"`.
|
| 209 |
+
|
| 210 |
+
Returns:
|
| 211 |
+
`torch.FloatTensor`: A tensor with the same shape as the input tensor, but with the number of channels equal to
|
| 212 |
+
`out_channels`.
|
| 213 |
+
"""
|
| 214 |
+
|
| 215 |
+
def __init__(self, in_channels: int, out_channels: int, act_fn: str):
|
| 216 |
+
super().__init__()
|
| 217 |
+
act_fn = get_activation(act_fn)
|
| 218 |
+
self.conv = nn.Sequential(
|
| 219 |
+
nn.Conv2d(in_channels, out_channels, kernel_size=3, padding=1),
|
| 220 |
+
act_fn,
|
| 221 |
+
nn.Conv2d(out_channels, out_channels, kernel_size=3, padding=1),
|
| 222 |
+
act_fn,
|
| 223 |
+
nn.Conv2d(out_channels, out_channels, kernel_size=3, padding=1),
|
| 224 |
+
)
|
| 225 |
+
self.skip = (
|
| 226 |
+
nn.Conv2d(in_channels, out_channels, kernel_size=1, bias=False)
|
| 227 |
+
if in_channels != out_channels
|
| 228 |
+
else nn.Identity()
|
| 229 |
+
)
|
| 230 |
+
self.fuse = nn.ReLU()
|
| 231 |
+
|
| 232 |
+
def forward(self, x: torch.FloatTensor) -> torch.FloatTensor:
|
| 233 |
+
return self.fuse(self.conv(x) + self.skip(x))
|
| 234 |
+
|
| 235 |
+
|
| 236 |
+
class UNetMidBlock2D(nn.Module):
|
| 237 |
+
"""
|
| 238 |
+
A 2D UNet mid-block [`UNetMidBlock2D`] with multiple residual blocks and optional attention blocks.
|
| 239 |
+
|
| 240 |
+
Args:
|
| 241 |
+
in_channels (`int`): The number of input channels.
|
| 242 |
+
temb_channels (`int`): The number of temporal embedding channels.
|
| 243 |
+
dropout (`float`, *optional*, defaults to 0.0): The dropout rate.
|
| 244 |
+
num_layers (`int`, *optional*, defaults to 1): The number of residual blocks.
|
| 245 |
+
resnet_eps (`float`, *optional*, 1e-6 ): The epsilon value for the resnet blocks.
|
| 246 |
+
resnet_time_scale_shift (`str`, *optional*, defaults to `default`):
|
| 247 |
+
The type of normalization to apply to the time embeddings. This can help to improve the performance of the
|
| 248 |
+
model on tasks with long-range temporal dependencies.
|
| 249 |
+
resnet_act_fn (`str`, *optional*, defaults to `swish`): The activation function for the resnet blocks.
|
| 250 |
+
resnet_groups (`int`, *optional*, defaults to 32):
|
| 251 |
+
The number of groups to use in the group normalization layers of the resnet blocks.
|
| 252 |
+
attn_groups (`Optional[int]`, *optional*, defaults to None): The number of groups for the attention blocks.
|
| 253 |
+
resnet_pre_norm (`bool`, *optional*, defaults to `True`):
|
| 254 |
+
Whether to use pre-normalization for the resnet blocks.
|
| 255 |
+
add_attention (`bool`, *optional*, defaults to `True`): Whether to add attention blocks.
|
| 256 |
+
attention_head_dim (`int`, *optional*, defaults to 1):
|
| 257 |
+
Dimension of a single attention head. The number of attention heads is determined based on this value and
|
| 258 |
+
the number of input channels.
|
| 259 |
+
output_scale_factor (`float`, *optional*, defaults to 1.0): The output scale factor.
|
| 260 |
+
|
| 261 |
+
Returns:
|
| 262 |
+
`torch.FloatTensor`: The output of the last residual block, which is a tensor of shape `(batch_size,
|
| 263 |
+
in_channels, height, width)`.
|
| 264 |
+
|
| 265 |
+
"""
|
| 266 |
+
|
| 267 |
+
def __init__(
|
| 268 |
+
self,
|
| 269 |
+
in_channels: int,
|
| 270 |
+
temb_channels: int,
|
| 271 |
+
dropout: float = 0.0,
|
| 272 |
+
num_layers: int = 1,
|
| 273 |
+
resnet_eps: float = 1e-6,
|
| 274 |
+
resnet_time_scale_shift: str = "default", # default, spatial
|
| 275 |
+
resnet_act_fn: str = "swish",
|
| 276 |
+
resnet_groups: int = 32,
|
| 277 |
+
attn_groups: Optional[int] = None,
|
| 278 |
+
resnet_pre_norm: bool = True,
|
| 279 |
+
add_attention: bool = True,
|
| 280 |
+
attention_head_dim: int = 1,
|
| 281 |
+
output_scale_factor: float = 1.0,
|
| 282 |
+
):
|
| 283 |
+
super().__init__()
|
| 284 |
+
resnet_groups = (
|
| 285 |
+
resnet_groups if resnet_groups is not None else min(in_channels // 4, 32)
|
| 286 |
+
)
|
| 287 |
+
self.add_attention = add_attention
|
| 288 |
+
|
| 289 |
+
if attn_groups is None:
|
| 290 |
+
attn_groups = (
|
| 291 |
+
resnet_groups if resnet_time_scale_shift == "default" else None
|
| 292 |
+
)
|
| 293 |
+
|
| 294 |
+
# there is always at least one resnet
|
| 295 |
+
resnets = [
|
| 296 |
+
ResnetBlock2D(
|
| 297 |
+
in_channels=in_channels,
|
| 298 |
+
out_channels=in_channels,
|
| 299 |
+
temb_channels=temb_channels,
|
| 300 |
+
eps=resnet_eps,
|
| 301 |
+
groups=resnet_groups,
|
| 302 |
+
dropout=dropout,
|
| 303 |
+
time_embedding_norm=resnet_time_scale_shift,
|
| 304 |
+
non_linearity=resnet_act_fn,
|
| 305 |
+
output_scale_factor=output_scale_factor,
|
| 306 |
+
pre_norm=resnet_pre_norm,
|
| 307 |
+
)
|
| 308 |
+
]
|
| 309 |
+
attentions = []
|
| 310 |
+
|
| 311 |
+
if attention_head_dim is None:
|
| 312 |
+
logger.warn(
|
| 313 |
+
f"It is not recommend to pass `attention_head_dim=None`. Defaulting `attention_head_dim` to `in_channels`: {in_channels}."
|
| 314 |
+
)
|
| 315 |
+
attention_head_dim = in_channels
|
| 316 |
+
|
| 317 |
+
for _ in range(num_layers):
|
| 318 |
+
if self.add_attention:
|
| 319 |
+
attentions.append(
|
| 320 |
+
Attention(
|
| 321 |
+
in_channels,
|
| 322 |
+
heads=in_channels // attention_head_dim,
|
| 323 |
+
dim_head=attention_head_dim,
|
| 324 |
+
rescale_output_factor=output_scale_factor,
|
| 325 |
+
eps=resnet_eps,
|
| 326 |
+
norm_num_groups=attn_groups,
|
| 327 |
+
spatial_norm_dim=temb_channels
|
| 328 |
+
if resnet_time_scale_shift == "spatial"
|
| 329 |
+
else None,
|
| 330 |
+
residual_connection=True,
|
| 331 |
+
bias=True,
|
| 332 |
+
upcast_softmax=True,
|
| 333 |
+
_from_deprecated_attn_block=True,
|
| 334 |
+
)
|
| 335 |
+
)
|
| 336 |
+
else:
|
| 337 |
+
attentions.append(None)
|
| 338 |
+
|
| 339 |
+
resnets.append(
|
| 340 |
+
ResnetBlock2D(
|
| 341 |
+
in_channels=in_channels,
|
| 342 |
+
out_channels=in_channels,
|
| 343 |
+
temb_channels=temb_channels,
|
| 344 |
+
eps=resnet_eps,
|
| 345 |
+
groups=resnet_groups,
|
| 346 |
+
dropout=dropout,
|
| 347 |
+
time_embedding_norm=resnet_time_scale_shift,
|
| 348 |
+
non_linearity=resnet_act_fn,
|
| 349 |
+
output_scale_factor=output_scale_factor,
|
| 350 |
+
pre_norm=resnet_pre_norm,
|
| 351 |
+
)
|
| 352 |
+
)
|
| 353 |
+
|
| 354 |
+
self.attentions = nn.ModuleList(attentions)
|
| 355 |
+
self.resnets = nn.ModuleList(resnets)
|
| 356 |
+
|
| 357 |
+
def forward(
|
| 358 |
+
self, hidden_states: torch.FloatTensor, temb: Optional[torch.FloatTensor] = None
|
| 359 |
+
) -> torch.FloatTensor:
|
| 360 |
+
hidden_states = self.resnets[0](hidden_states, temb)
|
| 361 |
+
for attn, resnet in zip(self.attentions, self.resnets[1:]):
|
| 362 |
+
if attn is not None:
|
| 363 |
+
hidden_states = attn(hidden_states, temb=temb)
|
| 364 |
+
hidden_states = resnet(hidden_states, temb)
|
| 365 |
+
|
| 366 |
+
return hidden_states
|
| 367 |
+
|
| 368 |
+
|
| 369 |
+
class UNetMidBlock2DCrossAttn(nn.Module):
|
| 370 |
+
def __init__(
|
| 371 |
+
self,
|
| 372 |
+
in_channels: int,
|
| 373 |
+
temb_channels: int,
|
| 374 |
+
dropout: float = 0.0,
|
| 375 |
+
num_layers: int = 1,
|
| 376 |
+
transformer_layers_per_block: Union[int, Tuple[int]] = 1,
|
| 377 |
+
resnet_eps: float = 1e-6,
|
| 378 |
+
resnet_time_scale_shift: str = "default",
|
| 379 |
+
resnet_act_fn: str = "swish",
|
| 380 |
+
resnet_groups: int = 32,
|
| 381 |
+
resnet_pre_norm: bool = True,
|
| 382 |
+
num_attention_heads: int = 1,
|
| 383 |
+
output_scale_factor: float = 1.0,
|
| 384 |
+
cross_attention_dim: int = 1280,
|
| 385 |
+
dual_cross_attention: bool = False,
|
| 386 |
+
use_linear_projection: bool = False,
|
| 387 |
+
upcast_attention: bool = False,
|
| 388 |
+
attention_type: str = "default",
|
| 389 |
+
):
|
| 390 |
+
super().__init__()
|
| 391 |
+
|
| 392 |
+
self.has_cross_attention = True
|
| 393 |
+
self.num_attention_heads = num_attention_heads
|
| 394 |
+
resnet_groups = (
|
| 395 |
+
resnet_groups if resnet_groups is not None else min(in_channels // 4, 32)
|
| 396 |
+
)
|
| 397 |
+
|
| 398 |
+
# support for variable transformer layers per block
|
| 399 |
+
if isinstance(transformer_layers_per_block, int):
|
| 400 |
+
transformer_layers_per_block = [transformer_layers_per_block] * num_layers
|
| 401 |
+
|
| 402 |
+
# there is always at least one resnet
|
| 403 |
+
resnets = [
|
| 404 |
+
ResnetBlock2D(
|
| 405 |
+
in_channels=in_channels,
|
| 406 |
+
out_channels=in_channels,
|
| 407 |
+
temb_channels=temb_channels,
|
| 408 |
+
eps=resnet_eps,
|
| 409 |
+
groups=resnet_groups,
|
| 410 |
+
dropout=dropout,
|
| 411 |
+
time_embedding_norm=resnet_time_scale_shift,
|
| 412 |
+
non_linearity=resnet_act_fn,
|
| 413 |
+
output_scale_factor=output_scale_factor,
|
| 414 |
+
pre_norm=resnet_pre_norm,
|
| 415 |
+
)
|
| 416 |
+
]
|
| 417 |
+
attentions = []
|
| 418 |
+
|
| 419 |
+
for i in range(num_layers):
|
| 420 |
+
if not dual_cross_attention:
|
| 421 |
+
attentions.append(
|
| 422 |
+
Transformer2DModel(
|
| 423 |
+
num_attention_heads,
|
| 424 |
+
in_channels // num_attention_heads,
|
| 425 |
+
in_channels=in_channels,
|
| 426 |
+
num_layers=transformer_layers_per_block[i],
|
| 427 |
+
cross_attention_dim=cross_attention_dim,
|
| 428 |
+
norm_num_groups=resnet_groups,
|
| 429 |
+
use_linear_projection=use_linear_projection,
|
| 430 |
+
upcast_attention=upcast_attention,
|
| 431 |
+
attention_type=attention_type,
|
| 432 |
+
)
|
| 433 |
+
)
|
| 434 |
+
else:
|
| 435 |
+
attentions.append(
|
| 436 |
+
DualTransformer2DModel(
|
| 437 |
+
num_attention_heads,
|
| 438 |
+
in_channels // num_attention_heads,
|
| 439 |
+
in_channels=in_channels,
|
| 440 |
+
num_layers=1,
|
| 441 |
+
cross_attention_dim=cross_attention_dim,
|
| 442 |
+
norm_num_groups=resnet_groups,
|
| 443 |
+
)
|
| 444 |
+
)
|
| 445 |
+
resnets.append(
|
| 446 |
+
ResnetBlock2D(
|
| 447 |
+
in_channels=in_channels,
|
| 448 |
+
out_channels=in_channels,
|
| 449 |
+
temb_channels=temb_channels,
|
| 450 |
+
eps=resnet_eps,
|
| 451 |
+
groups=resnet_groups,
|
| 452 |
+
dropout=dropout,
|
| 453 |
+
time_embedding_norm=resnet_time_scale_shift,
|
| 454 |
+
non_linearity=resnet_act_fn,
|
| 455 |
+
output_scale_factor=output_scale_factor,
|
| 456 |
+
pre_norm=resnet_pre_norm,
|
| 457 |
+
)
|
| 458 |
+
)
|
| 459 |
+
|
| 460 |
+
self.attentions = nn.ModuleList(attentions)
|
| 461 |
+
self.resnets = nn.ModuleList(resnets)
|
| 462 |
+
|
| 463 |
+
self.gradient_checkpointing = False
|
| 464 |
+
|
| 465 |
+
def forward(
|
| 466 |
+
self,
|
| 467 |
+
hidden_states: torch.FloatTensor,
|
| 468 |
+
temb: Optional[torch.FloatTensor] = None,
|
| 469 |
+
encoder_hidden_states: Optional[torch.FloatTensor] = None,
|
| 470 |
+
attention_mask: Optional[torch.FloatTensor] = None,
|
| 471 |
+
cross_attention_kwargs: Optional[Dict[str, Any]] = None,
|
| 472 |
+
encoder_attention_mask: Optional[torch.FloatTensor] = None,
|
| 473 |
+
) -> torch.FloatTensor:
|
| 474 |
+
lora_scale = (
|
| 475 |
+
cross_attention_kwargs.get("scale", 1.0)
|
| 476 |
+
if cross_attention_kwargs is not None
|
| 477 |
+
else 1.0
|
| 478 |
+
)
|
| 479 |
+
hidden_states = self.resnets[0](hidden_states, temb, scale=lora_scale)
|
| 480 |
+
for attn, resnet in zip(self.attentions, self.resnets[1:]):
|
| 481 |
+
if self.training and self.gradient_checkpointing:
|
| 482 |
+
|
| 483 |
+
def create_custom_forward(module, return_dict=None):
|
| 484 |
+
def custom_forward(*inputs):
|
| 485 |
+
if return_dict is not None:
|
| 486 |
+
return module(*inputs, return_dict=return_dict)
|
| 487 |
+
else:
|
| 488 |
+
return module(*inputs)
|
| 489 |
+
|
| 490 |
+
return custom_forward
|
| 491 |
+
|
| 492 |
+
ckpt_kwargs: Dict[str, Any] = (
|
| 493 |
+
{"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
|
| 494 |
+
)
|
| 495 |
+
hidden_states, ref_feature = attn(
|
| 496 |
+
hidden_states,
|
| 497 |
+
encoder_hidden_states=encoder_hidden_states,
|
| 498 |
+
cross_attention_kwargs=cross_attention_kwargs,
|
| 499 |
+
attention_mask=attention_mask,
|
| 500 |
+
encoder_attention_mask=encoder_attention_mask,
|
| 501 |
+
return_dict=False,
|
| 502 |
+
)
|
| 503 |
+
hidden_states = torch.utils.checkpoint.checkpoint(
|
| 504 |
+
create_custom_forward(resnet),
|
| 505 |
+
hidden_states,
|
| 506 |
+
temb,
|
| 507 |
+
**ckpt_kwargs,
|
| 508 |
+
)
|
| 509 |
+
else:
|
| 510 |
+
hidden_states, ref_feature = attn(
|
| 511 |
+
hidden_states,
|
| 512 |
+
encoder_hidden_states=encoder_hidden_states,
|
| 513 |
+
cross_attention_kwargs=cross_attention_kwargs,
|
| 514 |
+
attention_mask=attention_mask,
|
| 515 |
+
encoder_attention_mask=encoder_attention_mask,
|
| 516 |
+
return_dict=False,
|
| 517 |
+
)
|
| 518 |
+
hidden_states = resnet(hidden_states, temb, scale=lora_scale)
|
| 519 |
+
|
| 520 |
+
return hidden_states
|
| 521 |
+
|
| 522 |
+
|
| 523 |
+
class CrossAttnDownBlock2D(nn.Module):
|
| 524 |
+
def __init__(
|
| 525 |
+
self,
|
| 526 |
+
in_channels: int,
|
| 527 |
+
out_channels: int,
|
| 528 |
+
temb_channels: int,
|
| 529 |
+
dropout: float = 0.0,
|
| 530 |
+
num_layers: int = 1,
|
| 531 |
+
transformer_layers_per_block: Union[int, Tuple[int]] = 1,
|
| 532 |
+
resnet_eps: float = 1e-6,
|
| 533 |
+
resnet_time_scale_shift: str = "default",
|
| 534 |
+
resnet_act_fn: str = "swish",
|
| 535 |
+
resnet_groups: int = 32,
|
| 536 |
+
resnet_pre_norm: bool = True,
|
| 537 |
+
num_attention_heads: int = 1,
|
| 538 |
+
cross_attention_dim: int = 1280,
|
| 539 |
+
output_scale_factor: float = 1.0,
|
| 540 |
+
downsample_padding: int = 1,
|
| 541 |
+
add_downsample: bool = True,
|
| 542 |
+
dual_cross_attention: bool = False,
|
| 543 |
+
use_linear_projection: bool = False,
|
| 544 |
+
only_cross_attention: bool = False,
|
| 545 |
+
upcast_attention: bool = False,
|
| 546 |
+
attention_type: str = "default",
|
| 547 |
+
):
|
| 548 |
+
super().__init__()
|
| 549 |
+
resnets = []
|
| 550 |
+
attentions = []
|
| 551 |
+
|
| 552 |
+
self.has_cross_attention = True
|
| 553 |
+
self.num_attention_heads = num_attention_heads
|
| 554 |
+
if isinstance(transformer_layers_per_block, int):
|
| 555 |
+
transformer_layers_per_block = [transformer_layers_per_block] * num_layers
|
| 556 |
+
|
| 557 |
+
for i in range(num_layers):
|
| 558 |
+
in_channels = in_channels if i == 0 else out_channels
|
| 559 |
+
resnets.append(
|
| 560 |
+
ResnetBlock2D(
|
| 561 |
+
in_channels=in_channels,
|
| 562 |
+
out_channels=out_channels,
|
| 563 |
+
temb_channels=temb_channels,
|
| 564 |
+
eps=resnet_eps,
|
| 565 |
+
groups=resnet_groups,
|
| 566 |
+
dropout=dropout,
|
| 567 |
+
time_embedding_norm=resnet_time_scale_shift,
|
| 568 |
+
non_linearity=resnet_act_fn,
|
| 569 |
+
output_scale_factor=output_scale_factor,
|
| 570 |
+
pre_norm=resnet_pre_norm,
|
| 571 |
+
)
|
| 572 |
+
)
|
| 573 |
+
if not dual_cross_attention:
|
| 574 |
+
attentions.append(
|
| 575 |
+
Transformer2DModel(
|
| 576 |
+
num_attention_heads,
|
| 577 |
+
out_channels // num_attention_heads,
|
| 578 |
+
in_channels=out_channels,
|
| 579 |
+
num_layers=transformer_layers_per_block[i],
|
| 580 |
+
cross_attention_dim=cross_attention_dim,
|
| 581 |
+
norm_num_groups=resnet_groups,
|
| 582 |
+
use_linear_projection=use_linear_projection,
|
| 583 |
+
only_cross_attention=only_cross_attention,
|
| 584 |
+
upcast_attention=upcast_attention,
|
| 585 |
+
attention_type=attention_type,
|
| 586 |
+
)
|
| 587 |
+
)
|
| 588 |
+
else:
|
| 589 |
+
attentions.append(
|
| 590 |
+
DualTransformer2DModel(
|
| 591 |
+
num_attention_heads,
|
| 592 |
+
out_channels // num_attention_heads,
|
| 593 |
+
in_channels=out_channels,
|
| 594 |
+
num_layers=1,
|
| 595 |
+
cross_attention_dim=cross_attention_dim,
|
| 596 |
+
norm_num_groups=resnet_groups,
|
| 597 |
+
)
|
| 598 |
+
)
|
| 599 |
+
self.attentions = nn.ModuleList(attentions)
|
| 600 |
+
self.resnets = nn.ModuleList(resnets)
|
| 601 |
+
|
| 602 |
+
if add_downsample:
|
| 603 |
+
self.downsamplers = nn.ModuleList(
|
| 604 |
+
[
|
| 605 |
+
Downsample2D(
|
| 606 |
+
out_channels,
|
| 607 |
+
use_conv=True,
|
| 608 |
+
out_channels=out_channels,
|
| 609 |
+
padding=downsample_padding,
|
| 610 |
+
name="op",
|
| 611 |
+
)
|
| 612 |
+
]
|
| 613 |
+
)
|
| 614 |
+
else:
|
| 615 |
+
self.downsamplers = None
|
| 616 |
+
|
| 617 |
+
self.gradient_checkpointing = False
|
| 618 |
+
|
| 619 |
+
def forward(
|
| 620 |
+
self,
|
| 621 |
+
hidden_states: torch.FloatTensor,
|
| 622 |
+
temb: Optional[torch.FloatTensor] = None,
|
| 623 |
+
encoder_hidden_states: Optional[torch.FloatTensor] = None,
|
| 624 |
+
attention_mask: Optional[torch.FloatTensor] = None,
|
| 625 |
+
cross_attention_kwargs: Optional[Dict[str, Any]] = None,
|
| 626 |
+
encoder_attention_mask: Optional[torch.FloatTensor] = None,
|
| 627 |
+
additional_residuals: Optional[torch.FloatTensor] = None,
|
| 628 |
+
) -> Tuple[torch.FloatTensor, Tuple[torch.FloatTensor, ...]]:
|
| 629 |
+
output_states = ()
|
| 630 |
+
|
| 631 |
+
lora_scale = (
|
| 632 |
+
cross_attention_kwargs.get("scale", 1.0)
|
| 633 |
+
if cross_attention_kwargs is not None
|
| 634 |
+
else 1.0
|
| 635 |
+
)
|
| 636 |
+
|
| 637 |
+
blocks = list(zip(self.resnets, self.attentions))
|
| 638 |
+
|
| 639 |
+
for i, (resnet, attn) in enumerate(blocks):
|
| 640 |
+
if self.training and self.gradient_checkpointing:
|
| 641 |
+
|
| 642 |
+
def create_custom_forward(module, return_dict=None):
|
| 643 |
+
def custom_forward(*inputs):
|
| 644 |
+
if return_dict is not None:
|
| 645 |
+
return module(*inputs, return_dict=return_dict)
|
| 646 |
+
else:
|
| 647 |
+
return module(*inputs)
|
| 648 |
+
|
| 649 |
+
return custom_forward
|
| 650 |
+
|
| 651 |
+
ckpt_kwargs: Dict[str, Any] = (
|
| 652 |
+
{"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
|
| 653 |
+
)
|
| 654 |
+
hidden_states = torch.utils.checkpoint.checkpoint(
|
| 655 |
+
create_custom_forward(resnet),
|
| 656 |
+
hidden_states,
|
| 657 |
+
temb,
|
| 658 |
+
**ckpt_kwargs,
|
| 659 |
+
)
|
| 660 |
+
hidden_states, ref_feature = attn(
|
| 661 |
+
hidden_states,
|
| 662 |
+
encoder_hidden_states=encoder_hidden_states,
|
| 663 |
+
cross_attention_kwargs=cross_attention_kwargs,
|
| 664 |
+
attention_mask=attention_mask,
|
| 665 |
+
encoder_attention_mask=encoder_attention_mask,
|
| 666 |
+
return_dict=False,
|
| 667 |
+
)
|
| 668 |
+
else:
|
| 669 |
+
hidden_states = resnet(hidden_states, temb, scale=lora_scale)
|
| 670 |
+
hidden_states, ref_feature = attn(
|
| 671 |
+
hidden_states,
|
| 672 |
+
encoder_hidden_states=encoder_hidden_states,
|
| 673 |
+
cross_attention_kwargs=cross_attention_kwargs,
|
| 674 |
+
attention_mask=attention_mask,
|
| 675 |
+
encoder_attention_mask=encoder_attention_mask,
|
| 676 |
+
return_dict=False,
|
| 677 |
+
)
|
| 678 |
+
|
| 679 |
+
# apply additional residuals to the output of the last pair of resnet and attention blocks
|
| 680 |
+
if i == len(blocks) - 1 and additional_residuals is not None:
|
| 681 |
+
hidden_states = hidden_states + additional_residuals
|
| 682 |
+
|
| 683 |
+
output_states = output_states + (hidden_states,)
|
| 684 |
+
|
| 685 |
+
if self.downsamplers is not None:
|
| 686 |
+
for downsampler in self.downsamplers:
|
| 687 |
+
hidden_states = downsampler(hidden_states, scale=lora_scale)
|
| 688 |
+
|
| 689 |
+
output_states = output_states + (hidden_states,)
|
| 690 |
+
|
| 691 |
+
return hidden_states, output_states
|
| 692 |
+
|
| 693 |
+
|
| 694 |
+
class DownBlock2D(nn.Module):
|
| 695 |
+
def __init__(
|
| 696 |
+
self,
|
| 697 |
+
in_channels: int,
|
| 698 |
+
out_channels: int,
|
| 699 |
+
temb_channels: int,
|
| 700 |
+
dropout: float = 0.0,
|
| 701 |
+
num_layers: int = 1,
|
| 702 |
+
resnet_eps: float = 1e-6,
|
| 703 |
+
resnet_time_scale_shift: str = "default",
|
| 704 |
+
resnet_act_fn: str = "swish",
|
| 705 |
+
resnet_groups: int = 32,
|
| 706 |
+
resnet_pre_norm: bool = True,
|
| 707 |
+
output_scale_factor: float = 1.0,
|
| 708 |
+
add_downsample: bool = True,
|
| 709 |
+
downsample_padding: int = 1,
|
| 710 |
+
):
|
| 711 |
+
super().__init__()
|
| 712 |
+
resnets = []
|
| 713 |
+
|
| 714 |
+
for i in range(num_layers):
|
| 715 |
+
in_channels = in_channels if i == 0 else out_channels
|
| 716 |
+
resnets.append(
|
| 717 |
+
ResnetBlock2D(
|
| 718 |
+
in_channels=in_channels,
|
| 719 |
+
out_channels=out_channels,
|
| 720 |
+
temb_channels=temb_channels,
|
| 721 |
+
eps=resnet_eps,
|
| 722 |
+
groups=resnet_groups,
|
| 723 |
+
dropout=dropout,
|
| 724 |
+
time_embedding_norm=resnet_time_scale_shift,
|
| 725 |
+
non_linearity=resnet_act_fn,
|
| 726 |
+
output_scale_factor=output_scale_factor,
|
| 727 |
+
pre_norm=resnet_pre_norm,
|
| 728 |
+
)
|
| 729 |
+
)
|
| 730 |
+
|
| 731 |
+
self.resnets = nn.ModuleList(resnets)
|
| 732 |
+
|
| 733 |
+
if add_downsample:
|
| 734 |
+
self.downsamplers = nn.ModuleList(
|
| 735 |
+
[
|
| 736 |
+
Downsample2D(
|
| 737 |
+
out_channels,
|
| 738 |
+
use_conv=True,
|
| 739 |
+
out_channels=out_channels,
|
| 740 |
+
padding=downsample_padding,
|
| 741 |
+
name="op",
|
| 742 |
+
)
|
| 743 |
+
]
|
| 744 |
+
)
|
| 745 |
+
else:
|
| 746 |
+
self.downsamplers = None
|
| 747 |
+
|
| 748 |
+
self.gradient_checkpointing = False
|
| 749 |
+
|
| 750 |
+
def forward(
|
| 751 |
+
self,
|
| 752 |
+
hidden_states: torch.FloatTensor,
|
| 753 |
+
temb: Optional[torch.FloatTensor] = None,
|
| 754 |
+
scale: float = 1.0,
|
| 755 |
+
) -> Tuple[torch.FloatTensor, Tuple[torch.FloatTensor, ...]]:
|
| 756 |
+
output_states = ()
|
| 757 |
+
|
| 758 |
+
for resnet in self.resnets:
|
| 759 |
+
if self.training and self.gradient_checkpointing:
|
| 760 |
+
|
| 761 |
+
def create_custom_forward(module):
|
| 762 |
+
def custom_forward(*inputs):
|
| 763 |
+
return module(*inputs)
|
| 764 |
+
|
| 765 |
+
return custom_forward
|
| 766 |
+
|
| 767 |
+
if is_torch_version(">=", "1.11.0"):
|
| 768 |
+
hidden_states = torch.utils.checkpoint.checkpoint(
|
| 769 |
+
create_custom_forward(resnet),
|
| 770 |
+
hidden_states,
|
| 771 |
+
temb,
|
| 772 |
+
use_reentrant=False,
|
| 773 |
+
)
|
| 774 |
+
else:
|
| 775 |
+
hidden_states = torch.utils.checkpoint.checkpoint(
|
| 776 |
+
create_custom_forward(resnet), hidden_states, temb
|
| 777 |
+
)
|
| 778 |
+
else:
|
| 779 |
+
hidden_states = resnet(hidden_states, temb, scale=scale)
|
| 780 |
+
|
| 781 |
+
output_states = output_states + (hidden_states,)
|
| 782 |
+
|
| 783 |
+
if self.downsamplers is not None:
|
| 784 |
+
for downsampler in self.downsamplers:
|
| 785 |
+
hidden_states = downsampler(hidden_states, scale=scale)
|
| 786 |
+
|
| 787 |
+
output_states = output_states + (hidden_states,)
|
| 788 |
+
|
| 789 |
+
return hidden_states, output_states
|
| 790 |
+
|
| 791 |
+
|
| 792 |
+
class CrossAttnUpBlock2D(nn.Module):
|
| 793 |
+
def __init__(
|
| 794 |
+
self,
|
| 795 |
+
in_channels: int,
|
| 796 |
+
out_channels: int,
|
| 797 |
+
prev_output_channel: int,
|
| 798 |
+
temb_channels: int,
|
| 799 |
+
resolution_idx: Optional[int] = None,
|
| 800 |
+
dropout: float = 0.0,
|
| 801 |
+
num_layers: int = 1,
|
| 802 |
+
transformer_layers_per_block: Union[int, Tuple[int]] = 1,
|
| 803 |
+
resnet_eps: float = 1e-6,
|
| 804 |
+
resnet_time_scale_shift: str = "default",
|
| 805 |
+
resnet_act_fn: str = "swish",
|
| 806 |
+
resnet_groups: int = 32,
|
| 807 |
+
resnet_pre_norm: bool = True,
|
| 808 |
+
num_attention_heads: int = 1,
|
| 809 |
+
cross_attention_dim: int = 1280,
|
| 810 |
+
output_scale_factor: float = 1.0,
|
| 811 |
+
add_upsample: bool = True,
|
| 812 |
+
dual_cross_attention: bool = False,
|
| 813 |
+
use_linear_projection: bool = False,
|
| 814 |
+
only_cross_attention: bool = False,
|
| 815 |
+
upcast_attention: bool = False,
|
| 816 |
+
attention_type: str = "default",
|
| 817 |
+
):
|
| 818 |
+
super().__init__()
|
| 819 |
+
resnets = []
|
| 820 |
+
attentions = []
|
| 821 |
+
|
| 822 |
+
self.has_cross_attention = True
|
| 823 |
+
self.num_attention_heads = num_attention_heads
|
| 824 |
+
|
| 825 |
+
if isinstance(transformer_layers_per_block, int):
|
| 826 |
+
transformer_layers_per_block = [transformer_layers_per_block] * num_layers
|
| 827 |
+
|
| 828 |
+
for i in range(num_layers):
|
| 829 |
+
res_skip_channels = in_channels if (i == num_layers - 1) else out_channels
|
| 830 |
+
resnet_in_channels = prev_output_channel if i == 0 else out_channels
|
| 831 |
+
|
| 832 |
+
resnets.append(
|
| 833 |
+
ResnetBlock2D(
|
| 834 |
+
in_channels=resnet_in_channels + res_skip_channels,
|
| 835 |
+
out_channels=out_channels,
|
| 836 |
+
temb_channels=temb_channels,
|
| 837 |
+
eps=resnet_eps,
|
| 838 |
+
groups=resnet_groups,
|
| 839 |
+
dropout=dropout,
|
| 840 |
+
time_embedding_norm=resnet_time_scale_shift,
|
| 841 |
+
non_linearity=resnet_act_fn,
|
| 842 |
+
output_scale_factor=output_scale_factor,
|
| 843 |
+
pre_norm=resnet_pre_norm,
|
| 844 |
+
)
|
| 845 |
+
)
|
| 846 |
+
if not dual_cross_attention:
|
| 847 |
+
attentions.append(
|
| 848 |
+
Transformer2DModel(
|
| 849 |
+
num_attention_heads,
|
| 850 |
+
out_channels // num_attention_heads,
|
| 851 |
+
in_channels=out_channels,
|
| 852 |
+
num_layers=transformer_layers_per_block[i],
|
| 853 |
+
cross_attention_dim=cross_attention_dim,
|
| 854 |
+
norm_num_groups=resnet_groups,
|
| 855 |
+
use_linear_projection=use_linear_projection,
|
| 856 |
+
only_cross_attention=only_cross_attention,
|
| 857 |
+
upcast_attention=upcast_attention,
|
| 858 |
+
attention_type=attention_type,
|
| 859 |
+
)
|
| 860 |
+
)
|
| 861 |
+
else:
|
| 862 |
+
attentions.append(
|
| 863 |
+
DualTransformer2DModel(
|
| 864 |
+
num_attention_heads,
|
| 865 |
+
out_channels // num_attention_heads,
|
| 866 |
+
in_channels=out_channels,
|
| 867 |
+
num_layers=1,
|
| 868 |
+
cross_attention_dim=cross_attention_dim,
|
| 869 |
+
norm_num_groups=resnet_groups,
|
| 870 |
+
)
|
| 871 |
+
)
|
| 872 |
+
self.attentions = nn.ModuleList(attentions)
|
| 873 |
+
self.resnets = nn.ModuleList(resnets)
|
| 874 |
+
|
| 875 |
+
if add_upsample:
|
| 876 |
+
self.upsamplers = nn.ModuleList(
|
| 877 |
+
[Upsample2D(out_channels, use_conv=True, out_channels=out_channels)]
|
| 878 |
+
)
|
| 879 |
+
else:
|
| 880 |
+
self.upsamplers = None
|
| 881 |
+
|
| 882 |
+
self.gradient_checkpointing = False
|
| 883 |
+
self.resolution_idx = resolution_idx
|
| 884 |
+
|
| 885 |
+
def forward(
|
| 886 |
+
self,
|
| 887 |
+
hidden_states: torch.FloatTensor,
|
| 888 |
+
res_hidden_states_tuple: Tuple[torch.FloatTensor, ...],
|
| 889 |
+
temb: Optional[torch.FloatTensor] = None,
|
| 890 |
+
encoder_hidden_states: Optional[torch.FloatTensor] = None,
|
| 891 |
+
cross_attention_kwargs: Optional[Dict[str, Any]] = None,
|
| 892 |
+
upsample_size: Optional[int] = None,
|
| 893 |
+
attention_mask: Optional[torch.FloatTensor] = None,
|
| 894 |
+
encoder_attention_mask: Optional[torch.FloatTensor] = None,
|
| 895 |
+
) -> torch.FloatTensor:
|
| 896 |
+
lora_scale = (
|
| 897 |
+
cross_attention_kwargs.get("scale", 1.0)
|
| 898 |
+
if cross_attention_kwargs is not None
|
| 899 |
+
else 1.0
|
| 900 |
+
)
|
| 901 |
+
is_freeu_enabled = (
|
| 902 |
+
getattr(self, "s1", None)
|
| 903 |
+
and getattr(self, "s2", None)
|
| 904 |
+
and getattr(self, "b1", None)
|
| 905 |
+
and getattr(self, "b2", None)
|
| 906 |
+
)
|
| 907 |
+
|
| 908 |
+
for resnet, attn in zip(self.resnets, self.attentions):
|
| 909 |
+
# pop res hidden states
|
| 910 |
+
res_hidden_states = res_hidden_states_tuple[-1]
|
| 911 |
+
res_hidden_states_tuple = res_hidden_states_tuple[:-1]
|
| 912 |
+
|
| 913 |
+
# FreeU: Only operate on the first two stages
|
| 914 |
+
if is_freeu_enabled:
|
| 915 |
+
hidden_states, res_hidden_states = apply_freeu(
|
| 916 |
+
self.resolution_idx,
|
| 917 |
+
hidden_states,
|
| 918 |
+
res_hidden_states,
|
| 919 |
+
s1=self.s1,
|
| 920 |
+
s2=self.s2,
|
| 921 |
+
b1=self.b1,
|
| 922 |
+
b2=self.b2,
|
| 923 |
+
)
|
| 924 |
+
|
| 925 |
+
hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)
|
| 926 |
+
|
| 927 |
+
if self.training and self.gradient_checkpointing:
|
| 928 |
+
|
| 929 |
+
def create_custom_forward(module, return_dict=None):
|
| 930 |
+
def custom_forward(*inputs):
|
| 931 |
+
if return_dict is not None:
|
| 932 |
+
return module(*inputs, return_dict=return_dict)
|
| 933 |
+
else:
|
| 934 |
+
return module(*inputs)
|
| 935 |
+
|
| 936 |
+
return custom_forward
|
| 937 |
+
|
| 938 |
+
ckpt_kwargs: Dict[str, Any] = (
|
| 939 |
+
{"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
|
| 940 |
+
)
|
| 941 |
+
hidden_states = torch.utils.checkpoint.checkpoint(
|
| 942 |
+
create_custom_forward(resnet),
|
| 943 |
+
hidden_states,
|
| 944 |
+
temb,
|
| 945 |
+
**ckpt_kwargs,
|
| 946 |
+
)
|
| 947 |
+
hidden_states, ref_feature = attn(
|
| 948 |
+
hidden_states,
|
| 949 |
+
encoder_hidden_states=encoder_hidden_states,
|
| 950 |
+
cross_attention_kwargs=cross_attention_kwargs,
|
| 951 |
+
attention_mask=attention_mask,
|
| 952 |
+
encoder_attention_mask=encoder_attention_mask,
|
| 953 |
+
return_dict=False,
|
| 954 |
+
)
|
| 955 |
+
else:
|
| 956 |
+
hidden_states = resnet(hidden_states, temb, scale=lora_scale)
|
| 957 |
+
hidden_states, ref_feature = attn(
|
| 958 |
+
hidden_states,
|
| 959 |
+
encoder_hidden_states=encoder_hidden_states,
|
| 960 |
+
cross_attention_kwargs=cross_attention_kwargs,
|
| 961 |
+
attention_mask=attention_mask,
|
| 962 |
+
encoder_attention_mask=encoder_attention_mask,
|
| 963 |
+
return_dict=False,
|
| 964 |
+
)
|
| 965 |
+
|
| 966 |
+
if self.upsamplers is not None:
|
| 967 |
+
for upsampler in self.upsamplers:
|
| 968 |
+
hidden_states = upsampler(
|
| 969 |
+
hidden_states, upsample_size, scale=lora_scale
|
| 970 |
+
)
|
| 971 |
+
|
| 972 |
+
return hidden_states
|
| 973 |
+
|
| 974 |
+
|
| 975 |
+
class UpBlock2D(nn.Module):
|
| 976 |
+
def __init__(
|
| 977 |
+
self,
|
| 978 |
+
in_channels: int,
|
| 979 |
+
prev_output_channel: int,
|
| 980 |
+
out_channels: int,
|
| 981 |
+
temb_channels: int,
|
| 982 |
+
resolution_idx: Optional[int] = None,
|
| 983 |
+
dropout: float = 0.0,
|
| 984 |
+
num_layers: int = 1,
|
| 985 |
+
resnet_eps: float = 1e-6,
|
| 986 |
+
resnet_time_scale_shift: str = "default",
|
| 987 |
+
resnet_act_fn: str = "swish",
|
| 988 |
+
resnet_groups: int = 32,
|
| 989 |
+
resnet_pre_norm: bool = True,
|
| 990 |
+
output_scale_factor: float = 1.0,
|
| 991 |
+
add_upsample: bool = True,
|
| 992 |
+
):
|
| 993 |
+
super().__init__()
|
| 994 |
+
resnets = []
|
| 995 |
+
|
| 996 |
+
for i in range(num_layers):
|
| 997 |
+
res_skip_channels = in_channels if (i == num_layers - 1) else out_channels
|
| 998 |
+
resnet_in_channels = prev_output_channel if i == 0 else out_channels
|
| 999 |
+
|
| 1000 |
+
resnets.append(
|
| 1001 |
+
ResnetBlock2D(
|
| 1002 |
+
in_channels=resnet_in_channels + res_skip_channels,
|
| 1003 |
+
out_channels=out_channels,
|
| 1004 |
+
temb_channels=temb_channels,
|
| 1005 |
+
eps=resnet_eps,
|
| 1006 |
+
groups=resnet_groups,
|
| 1007 |
+
dropout=dropout,
|
| 1008 |
+
time_embedding_norm=resnet_time_scale_shift,
|
| 1009 |
+
non_linearity=resnet_act_fn,
|
| 1010 |
+
output_scale_factor=output_scale_factor,
|
| 1011 |
+
pre_norm=resnet_pre_norm,
|
| 1012 |
+
)
|
| 1013 |
+
)
|
| 1014 |
+
|
| 1015 |
+
self.resnets = nn.ModuleList(resnets)
|
| 1016 |
+
|
| 1017 |
+
if add_upsample:
|
| 1018 |
+
self.upsamplers = nn.ModuleList(
|
| 1019 |
+
[Upsample2D(out_channels, use_conv=True, out_channels=out_channels)]
|
| 1020 |
+
)
|
| 1021 |
+
else:
|
| 1022 |
+
self.upsamplers = None
|
| 1023 |
+
|
| 1024 |
+
self.gradient_checkpointing = False
|
| 1025 |
+
self.resolution_idx = resolution_idx
|
| 1026 |
+
|
| 1027 |
+
def forward(
|
| 1028 |
+
self,
|
| 1029 |
+
hidden_states: torch.FloatTensor,
|
| 1030 |
+
res_hidden_states_tuple: Tuple[torch.FloatTensor, ...],
|
| 1031 |
+
temb: Optional[torch.FloatTensor] = None,
|
| 1032 |
+
upsample_size: Optional[int] = None,
|
| 1033 |
+
scale: float = 1.0,
|
| 1034 |
+
) -> torch.FloatTensor:
|
| 1035 |
+
is_freeu_enabled = (
|
| 1036 |
+
getattr(self, "s1", None)
|
| 1037 |
+
and getattr(self, "s2", None)
|
| 1038 |
+
and getattr(self, "b1", None)
|
| 1039 |
+
and getattr(self, "b2", None)
|
| 1040 |
+
)
|
| 1041 |
+
|
| 1042 |
+
for resnet in self.resnets:
|
| 1043 |
+
# pop res hidden states
|
| 1044 |
+
res_hidden_states = res_hidden_states_tuple[-1]
|
| 1045 |
+
res_hidden_states_tuple = res_hidden_states_tuple[:-1]
|
| 1046 |
+
|
| 1047 |
+
# FreeU: Only operate on the first two stages
|
| 1048 |
+
if is_freeu_enabled:
|
| 1049 |
+
hidden_states, res_hidden_states = apply_freeu(
|
| 1050 |
+
self.resolution_idx,
|
| 1051 |
+
hidden_states,
|
| 1052 |
+
res_hidden_states,
|
| 1053 |
+
s1=self.s1,
|
| 1054 |
+
s2=self.s2,
|
| 1055 |
+
b1=self.b1,
|
| 1056 |
+
b2=self.b2,
|
| 1057 |
+
)
|
| 1058 |
+
|
| 1059 |
+
hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)
|
| 1060 |
+
|
| 1061 |
+
if self.training and self.gradient_checkpointing:
|
| 1062 |
+
|
| 1063 |
+
def create_custom_forward(module):
|
| 1064 |
+
def custom_forward(*inputs):
|
| 1065 |
+
return module(*inputs)
|
| 1066 |
+
|
| 1067 |
+
return custom_forward
|
| 1068 |
+
|
| 1069 |
+
if is_torch_version(">=", "1.11.0"):
|
| 1070 |
+
hidden_states = torch.utils.checkpoint.checkpoint(
|
| 1071 |
+
create_custom_forward(resnet),
|
| 1072 |
+
hidden_states,
|
| 1073 |
+
temb,
|
| 1074 |
+
use_reentrant=False,
|
| 1075 |
+
)
|
| 1076 |
+
else:
|
| 1077 |
+
hidden_states = torch.utils.checkpoint.checkpoint(
|
| 1078 |
+
create_custom_forward(resnet), hidden_states, temb
|
| 1079 |
+
)
|
| 1080 |
+
else:
|
| 1081 |
+
hidden_states = resnet(hidden_states, temb, scale=scale)
|
| 1082 |
+
|
| 1083 |
+
if self.upsamplers is not None:
|
| 1084 |
+
for upsampler in self.upsamplers:
|
| 1085 |
+
hidden_states = upsampler(hidden_states, upsample_size, scale=scale)
|
| 1086 |
+
|
| 1087 |
+
return hidden_states
|
genwarp/models/unet_2d_condition.py
ADDED
|
@@ -0,0 +1,1324 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# This code is adapted from below and then modified.
|
| 2 |
+
# -----------------------------------------------------------------------------
|
| 3 |
+
# Moore-AnimateAnyone
|
| 4 |
+
# Apache License, Version 2.0
|
| 5 |
+
# Copyright @2023-2024 Moore Threads Technology Co., Ltd.
|
| 6 |
+
# https://github.com/MooreThreads/Moore-AnimateAnyone
|
| 7 |
+
# -----------------------------------------------------------------------------
|
| 8 |
+
# Diffusers
|
| 9 |
+
# Apache License, Version 2.0
|
| 10 |
+
# Copyright (c) Hugging Face Inc.
|
| 11 |
+
# https://github.com/huggingface/diffusers
|
| 12 |
+
# ==============================================================================
|
| 13 |
+
|
| 14 |
+
# Adapted from https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/unet_2d_condition.py
|
| 15 |
+
from dataclasses import dataclass
|
| 16 |
+
from typing import Any, Dict, List, Optional, Tuple, Union
|
| 17 |
+
|
| 18 |
+
import torch
|
| 19 |
+
import torch.nn as nn
|
| 20 |
+
import torch.utils.checkpoint
|
| 21 |
+
from diffusers.configuration_utils import ConfigMixin, register_to_config
|
| 22 |
+
from diffusers.loaders import UNet2DConditionLoadersMixin
|
| 23 |
+
from diffusers.models.activations import get_activation
|
| 24 |
+
from diffusers.models.attention_processor import (
|
| 25 |
+
ADDED_KV_ATTENTION_PROCESSORS,
|
| 26 |
+
CROSS_ATTENTION_PROCESSORS,
|
| 27 |
+
AttentionProcessor,
|
| 28 |
+
AttnAddedKVProcessor,
|
| 29 |
+
AttnProcessor,
|
| 30 |
+
)
|
| 31 |
+
from diffusers.models.embeddings import (
|
| 32 |
+
GaussianFourierProjection,
|
| 33 |
+
ImageHintTimeEmbedding,
|
| 34 |
+
ImageProjection,
|
| 35 |
+
ImageTimeEmbedding,
|
| 36 |
+
# PositionNet,
|
| 37 |
+
TextImageProjection,
|
| 38 |
+
TextImageTimeEmbedding,
|
| 39 |
+
TextTimeEmbedding,
|
| 40 |
+
TimestepEmbedding,
|
| 41 |
+
Timesteps,
|
| 42 |
+
)
|
| 43 |
+
from diffusers.models.modeling_utils import ModelMixin
|
| 44 |
+
from diffusers.utils import (
|
| 45 |
+
USE_PEFT_BACKEND,
|
| 46 |
+
BaseOutput,
|
| 47 |
+
deprecate,
|
| 48 |
+
logging,
|
| 49 |
+
scale_lora_layers,
|
| 50 |
+
unscale_lora_layers,
|
| 51 |
+
)
|
| 52 |
+
|
| 53 |
+
from .unet_2d_blocks import (
|
| 54 |
+
UNetMidBlock2D,
|
| 55 |
+
UNetMidBlock2DCrossAttn,
|
| 56 |
+
get_down_block,
|
| 57 |
+
get_up_block,
|
| 58 |
+
)
|
| 59 |
+
|
| 60 |
+
logger = logging.get_logger(__name__) # pylint: disable=invalid-name
|
| 61 |
+
|
| 62 |
+
|
| 63 |
+
@dataclass
|
| 64 |
+
class UNet2DConditionOutput(BaseOutput):
|
| 65 |
+
"""
|
| 66 |
+
The output of [`UNet2DConditionModel`].
|
| 67 |
+
|
| 68 |
+
Args:
|
| 69 |
+
sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`):
|
| 70 |
+
The hidden states output conditioned on `encoder_hidden_states` input. Output of last layer of model.
|
| 71 |
+
"""
|
| 72 |
+
|
| 73 |
+
sample: torch.FloatTensor = None
|
| 74 |
+
ref_features: Tuple[torch.FloatTensor] = None
|
| 75 |
+
|
| 76 |
+
|
| 77 |
+
class UNet2DConditionModel(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin):
|
| 78 |
+
r"""
|
| 79 |
+
A conditional 2D UNet model that takes a noisy sample, conditional state, and a timestep and returns a sample
|
| 80 |
+
shaped output.
|
| 81 |
+
|
| 82 |
+
This model inherits from [`ModelMixin`]. Check the superclass documentation for it's generic methods implemented
|
| 83 |
+
for all models (such as downloading or saving).
|
| 84 |
+
|
| 85 |
+
Parameters:
|
| 86 |
+
sample_size (`int` or `Tuple[int, int]`, *optional*, defaults to `None`):
|
| 87 |
+
Height and width of input/output sample.
|
| 88 |
+
in_channels (`int`, *optional*, defaults to 4): Number of channels in the input sample.
|
| 89 |
+
out_channels (`int`, *optional*, defaults to 4): Number of channels in the output.
|
| 90 |
+
center_input_sample (`bool`, *optional*, defaults to `False`): Whether to center the input sample.
|
| 91 |
+
flip_sin_to_cos (`bool`, *optional*, defaults to `False`):
|
| 92 |
+
Whether to flip the sin to cos in the time embedding.
|
| 93 |
+
freq_shift (`int`, *optional*, defaults to 0): The frequency shift to apply to the time embedding.
|
| 94 |
+
down_block_types (`Tuple[str]`, *optional*, defaults to `("CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "DownBlock2D")`):
|
| 95 |
+
The tuple of downsample blocks to use.
|
| 96 |
+
mid_block_type (`str`, *optional*, defaults to `"UNetMidBlock2DCrossAttn"`):
|
| 97 |
+
Block type for middle of UNet, it can be one of `UNetMidBlock2DCrossAttn`, `UNetMidBlock2D`, or
|
| 98 |
+
`UNetMidBlock2DSimpleCrossAttn`. If `None`, the mid block layer is skipped.
|
| 99 |
+
up_block_types (`Tuple[str]`, *optional*, defaults to `("UpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D")`):
|
| 100 |
+
The tuple of upsample blocks to use.
|
| 101 |
+
only_cross_attention(`bool` or `Tuple[bool]`, *optional*, default to `False`):
|
| 102 |
+
Whether to include self-attention in the basic transformer blocks, see
|
| 103 |
+
[`~models.attention.BasicTransformerBlock`].
|
| 104 |
+
block_out_channels (`Tuple[int]`, *optional*, defaults to `(320, 640, 1280, 1280)`):
|
| 105 |
+
The tuple of output channels for each block.
|
| 106 |
+
layers_per_block (`int`, *optional*, defaults to 2): The number of layers per block.
|
| 107 |
+
downsample_padding (`int`, *optional*, defaults to 1): The padding to use for the downsampling convolution.
|
| 108 |
+
mid_block_scale_factor (`float`, *optional*, defaults to 1.0): The scale factor to use for the mid block.
|
| 109 |
+
dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
|
| 110 |
+
act_fn (`str`, *optional*, defaults to `"silu"`): The activation function to use.
|
| 111 |
+
norm_num_groups (`int`, *optional*, defaults to 32): The number of groups to use for the normalization.
|
| 112 |
+
If `None`, normalization and activation layers is skipped in post-processing.
|
| 113 |
+
norm_eps (`float`, *optional*, defaults to 1e-5): The epsilon to use for the normalization.
|
| 114 |
+
cross_attention_dim (`int` or `Tuple[int]`, *optional*, defaults to 1280):
|
| 115 |
+
The dimension of the cross attention features.
|
| 116 |
+
transformer_layers_per_block (`int`, `Tuple[int]`, or `Tuple[Tuple]` , *optional*, defaults to 1):
|
| 117 |
+
The number of transformer blocks of type [`~models.attention.BasicTransformerBlock`]. Only relevant for
|
| 118 |
+
[`~models.unet_2d_blocks.CrossAttnDownBlock2D`], [`~models.unet_2d_blocks.CrossAttnUpBlock2D`],
|
| 119 |
+
[`~models.unet_2d_blocks.UNetMidBlock2DCrossAttn`].
|
| 120 |
+
reverse_transformer_layers_per_block : (`Tuple[Tuple]`, *optional*, defaults to None):
|
| 121 |
+
The number of transformer blocks of type [`~models.attention.BasicTransformerBlock`], in the upsampling
|
| 122 |
+
blocks of the U-Net. Only relevant if `transformer_layers_per_block` is of type `Tuple[Tuple]` and for
|
| 123 |
+
[`~models.unet_2d_blocks.CrossAttnDownBlock2D`], [`~models.unet_2d_blocks.CrossAttnUpBlock2D`],
|
| 124 |
+
[`~models.unet_2d_blocks.UNetMidBlock2DCrossAttn`].
|
| 125 |
+
encoder_hid_dim (`int`, *optional*, defaults to None):
|
| 126 |
+
If `encoder_hid_dim_type` is defined, `encoder_hidden_states` will be projected from `encoder_hid_dim`
|
| 127 |
+
dimension to `cross_attention_dim`.
|
| 128 |
+
encoder_hid_dim_type (`str`, *optional*, defaults to `None`):
|
| 129 |
+
If given, the `encoder_hidden_states` and potentially other embeddings are down-projected to text
|
| 130 |
+
embeddings of dimension `cross_attention` according to `encoder_hid_dim_type`.
|
| 131 |
+
attention_head_dim (`int`, *optional*, defaults to 8): The dimension of the attention heads.
|
| 132 |
+
num_attention_heads (`int`, *optional*):
|
| 133 |
+
The number of attention heads. If not defined, defaults to `attention_head_dim`
|
| 134 |
+
resnet_time_scale_shift (`str`, *optional*, defaults to `"default"`): Time scale shift config
|
| 135 |
+
for ResNet blocks (see [`~models.resnet.ResnetBlock2D`]). Choose from `default` or `scale_shift`.
|
| 136 |
+
class_embed_type (`str`, *optional*, defaults to `None`):
|
| 137 |
+
The type of class embedding to use which is ultimately summed with the time embeddings. Choose from `None`,
|
| 138 |
+
`"timestep"`, `"identity"`, `"projection"`, or `"simple_projection"`.
|
| 139 |
+
addition_embed_type (`str`, *optional*, defaults to `None`):
|
| 140 |
+
Configures an optional embedding which will be summed with the time embeddings. Choose from `None` or
|
| 141 |
+
"text". "text" will use the `TextTimeEmbedding` layer.
|
| 142 |
+
addition_time_embed_dim: (`int`, *optional*, defaults to `None`):
|
| 143 |
+
Dimension for the timestep embeddings.
|
| 144 |
+
num_class_embeds (`int`, *optional*, defaults to `None`):
|
| 145 |
+
Input dimension of the learnable embedding matrix to be projected to `time_embed_dim`, when performing
|
| 146 |
+
class conditioning with `class_embed_type` equal to `None`.
|
| 147 |
+
time_embedding_type (`str`, *optional*, defaults to `positional`):
|
| 148 |
+
The type of position embedding to use for timesteps. Choose from `positional` or `fourier`.
|
| 149 |
+
time_embedding_dim (`int`, *optional*, defaults to `None`):
|
| 150 |
+
An optional override for the dimension of the projected time embedding.
|
| 151 |
+
time_embedding_act_fn (`str`, *optional*, defaults to `None`):
|
| 152 |
+
Optional activation function to use only once on the time embeddings before they are passed to the rest of
|
| 153 |
+
the UNet. Choose from `silu`, `mish`, `gelu`, and `swish`.
|
| 154 |
+
timestep_post_act (`str`, *optional*, defaults to `None`):
|
| 155 |
+
The second activation function to use in timestep embedding. Choose from `silu`, `mish` and `gelu`.
|
| 156 |
+
time_cond_proj_dim (`int`, *optional*, defaults to `None`):
|
| 157 |
+
The dimension of `cond_proj` layer in the timestep embedding.
|
| 158 |
+
conv_in_kernel (`int`, *optional*, default to `3`): The kernel size of `conv_in` layer. conv_out_kernel (`int`,
|
| 159 |
+
*optional*, default to `3`): The kernel size of `conv_out` layer. projection_class_embeddings_input_dim (`int`,
|
| 160 |
+
*optional*): The dimension of the `class_labels` input when
|
| 161 |
+
`class_embed_type="projection"`. Required when `class_embed_type="projection"`.
|
| 162 |
+
class_embeddings_concat (`bool`, *optional*, defaults to `False`): Whether to concatenate the time
|
| 163 |
+
embeddings with the class embeddings.
|
| 164 |
+
mid_block_only_cross_attention (`bool`, *optional*, defaults to `None`):
|
| 165 |
+
Whether to use cross attention with the mid block when using the `UNetMidBlock2DSimpleCrossAttn`. If
|
| 166 |
+
`only_cross_attention` is given as a single boolean and `mid_block_only_cross_attention` is `None`, the
|
| 167 |
+
`only_cross_attention` value is used as the value for `mid_block_only_cross_attention`. Default to `False`
|
| 168 |
+
otherwise.
|
| 169 |
+
"""
|
| 170 |
+
|
| 171 |
+
_supports_gradient_checkpointing = True
|
| 172 |
+
|
| 173 |
+
@register_to_config
|
| 174 |
+
def __init__(
|
| 175 |
+
self,
|
| 176 |
+
sample_size: Optional[int] = None,
|
| 177 |
+
in_channels: int = 4,
|
| 178 |
+
out_channels: int = 4,
|
| 179 |
+
center_input_sample: bool = False,
|
| 180 |
+
flip_sin_to_cos: bool = True,
|
| 181 |
+
freq_shift: int = 0,
|
| 182 |
+
down_block_types: Tuple[str] = (
|
| 183 |
+
"CrossAttnDownBlock2D",
|
| 184 |
+
"CrossAttnDownBlock2D",
|
| 185 |
+
"CrossAttnDownBlock2D",
|
| 186 |
+
"DownBlock2D",
|
| 187 |
+
),
|
| 188 |
+
mid_block_type: Optional[str] = "UNetMidBlock2DCrossAttn",
|
| 189 |
+
up_block_types: Tuple[str] = (
|
| 190 |
+
"UpBlock2D",
|
| 191 |
+
"CrossAttnUpBlock2D",
|
| 192 |
+
"CrossAttnUpBlock2D",
|
| 193 |
+
"CrossAttnUpBlock2D",
|
| 194 |
+
),
|
| 195 |
+
only_cross_attention: Union[bool, Tuple[bool]] = False,
|
| 196 |
+
block_out_channels: Tuple[int] = (320, 640, 1280, 1280),
|
| 197 |
+
layers_per_block: Union[int, Tuple[int]] = 2,
|
| 198 |
+
downsample_padding: int = 1,
|
| 199 |
+
mid_block_scale_factor: float = 1,
|
| 200 |
+
dropout: float = 0.0,
|
| 201 |
+
act_fn: str = "silu",
|
| 202 |
+
norm_num_groups: Optional[int] = 32,
|
| 203 |
+
norm_eps: float = 1e-5,
|
| 204 |
+
cross_attention_dim: Union[int, Tuple[int]] = 1280,
|
| 205 |
+
transformer_layers_per_block: Union[int, Tuple[int], Tuple[Tuple]] = 1,
|
| 206 |
+
reverse_transformer_layers_per_block: Optional[Tuple[Tuple[int]]] = None,
|
| 207 |
+
encoder_hid_dim: Optional[int] = None,
|
| 208 |
+
encoder_hid_dim_type: Optional[str] = None,
|
| 209 |
+
attention_head_dim: Union[int, Tuple[int]] = 8,
|
| 210 |
+
num_attention_heads: Optional[Union[int, Tuple[int]]] = None,
|
| 211 |
+
dual_cross_attention: bool = False,
|
| 212 |
+
use_linear_projection: bool = False,
|
| 213 |
+
class_embed_type: Optional[str] = None,
|
| 214 |
+
addition_embed_type: Optional[str] = None,
|
| 215 |
+
addition_time_embed_dim: Optional[int] = None,
|
| 216 |
+
num_class_embeds: Optional[int] = None,
|
| 217 |
+
upcast_attention: bool = False,
|
| 218 |
+
resnet_time_scale_shift: str = "default",
|
| 219 |
+
resnet_skip_time_act: bool = False,
|
| 220 |
+
resnet_out_scale_factor: int = 1.0,
|
| 221 |
+
time_embedding_type: str = "positional",
|
| 222 |
+
time_embedding_dim: Optional[int] = None,
|
| 223 |
+
time_embedding_act_fn: Optional[str] = None,
|
| 224 |
+
timestep_post_act: Optional[str] = None,
|
| 225 |
+
time_cond_proj_dim: Optional[int] = None,
|
| 226 |
+
conv_in_kernel: int = 3,
|
| 227 |
+
conv_out_kernel: int = 3,
|
| 228 |
+
projection_class_embeddings_input_dim: Optional[int] = None,
|
| 229 |
+
attention_type: str = "default",
|
| 230 |
+
class_embeddings_concat: bool = False,
|
| 231 |
+
mid_block_only_cross_attention: Optional[bool] = None,
|
| 232 |
+
cross_attention_norm: Optional[str] = None,
|
| 233 |
+
addition_embed_type_num_heads=64,
|
| 234 |
+
):
|
| 235 |
+
super().__init__()
|
| 236 |
+
|
| 237 |
+
self.sample_size = sample_size
|
| 238 |
+
|
| 239 |
+
if num_attention_heads is not None:
|
| 240 |
+
raise ValueError(
|
| 241 |
+
"At the moment it is not possible to define the number of attention heads via `num_attention_heads` because of a naming issue as described in https://github.com/huggingface/diffusers/issues/2011#issuecomment-1547958131. Passing `num_attention_heads` will only be supported in diffusers v0.19."
|
| 242 |
+
)
|
| 243 |
+
|
| 244 |
+
# If `num_attention_heads` is not defined (which is the case for most models)
|
| 245 |
+
# it will default to `attention_head_dim`. This looks weird upon first reading it and it is.
|
| 246 |
+
# The reason for this behavior is to correct for incorrectly named variables that were introduced
|
| 247 |
+
# when this library was created. The incorrect naming was only discovered much later in https://github.com/huggingface/diffusers/issues/2011#issuecomment-1547958131
|
| 248 |
+
# Changing `attention_head_dim` to `num_attention_heads` for 40,000+ configurations is too backwards breaking
|
| 249 |
+
# which is why we correct for the naming here.
|
| 250 |
+
num_attention_heads = num_attention_heads or attention_head_dim
|
| 251 |
+
|
| 252 |
+
# Check inputs
|
| 253 |
+
if len(down_block_types) != len(up_block_types):
|
| 254 |
+
raise ValueError(
|
| 255 |
+
f"Must provide the same number of `down_block_types` as `up_block_types`. `down_block_types`: {down_block_types}. `up_block_types`: {up_block_types}."
|
| 256 |
+
)
|
| 257 |
+
|
| 258 |
+
if len(block_out_channels) != len(down_block_types):
|
| 259 |
+
raise ValueError(
|
| 260 |
+
f"Must provide the same number of `block_out_channels` as `down_block_types`. `block_out_channels`: {block_out_channels}. `down_block_types`: {down_block_types}."
|
| 261 |
+
)
|
| 262 |
+
|
| 263 |
+
if not isinstance(only_cross_attention, bool) and len(
|
| 264 |
+
only_cross_attention
|
| 265 |
+
) != len(down_block_types):
|
| 266 |
+
raise ValueError(
|
| 267 |
+
f"Must provide the same number of `only_cross_attention` as `down_block_types`. `only_cross_attention`: {only_cross_attention}. `down_block_types`: {down_block_types}."
|
| 268 |
+
)
|
| 269 |
+
|
| 270 |
+
if not isinstance(num_attention_heads, int) and len(num_attention_heads) != len(
|
| 271 |
+
down_block_types
|
| 272 |
+
):
|
| 273 |
+
raise ValueError(
|
| 274 |
+
f"Must provide the same number of `num_attention_heads` as `down_block_types`. `num_attention_heads`: {num_attention_heads}. `down_block_types`: {down_block_types}."
|
| 275 |
+
)
|
| 276 |
+
|
| 277 |
+
if not isinstance(attention_head_dim, int) and len(attention_head_dim) != len(
|
| 278 |
+
down_block_types
|
| 279 |
+
):
|
| 280 |
+
raise ValueError(
|
| 281 |
+
f"Must provide the same number of `attention_head_dim` as `down_block_types`. `attention_head_dim`: {attention_head_dim}. `down_block_types`: {down_block_types}."
|
| 282 |
+
)
|
| 283 |
+
|
| 284 |
+
if isinstance(cross_attention_dim, list) and len(cross_attention_dim) != len(
|
| 285 |
+
down_block_types
|
| 286 |
+
):
|
| 287 |
+
raise ValueError(
|
| 288 |
+
f"Must provide the same number of `cross_attention_dim` as `down_block_types`. `cross_attention_dim`: {cross_attention_dim}. `down_block_types`: {down_block_types}."
|
| 289 |
+
)
|
| 290 |
+
|
| 291 |
+
if not isinstance(layers_per_block, int) and len(layers_per_block) != len(
|
| 292 |
+
down_block_types
|
| 293 |
+
):
|
| 294 |
+
raise ValueError(
|
| 295 |
+
f"Must provide the same number of `layers_per_block` as `down_block_types`. `layers_per_block`: {layers_per_block}. `down_block_types`: {down_block_types}."
|
| 296 |
+
)
|
| 297 |
+
if (
|
| 298 |
+
isinstance(transformer_layers_per_block, list)
|
| 299 |
+
and reverse_transformer_layers_per_block is None
|
| 300 |
+
):
|
| 301 |
+
for layer_number_per_block in transformer_layers_per_block:
|
| 302 |
+
if isinstance(layer_number_per_block, list):
|
| 303 |
+
raise ValueError(
|
| 304 |
+
"Must provide 'reverse_transformer_layers_per_block` if using asymmetrical UNet."
|
| 305 |
+
)
|
| 306 |
+
|
| 307 |
+
# input
|
| 308 |
+
conv_in_padding = (conv_in_kernel - 1) // 2
|
| 309 |
+
self.conv_in = nn.Conv2d(
|
| 310 |
+
in_channels,
|
| 311 |
+
block_out_channels[0],
|
| 312 |
+
kernel_size=conv_in_kernel,
|
| 313 |
+
padding=conv_in_padding,
|
| 314 |
+
)
|
| 315 |
+
|
| 316 |
+
# time
|
| 317 |
+
if time_embedding_type == "fourier":
|
| 318 |
+
time_embed_dim = time_embedding_dim or block_out_channels[0] * 2
|
| 319 |
+
if time_embed_dim % 2 != 0:
|
| 320 |
+
raise ValueError(
|
| 321 |
+
f"`time_embed_dim` should be divisible by 2, but is {time_embed_dim}."
|
| 322 |
+
)
|
| 323 |
+
self.time_proj = GaussianFourierProjection(
|
| 324 |
+
time_embed_dim // 2,
|
| 325 |
+
set_W_to_weight=False,
|
| 326 |
+
log=False,
|
| 327 |
+
flip_sin_to_cos=flip_sin_to_cos,
|
| 328 |
+
)
|
| 329 |
+
timestep_input_dim = time_embed_dim
|
| 330 |
+
elif time_embedding_type == "positional":
|
| 331 |
+
time_embed_dim = time_embedding_dim or block_out_channels[0] * 4
|
| 332 |
+
|
| 333 |
+
self.time_proj = Timesteps(
|
| 334 |
+
block_out_channels[0], flip_sin_to_cos, freq_shift
|
| 335 |
+
)
|
| 336 |
+
timestep_input_dim = block_out_channels[0]
|
| 337 |
+
else:
|
| 338 |
+
raise ValueError(
|
| 339 |
+
f"{time_embedding_type} does not exist. Please make sure to use one of `fourier` or `positional`."
|
| 340 |
+
)
|
| 341 |
+
|
| 342 |
+
self.time_embedding = TimestepEmbedding(
|
| 343 |
+
timestep_input_dim,
|
| 344 |
+
time_embed_dim,
|
| 345 |
+
act_fn=act_fn,
|
| 346 |
+
post_act_fn=timestep_post_act,
|
| 347 |
+
cond_proj_dim=time_cond_proj_dim,
|
| 348 |
+
)
|
| 349 |
+
|
| 350 |
+
if encoder_hid_dim_type is None and encoder_hid_dim is not None:
|
| 351 |
+
encoder_hid_dim_type = "text_proj"
|
| 352 |
+
self.register_to_config(encoder_hid_dim_type=encoder_hid_dim_type)
|
| 353 |
+
logger.info(
|
| 354 |
+
"encoder_hid_dim_type defaults to 'text_proj' as `encoder_hid_dim` is defined."
|
| 355 |
+
)
|
| 356 |
+
|
| 357 |
+
if encoder_hid_dim is None and encoder_hid_dim_type is not None:
|
| 358 |
+
raise ValueError(
|
| 359 |
+
f"`encoder_hid_dim` has to be defined when `encoder_hid_dim_type` is set to {encoder_hid_dim_type}."
|
| 360 |
+
)
|
| 361 |
+
|
| 362 |
+
if encoder_hid_dim_type == "text_proj":
|
| 363 |
+
self.encoder_hid_proj = nn.Linear(encoder_hid_dim, cross_attention_dim)
|
| 364 |
+
elif encoder_hid_dim_type == "text_image_proj":
|
| 365 |
+
# image_embed_dim DOESN'T have to be `cross_attention_dim`. To not clutter the __init__ too much
|
| 366 |
+
# they are set to `cross_attention_dim` here as this is exactly the required dimension for the currently only use
|
| 367 |
+
# case when `addition_embed_type == "text_image_proj"` (Kadinsky 2.1)`
|
| 368 |
+
self.encoder_hid_proj = TextImageProjection(
|
| 369 |
+
text_embed_dim=encoder_hid_dim,
|
| 370 |
+
image_embed_dim=cross_attention_dim,
|
| 371 |
+
cross_attention_dim=cross_attention_dim,
|
| 372 |
+
)
|
| 373 |
+
elif encoder_hid_dim_type == "image_proj":
|
| 374 |
+
# Kandinsky 2.2
|
| 375 |
+
self.encoder_hid_proj = ImageProjection(
|
| 376 |
+
image_embed_dim=encoder_hid_dim,
|
| 377 |
+
cross_attention_dim=cross_attention_dim,
|
| 378 |
+
)
|
| 379 |
+
elif encoder_hid_dim_type is not None:
|
| 380 |
+
raise ValueError(
|
| 381 |
+
f"encoder_hid_dim_type: {encoder_hid_dim_type} must be None, 'text_proj' or 'text_image_proj'."
|
| 382 |
+
)
|
| 383 |
+
else:
|
| 384 |
+
self.encoder_hid_proj = None
|
| 385 |
+
|
| 386 |
+
# class embedding
|
| 387 |
+
if class_embed_type is None and num_class_embeds is not None:
|
| 388 |
+
self.class_embedding = nn.Embedding(num_class_embeds, time_embed_dim)
|
| 389 |
+
elif class_embed_type == "timestep":
|
| 390 |
+
self.class_embedding = TimestepEmbedding(
|
| 391 |
+
timestep_input_dim, time_embed_dim, act_fn=act_fn
|
| 392 |
+
)
|
| 393 |
+
elif class_embed_type == "identity":
|
| 394 |
+
self.class_embedding = nn.Identity(time_embed_dim, time_embed_dim)
|
| 395 |
+
elif class_embed_type == "projection":
|
| 396 |
+
if projection_class_embeddings_input_dim is None:
|
| 397 |
+
raise ValueError(
|
| 398 |
+
"`class_embed_type`: 'projection' requires `projection_class_embeddings_input_dim` be set"
|
| 399 |
+
)
|
| 400 |
+
# The projection `class_embed_type` is the same as the timestep `class_embed_type` except
|
| 401 |
+
# 1. the `class_labels` inputs are not first converted to sinusoidal embeddings
|
| 402 |
+
# 2. it projects from an arbitrary input dimension.
|
| 403 |
+
#
|
| 404 |
+
# Note that `TimestepEmbedding` is quite general, being mainly linear layers and activations.
|
| 405 |
+
# When used for embedding actual timesteps, the timesteps are first converted to sinusoidal embeddings.
|
| 406 |
+
# As a result, `TimestepEmbedding` can be passed arbitrary vectors.
|
| 407 |
+
self.class_embedding = TimestepEmbedding(
|
| 408 |
+
projection_class_embeddings_input_dim, time_embed_dim
|
| 409 |
+
)
|
| 410 |
+
elif class_embed_type == "simple_projection":
|
| 411 |
+
if projection_class_embeddings_input_dim is None:
|
| 412 |
+
raise ValueError(
|
| 413 |
+
"`class_embed_type`: 'simple_projection' requires `projection_class_embeddings_input_dim` be set"
|
| 414 |
+
)
|
| 415 |
+
self.class_embedding = nn.Linear(
|
| 416 |
+
projection_class_embeddings_input_dim, time_embed_dim
|
| 417 |
+
)
|
| 418 |
+
else:
|
| 419 |
+
self.class_embedding = None
|
| 420 |
+
|
| 421 |
+
if addition_embed_type == "text":
|
| 422 |
+
if encoder_hid_dim is not None:
|
| 423 |
+
text_time_embedding_from_dim = encoder_hid_dim
|
| 424 |
+
else:
|
| 425 |
+
text_time_embedding_from_dim = cross_attention_dim
|
| 426 |
+
|
| 427 |
+
self.add_embedding = TextTimeEmbedding(
|
| 428 |
+
text_time_embedding_from_dim,
|
| 429 |
+
time_embed_dim,
|
| 430 |
+
num_heads=addition_embed_type_num_heads,
|
| 431 |
+
)
|
| 432 |
+
elif addition_embed_type == "text_image":
|
| 433 |
+
# text_embed_dim and image_embed_dim DON'T have to be `cross_attention_dim`. To not clutter the __init__ too much
|
| 434 |
+
# they are set to `cross_attention_dim` here as this is exactly the required dimension for the currently only use
|
| 435 |
+
# case when `addition_embed_type == "text_image"` (Kadinsky 2.1)`
|
| 436 |
+
self.add_embedding = TextImageTimeEmbedding(
|
| 437 |
+
text_embed_dim=cross_attention_dim,
|
| 438 |
+
image_embed_dim=cross_attention_dim,
|
| 439 |
+
time_embed_dim=time_embed_dim,
|
| 440 |
+
)
|
| 441 |
+
elif addition_embed_type == "text_time":
|
| 442 |
+
self.add_time_proj = Timesteps(
|
| 443 |
+
addition_time_embed_dim, flip_sin_to_cos, freq_shift
|
| 444 |
+
)
|
| 445 |
+
self.add_embedding = TimestepEmbedding(
|
| 446 |
+
projection_class_embeddings_input_dim, time_embed_dim
|
| 447 |
+
)
|
| 448 |
+
elif addition_embed_type == "image":
|
| 449 |
+
# Kandinsky 2.2
|
| 450 |
+
self.add_embedding = ImageTimeEmbedding(
|
| 451 |
+
image_embed_dim=encoder_hid_dim, time_embed_dim=time_embed_dim
|
| 452 |
+
)
|
| 453 |
+
elif addition_embed_type == "image_hint":
|
| 454 |
+
# Kandinsky 2.2 ControlNet
|
| 455 |
+
self.add_embedding = ImageHintTimeEmbedding(
|
| 456 |
+
image_embed_dim=encoder_hid_dim, time_embed_dim=time_embed_dim
|
| 457 |
+
)
|
| 458 |
+
elif addition_embed_type is not None:
|
| 459 |
+
raise ValueError(
|
| 460 |
+
f"addition_embed_type: {addition_embed_type} must be None, 'text' or 'text_image'."
|
| 461 |
+
)
|
| 462 |
+
|
| 463 |
+
if time_embedding_act_fn is None:
|
| 464 |
+
self.time_embed_act = None
|
| 465 |
+
else:
|
| 466 |
+
self.time_embed_act = get_activation(time_embedding_act_fn)
|
| 467 |
+
|
| 468 |
+
self.down_blocks = nn.ModuleList([])
|
| 469 |
+
self.up_blocks = nn.ModuleList([])
|
| 470 |
+
|
| 471 |
+
if isinstance(only_cross_attention, bool):
|
| 472 |
+
if mid_block_only_cross_attention is None:
|
| 473 |
+
mid_block_only_cross_attention = only_cross_attention
|
| 474 |
+
|
| 475 |
+
only_cross_attention = [only_cross_attention] * len(down_block_types)
|
| 476 |
+
|
| 477 |
+
if mid_block_only_cross_attention is None:
|
| 478 |
+
mid_block_only_cross_attention = False
|
| 479 |
+
|
| 480 |
+
if isinstance(num_attention_heads, int):
|
| 481 |
+
num_attention_heads = (num_attention_heads,) * len(down_block_types)
|
| 482 |
+
|
| 483 |
+
if isinstance(attention_head_dim, int):
|
| 484 |
+
attention_head_dim = (attention_head_dim,) * len(down_block_types)
|
| 485 |
+
|
| 486 |
+
if isinstance(cross_attention_dim, int):
|
| 487 |
+
cross_attention_dim = (cross_attention_dim,) * len(down_block_types)
|
| 488 |
+
|
| 489 |
+
if isinstance(layers_per_block, int):
|
| 490 |
+
layers_per_block = [layers_per_block] * len(down_block_types)
|
| 491 |
+
|
| 492 |
+
if isinstance(transformer_layers_per_block, int):
|
| 493 |
+
transformer_layers_per_block = [transformer_layers_per_block] * len(
|
| 494 |
+
down_block_types
|
| 495 |
+
)
|
| 496 |
+
|
| 497 |
+
if class_embeddings_concat:
|
| 498 |
+
# The time embeddings are concatenated with the class embeddings. The dimension of the
|
| 499 |
+
# time embeddings passed to the down, middle, and up blocks is twice the dimension of the
|
| 500 |
+
# regular time embeddings
|
| 501 |
+
blocks_time_embed_dim = time_embed_dim * 2
|
| 502 |
+
else:
|
| 503 |
+
blocks_time_embed_dim = time_embed_dim
|
| 504 |
+
|
| 505 |
+
# down
|
| 506 |
+
output_channel = block_out_channels[0]
|
| 507 |
+
for i, down_block_type in enumerate(down_block_types):
|
| 508 |
+
input_channel = output_channel
|
| 509 |
+
output_channel = block_out_channels[i]
|
| 510 |
+
is_final_block = i == len(block_out_channels) - 1
|
| 511 |
+
|
| 512 |
+
down_block = get_down_block(
|
| 513 |
+
down_block_type,
|
| 514 |
+
num_layers=layers_per_block[i],
|
| 515 |
+
transformer_layers_per_block=transformer_layers_per_block[i],
|
| 516 |
+
in_channels=input_channel,
|
| 517 |
+
out_channels=output_channel,
|
| 518 |
+
temb_channels=blocks_time_embed_dim,
|
| 519 |
+
add_downsample=not is_final_block,
|
| 520 |
+
resnet_eps=norm_eps,
|
| 521 |
+
resnet_act_fn=act_fn,
|
| 522 |
+
resnet_groups=norm_num_groups,
|
| 523 |
+
cross_attention_dim=cross_attention_dim[i],
|
| 524 |
+
num_attention_heads=num_attention_heads[i],
|
| 525 |
+
downsample_padding=downsample_padding,
|
| 526 |
+
dual_cross_attention=dual_cross_attention,
|
| 527 |
+
use_linear_projection=use_linear_projection,
|
| 528 |
+
only_cross_attention=only_cross_attention[i],
|
| 529 |
+
upcast_attention=upcast_attention,
|
| 530 |
+
resnet_time_scale_shift=resnet_time_scale_shift,
|
| 531 |
+
attention_type=attention_type,
|
| 532 |
+
resnet_skip_time_act=resnet_skip_time_act,
|
| 533 |
+
resnet_out_scale_factor=resnet_out_scale_factor,
|
| 534 |
+
cross_attention_norm=cross_attention_norm,
|
| 535 |
+
attention_head_dim=attention_head_dim[i]
|
| 536 |
+
if attention_head_dim[i] is not None
|
| 537 |
+
else output_channel,
|
| 538 |
+
dropout=dropout,
|
| 539 |
+
)
|
| 540 |
+
self.down_blocks.append(down_block)
|
| 541 |
+
|
| 542 |
+
# mid
|
| 543 |
+
if mid_block_type == "UNetMidBlock2DCrossAttn":
|
| 544 |
+
self.mid_block = UNetMidBlock2DCrossAttn(
|
| 545 |
+
transformer_layers_per_block=transformer_layers_per_block[-1],
|
| 546 |
+
in_channels=block_out_channels[-1],
|
| 547 |
+
temb_channels=blocks_time_embed_dim,
|
| 548 |
+
dropout=dropout,
|
| 549 |
+
resnet_eps=norm_eps,
|
| 550 |
+
resnet_act_fn=act_fn,
|
| 551 |
+
output_scale_factor=mid_block_scale_factor,
|
| 552 |
+
resnet_time_scale_shift=resnet_time_scale_shift,
|
| 553 |
+
cross_attention_dim=cross_attention_dim[-1],
|
| 554 |
+
num_attention_heads=num_attention_heads[-1],
|
| 555 |
+
resnet_groups=norm_num_groups,
|
| 556 |
+
dual_cross_attention=dual_cross_attention,
|
| 557 |
+
use_linear_projection=use_linear_projection,
|
| 558 |
+
upcast_attention=upcast_attention,
|
| 559 |
+
attention_type=attention_type,
|
| 560 |
+
)
|
| 561 |
+
elif mid_block_type == "UNetMidBlock2DSimpleCrossAttn":
|
| 562 |
+
raise NotImplementedError(f"Unsupport mid_block_type: {mid_block_type}")
|
| 563 |
+
elif mid_block_type == "UNetMidBlock2D":
|
| 564 |
+
self.mid_block = UNetMidBlock2D(
|
| 565 |
+
in_channels=block_out_channels[-1],
|
| 566 |
+
temb_channels=blocks_time_embed_dim,
|
| 567 |
+
dropout=dropout,
|
| 568 |
+
num_layers=0,
|
| 569 |
+
resnet_eps=norm_eps,
|
| 570 |
+
resnet_act_fn=act_fn,
|
| 571 |
+
output_scale_factor=mid_block_scale_factor,
|
| 572 |
+
resnet_groups=norm_num_groups,
|
| 573 |
+
resnet_time_scale_shift=resnet_time_scale_shift,
|
| 574 |
+
add_attention=False,
|
| 575 |
+
)
|
| 576 |
+
elif mid_block_type is None:
|
| 577 |
+
self.mid_block = None
|
| 578 |
+
else:
|
| 579 |
+
raise ValueError(f"unknown mid_block_type : {mid_block_type}")
|
| 580 |
+
|
| 581 |
+
# count how many layers upsample the images
|
| 582 |
+
self.num_upsamplers = 0
|
| 583 |
+
|
| 584 |
+
# up
|
| 585 |
+
reversed_block_out_channels = list(reversed(block_out_channels))
|
| 586 |
+
reversed_num_attention_heads = list(reversed(num_attention_heads))
|
| 587 |
+
reversed_layers_per_block = list(reversed(layers_per_block))
|
| 588 |
+
reversed_cross_attention_dim = list(reversed(cross_attention_dim))
|
| 589 |
+
reversed_transformer_layers_per_block = (
|
| 590 |
+
list(reversed(transformer_layers_per_block))
|
| 591 |
+
if reverse_transformer_layers_per_block is None
|
| 592 |
+
else reverse_transformer_layers_per_block
|
| 593 |
+
)
|
| 594 |
+
only_cross_attention = list(reversed(only_cross_attention))
|
| 595 |
+
|
| 596 |
+
output_channel = reversed_block_out_channels[0]
|
| 597 |
+
for i, up_block_type in enumerate(up_block_types):
|
| 598 |
+
is_final_block = i == len(block_out_channels) - 1
|
| 599 |
+
|
| 600 |
+
prev_output_channel = output_channel
|
| 601 |
+
output_channel = reversed_block_out_channels[i]
|
| 602 |
+
input_channel = reversed_block_out_channels[
|
| 603 |
+
min(i + 1, len(block_out_channels) - 1)
|
| 604 |
+
]
|
| 605 |
+
|
| 606 |
+
# add upsample block for all BUT final layer
|
| 607 |
+
if not is_final_block:
|
| 608 |
+
add_upsample = True
|
| 609 |
+
self.num_upsamplers += 1
|
| 610 |
+
else:
|
| 611 |
+
add_upsample = False
|
| 612 |
+
|
| 613 |
+
up_block = get_up_block(
|
| 614 |
+
up_block_type,
|
| 615 |
+
num_layers=reversed_layers_per_block[i] + 1,
|
| 616 |
+
transformer_layers_per_block=reversed_transformer_layers_per_block[i],
|
| 617 |
+
in_channels=input_channel,
|
| 618 |
+
out_channels=output_channel,
|
| 619 |
+
prev_output_channel=prev_output_channel,
|
| 620 |
+
temb_channels=blocks_time_embed_dim,
|
| 621 |
+
add_upsample=add_upsample,
|
| 622 |
+
resnet_eps=norm_eps,
|
| 623 |
+
resnet_act_fn=act_fn,
|
| 624 |
+
resolution_idx=i,
|
| 625 |
+
resnet_groups=norm_num_groups,
|
| 626 |
+
cross_attention_dim=reversed_cross_attention_dim[i],
|
| 627 |
+
num_attention_heads=reversed_num_attention_heads[i],
|
| 628 |
+
dual_cross_attention=dual_cross_attention,
|
| 629 |
+
use_linear_projection=use_linear_projection,
|
| 630 |
+
only_cross_attention=only_cross_attention[i],
|
| 631 |
+
upcast_attention=upcast_attention,
|
| 632 |
+
resnet_time_scale_shift=resnet_time_scale_shift,
|
| 633 |
+
attention_type=attention_type,
|
| 634 |
+
resnet_skip_time_act=resnet_skip_time_act,
|
| 635 |
+
resnet_out_scale_factor=resnet_out_scale_factor,
|
| 636 |
+
cross_attention_norm=cross_attention_norm,
|
| 637 |
+
attention_head_dim=attention_head_dim[i]
|
| 638 |
+
if attention_head_dim[i] is not None
|
| 639 |
+
else output_channel,
|
| 640 |
+
dropout=dropout,
|
| 641 |
+
)
|
| 642 |
+
self.up_blocks.append(up_block)
|
| 643 |
+
prev_output_channel = output_channel
|
| 644 |
+
|
| 645 |
+
# out
|
| 646 |
+
if norm_num_groups is not None:
|
| 647 |
+
self.conv_norm_out = nn.GroupNorm(
|
| 648 |
+
num_channels=block_out_channels[0],
|
| 649 |
+
num_groups=norm_num_groups,
|
| 650 |
+
eps=norm_eps,
|
| 651 |
+
)
|
| 652 |
+
|
| 653 |
+
self.conv_act = get_activation(act_fn)
|
| 654 |
+
|
| 655 |
+
else:
|
| 656 |
+
self.conv_norm_out = None
|
| 657 |
+
self.conv_act = None
|
| 658 |
+
self.conv_norm_out = None
|
| 659 |
+
|
| 660 |
+
conv_out_padding = (conv_out_kernel - 1) // 2
|
| 661 |
+
# self.conv_out = nn.Conv2d(
|
| 662 |
+
# block_out_channels[0],
|
| 663 |
+
# out_channels,
|
| 664 |
+
# kernel_size=conv_out_kernel,
|
| 665 |
+
# padding=conv_out_padding,
|
| 666 |
+
# )
|
| 667 |
+
|
| 668 |
+
if attention_type in ["gated", "gated-text-image"]:
|
| 669 |
+
positive_len = 768
|
| 670 |
+
if isinstance(cross_attention_dim, int):
|
| 671 |
+
positive_len = cross_attention_dim
|
| 672 |
+
elif isinstance(cross_attention_dim, tuple) or isinstance(
|
| 673 |
+
cross_attention_dim, list
|
| 674 |
+
):
|
| 675 |
+
positive_len = cross_attention_dim[0]
|
| 676 |
+
|
| 677 |
+
feature_type = "text-only" if attention_type == "gated" else "text-image"
|
| 678 |
+
# self.position_net = PositionNet(
|
| 679 |
+
# positive_len=positive_len,
|
| 680 |
+
# out_dim=cross_attention_dim,
|
| 681 |
+
# feature_type=feature_type,
|
| 682 |
+
# )
|
| 683 |
+
|
| 684 |
+
@property
|
| 685 |
+
def attn_processors(self) -> Dict[str, AttentionProcessor]:
|
| 686 |
+
r"""
|
| 687 |
+
Returns:
|
| 688 |
+
`dict` of attention processors: A dictionary containing all attention processors used in the model with
|
| 689 |
+
indexed by its weight name.
|
| 690 |
+
"""
|
| 691 |
+
# set recursively
|
| 692 |
+
processors = {}
|
| 693 |
+
|
| 694 |
+
def fn_recursive_add_processors(
|
| 695 |
+
name: str,
|
| 696 |
+
module: torch.nn.Module,
|
| 697 |
+
processors: Dict[str, AttentionProcessor],
|
| 698 |
+
):
|
| 699 |
+
if hasattr(module, "get_processor"):
|
| 700 |
+
processors[f"{name}.processor"] = module.get_processor(
|
| 701 |
+
return_deprecated_lora=True
|
| 702 |
+
)
|
| 703 |
+
|
| 704 |
+
for sub_name, child in module.named_children():
|
| 705 |
+
fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)
|
| 706 |
+
|
| 707 |
+
return processors
|
| 708 |
+
|
| 709 |
+
for name, module in self.named_children():
|
| 710 |
+
fn_recursive_add_processors(name, module, processors)
|
| 711 |
+
|
| 712 |
+
return processors
|
| 713 |
+
|
| 714 |
+
def set_attn_processor(
|
| 715 |
+
self,
|
| 716 |
+
processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]],
|
| 717 |
+
_remove_lora=False,
|
| 718 |
+
):
|
| 719 |
+
r"""
|
| 720 |
+
Sets the attention processor to use to compute attention.
|
| 721 |
+
|
| 722 |
+
Parameters:
|
| 723 |
+
processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
|
| 724 |
+
The instantiated processor class or a dictionary of processor classes that will be set as the processor
|
| 725 |
+
for **all** `Attention` layers.
|
| 726 |
+
|
| 727 |
+
If `processor` is a dict, the key needs to define the path to the corresponding cross attention
|
| 728 |
+
processor. This is strongly recommended when setting trainable attention processors.
|
| 729 |
+
|
| 730 |
+
"""
|
| 731 |
+
count = len(self.attn_processors.keys())
|
| 732 |
+
|
| 733 |
+
if isinstance(processor, dict) and len(processor) != count:
|
| 734 |
+
raise ValueError(
|
| 735 |
+
f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
|
| 736 |
+
f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
|
| 737 |
+
)
|
| 738 |
+
|
| 739 |
+
def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor):
|
| 740 |
+
if hasattr(module, "set_processor"):
|
| 741 |
+
if not isinstance(processor, dict):
|
| 742 |
+
module.set_processor(processor, _remove_lora=_remove_lora)
|
| 743 |
+
else:
|
| 744 |
+
module.set_processor(
|
| 745 |
+
processor.pop(f"{name}.processor"), _remove_lora=_remove_lora
|
| 746 |
+
)
|
| 747 |
+
|
| 748 |
+
for sub_name, child in module.named_children():
|
| 749 |
+
fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)
|
| 750 |
+
|
| 751 |
+
for name, module in self.named_children():
|
| 752 |
+
fn_recursive_attn_processor(name, module, processor)
|
| 753 |
+
|
| 754 |
+
def set_default_attn_processor(self):
|
| 755 |
+
"""
|
| 756 |
+
Disables custom attention processors and sets the default attention implementation.
|
| 757 |
+
"""
|
| 758 |
+
if all(
|
| 759 |
+
proc.__class__ in ADDED_KV_ATTENTION_PROCESSORS
|
| 760 |
+
for proc in self.attn_processors.values()
|
| 761 |
+
):
|
| 762 |
+
processor = AttnAddedKVProcessor()
|
| 763 |
+
elif all(
|
| 764 |
+
proc.__class__ in CROSS_ATTENTION_PROCESSORS
|
| 765 |
+
for proc in self.attn_processors.values()
|
| 766 |
+
):
|
| 767 |
+
processor = AttnProcessor()
|
| 768 |
+
else:
|
| 769 |
+
raise ValueError(
|
| 770 |
+
f"Cannot call `set_default_attn_processor` when attention processors are of type {next(iter(self.attn_processors.values()))}"
|
| 771 |
+
)
|
| 772 |
+
|
| 773 |
+
self.set_attn_processor(processor, _remove_lora=True)
|
| 774 |
+
|
| 775 |
+
def set_attention_slice(self, slice_size):
|
| 776 |
+
r"""
|
| 777 |
+
Enable sliced attention computation.
|
| 778 |
+
|
| 779 |
+
When this option is enabled, the attention module splits the input tensor in slices to compute attention in
|
| 780 |
+
several steps. This is useful for saving some memory in exchange for a small decrease in speed.
|
| 781 |
+
|
| 782 |
+
Args:
|
| 783 |
+
slice_size (`str` or `int` or `list(int)`, *optional*, defaults to `"auto"`):
|
| 784 |
+
When `"auto"`, input to the attention heads is halved, so attention is computed in two steps. If
|
| 785 |
+
`"max"`, maximum amount of memory is saved by running only one slice at a time. If a number is
|
| 786 |
+
provided, uses as many slices as `attention_head_dim // slice_size`. In this case, `attention_head_dim`
|
| 787 |
+
must be a multiple of `slice_size`.
|
| 788 |
+
"""
|
| 789 |
+
sliceable_head_dims = []
|
| 790 |
+
|
| 791 |
+
def fn_recursive_retrieve_sliceable_dims(module: torch.nn.Module):
|
| 792 |
+
if hasattr(module, "set_attention_slice"):
|
| 793 |
+
sliceable_head_dims.append(module.sliceable_head_dim)
|
| 794 |
+
|
| 795 |
+
for child in module.children():
|
| 796 |
+
fn_recursive_retrieve_sliceable_dims(child)
|
| 797 |
+
|
| 798 |
+
# retrieve number of attention layers
|
| 799 |
+
for module in self.children():
|
| 800 |
+
fn_recursive_retrieve_sliceable_dims(module)
|
| 801 |
+
|
| 802 |
+
num_sliceable_layers = len(sliceable_head_dims)
|
| 803 |
+
|
| 804 |
+
if slice_size == "auto":
|
| 805 |
+
# half the attention head size is usually a good trade-off between
|
| 806 |
+
# speed and memory
|
| 807 |
+
slice_size = [dim // 2 for dim in sliceable_head_dims]
|
| 808 |
+
elif slice_size == "max":
|
| 809 |
+
# make smallest slice possible
|
| 810 |
+
slice_size = num_sliceable_layers * [1]
|
| 811 |
+
|
| 812 |
+
slice_size = (
|
| 813 |
+
num_sliceable_layers * [slice_size]
|
| 814 |
+
if not isinstance(slice_size, list)
|
| 815 |
+
else slice_size
|
| 816 |
+
)
|
| 817 |
+
|
| 818 |
+
if len(slice_size) != len(sliceable_head_dims):
|
| 819 |
+
raise ValueError(
|
| 820 |
+
f"You have provided {len(slice_size)}, but {self.config} has {len(sliceable_head_dims)} different"
|
| 821 |
+
f" attention layers. Make sure to match `len(slice_size)` to be {len(sliceable_head_dims)}."
|
| 822 |
+
)
|
| 823 |
+
|
| 824 |
+
for i in range(len(slice_size)):
|
| 825 |
+
size = slice_size[i]
|
| 826 |
+
dim = sliceable_head_dims[i]
|
| 827 |
+
if size is not None and size > dim:
|
| 828 |
+
raise ValueError(f"size {size} has to be smaller or equal to {dim}.")
|
| 829 |
+
|
| 830 |
+
# Recursively walk through all the children.
|
| 831 |
+
# Any children which exposes the set_attention_slice method
|
| 832 |
+
# gets the message
|
| 833 |
+
def fn_recursive_set_attention_slice(
|
| 834 |
+
module: torch.nn.Module, slice_size: List[int]
|
| 835 |
+
):
|
| 836 |
+
if hasattr(module, "set_attention_slice"):
|
| 837 |
+
module.set_attention_slice(slice_size.pop())
|
| 838 |
+
|
| 839 |
+
for child in module.children():
|
| 840 |
+
fn_recursive_set_attention_slice(child, slice_size)
|
| 841 |
+
|
| 842 |
+
reversed_slice_size = list(reversed(slice_size))
|
| 843 |
+
for module in self.children():
|
| 844 |
+
fn_recursive_set_attention_slice(module, reversed_slice_size)
|
| 845 |
+
|
| 846 |
+
def _set_gradient_checkpointing(self, module, value=False):
|
| 847 |
+
if hasattr(module, "gradient_checkpointing"):
|
| 848 |
+
module.gradient_checkpointing = value
|
| 849 |
+
|
| 850 |
+
def enable_freeu(self, s1, s2, b1, b2):
|
| 851 |
+
r"""Enables the FreeU mechanism from https://arxiv.org/abs/2309.11497.
|
| 852 |
+
|
| 853 |
+
The suffixes after the scaling factors represent the stage blocks where they are being applied.
|
| 854 |
+
|
| 855 |
+
Please refer to the [official repository](https://github.com/ChenyangSi/FreeU) for combinations of values that
|
| 856 |
+
are known to work well for different pipelines such as Stable Diffusion v1, v2, and Stable Diffusion XL.
|
| 857 |
+
|
| 858 |
+
Args:
|
| 859 |
+
s1 (`float`):
|
| 860 |
+
Scaling factor for stage 1 to attenuate the contributions of the skip features. This is done to
|
| 861 |
+
mitigate the "oversmoothing effect" in the enhanced denoising process.
|
| 862 |
+
s2 (`float`):
|
| 863 |
+
Scaling factor for stage 2 to attenuate the contributions of the skip features. This is done to
|
| 864 |
+
mitigate the "oversmoothing effect" in the enhanced denoising process.
|
| 865 |
+
b1 (`float`): Scaling factor for stage 1 to amplify the contributions of backbone features.
|
| 866 |
+
b2 (`float`): Scaling factor for stage 2 to amplify the contributions of backbone features.
|
| 867 |
+
"""
|
| 868 |
+
for i, upsample_block in enumerate(self.up_blocks):
|
| 869 |
+
setattr(upsample_block, "s1", s1)
|
| 870 |
+
setattr(upsample_block, "s2", s2)
|
| 871 |
+
setattr(upsample_block, "b1", b1)
|
| 872 |
+
setattr(upsample_block, "b2", b2)
|
| 873 |
+
|
| 874 |
+
def disable_freeu(self):
|
| 875 |
+
"""Disables the FreeU mechanism."""
|
| 876 |
+
freeu_keys = {"s1", "s2", "b1", "b2"}
|
| 877 |
+
for i, upsample_block in enumerate(self.up_blocks):
|
| 878 |
+
for k in freeu_keys:
|
| 879 |
+
if (
|
| 880 |
+
hasattr(upsample_block, k)
|
| 881 |
+
or getattr(upsample_block, k, None) is not None
|
| 882 |
+
):
|
| 883 |
+
setattr(upsample_block, k, None)
|
| 884 |
+
|
| 885 |
+
def forward(
|
| 886 |
+
self,
|
| 887 |
+
sample: torch.FloatTensor,
|
| 888 |
+
timestep: Union[torch.Tensor, float, int],
|
| 889 |
+
encoder_hidden_states: torch.Tensor,
|
| 890 |
+
class_labels: Optional[torch.Tensor] = None,
|
| 891 |
+
timestep_cond: Optional[torch.Tensor] = None,
|
| 892 |
+
attention_mask: Optional[torch.Tensor] = None,
|
| 893 |
+
cross_attention_kwargs: Optional[Dict[str, Any]] = None,
|
| 894 |
+
added_cond_kwargs: Optional[Dict[str, torch.Tensor]] = None,
|
| 895 |
+
down_block_additional_residuals: Optional[Tuple[torch.Tensor]] = None,
|
| 896 |
+
mid_block_additional_residual: Optional[torch.Tensor] = None,
|
| 897 |
+
down_intrablock_additional_residuals: Optional[Tuple[torch.Tensor]] = None,
|
| 898 |
+
encoder_attention_mask: Optional[torch.Tensor] = None,
|
| 899 |
+
pose_cond_fea: Optional[torch.Tensor] = None,
|
| 900 |
+
return_dict: bool = True,
|
| 901 |
+
) -> Union[UNet2DConditionOutput, Tuple]:
|
| 902 |
+
r"""
|
| 903 |
+
The [`UNet2DConditionModel`] forward method.
|
| 904 |
+
|
| 905 |
+
Args:
|
| 906 |
+
sample (`torch.FloatTensor`):
|
| 907 |
+
The noisy input tensor with the following shape `(batch, channel, height, width)`.
|
| 908 |
+
timestep (`torch.FloatTensor` or `float` or `int`): The number of timesteps to denoise an input.
|
| 909 |
+
encoder_hidden_states (`torch.FloatTensor`):
|
| 910 |
+
The encoder hidden states with shape `(batch, sequence_length, feature_dim)`.
|
| 911 |
+
class_labels (`torch.Tensor`, *optional*, defaults to `None`):
|
| 912 |
+
Optional class labels for conditioning. Their embeddings will be summed with the timestep embeddings.
|
| 913 |
+
timestep_cond: (`torch.Tensor`, *optional*, defaults to `None`):
|
| 914 |
+
Conditional embeddings for timestep. If provided, the embeddings will be summed with the samples passed
|
| 915 |
+
through the `self.time_embedding` layer to obtain the timestep embeddings.
|
| 916 |
+
attention_mask (`torch.Tensor`, *optional*, defaults to `None`):
|
| 917 |
+
An attention mask of shape `(batch, key_tokens)` is applied to `encoder_hidden_states`. If `1` the mask
|
| 918 |
+
is kept, otherwise if `0` it is discarded. Mask will be converted into a bias, which adds large
|
| 919 |
+
negative values to the attention scores corresponding to "discard" tokens.
|
| 920 |
+
cross_attention_kwargs (`dict`, *optional*):
|
| 921 |
+
A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
|
| 922 |
+
`self.processor` in
|
| 923 |
+
[diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
|
| 924 |
+
added_cond_kwargs: (`dict`, *optional*):
|
| 925 |
+
A kwargs dictionary containing additional embeddings that if specified are added to the embeddings that
|
| 926 |
+
are passed along to the UNet blocks.
|
| 927 |
+
down_block_additional_residuals: (`tuple` of `torch.Tensor`, *optional*):
|
| 928 |
+
A tuple of tensors that if specified are added to the residuals of down unet blocks.
|
| 929 |
+
mid_block_additional_residual: (`torch.Tensor`, *optional*):
|
| 930 |
+
A tensor that if specified is added to the residual of the middle unet block.
|
| 931 |
+
encoder_attention_mask (`torch.Tensor`):
|
| 932 |
+
A cross-attention mask of shape `(batch, sequence_length)` is applied to `encoder_hidden_states`. If
|
| 933 |
+
`True` the mask is kept, otherwise if `False` it is discarded. Mask will be converted into a bias,
|
| 934 |
+
which adds large negative values to the attention scores corresponding to "discard" tokens.
|
| 935 |
+
return_dict (`bool`, *optional*, defaults to `True`):
|
| 936 |
+
Whether or not to return a [`~models.unet_2d_condition.UNet2DConditionOutput`] instead of a plain
|
| 937 |
+
tuple.
|
| 938 |
+
cross_attention_kwargs (`dict`, *optional*):
|
| 939 |
+
A kwargs dictionary that if specified is passed along to the [`AttnProcessor`].
|
| 940 |
+
added_cond_kwargs: (`dict`, *optional*):
|
| 941 |
+
A kwargs dictionary containin additional embeddings that if specified are added to the embeddings that
|
| 942 |
+
are passed along to the UNet blocks.
|
| 943 |
+
down_block_additional_residuals (`tuple` of `torch.Tensor`, *optional*):
|
| 944 |
+
additional residuals to be added to UNet long skip connections from down blocks to up blocks for
|
| 945 |
+
example from ControlNet side model(s)
|
| 946 |
+
mid_block_additional_residual (`torch.Tensor`, *optional*):
|
| 947 |
+
additional residual to be added to UNet mid block output, for example from ControlNet side model
|
| 948 |
+
down_intrablock_additional_residuals (`tuple` of `torch.Tensor`, *optional*):
|
| 949 |
+
additional residuals to be added within UNet down blocks, for example from T2I-Adapter side model(s)
|
| 950 |
+
|
| 951 |
+
Returns:
|
| 952 |
+
[`~models.unet_2d_condition.UNet2DConditionOutput`] or `tuple`:
|
| 953 |
+
If `return_dict` is True, an [`~models.unet_2d_condition.UNet2DConditionOutput`] is returned, otherwise
|
| 954 |
+
a `tuple` is returned where the first element is the sample tensor.
|
| 955 |
+
"""
|
| 956 |
+
# By default samples have to be AT least a multiple of the overall upsampling factor.
|
| 957 |
+
# The overall upsampling factor is equal to 2 ** (# num of upsampling layers).
|
| 958 |
+
# However, the upsampling interpolation output size can be forced to fit any upsampling size
|
| 959 |
+
# on the fly if necessary.
|
| 960 |
+
default_overall_up_factor = 2**self.num_upsamplers
|
| 961 |
+
|
| 962 |
+
# upsample size should be forwarded when sample is not a multiple of `default_overall_up_factor`
|
| 963 |
+
forward_upsample_size = False
|
| 964 |
+
upsample_size = None
|
| 965 |
+
|
| 966 |
+
for dim in sample.shape[-2:]:
|
| 967 |
+
if dim % default_overall_up_factor != 0:
|
| 968 |
+
# Forward upsample size to force interpolation output size.
|
| 969 |
+
forward_upsample_size = True
|
| 970 |
+
break
|
| 971 |
+
|
| 972 |
+
# ensure attention_mask is a bias, and give it a singleton query_tokens dimension
|
| 973 |
+
# expects mask of shape:
|
| 974 |
+
# [batch, key_tokens]
|
| 975 |
+
# adds singleton query_tokens dimension:
|
| 976 |
+
# [batch, 1, key_tokens]
|
| 977 |
+
# this helps to broadcast it as a bias over attention scores, which will be in one of the following shapes:
|
| 978 |
+
# [batch, heads, query_tokens, key_tokens] (e.g. torch sdp attn)
|
| 979 |
+
# [batch * heads, query_tokens, key_tokens] (e.g. xformers or classic attn)
|
| 980 |
+
if attention_mask is not None:
|
| 981 |
+
# assume that mask is expressed as:
|
| 982 |
+
# (1 = keep, 0 = discard)
|
| 983 |
+
# convert mask into a bias that can be added to attention scores:
|
| 984 |
+
# (keep = +0, discard = -10000.0)
|
| 985 |
+
attention_mask = (1 - attention_mask.to(sample.dtype)) * -10000.0
|
| 986 |
+
attention_mask = attention_mask.unsqueeze(1)
|
| 987 |
+
|
| 988 |
+
# convert encoder_attention_mask to a bias the same way we do for attention_mask
|
| 989 |
+
if encoder_attention_mask is not None:
|
| 990 |
+
encoder_attention_mask = (
|
| 991 |
+
1 - encoder_attention_mask.to(sample.dtype)
|
| 992 |
+
) * -10000.0
|
| 993 |
+
encoder_attention_mask = encoder_attention_mask.unsqueeze(1)
|
| 994 |
+
|
| 995 |
+
# 0. center input if necessary
|
| 996 |
+
if self.config.center_input_sample:
|
| 997 |
+
sample = 2 * sample - 1.0
|
| 998 |
+
|
| 999 |
+
# 1. time
|
| 1000 |
+
timesteps = timestep
|
| 1001 |
+
if not torch.is_tensor(timesteps):
|
| 1002 |
+
# TODO: this requires sync between CPU and GPU. So try to pass timesteps as tensors if you can
|
| 1003 |
+
# This would be a good case for the `match` statement (Python 3.10+)
|
| 1004 |
+
is_mps = sample.device.type == "mps"
|
| 1005 |
+
if isinstance(timestep, float):
|
| 1006 |
+
dtype = torch.float32 if is_mps else torch.float64
|
| 1007 |
+
else:
|
| 1008 |
+
dtype = torch.int32 if is_mps else torch.int64
|
| 1009 |
+
timesteps = torch.tensor([timesteps], dtype=dtype, device=sample.device)
|
| 1010 |
+
elif len(timesteps.shape) == 0:
|
| 1011 |
+
timesteps = timesteps[None].to(sample.device)
|
| 1012 |
+
|
| 1013 |
+
# broadcast to batch dimension in a way that's compatible with ONNX/Core ML
|
| 1014 |
+
timesteps = timesteps.expand(sample.shape[0])
|
| 1015 |
+
|
| 1016 |
+
t_emb = self.time_proj(timesteps)
|
| 1017 |
+
|
| 1018 |
+
# `Timesteps` does not contain any weights and will always return f32 tensors
|
| 1019 |
+
# but time_embedding might actually be running in fp16. so we need to cast here.
|
| 1020 |
+
# there might be better ways to encapsulate this.
|
| 1021 |
+
t_emb = t_emb.to(dtype=sample.dtype)
|
| 1022 |
+
|
| 1023 |
+
emb = self.time_embedding(t_emb, timestep_cond)
|
| 1024 |
+
aug_emb = None
|
| 1025 |
+
|
| 1026 |
+
if self.class_embedding is not None:
|
| 1027 |
+
if class_labels is None:
|
| 1028 |
+
raise ValueError(
|
| 1029 |
+
"class_labels should be provided when num_class_embeds > 0"
|
| 1030 |
+
)
|
| 1031 |
+
|
| 1032 |
+
if self.config.class_embed_type == "timestep":
|
| 1033 |
+
class_labels = self.time_proj(class_labels)
|
| 1034 |
+
|
| 1035 |
+
# `Timesteps` does not contain any weights and will always return f32 tensors
|
| 1036 |
+
# there might be better ways to encapsulate this.
|
| 1037 |
+
class_labels = class_labels.to(dtype=sample.dtype)
|
| 1038 |
+
|
| 1039 |
+
class_emb = self.class_embedding(class_labels).to(dtype=sample.dtype)
|
| 1040 |
+
|
| 1041 |
+
if self.config.class_embeddings_concat:
|
| 1042 |
+
emb = torch.cat([emb, class_emb], dim=-1)
|
| 1043 |
+
else:
|
| 1044 |
+
emb = emb + class_emb
|
| 1045 |
+
|
| 1046 |
+
if self.config.addition_embed_type == "text":
|
| 1047 |
+
aug_emb = self.add_embedding(encoder_hidden_states)
|
| 1048 |
+
elif self.config.addition_embed_type == "text_image":
|
| 1049 |
+
# Kandinsky 2.1 - style
|
| 1050 |
+
if "image_embeds" not in added_cond_kwargs:
|
| 1051 |
+
raise ValueError(
|
| 1052 |
+
f"{self.__class__} has the config param `addition_embed_type` set to 'text_image' which requires the keyword argument `image_embeds` to be passed in `added_cond_kwargs`"
|
| 1053 |
+
)
|
| 1054 |
+
|
| 1055 |
+
image_embs = added_cond_kwargs.get("image_embeds")
|
| 1056 |
+
text_embs = added_cond_kwargs.get("text_embeds", encoder_hidden_states)
|
| 1057 |
+
aug_emb = self.add_embedding(text_embs, image_embs)
|
| 1058 |
+
elif self.config.addition_embed_type == "text_time":
|
| 1059 |
+
# SDXL - style
|
| 1060 |
+
if "text_embeds" not in added_cond_kwargs:
|
| 1061 |
+
raise ValueError(
|
| 1062 |
+
f"{self.__class__} has the config param `addition_embed_type` set to 'text_time' which requires the keyword argument `text_embeds` to be passed in `added_cond_kwargs`"
|
| 1063 |
+
)
|
| 1064 |
+
text_embeds = added_cond_kwargs.get("text_embeds")
|
| 1065 |
+
if "time_ids" not in added_cond_kwargs:
|
| 1066 |
+
raise ValueError(
|
| 1067 |
+
f"{self.__class__} has the config param `addition_embed_type` set to 'text_time' which requires the keyword argument `time_ids` to be passed in `added_cond_kwargs`"
|
| 1068 |
+
)
|
| 1069 |
+
time_ids = added_cond_kwargs.get("time_ids")
|
| 1070 |
+
time_embeds = self.add_time_proj(time_ids.flatten())
|
| 1071 |
+
time_embeds = time_embeds.reshape((text_embeds.shape[0], -1))
|
| 1072 |
+
add_embeds = torch.concat([text_embeds, time_embeds], dim=-1)
|
| 1073 |
+
add_embeds = add_embeds.to(emb.dtype)
|
| 1074 |
+
aug_emb = self.add_embedding(add_embeds)
|
| 1075 |
+
elif self.config.addition_embed_type == "image":
|
| 1076 |
+
# Kandinsky 2.2 - style
|
| 1077 |
+
if "image_embeds" not in added_cond_kwargs:
|
| 1078 |
+
raise ValueError(
|
| 1079 |
+
f"{self.__class__} has the config param `addition_embed_type` set to 'image' which requires the keyword argument `image_embeds` to be passed in `added_cond_kwargs`"
|
| 1080 |
+
)
|
| 1081 |
+
image_embs = added_cond_kwargs.get("image_embeds")
|
| 1082 |
+
aug_emb = self.add_embedding(image_embs)
|
| 1083 |
+
elif self.config.addition_embed_type == "image_hint":
|
| 1084 |
+
# Kandinsky 2.2 - style
|
| 1085 |
+
if (
|
| 1086 |
+
"image_embeds" not in added_cond_kwargs
|
| 1087 |
+
or "hint" not in added_cond_kwargs
|
| 1088 |
+
):
|
| 1089 |
+
raise ValueError(
|
| 1090 |
+
f"{self.__class__} has the config param `addition_embed_type` set to 'image_hint' which requires the keyword arguments `image_embeds` and `hint` to be passed in `added_cond_kwargs`"
|
| 1091 |
+
)
|
| 1092 |
+
image_embs = added_cond_kwargs.get("image_embeds")
|
| 1093 |
+
hint = added_cond_kwargs.get("hint")
|
| 1094 |
+
aug_emb, hint = self.add_embedding(image_embs, hint)
|
| 1095 |
+
sample = torch.cat([sample, hint], dim=1)
|
| 1096 |
+
|
| 1097 |
+
emb = emb + aug_emb if aug_emb is not None else emb
|
| 1098 |
+
|
| 1099 |
+
if self.time_embed_act is not None:
|
| 1100 |
+
emb = self.time_embed_act(emb)
|
| 1101 |
+
|
| 1102 |
+
if (
|
| 1103 |
+
self.encoder_hid_proj is not None
|
| 1104 |
+
and self.config.encoder_hid_dim_type == "text_proj"
|
| 1105 |
+
):
|
| 1106 |
+
encoder_hidden_states = self.encoder_hid_proj(encoder_hidden_states)
|
| 1107 |
+
elif (
|
| 1108 |
+
self.encoder_hid_proj is not None
|
| 1109 |
+
and self.config.encoder_hid_dim_type == "text_image_proj"
|
| 1110 |
+
):
|
| 1111 |
+
# Kadinsky 2.1 - style
|
| 1112 |
+
if "image_embeds" not in added_cond_kwargs:
|
| 1113 |
+
raise ValueError(
|
| 1114 |
+
f"{self.__class__} has the config param `encoder_hid_dim_type` set to 'text_image_proj' which requires the keyword argument `image_embeds` to be passed in `added_conditions`"
|
| 1115 |
+
)
|
| 1116 |
+
|
| 1117 |
+
image_embeds = added_cond_kwargs.get("image_embeds")
|
| 1118 |
+
encoder_hidden_states = self.encoder_hid_proj(
|
| 1119 |
+
encoder_hidden_states, image_embeds
|
| 1120 |
+
)
|
| 1121 |
+
elif (
|
| 1122 |
+
self.encoder_hid_proj is not None
|
| 1123 |
+
and self.config.encoder_hid_dim_type == "image_proj"
|
| 1124 |
+
):
|
| 1125 |
+
# Kandinsky 2.2 - style
|
| 1126 |
+
if "image_embeds" not in added_cond_kwargs:
|
| 1127 |
+
raise ValueError(
|
| 1128 |
+
f"{self.__class__} has the config param `encoder_hid_dim_type` set to 'image_proj' which requires the keyword argument `image_embeds` to be passed in `added_conditions`"
|
| 1129 |
+
)
|
| 1130 |
+
image_embeds = added_cond_kwargs.get("image_embeds")
|
| 1131 |
+
encoder_hidden_states = self.encoder_hid_proj(image_embeds)
|
| 1132 |
+
elif (
|
| 1133 |
+
self.encoder_hid_proj is not None
|
| 1134 |
+
and self.config.encoder_hid_dim_type == "ip_image_proj"
|
| 1135 |
+
):
|
| 1136 |
+
if "image_embeds" not in added_cond_kwargs:
|
| 1137 |
+
raise ValueError(
|
| 1138 |
+
f"{self.__class__} has the config param `encoder_hid_dim_type` set to 'ip_image_proj' which requires the keyword argument `image_embeds` to be passed in `added_conditions`"
|
| 1139 |
+
)
|
| 1140 |
+
image_embeds = added_cond_kwargs.get("image_embeds")
|
| 1141 |
+
image_embeds = self.encoder_hid_proj(image_embeds).to(
|
| 1142 |
+
encoder_hidden_states.dtype
|
| 1143 |
+
)
|
| 1144 |
+
encoder_hidden_states = torch.cat(
|
| 1145 |
+
[encoder_hidden_states, image_embeds], dim=1
|
| 1146 |
+
)
|
| 1147 |
+
|
| 1148 |
+
# 2. pre-process
|
| 1149 |
+
sample = self.conv_in(sample)
|
| 1150 |
+
if pose_cond_fea is not None:
|
| 1151 |
+
sample = sample + pose_cond_fea
|
| 1152 |
+
|
| 1153 |
+
# 2.5 GLIGEN position net
|
| 1154 |
+
# if (
|
| 1155 |
+
# cross_attention_kwargs is not None
|
| 1156 |
+
# and cross_attention_kwargs.get("gligen", None) is not None
|
| 1157 |
+
# ):
|
| 1158 |
+
# cross_attention_kwargs = cross_attention_kwargs.copy()
|
| 1159 |
+
# gligen_args = cross_attention_kwargs.pop("gligen")
|
| 1160 |
+
# cross_attention_kwargs["gligen"] = {
|
| 1161 |
+
# "objs": self.position_net(**gligen_args)
|
| 1162 |
+
# }
|
| 1163 |
+
|
| 1164 |
+
# 3. down
|
| 1165 |
+
lora_scale = (
|
| 1166 |
+
cross_attention_kwargs.get("scale", 1.0)
|
| 1167 |
+
if cross_attention_kwargs is not None
|
| 1168 |
+
else 1.0
|
| 1169 |
+
)
|
| 1170 |
+
if USE_PEFT_BACKEND:
|
| 1171 |
+
# weight the lora layers by setting `lora_scale` for each PEFT layer
|
| 1172 |
+
scale_lora_layers(self, lora_scale)
|
| 1173 |
+
|
| 1174 |
+
is_controlnet = (
|
| 1175 |
+
mid_block_additional_residual is not None
|
| 1176 |
+
and down_block_additional_residuals is not None
|
| 1177 |
+
)
|
| 1178 |
+
# using new arg down_intrablock_additional_residuals for T2I-Adapters, to distinguish from controlnets
|
| 1179 |
+
is_adapter = down_intrablock_additional_residuals is not None
|
| 1180 |
+
# maintain backward compatibility for legacy usage, where
|
| 1181 |
+
# T2I-Adapter and ControlNet both use down_block_additional_residuals arg
|
| 1182 |
+
# but can only use one or the other
|
| 1183 |
+
if (
|
| 1184 |
+
not is_adapter
|
| 1185 |
+
and mid_block_additional_residual is None
|
| 1186 |
+
and down_block_additional_residuals is not None
|
| 1187 |
+
):
|
| 1188 |
+
deprecate(
|
| 1189 |
+
"T2I should not use down_block_additional_residuals",
|
| 1190 |
+
"1.3.0",
|
| 1191 |
+
"Passing intrablock residual connections with `down_block_additional_residuals` is deprecated \
|
| 1192 |
+
and will be removed in diffusers 1.3.0. `down_block_additional_residuals` should only be used \
|
| 1193 |
+
for ControlNet. Please make sure use `down_intrablock_additional_residuals` instead. ",
|
| 1194 |
+
standard_warn=False,
|
| 1195 |
+
)
|
| 1196 |
+
down_intrablock_additional_residuals = down_block_additional_residuals
|
| 1197 |
+
is_adapter = True
|
| 1198 |
+
|
| 1199 |
+
down_block_res_samples = (sample,)
|
| 1200 |
+
tot_referece_features = ()
|
| 1201 |
+
for downsample_block in self.down_blocks:
|
| 1202 |
+
if (
|
| 1203 |
+
hasattr(downsample_block, "has_cross_attention")
|
| 1204 |
+
and downsample_block.has_cross_attention
|
| 1205 |
+
):
|
| 1206 |
+
# For t2i-adapter CrossAttnDownBlock2D
|
| 1207 |
+
additional_residuals = {}
|
| 1208 |
+
if is_adapter and len(down_intrablock_additional_residuals) > 0:
|
| 1209 |
+
additional_residuals[
|
| 1210 |
+
"additional_residuals"
|
| 1211 |
+
] = down_intrablock_additional_residuals.pop(0)
|
| 1212 |
+
|
| 1213 |
+
sample, res_samples = downsample_block(
|
| 1214 |
+
hidden_states=sample,
|
| 1215 |
+
temb=emb,
|
| 1216 |
+
encoder_hidden_states=encoder_hidden_states,
|
| 1217 |
+
attention_mask=attention_mask,
|
| 1218 |
+
cross_attention_kwargs=cross_attention_kwargs,
|
| 1219 |
+
encoder_attention_mask=encoder_attention_mask,
|
| 1220 |
+
**additional_residuals,
|
| 1221 |
+
)
|
| 1222 |
+
else:
|
| 1223 |
+
sample, res_samples = downsample_block(
|
| 1224 |
+
hidden_states=sample, temb=emb, scale=lora_scale
|
| 1225 |
+
)
|
| 1226 |
+
if is_adapter and len(down_intrablock_additional_residuals) > 0:
|
| 1227 |
+
sample += down_intrablock_additional_residuals.pop(0)
|
| 1228 |
+
|
| 1229 |
+
down_block_res_samples += res_samples
|
| 1230 |
+
|
| 1231 |
+
if is_controlnet:
|
| 1232 |
+
new_down_block_res_samples = ()
|
| 1233 |
+
|
| 1234 |
+
for down_block_res_sample, down_block_additional_residual in zip(
|
| 1235 |
+
down_block_res_samples, down_block_additional_residuals
|
| 1236 |
+
):
|
| 1237 |
+
down_block_res_sample = (
|
| 1238 |
+
down_block_res_sample + down_block_additional_residual
|
| 1239 |
+
)
|
| 1240 |
+
new_down_block_res_samples = new_down_block_res_samples + (
|
| 1241 |
+
down_block_res_sample,
|
| 1242 |
+
)
|
| 1243 |
+
|
| 1244 |
+
down_block_res_samples = new_down_block_res_samples
|
| 1245 |
+
|
| 1246 |
+
# 4. mid
|
| 1247 |
+
if self.mid_block is not None:
|
| 1248 |
+
if (
|
| 1249 |
+
hasattr(self.mid_block, "has_cross_attention")
|
| 1250 |
+
and self.mid_block.has_cross_attention
|
| 1251 |
+
):
|
| 1252 |
+
sample = self.mid_block(
|
| 1253 |
+
sample,
|
| 1254 |
+
emb,
|
| 1255 |
+
encoder_hidden_states=encoder_hidden_states,
|
| 1256 |
+
attention_mask=attention_mask,
|
| 1257 |
+
cross_attention_kwargs=cross_attention_kwargs,
|
| 1258 |
+
encoder_attention_mask=encoder_attention_mask,
|
| 1259 |
+
)
|
| 1260 |
+
else:
|
| 1261 |
+
sample = self.mid_block(sample, emb)
|
| 1262 |
+
|
| 1263 |
+
# To support T2I-Adapter-XL
|
| 1264 |
+
if (
|
| 1265 |
+
is_adapter
|
| 1266 |
+
and len(down_intrablock_additional_residuals) > 0
|
| 1267 |
+
and sample.shape == down_intrablock_additional_residuals[0].shape
|
| 1268 |
+
):
|
| 1269 |
+
sample += down_intrablock_additional_residuals.pop(0)
|
| 1270 |
+
|
| 1271 |
+
if is_controlnet:
|
| 1272 |
+
sample = sample + mid_block_additional_residual
|
| 1273 |
+
|
| 1274 |
+
# 5. up
|
| 1275 |
+
for i, upsample_block in enumerate(self.up_blocks):
|
| 1276 |
+
is_final_block = i == len(self.up_blocks) - 1
|
| 1277 |
+
|
| 1278 |
+
res_samples = down_block_res_samples[-len(upsample_block.resnets) :]
|
| 1279 |
+
down_block_res_samples = down_block_res_samples[
|
| 1280 |
+
: -len(upsample_block.resnets)
|
| 1281 |
+
]
|
| 1282 |
+
|
| 1283 |
+
# if we have not reached the final block and need to forward the
|
| 1284 |
+
# upsample size, we do it here
|
| 1285 |
+
if not is_final_block and forward_upsample_size:
|
| 1286 |
+
upsample_size = down_block_res_samples[-1].shape[2:]
|
| 1287 |
+
|
| 1288 |
+
if (
|
| 1289 |
+
hasattr(upsample_block, "has_cross_attention")
|
| 1290 |
+
and upsample_block.has_cross_attention
|
| 1291 |
+
):
|
| 1292 |
+
sample = upsample_block(
|
| 1293 |
+
hidden_states=sample,
|
| 1294 |
+
temb=emb,
|
| 1295 |
+
res_hidden_states_tuple=res_samples,
|
| 1296 |
+
encoder_hidden_states=encoder_hidden_states,
|
| 1297 |
+
cross_attention_kwargs=cross_attention_kwargs,
|
| 1298 |
+
upsample_size=upsample_size,
|
| 1299 |
+
attention_mask=attention_mask,
|
| 1300 |
+
encoder_attention_mask=encoder_attention_mask,
|
| 1301 |
+
)
|
| 1302 |
+
else:
|
| 1303 |
+
sample = upsample_block(
|
| 1304 |
+
hidden_states=sample,
|
| 1305 |
+
temb=emb,
|
| 1306 |
+
res_hidden_states_tuple=res_samples,
|
| 1307 |
+
upsample_size=upsample_size,
|
| 1308 |
+
scale=lora_scale,
|
| 1309 |
+
)
|
| 1310 |
+
|
| 1311 |
+
# 6. post-process
|
| 1312 |
+
# if self.conv_norm_out:
|
| 1313 |
+
# sample = self.conv_norm_out(sample)
|
| 1314 |
+
# sample = self.conv_act(sample)
|
| 1315 |
+
# sample = self.conv_out(sample)
|
| 1316 |
+
|
| 1317 |
+
if USE_PEFT_BACKEND:
|
| 1318 |
+
# remove `lora_scale` from each PEFT layer
|
| 1319 |
+
unscale_lora_layers(self, lora_scale)
|
| 1320 |
+
|
| 1321 |
+
if not return_dict:
|
| 1322 |
+
return (sample,)
|
| 1323 |
+
|
| 1324 |
+
return UNet2DConditionOutput(sample=sample)
|
genwarp/models/unet_3d.py
ADDED
|
@@ -0,0 +1,645 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# This code is adapted from below and then modified.
|
| 2 |
+
# -----------------------------------------------------------------------------
|
| 3 |
+
# Moore-AnimateAnyone
|
| 4 |
+
# Apache License, Version 2.0
|
| 5 |
+
# Copyright @2023-2024 Moore Threads Technology Co., Ltd.
|
| 6 |
+
# https://github.com/MooreThreads/Moore-AnimateAnyone
|
| 7 |
+
# -----------------------------------------------------------------------------
|
| 8 |
+
# AnimateDiff
|
| 9 |
+
# Apache License, Version 2.0
|
| 10 |
+
# https://github.com/guoyww/AnimateDiff
|
| 11 |
+
# ==============================================================================
|
| 12 |
+
|
| 13 |
+
# Adapted from https://github.com/guoyww/AnimateDiff/blob/main/animatediff/models/unet_blocks.py
|
| 14 |
+
|
| 15 |
+
from collections import OrderedDict
|
| 16 |
+
from dataclasses import dataclass
|
| 17 |
+
from os import PathLike
|
| 18 |
+
from pathlib import Path
|
| 19 |
+
from typing import Dict, List, Optional, Tuple, Union
|
| 20 |
+
|
| 21 |
+
import torch
|
| 22 |
+
import torch.nn as nn
|
| 23 |
+
import torch.utils.checkpoint
|
| 24 |
+
from diffusers.configuration_utils import ConfigMixin, register_to_config
|
| 25 |
+
from diffusers.models.attention_processor import AttentionProcessor
|
| 26 |
+
from diffusers.models.embeddings import TimestepEmbedding, Timesteps
|
| 27 |
+
from diffusers.models.modeling_utils import ModelMixin
|
| 28 |
+
from diffusers.utils import SAFETENSORS_WEIGHTS_NAME, WEIGHTS_NAME, BaseOutput, logging
|
| 29 |
+
from safetensors.torch import load_file
|
| 30 |
+
|
| 31 |
+
from .resnet import InflatedConv3d, InflatedGroupNorm
|
| 32 |
+
from .unet_3d_blocks import UNetMidBlock3DCrossAttn, get_down_block, get_up_block
|
| 33 |
+
|
| 34 |
+
logger = logging.get_logger(__name__) # pylint: disable=invalid-name
|
| 35 |
+
|
| 36 |
+
|
| 37 |
+
@dataclass
|
| 38 |
+
class UNet3DConditionOutput(BaseOutput):
|
| 39 |
+
sample: torch.FloatTensor
|
| 40 |
+
|
| 41 |
+
|
| 42 |
+
class UNet3DConditionModel(ModelMixin, ConfigMixin):
|
| 43 |
+
_supports_gradient_checkpointing = True
|
| 44 |
+
|
| 45 |
+
@register_to_config
|
| 46 |
+
def __init__(
|
| 47 |
+
self,
|
| 48 |
+
sample_size: Optional[int] = None,
|
| 49 |
+
in_channels: int = 4,
|
| 50 |
+
out_channels: int = 4,
|
| 51 |
+
center_input_sample: bool = False,
|
| 52 |
+
flip_sin_to_cos: bool = True,
|
| 53 |
+
freq_shift: int = 0,
|
| 54 |
+
down_block_types: Tuple[str] = (
|
| 55 |
+
"CrossAttnDownBlock3D",
|
| 56 |
+
"CrossAttnDownBlock3D",
|
| 57 |
+
"CrossAttnDownBlock3D",
|
| 58 |
+
"DownBlock3D",
|
| 59 |
+
),
|
| 60 |
+
mid_block_type: str = "UNetMidBlock3DCrossAttn",
|
| 61 |
+
up_block_types: Tuple[str] = (
|
| 62 |
+
"UpBlock3D",
|
| 63 |
+
"CrossAttnUpBlock3D",
|
| 64 |
+
"CrossAttnUpBlock3D",
|
| 65 |
+
"CrossAttnUpBlock3D",
|
| 66 |
+
),
|
| 67 |
+
only_cross_attention: Union[bool, Tuple[bool]] = False,
|
| 68 |
+
block_out_channels: Tuple[int] = (320, 640, 1280, 1280),
|
| 69 |
+
layers_per_block: int = 2,
|
| 70 |
+
downsample_padding: int = 1,
|
| 71 |
+
mid_block_scale_factor: float = 1,
|
| 72 |
+
act_fn: str = "silu",
|
| 73 |
+
norm_num_groups: int = 32,
|
| 74 |
+
norm_eps: float = 1e-5,
|
| 75 |
+
cross_attention_dim: int = 1280,
|
| 76 |
+
attention_head_dim: Union[int, Tuple[int]] = 8,
|
| 77 |
+
dual_cross_attention: bool = False,
|
| 78 |
+
use_linear_projection: bool = False,
|
| 79 |
+
class_embed_type: Optional[str] = None,
|
| 80 |
+
num_class_embeds: Optional[int] = None,
|
| 81 |
+
upcast_attention: bool = False,
|
| 82 |
+
resnet_time_scale_shift: str = "default",
|
| 83 |
+
use_inflated_groupnorm=False,
|
| 84 |
+
# Additional
|
| 85 |
+
use_motion_module=False,
|
| 86 |
+
motion_module_resolutions=(1, 2, 4, 8),
|
| 87 |
+
motion_module_mid_block=False,
|
| 88 |
+
motion_module_decoder_only=False,
|
| 89 |
+
motion_module_type=None,
|
| 90 |
+
motion_module_kwargs={},
|
| 91 |
+
unet_use_cross_frame_attention=None,
|
| 92 |
+
unet_use_temporal_attention=None,
|
| 93 |
+
use_zero_convs=False,
|
| 94 |
+
):
|
| 95 |
+
super().__init__()
|
| 96 |
+
|
| 97 |
+
self.sample_size = sample_size
|
| 98 |
+
time_embed_dim = block_out_channels[0] * 4
|
| 99 |
+
|
| 100 |
+
# input
|
| 101 |
+
self.conv_in = InflatedConv3d(
|
| 102 |
+
in_channels, block_out_channels[0], kernel_size=3, padding=(1, 1)
|
| 103 |
+
)
|
| 104 |
+
|
| 105 |
+
# time
|
| 106 |
+
self.time_proj = Timesteps(block_out_channels[0], flip_sin_to_cos, freq_shift)
|
| 107 |
+
timestep_input_dim = block_out_channels[0]
|
| 108 |
+
|
| 109 |
+
self.time_embedding = TimestepEmbedding(timestep_input_dim, time_embed_dim)
|
| 110 |
+
|
| 111 |
+
# class embedding
|
| 112 |
+
if class_embed_type is None and num_class_embeds is not None:
|
| 113 |
+
self.class_embedding = nn.Embedding(num_class_embeds, time_embed_dim)
|
| 114 |
+
elif class_embed_type == "timestep":
|
| 115 |
+
self.class_embedding = TimestepEmbedding(timestep_input_dim, time_embed_dim)
|
| 116 |
+
elif class_embed_type == "identity":
|
| 117 |
+
self.class_embedding = nn.Identity(time_embed_dim, time_embed_dim)
|
| 118 |
+
else:
|
| 119 |
+
self.class_embedding = None
|
| 120 |
+
|
| 121 |
+
self.down_blocks = nn.ModuleList([])
|
| 122 |
+
self.mid_block = None
|
| 123 |
+
self.up_blocks = nn.ModuleList([])
|
| 124 |
+
|
| 125 |
+
if isinstance(only_cross_attention, bool):
|
| 126 |
+
only_cross_attention = [only_cross_attention] * len(down_block_types)
|
| 127 |
+
|
| 128 |
+
if isinstance(attention_head_dim, int):
|
| 129 |
+
attention_head_dim = (attention_head_dim,) * len(down_block_types)
|
| 130 |
+
|
| 131 |
+
# down
|
| 132 |
+
output_channel = block_out_channels[0]
|
| 133 |
+
for i, down_block_type in enumerate(down_block_types):
|
| 134 |
+
res = 2**i
|
| 135 |
+
input_channel = output_channel
|
| 136 |
+
output_channel = block_out_channels[i]
|
| 137 |
+
is_final_block = i == len(block_out_channels) - 1
|
| 138 |
+
|
| 139 |
+
down_block = get_down_block(
|
| 140 |
+
down_block_type,
|
| 141 |
+
num_layers=layers_per_block,
|
| 142 |
+
in_channels=input_channel,
|
| 143 |
+
out_channels=output_channel,
|
| 144 |
+
temb_channels=time_embed_dim,
|
| 145 |
+
add_downsample=not is_final_block,
|
| 146 |
+
resnet_eps=norm_eps,
|
| 147 |
+
resnet_act_fn=act_fn,
|
| 148 |
+
resnet_groups=norm_num_groups,
|
| 149 |
+
cross_attention_dim=cross_attention_dim,
|
| 150 |
+
attn_num_head_channels=attention_head_dim[i],
|
| 151 |
+
downsample_padding=downsample_padding,
|
| 152 |
+
dual_cross_attention=dual_cross_attention,
|
| 153 |
+
use_linear_projection=use_linear_projection,
|
| 154 |
+
only_cross_attention=only_cross_attention[i],
|
| 155 |
+
upcast_attention=upcast_attention,
|
| 156 |
+
resnet_time_scale_shift=resnet_time_scale_shift,
|
| 157 |
+
unet_use_cross_frame_attention=unet_use_cross_frame_attention,
|
| 158 |
+
unet_use_temporal_attention=unet_use_temporal_attention,
|
| 159 |
+
use_inflated_groupnorm=use_inflated_groupnorm,
|
| 160 |
+
use_motion_module=use_motion_module
|
| 161 |
+
and (res in motion_module_resolutions)
|
| 162 |
+
and (not motion_module_decoder_only),
|
| 163 |
+
motion_module_type=motion_module_type,
|
| 164 |
+
motion_module_kwargs=motion_module_kwargs,
|
| 165 |
+
use_zero_convs=use_zero_convs,
|
| 166 |
+
)
|
| 167 |
+
self.down_blocks.append(down_block)
|
| 168 |
+
|
| 169 |
+
# mid
|
| 170 |
+
if mid_block_type == "UNetMidBlock3DCrossAttn":
|
| 171 |
+
self.mid_block = UNetMidBlock3DCrossAttn(
|
| 172 |
+
in_channels=block_out_channels[-1],
|
| 173 |
+
temb_channels=time_embed_dim,
|
| 174 |
+
resnet_eps=norm_eps,
|
| 175 |
+
resnet_act_fn=act_fn,
|
| 176 |
+
output_scale_factor=mid_block_scale_factor,
|
| 177 |
+
resnet_time_scale_shift=resnet_time_scale_shift,
|
| 178 |
+
cross_attention_dim=cross_attention_dim,
|
| 179 |
+
attn_num_head_channels=attention_head_dim[-1],
|
| 180 |
+
resnet_groups=norm_num_groups,
|
| 181 |
+
dual_cross_attention=dual_cross_attention,
|
| 182 |
+
use_linear_projection=use_linear_projection,
|
| 183 |
+
upcast_attention=upcast_attention,
|
| 184 |
+
unet_use_cross_frame_attention=unet_use_cross_frame_attention,
|
| 185 |
+
unet_use_temporal_attention=unet_use_temporal_attention,
|
| 186 |
+
use_inflated_groupnorm=use_inflated_groupnorm,
|
| 187 |
+
use_motion_module=use_motion_module and motion_module_mid_block,
|
| 188 |
+
motion_module_type=motion_module_type,
|
| 189 |
+
motion_module_kwargs=motion_module_kwargs,
|
| 190 |
+
use_zero_convs=use_zero_convs,
|
| 191 |
+
)
|
| 192 |
+
else:
|
| 193 |
+
raise ValueError(f"unknown mid_block_type : {mid_block_type}")
|
| 194 |
+
|
| 195 |
+
# count how many layers upsample the videos
|
| 196 |
+
self.num_upsamplers = 0
|
| 197 |
+
|
| 198 |
+
# up
|
| 199 |
+
reversed_block_out_channels = list(reversed(block_out_channels))
|
| 200 |
+
reversed_attention_head_dim = list(reversed(attention_head_dim))
|
| 201 |
+
only_cross_attention = list(reversed(only_cross_attention))
|
| 202 |
+
output_channel = reversed_block_out_channels[0]
|
| 203 |
+
for i, up_block_type in enumerate(up_block_types):
|
| 204 |
+
res = 2 ** (3 - i)
|
| 205 |
+
is_final_block = i == len(block_out_channels) - 1
|
| 206 |
+
|
| 207 |
+
prev_output_channel = output_channel
|
| 208 |
+
output_channel = reversed_block_out_channels[i]
|
| 209 |
+
input_channel = reversed_block_out_channels[
|
| 210 |
+
min(i + 1, len(block_out_channels) - 1)
|
| 211 |
+
]
|
| 212 |
+
|
| 213 |
+
# add upsample block for all BUT final layer
|
| 214 |
+
if not is_final_block:
|
| 215 |
+
add_upsample = True
|
| 216 |
+
self.num_upsamplers += 1
|
| 217 |
+
else:
|
| 218 |
+
add_upsample = False
|
| 219 |
+
|
| 220 |
+
up_block = get_up_block(
|
| 221 |
+
up_block_type,
|
| 222 |
+
num_layers=layers_per_block + 1,
|
| 223 |
+
in_channels=input_channel,
|
| 224 |
+
out_channels=output_channel,
|
| 225 |
+
prev_output_channel=prev_output_channel,
|
| 226 |
+
temb_channels=time_embed_dim,
|
| 227 |
+
add_upsample=add_upsample,
|
| 228 |
+
resnet_eps=norm_eps,
|
| 229 |
+
resnet_act_fn=act_fn,
|
| 230 |
+
resnet_groups=norm_num_groups,
|
| 231 |
+
cross_attention_dim=cross_attention_dim,
|
| 232 |
+
attn_num_head_channels=reversed_attention_head_dim[i],
|
| 233 |
+
dual_cross_attention=dual_cross_attention,
|
| 234 |
+
use_linear_projection=use_linear_projection,
|
| 235 |
+
only_cross_attention=only_cross_attention[i],
|
| 236 |
+
upcast_attention=upcast_attention,
|
| 237 |
+
resnet_time_scale_shift=resnet_time_scale_shift,
|
| 238 |
+
unet_use_cross_frame_attention=unet_use_cross_frame_attention,
|
| 239 |
+
unet_use_temporal_attention=unet_use_temporal_attention,
|
| 240 |
+
use_inflated_groupnorm=use_inflated_groupnorm,
|
| 241 |
+
use_motion_module=use_motion_module
|
| 242 |
+
and (res in motion_module_resolutions),
|
| 243 |
+
motion_module_type=motion_module_type,
|
| 244 |
+
motion_module_kwargs=motion_module_kwargs,
|
| 245 |
+
use_zero_convs=use_zero_convs,
|
| 246 |
+
)
|
| 247 |
+
self.up_blocks.append(up_block)
|
| 248 |
+
prev_output_channel = output_channel
|
| 249 |
+
|
| 250 |
+
# out
|
| 251 |
+
if use_inflated_groupnorm:
|
| 252 |
+
self.conv_norm_out = InflatedGroupNorm(
|
| 253 |
+
num_channels=block_out_channels[0],
|
| 254 |
+
num_groups=norm_num_groups,
|
| 255 |
+
eps=norm_eps,
|
| 256 |
+
)
|
| 257 |
+
else:
|
| 258 |
+
self.conv_norm_out = nn.GroupNorm(
|
| 259 |
+
num_channels=block_out_channels[0],
|
| 260 |
+
num_groups=norm_num_groups,
|
| 261 |
+
eps=norm_eps,
|
| 262 |
+
)
|
| 263 |
+
self.conv_act = nn.SiLU()
|
| 264 |
+
self.conv_out = InflatedConv3d(
|
| 265 |
+
block_out_channels[0], out_channels, kernel_size=3, padding=1
|
| 266 |
+
)
|
| 267 |
+
|
| 268 |
+
|
| 269 |
+
|
| 270 |
+
@property
|
| 271 |
+
# Copied from diffusers.models.unet_2d_condition.UNet2DConditionModel.attn_processors
|
| 272 |
+
def attn_processors(self) -> Dict[str, AttentionProcessor]:
|
| 273 |
+
r"""
|
| 274 |
+
Returns:
|
| 275 |
+
`dict` of attention processors: A dictionary containing all attention processors used in the model with
|
| 276 |
+
indexed by its weight name.
|
| 277 |
+
"""
|
| 278 |
+
# set recursively
|
| 279 |
+
processors = {}
|
| 280 |
+
|
| 281 |
+
def fn_recursive_add_processors(
|
| 282 |
+
name: str,
|
| 283 |
+
module: torch.nn.Module,
|
| 284 |
+
processors: Dict[str, AttentionProcessor],
|
| 285 |
+
):
|
| 286 |
+
if hasattr(module, "set_processor"):
|
| 287 |
+
processors[f"{name}.processor"] = module.processor
|
| 288 |
+
|
| 289 |
+
for sub_name, child in module.named_children():
|
| 290 |
+
if "temporal_transformer" not in sub_name:
|
| 291 |
+
fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)
|
| 292 |
+
|
| 293 |
+
return processors
|
| 294 |
+
|
| 295 |
+
for name, module in self.named_children():
|
| 296 |
+
if "temporal_transformer" not in name:
|
| 297 |
+
fn_recursive_add_processors(name, module, processors)
|
| 298 |
+
|
| 299 |
+
return processors
|
| 300 |
+
|
| 301 |
+
def set_attention_slice(self, slice_size):
|
| 302 |
+
r"""
|
| 303 |
+
Enable sliced attention computation.
|
| 304 |
+
|
| 305 |
+
When this option is enabled, the attention module will split the input tensor in slices, to compute attention
|
| 306 |
+
in several steps. This is useful to save some memory in exchange for a small speed decrease.
|
| 307 |
+
|
| 308 |
+
Args:
|
| 309 |
+
slice_size (`str` or `int` or `list(int)`, *optional*, defaults to `"auto"`):
|
| 310 |
+
When `"auto"`, halves the input to the attention heads, so attention will be computed in two steps. If
|
| 311 |
+
`"max"`, maxium amount of memory will be saved by running only one slice at a time. If a number is
|
| 312 |
+
provided, uses as many slices as `attention_head_dim // slice_size`. In this case, `attention_head_dim`
|
| 313 |
+
must be a multiple of `slice_size`.
|
| 314 |
+
"""
|
| 315 |
+
sliceable_head_dims = []
|
| 316 |
+
|
| 317 |
+
def fn_recursive_retrieve_slicable_dims(module: torch.nn.Module):
|
| 318 |
+
if hasattr(module, "set_attention_slice"):
|
| 319 |
+
sliceable_head_dims.append(module.sliceable_head_dim)
|
| 320 |
+
|
| 321 |
+
for child in module.children():
|
| 322 |
+
fn_recursive_retrieve_slicable_dims(child)
|
| 323 |
+
|
| 324 |
+
# retrieve number of attention layers
|
| 325 |
+
for module in self.children():
|
| 326 |
+
fn_recursive_retrieve_slicable_dims(module)
|
| 327 |
+
|
| 328 |
+
num_slicable_layers = len(sliceable_head_dims)
|
| 329 |
+
|
| 330 |
+
if slice_size == "auto":
|
| 331 |
+
# half the attention head size is usually a good trade-off between
|
| 332 |
+
# speed and memory
|
| 333 |
+
slice_size = [dim // 2 for dim in sliceable_head_dims]
|
| 334 |
+
elif slice_size == "max":
|
| 335 |
+
# make smallest slice possible
|
| 336 |
+
slice_size = num_slicable_layers * [1]
|
| 337 |
+
|
| 338 |
+
slice_size = (
|
| 339 |
+
num_slicable_layers * [slice_size]
|
| 340 |
+
if not isinstance(slice_size, list)
|
| 341 |
+
else slice_size
|
| 342 |
+
)
|
| 343 |
+
|
| 344 |
+
if len(slice_size) != len(sliceable_head_dims):
|
| 345 |
+
raise ValueError(
|
| 346 |
+
f"You have provided {len(slice_size)}, but {self.config} has {len(sliceable_head_dims)} different"
|
| 347 |
+
f" attention layers. Make sure to match `len(slice_size)` to be {len(sliceable_head_dims)}."
|
| 348 |
+
)
|
| 349 |
+
|
| 350 |
+
for i in range(len(slice_size)):
|
| 351 |
+
size = slice_size[i]
|
| 352 |
+
dim = sliceable_head_dims[i]
|
| 353 |
+
if size is not None and size > dim:
|
| 354 |
+
raise ValueError(f"size {size} has to be smaller or equal to {dim}.")
|
| 355 |
+
|
| 356 |
+
# Recursively walk through all the children.
|
| 357 |
+
# Any children which exposes the set_attention_slice method
|
| 358 |
+
# gets the message
|
| 359 |
+
def fn_recursive_set_attention_slice(
|
| 360 |
+
module: torch.nn.Module, slice_size: List[int]
|
| 361 |
+
):
|
| 362 |
+
if hasattr(module, "set_attention_slice"):
|
| 363 |
+
module.set_attention_slice(slice_size.pop())
|
| 364 |
+
|
| 365 |
+
for child in module.children():
|
| 366 |
+
fn_recursive_set_attention_slice(child, slice_size)
|
| 367 |
+
|
| 368 |
+
reversed_slice_size = list(reversed(slice_size))
|
| 369 |
+
for module in self.children():
|
| 370 |
+
fn_recursive_set_attention_slice(module, reversed_slice_size)
|
| 371 |
+
|
| 372 |
+
def _set_gradient_checkpointing(self, module, value=False):
|
| 373 |
+
if hasattr(module, "gradient_checkpointing"):
|
| 374 |
+
module.gradient_checkpointing = value
|
| 375 |
+
|
| 376 |
+
# Copied from diffusers.models.unet_2d_condition.UNet2DConditionModel.set_attn_processor
|
| 377 |
+
def set_attn_processor(
|
| 378 |
+
self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]
|
| 379 |
+
):
|
| 380 |
+
r"""
|
| 381 |
+
Sets the attention processor to use to compute attention.
|
| 382 |
+
|
| 383 |
+
Parameters:
|
| 384 |
+
processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
|
| 385 |
+
The instantiated processor class or a dictionary of processor classes that will be set as the processor
|
| 386 |
+
for **all** `Attention` layers.
|
| 387 |
+
|
| 388 |
+
If `processor` is a dict, the key needs to define the path to the corresponding cross attention
|
| 389 |
+
processor. This is strongly recommended when setting trainable attention processors.
|
| 390 |
+
|
| 391 |
+
"""
|
| 392 |
+
count = len(self.attn_processors.keys())
|
| 393 |
+
|
| 394 |
+
if isinstance(processor, dict) and len(processor) != count:
|
| 395 |
+
raise ValueError(
|
| 396 |
+
f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
|
| 397 |
+
f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
|
| 398 |
+
)
|
| 399 |
+
|
| 400 |
+
def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor):
|
| 401 |
+
if hasattr(module, "set_processor"):
|
| 402 |
+
if not isinstance(processor, dict):
|
| 403 |
+
module.set_processor(processor)
|
| 404 |
+
else:
|
| 405 |
+
module.set_processor(processor.pop(f"{name}.processor"))
|
| 406 |
+
|
| 407 |
+
for sub_name, child in module.named_children():
|
| 408 |
+
if "temporal_transformer" not in sub_name:
|
| 409 |
+
fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)
|
| 410 |
+
|
| 411 |
+
for name, module in self.named_children():
|
| 412 |
+
if "temporal_transformer" not in name:
|
| 413 |
+
fn_recursive_attn_processor(name, module, processor)
|
| 414 |
+
|
| 415 |
+
def forward(
|
| 416 |
+
self,
|
| 417 |
+
sample: torch.FloatTensor,
|
| 418 |
+
timestep: Union[torch.Tensor, float, int],
|
| 419 |
+
encoder_hidden_states: torch.Tensor,
|
| 420 |
+
class_labels: Optional[torch.Tensor] = None,
|
| 421 |
+
pose_cond_fea: Optional[torch.Tensor] = None,
|
| 422 |
+
attention_mask: Optional[torch.Tensor] = None,
|
| 423 |
+
down_block_additional_residuals: Optional[Tuple[torch.Tensor]] = None,
|
| 424 |
+
mid_block_additional_residual: Optional[torch.Tensor] = None,
|
| 425 |
+
return_dict: bool = True,
|
| 426 |
+
) -> Union[UNet3DConditionOutput, Tuple]:
|
| 427 |
+
r"""
|
| 428 |
+
Args:
|
| 429 |
+
sample (`torch.FloatTensor`): (batch, channel, height, width) noisy inputs tensor
|
| 430 |
+
timestep (`torch.FloatTensor` or `float` or `int`): (batch) timesteps
|
| 431 |
+
encoder_hidden_states (`torch.FloatTensor`): (batch, sequence_length, feature_dim) encoder hidden states
|
| 432 |
+
return_dict (`bool`, *optional*, defaults to `True`):
|
| 433 |
+
Whether or not to return a [`models.unet_2d_condition.UNet2DConditionOutput`] instead of a plain tuple.
|
| 434 |
+
|
| 435 |
+
Returns:
|
| 436 |
+
[`~models.unet_2d_condition.UNet2DConditionOutput`] or `tuple`:
|
| 437 |
+
[`~models.unet_2d_condition.UNet2DConditionOutput`] if `return_dict` is True, otherwise a `tuple`. When
|
| 438 |
+
returning a tuple, the first element is the sample tensor.
|
| 439 |
+
"""
|
| 440 |
+
# By default samples have to be AT least a multiple of the overall upsampling factor.
|
| 441 |
+
# The overall upsampling factor is equal to 2 ** (# num of upsampling layears).
|
| 442 |
+
# However, the upsampling interpolation output size can be forced to fit any upsampling size
|
| 443 |
+
# on the fly if necessary.
|
| 444 |
+
default_overall_up_factor = 2**self.num_upsamplers
|
| 445 |
+
|
| 446 |
+
# upsample size should be forwarded when sample is not a multiple of `default_overall_up_factor`
|
| 447 |
+
forward_upsample_size = False
|
| 448 |
+
upsample_size = None
|
| 449 |
+
|
| 450 |
+
if any(s % default_overall_up_factor != 0 for s in sample.shape[-2:]):
|
| 451 |
+
logger.info("Forward upsample size to force interpolation output size.")
|
| 452 |
+
forward_upsample_size = True
|
| 453 |
+
|
| 454 |
+
# prepare attention_mask
|
| 455 |
+
if attention_mask is not None:
|
| 456 |
+
attention_mask = (1 - attention_mask.to(sample.dtype)) * -10000.0
|
| 457 |
+
attention_mask = attention_mask.unsqueeze(1)
|
| 458 |
+
|
| 459 |
+
# center input if necessary
|
| 460 |
+
if self.config.center_input_sample:
|
| 461 |
+
sample = 2 * sample - 1.0
|
| 462 |
+
|
| 463 |
+
# time
|
| 464 |
+
timesteps = timestep
|
| 465 |
+
if not torch.is_tensor(timesteps):
|
| 466 |
+
# This would be a good case for the `match` statement (Python 3.10+)
|
| 467 |
+
is_mps = sample.device.type == "mps"
|
| 468 |
+
if isinstance(timestep, float):
|
| 469 |
+
dtype = torch.float32 if is_mps else torch.float64
|
| 470 |
+
else:
|
| 471 |
+
dtype = torch.int32 if is_mps else torch.int64
|
| 472 |
+
timesteps = torch.tensor([timesteps], dtype=dtype, device=sample.device)
|
| 473 |
+
elif len(timesteps.shape) == 0:
|
| 474 |
+
timesteps = timesteps[None].to(sample.device)
|
| 475 |
+
|
| 476 |
+
# broadcast to batch dimension in a way that's compatible with ONNX/Core ML
|
| 477 |
+
timesteps = timesteps.expand(sample.shape[0])
|
| 478 |
+
|
| 479 |
+
t_emb = self.time_proj(timesteps)
|
| 480 |
+
|
| 481 |
+
# timesteps does not contain any weights and will always return f32 tensors
|
| 482 |
+
# but time_embedding might actually be running in fp16. so we need to cast here.
|
| 483 |
+
# there might be better ways to encapsulate this.
|
| 484 |
+
t_emb = t_emb.to(dtype=self.dtype)
|
| 485 |
+
emb = self.time_embedding(t_emb)
|
| 486 |
+
|
| 487 |
+
if self.class_embedding is not None:
|
| 488 |
+
if class_labels is None:
|
| 489 |
+
raise ValueError(
|
| 490 |
+
"class_labels should be provided when num_class_embeds > 0"
|
| 491 |
+
)
|
| 492 |
+
|
| 493 |
+
if self.config.class_embed_type == "timestep":
|
| 494 |
+
class_labels = self.time_proj(class_labels)
|
| 495 |
+
|
| 496 |
+
class_emb = self.class_embedding(class_labels).to(dtype=self.dtype)
|
| 497 |
+
emb = emb + class_emb
|
| 498 |
+
|
| 499 |
+
# pre-process
|
| 500 |
+
sample = self.conv_in(sample)
|
| 501 |
+
if pose_cond_fea is not None:
|
| 502 |
+
sample = sample + pose_cond_fea
|
| 503 |
+
|
| 504 |
+
# down
|
| 505 |
+
down_block_res_samples = (sample,)
|
| 506 |
+
for downsample_block in self.down_blocks:
|
| 507 |
+
if (
|
| 508 |
+
hasattr(downsample_block, "has_cross_attention")
|
| 509 |
+
and downsample_block.has_cross_attention
|
| 510 |
+
):
|
| 511 |
+
sample, res_samples = downsample_block(
|
| 512 |
+
hidden_states=sample,
|
| 513 |
+
temb=emb,
|
| 514 |
+
encoder_hidden_states=encoder_hidden_states,
|
| 515 |
+
attention_mask=attention_mask,
|
| 516 |
+
)
|
| 517 |
+
else:
|
| 518 |
+
sample, res_samples = downsample_block(
|
| 519 |
+
hidden_states=sample,
|
| 520 |
+
temb=emb,
|
| 521 |
+
encoder_hidden_states=encoder_hidden_states,
|
| 522 |
+
)
|
| 523 |
+
|
| 524 |
+
down_block_res_samples += res_samples
|
| 525 |
+
|
| 526 |
+
if down_block_additional_residuals is not None:
|
| 527 |
+
new_down_block_res_samples = ()
|
| 528 |
+
|
| 529 |
+
for down_block_res_sample, down_block_additional_residual in zip(
|
| 530 |
+
down_block_res_samples, down_block_additional_residuals
|
| 531 |
+
):
|
| 532 |
+
down_block_res_sample = (
|
| 533 |
+
down_block_res_sample + down_block_additional_residual
|
| 534 |
+
)
|
| 535 |
+
new_down_block_res_samples += (down_block_res_sample,)
|
| 536 |
+
|
| 537 |
+
down_block_res_samples = new_down_block_res_samples
|
| 538 |
+
|
| 539 |
+
# mid
|
| 540 |
+
sample = self.mid_block(
|
| 541 |
+
sample,
|
| 542 |
+
emb,
|
| 543 |
+
encoder_hidden_states=encoder_hidden_states,
|
| 544 |
+
attention_mask=attention_mask,
|
| 545 |
+
)
|
| 546 |
+
|
| 547 |
+
if mid_block_additional_residual is not None:
|
| 548 |
+
sample = sample + mid_block_additional_residual
|
| 549 |
+
|
| 550 |
+
# up
|
| 551 |
+
for i, upsample_block in enumerate(self.up_blocks):
|
| 552 |
+
is_final_block = i == len(self.up_blocks) - 1
|
| 553 |
+
|
| 554 |
+
res_samples = down_block_res_samples[-len(upsample_block.resnets) :]
|
| 555 |
+
down_block_res_samples = down_block_res_samples[
|
| 556 |
+
: -len(upsample_block.resnets)
|
| 557 |
+
]
|
| 558 |
+
|
| 559 |
+
# if we have not reached the final block and need to forward the
|
| 560 |
+
# upsample size, we do it here
|
| 561 |
+
if not is_final_block and forward_upsample_size:
|
| 562 |
+
upsample_size = down_block_res_samples[-1].shape[2:]
|
| 563 |
+
|
| 564 |
+
if (
|
| 565 |
+
hasattr(upsample_block, "has_cross_attention")
|
| 566 |
+
and upsample_block.has_cross_attention
|
| 567 |
+
):
|
| 568 |
+
sample = upsample_block(
|
| 569 |
+
hidden_states=sample,
|
| 570 |
+
temb=emb,
|
| 571 |
+
res_hidden_states_tuple=res_samples,
|
| 572 |
+
encoder_hidden_states=encoder_hidden_states,
|
| 573 |
+
upsample_size=upsample_size,
|
| 574 |
+
attention_mask=attention_mask,
|
| 575 |
+
)
|
| 576 |
+
else:
|
| 577 |
+
sample = upsample_block(
|
| 578 |
+
hidden_states=sample,
|
| 579 |
+
temb=emb,
|
| 580 |
+
res_hidden_states_tuple=res_samples,
|
| 581 |
+
upsample_size=upsample_size,
|
| 582 |
+
encoder_hidden_states=encoder_hidden_states,
|
| 583 |
+
)
|
| 584 |
+
|
| 585 |
+
# post-process
|
| 586 |
+
sample = self.conv_norm_out(sample)
|
| 587 |
+
sample = self.conv_act(sample)
|
| 588 |
+
sample = self.conv_out(sample)
|
| 589 |
+
|
| 590 |
+
if not return_dict:
|
| 591 |
+
return (sample,)
|
| 592 |
+
|
| 593 |
+
return UNet3DConditionOutput(sample=sample)
|
| 594 |
+
|
| 595 |
+
@classmethod
|
| 596 |
+
def from_pretrained_2d(
|
| 597 |
+
cls,
|
| 598 |
+
config_file: PathLike,
|
| 599 |
+
ckpt_file: PathLike
|
| 600 |
+
):
|
| 601 |
+
unet_additional_kwargs={
|
| 602 |
+
"use_motion_module": False,
|
| 603 |
+
"unet_use_temporal_attention": False,
|
| 604 |
+
"use_zero_convs": False
|
| 605 |
+
}
|
| 606 |
+
|
| 607 |
+
config_file = Path(config_file)
|
| 608 |
+
ckpt_file = Path(ckpt_file)
|
| 609 |
+
|
| 610 |
+
if not (config_file.exists() and config_file.is_file()):
|
| 611 |
+
raise RuntimeError(f"{config_file} does not exist or is not a file")
|
| 612 |
+
if not (config_file.exists() and config_file.is_file()):
|
| 613 |
+
raise RuntimeError(f"{ckpt_file} does not exist or is not a file")
|
| 614 |
+
|
| 615 |
+
unet_config = cls.load_config(config_file)
|
| 616 |
+
unet_config["_class_name"] = cls.__name__
|
| 617 |
+
unet_config["down_block_types"] = [
|
| 618 |
+
"CrossAttnDownBlock3D",
|
| 619 |
+
"CrossAttnDownBlock3D",
|
| 620 |
+
"CrossAttnDownBlock3D",
|
| 621 |
+
"DownBlock3D",
|
| 622 |
+
]
|
| 623 |
+
unet_config["up_block_types"] = [
|
| 624 |
+
"UpBlock3D",
|
| 625 |
+
"CrossAttnUpBlock3D",
|
| 626 |
+
"CrossAttnUpBlock3D",
|
| 627 |
+
"CrossAttnUpBlock3D",
|
| 628 |
+
]
|
| 629 |
+
unet_config["mid_block_type"] = "UNetMidBlock3DCrossAttn"
|
| 630 |
+
|
| 631 |
+
model = cls.from_config(unet_config, **unet_additional_kwargs)
|
| 632 |
+
state_dict = torch.load(
|
| 633 |
+
ckpt_file, map_location="cpu", weights_only=True,
|
| 634 |
+
)
|
| 635 |
+
|
| 636 |
+
# load the weights into the model
|
| 637 |
+
m, u = model.load_state_dict(state_dict, strict=False)
|
| 638 |
+
logger.debug(f"### missing keys: {len(m)}; \n### unexpected keys: {len(u)};")
|
| 639 |
+
|
| 640 |
+
params = [
|
| 641 |
+
p.numel() if "temporal" in n else 0 for n, p in model.named_parameters()
|
| 642 |
+
]
|
| 643 |
+
logger.info(f"Loaded {sum(params) / 1e6}M-parameter motion module")
|
| 644 |
+
|
| 645 |
+
return model
|
genwarp/models/unet_3d_blocks.py
ADDED
|
@@ -0,0 +1,885 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# This code is adapted from below and then modified.
|
| 2 |
+
# -----------------------------------------------------------------------------
|
| 3 |
+
# Moore-AnimateAnyone
|
| 4 |
+
# Apache License, Version 2.0
|
| 5 |
+
# Copyright @2023-2024 Moore Threads Technology Co., Ltd.
|
| 6 |
+
# https://github.com/MooreThreads/Moore-AnimateAnyone
|
| 7 |
+
# -----------------------------------------------------------------------------
|
| 8 |
+
# Diffusers
|
| 9 |
+
# Apache License, Version 2.0
|
| 10 |
+
# Copyright (c) Hugging Face Inc.
|
| 11 |
+
# https://github.com/huggingface/diffusers
|
| 12 |
+
# ==============================================================================
|
| 13 |
+
|
| 14 |
+
# Adapted from https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/unet_2d_blocks.py
|
| 15 |
+
|
| 16 |
+
import pdb
|
| 17 |
+
|
| 18 |
+
import torch
|
| 19 |
+
from torch import nn
|
| 20 |
+
|
| 21 |
+
from .motion_module import get_motion_module
|
| 22 |
+
|
| 23 |
+
# from .motion_module import get_motion_module
|
| 24 |
+
from .resnet import Downsample3D, ResnetBlock3D, Upsample3D
|
| 25 |
+
from .transformer_3d import Transformer3DModel
|
| 26 |
+
|
| 27 |
+
|
| 28 |
+
def get_down_block(
|
| 29 |
+
down_block_type,
|
| 30 |
+
num_layers,
|
| 31 |
+
in_channels,
|
| 32 |
+
out_channels,
|
| 33 |
+
temb_channels,
|
| 34 |
+
add_downsample,
|
| 35 |
+
resnet_eps,
|
| 36 |
+
resnet_act_fn,
|
| 37 |
+
attn_num_head_channels,
|
| 38 |
+
resnet_groups=None,
|
| 39 |
+
cross_attention_dim=None,
|
| 40 |
+
downsample_padding=None,
|
| 41 |
+
dual_cross_attention=False,
|
| 42 |
+
use_linear_projection=False,
|
| 43 |
+
only_cross_attention=False,
|
| 44 |
+
upcast_attention=False,
|
| 45 |
+
resnet_time_scale_shift="default",
|
| 46 |
+
unet_use_cross_frame_attention=None,
|
| 47 |
+
unet_use_temporal_attention=None,
|
| 48 |
+
use_inflated_groupnorm=None,
|
| 49 |
+
use_motion_module=None,
|
| 50 |
+
motion_module_type=None,
|
| 51 |
+
motion_module_kwargs=None,
|
| 52 |
+
use_zero_convs=False,
|
| 53 |
+
):
|
| 54 |
+
down_block_type = (
|
| 55 |
+
down_block_type[7:]
|
| 56 |
+
if down_block_type.startswith("UNetRes")
|
| 57 |
+
else down_block_type
|
| 58 |
+
)
|
| 59 |
+
if down_block_type == "DownBlock3D":
|
| 60 |
+
return DownBlock3D(
|
| 61 |
+
num_layers=num_layers,
|
| 62 |
+
in_channels=in_channels,
|
| 63 |
+
out_channels=out_channels,
|
| 64 |
+
temb_channels=temb_channels,
|
| 65 |
+
add_downsample=add_downsample,
|
| 66 |
+
resnet_eps=resnet_eps,
|
| 67 |
+
resnet_act_fn=resnet_act_fn,
|
| 68 |
+
resnet_groups=resnet_groups,
|
| 69 |
+
downsample_padding=downsample_padding,
|
| 70 |
+
resnet_time_scale_shift=resnet_time_scale_shift,
|
| 71 |
+
use_inflated_groupnorm=use_inflated_groupnorm,
|
| 72 |
+
use_motion_module=use_motion_module,
|
| 73 |
+
motion_module_type=motion_module_type,
|
| 74 |
+
motion_module_kwargs=motion_module_kwargs,
|
| 75 |
+
)
|
| 76 |
+
elif down_block_type == "CrossAttnDownBlock3D":
|
| 77 |
+
if cross_attention_dim is None:
|
| 78 |
+
raise ValueError(
|
| 79 |
+
"cross_attention_dim must be specified for CrossAttnDownBlock3D"
|
| 80 |
+
)
|
| 81 |
+
return CrossAttnDownBlock3D(
|
| 82 |
+
num_layers=num_layers,
|
| 83 |
+
in_channels=in_channels,
|
| 84 |
+
out_channels=out_channels,
|
| 85 |
+
temb_channels=temb_channels,
|
| 86 |
+
add_downsample=add_downsample,
|
| 87 |
+
resnet_eps=resnet_eps,
|
| 88 |
+
resnet_act_fn=resnet_act_fn,
|
| 89 |
+
resnet_groups=resnet_groups,
|
| 90 |
+
downsample_padding=downsample_padding,
|
| 91 |
+
cross_attention_dim=cross_attention_dim,
|
| 92 |
+
attn_num_head_channels=attn_num_head_channels,
|
| 93 |
+
dual_cross_attention=dual_cross_attention,
|
| 94 |
+
use_linear_projection=use_linear_projection,
|
| 95 |
+
only_cross_attention=only_cross_attention,
|
| 96 |
+
upcast_attention=upcast_attention,
|
| 97 |
+
resnet_time_scale_shift=resnet_time_scale_shift,
|
| 98 |
+
unet_use_cross_frame_attention=unet_use_cross_frame_attention,
|
| 99 |
+
unet_use_temporal_attention=unet_use_temporal_attention,
|
| 100 |
+
use_inflated_groupnorm=use_inflated_groupnorm,
|
| 101 |
+
use_motion_module=use_motion_module,
|
| 102 |
+
motion_module_type=motion_module_type,
|
| 103 |
+
motion_module_kwargs=motion_module_kwargs,
|
| 104 |
+
use_zero_convs=use_zero_convs,
|
| 105 |
+
)
|
| 106 |
+
raise ValueError(f"{down_block_type} does not exist.")
|
| 107 |
+
|
| 108 |
+
|
| 109 |
+
def get_up_block(
|
| 110 |
+
up_block_type,
|
| 111 |
+
num_layers,
|
| 112 |
+
in_channels,
|
| 113 |
+
out_channels,
|
| 114 |
+
prev_output_channel,
|
| 115 |
+
temb_channels,
|
| 116 |
+
add_upsample,
|
| 117 |
+
resnet_eps,
|
| 118 |
+
resnet_act_fn,
|
| 119 |
+
attn_num_head_channels,
|
| 120 |
+
resnet_groups=None,
|
| 121 |
+
cross_attention_dim=None,
|
| 122 |
+
dual_cross_attention=False,
|
| 123 |
+
use_linear_projection=False,
|
| 124 |
+
only_cross_attention=False,
|
| 125 |
+
upcast_attention=False,
|
| 126 |
+
resnet_time_scale_shift="default",
|
| 127 |
+
unet_use_cross_frame_attention=None,
|
| 128 |
+
unet_use_temporal_attention=None,
|
| 129 |
+
use_inflated_groupnorm=None,
|
| 130 |
+
use_motion_module=None,
|
| 131 |
+
motion_module_type=None,
|
| 132 |
+
motion_module_kwargs=None,
|
| 133 |
+
use_zero_convs=False,
|
| 134 |
+
):
|
| 135 |
+
up_block_type = (
|
| 136 |
+
up_block_type[7:] if up_block_type.startswith("UNetRes") else up_block_type
|
| 137 |
+
)
|
| 138 |
+
if up_block_type == "UpBlock3D":
|
| 139 |
+
return UpBlock3D(
|
| 140 |
+
num_layers=num_layers,
|
| 141 |
+
in_channels=in_channels,
|
| 142 |
+
out_channels=out_channels,
|
| 143 |
+
prev_output_channel=prev_output_channel,
|
| 144 |
+
temb_channels=temb_channels,
|
| 145 |
+
add_upsample=add_upsample,
|
| 146 |
+
resnet_eps=resnet_eps,
|
| 147 |
+
resnet_act_fn=resnet_act_fn,
|
| 148 |
+
resnet_groups=resnet_groups,
|
| 149 |
+
resnet_time_scale_shift=resnet_time_scale_shift,
|
| 150 |
+
use_inflated_groupnorm=use_inflated_groupnorm,
|
| 151 |
+
use_motion_module=use_motion_module,
|
| 152 |
+
motion_module_type=motion_module_type,
|
| 153 |
+
motion_module_kwargs=motion_module_kwargs,
|
| 154 |
+
)
|
| 155 |
+
elif up_block_type == "CrossAttnUpBlock3D":
|
| 156 |
+
if cross_attention_dim is None:
|
| 157 |
+
raise ValueError(
|
| 158 |
+
"cross_attention_dim must be specified for CrossAttnUpBlock3D"
|
| 159 |
+
)
|
| 160 |
+
return CrossAttnUpBlock3D(
|
| 161 |
+
num_layers=num_layers,
|
| 162 |
+
in_channels=in_channels,
|
| 163 |
+
out_channels=out_channels,
|
| 164 |
+
prev_output_channel=prev_output_channel,
|
| 165 |
+
temb_channels=temb_channels,
|
| 166 |
+
add_upsample=add_upsample,
|
| 167 |
+
resnet_eps=resnet_eps,
|
| 168 |
+
resnet_act_fn=resnet_act_fn,
|
| 169 |
+
resnet_groups=resnet_groups,
|
| 170 |
+
cross_attention_dim=cross_attention_dim,
|
| 171 |
+
attn_num_head_channels=attn_num_head_channels,
|
| 172 |
+
dual_cross_attention=dual_cross_attention,
|
| 173 |
+
use_linear_projection=use_linear_projection,
|
| 174 |
+
only_cross_attention=only_cross_attention,
|
| 175 |
+
upcast_attention=upcast_attention,
|
| 176 |
+
resnet_time_scale_shift=resnet_time_scale_shift,
|
| 177 |
+
unet_use_cross_frame_attention=unet_use_cross_frame_attention,
|
| 178 |
+
unet_use_temporal_attention=unet_use_temporal_attention,
|
| 179 |
+
use_inflated_groupnorm=use_inflated_groupnorm,
|
| 180 |
+
use_motion_module=use_motion_module,
|
| 181 |
+
motion_module_type=motion_module_type,
|
| 182 |
+
motion_module_kwargs=motion_module_kwargs,
|
| 183 |
+
use_zero_convs=use_zero_convs,
|
| 184 |
+
)
|
| 185 |
+
raise ValueError(f"{up_block_type} does not exist.")
|
| 186 |
+
|
| 187 |
+
|
| 188 |
+
class UNetMidBlock3DCrossAttn(nn.Module):
|
| 189 |
+
def __init__(
|
| 190 |
+
self,
|
| 191 |
+
in_channels: int,
|
| 192 |
+
temb_channels: int,
|
| 193 |
+
dropout: float = 0.0,
|
| 194 |
+
num_layers: int = 1,
|
| 195 |
+
resnet_eps: float = 1e-6,
|
| 196 |
+
resnet_time_scale_shift: str = "default",
|
| 197 |
+
resnet_act_fn: str = "swish",
|
| 198 |
+
resnet_groups: int = 32,
|
| 199 |
+
resnet_pre_norm: bool = True,
|
| 200 |
+
attn_num_head_channels=1,
|
| 201 |
+
output_scale_factor=1.0,
|
| 202 |
+
cross_attention_dim=1280,
|
| 203 |
+
dual_cross_attention=False,
|
| 204 |
+
use_linear_projection=False,
|
| 205 |
+
upcast_attention=False,
|
| 206 |
+
unet_use_cross_frame_attention=None,
|
| 207 |
+
unet_use_temporal_attention=None,
|
| 208 |
+
use_inflated_groupnorm=None,
|
| 209 |
+
use_motion_module=None,
|
| 210 |
+
motion_module_type=None,
|
| 211 |
+
motion_module_kwargs=None,
|
| 212 |
+
use_zero_convs=False,
|
| 213 |
+
):
|
| 214 |
+
super().__init__()
|
| 215 |
+
|
| 216 |
+
self.has_cross_attention = True
|
| 217 |
+
self.attn_num_head_channels = attn_num_head_channels
|
| 218 |
+
resnet_groups = (
|
| 219 |
+
resnet_groups if resnet_groups is not None else min(in_channels // 4, 32)
|
| 220 |
+
)
|
| 221 |
+
|
| 222 |
+
# there is always at least one resnet
|
| 223 |
+
resnets = [
|
| 224 |
+
ResnetBlock3D(
|
| 225 |
+
in_channels=in_channels,
|
| 226 |
+
out_channels=in_channels,
|
| 227 |
+
temb_channels=temb_channels,
|
| 228 |
+
eps=resnet_eps,
|
| 229 |
+
groups=resnet_groups,
|
| 230 |
+
dropout=dropout,
|
| 231 |
+
time_embedding_norm=resnet_time_scale_shift,
|
| 232 |
+
non_linearity=resnet_act_fn,
|
| 233 |
+
output_scale_factor=output_scale_factor,
|
| 234 |
+
pre_norm=resnet_pre_norm,
|
| 235 |
+
use_inflated_groupnorm=use_inflated_groupnorm,
|
| 236 |
+
)
|
| 237 |
+
]
|
| 238 |
+
attentions = []
|
| 239 |
+
motion_modules = []
|
| 240 |
+
|
| 241 |
+
for _ in range(num_layers):
|
| 242 |
+
if dual_cross_attention:
|
| 243 |
+
raise NotImplementedError
|
| 244 |
+
attentions.append(
|
| 245 |
+
Transformer3DModel(
|
| 246 |
+
attn_num_head_channels,
|
| 247 |
+
in_channels // attn_num_head_channels,
|
| 248 |
+
in_channels=in_channels,
|
| 249 |
+
num_layers=1,
|
| 250 |
+
cross_attention_dim=cross_attention_dim,
|
| 251 |
+
norm_num_groups=resnet_groups,
|
| 252 |
+
use_linear_projection=use_linear_projection,
|
| 253 |
+
upcast_attention=upcast_attention,
|
| 254 |
+
unet_use_cross_frame_attention=unet_use_cross_frame_attention,
|
| 255 |
+
unet_use_temporal_attention=unet_use_temporal_attention,
|
| 256 |
+
use_zero_convs=use_zero_convs,
|
| 257 |
+
)
|
| 258 |
+
)
|
| 259 |
+
motion_modules.append(
|
| 260 |
+
get_motion_module(
|
| 261 |
+
in_channels=in_channels,
|
| 262 |
+
motion_module_type=motion_module_type,
|
| 263 |
+
motion_module_kwargs=motion_module_kwargs,
|
| 264 |
+
)
|
| 265 |
+
if use_motion_module
|
| 266 |
+
else None
|
| 267 |
+
)
|
| 268 |
+
resnets.append(
|
| 269 |
+
ResnetBlock3D(
|
| 270 |
+
in_channels=in_channels,
|
| 271 |
+
out_channels=in_channels,
|
| 272 |
+
temb_channels=temb_channels,
|
| 273 |
+
eps=resnet_eps,
|
| 274 |
+
groups=resnet_groups,
|
| 275 |
+
dropout=dropout,
|
| 276 |
+
time_embedding_norm=resnet_time_scale_shift,
|
| 277 |
+
non_linearity=resnet_act_fn,
|
| 278 |
+
output_scale_factor=output_scale_factor,
|
| 279 |
+
pre_norm=resnet_pre_norm,
|
| 280 |
+
use_inflated_groupnorm=use_inflated_groupnorm,
|
| 281 |
+
)
|
| 282 |
+
)
|
| 283 |
+
|
| 284 |
+
self.attentions = nn.ModuleList(attentions)
|
| 285 |
+
self.resnets = nn.ModuleList(resnets)
|
| 286 |
+
self.motion_modules = nn.ModuleList(motion_modules)
|
| 287 |
+
|
| 288 |
+
def forward(
|
| 289 |
+
self,
|
| 290 |
+
hidden_states,
|
| 291 |
+
temb=None,
|
| 292 |
+
encoder_hidden_states=None,
|
| 293 |
+
attention_mask=None,
|
| 294 |
+
):
|
| 295 |
+
hidden_states = self.resnets[0](hidden_states, temb)
|
| 296 |
+
for attn, resnet, motion_module in zip(
|
| 297 |
+
self.attentions, self.resnets[1:], self.motion_modules
|
| 298 |
+
):
|
| 299 |
+
hidden_states = attn(
|
| 300 |
+
hidden_states,
|
| 301 |
+
encoder_hidden_states=encoder_hidden_states,
|
| 302 |
+
).sample
|
| 303 |
+
hidden_states = (
|
| 304 |
+
motion_module(
|
| 305 |
+
hidden_states, temb, encoder_hidden_states=encoder_hidden_states
|
| 306 |
+
)
|
| 307 |
+
if motion_module is not None
|
| 308 |
+
else hidden_states
|
| 309 |
+
)
|
| 310 |
+
hidden_states = resnet(hidden_states, temb)
|
| 311 |
+
|
| 312 |
+
return hidden_states
|
| 313 |
+
|
| 314 |
+
|
| 315 |
+
class CrossAttnDownBlock3D(nn.Module):
|
| 316 |
+
def __init__(
|
| 317 |
+
self,
|
| 318 |
+
in_channels: int,
|
| 319 |
+
out_channels: int,
|
| 320 |
+
temb_channels: int,
|
| 321 |
+
dropout: float = 0.0,
|
| 322 |
+
num_layers: int = 1,
|
| 323 |
+
resnet_eps: float = 1e-6,
|
| 324 |
+
resnet_time_scale_shift: str = "default",
|
| 325 |
+
resnet_act_fn: str = "swish",
|
| 326 |
+
resnet_groups: int = 32,
|
| 327 |
+
resnet_pre_norm: bool = True,
|
| 328 |
+
attn_num_head_channels=1,
|
| 329 |
+
cross_attention_dim=1280,
|
| 330 |
+
output_scale_factor=1.0,
|
| 331 |
+
downsample_padding=1,
|
| 332 |
+
add_downsample=True,
|
| 333 |
+
dual_cross_attention=False,
|
| 334 |
+
use_linear_projection=False,
|
| 335 |
+
only_cross_attention=False,
|
| 336 |
+
upcast_attention=False,
|
| 337 |
+
unet_use_cross_frame_attention=None,
|
| 338 |
+
unet_use_temporal_attention=None,
|
| 339 |
+
use_inflated_groupnorm=None,
|
| 340 |
+
use_motion_module=None,
|
| 341 |
+
motion_module_type=None,
|
| 342 |
+
motion_module_kwargs=None,
|
| 343 |
+
use_zero_convs=False,
|
| 344 |
+
):
|
| 345 |
+
super().__init__()
|
| 346 |
+
resnets = []
|
| 347 |
+
attentions = []
|
| 348 |
+
motion_modules = []
|
| 349 |
+
|
| 350 |
+
self.has_cross_attention = True
|
| 351 |
+
self.attn_num_head_channels = attn_num_head_channels
|
| 352 |
+
|
| 353 |
+
for i in range(num_layers):
|
| 354 |
+
in_channels = in_channels if i == 0 else out_channels
|
| 355 |
+
resnets.append(
|
| 356 |
+
ResnetBlock3D(
|
| 357 |
+
in_channels=in_channels,
|
| 358 |
+
out_channels=out_channels,
|
| 359 |
+
temb_channels=temb_channels,
|
| 360 |
+
eps=resnet_eps,
|
| 361 |
+
groups=resnet_groups,
|
| 362 |
+
dropout=dropout,
|
| 363 |
+
time_embedding_norm=resnet_time_scale_shift,
|
| 364 |
+
non_linearity=resnet_act_fn,
|
| 365 |
+
output_scale_factor=output_scale_factor,
|
| 366 |
+
pre_norm=resnet_pre_norm,
|
| 367 |
+
use_inflated_groupnorm=use_inflated_groupnorm,
|
| 368 |
+
)
|
| 369 |
+
)
|
| 370 |
+
if dual_cross_attention:
|
| 371 |
+
raise NotImplementedError
|
| 372 |
+
attentions.append(
|
| 373 |
+
Transformer3DModel(
|
| 374 |
+
attn_num_head_channels,
|
| 375 |
+
out_channels // attn_num_head_channels,
|
| 376 |
+
in_channels=out_channels,
|
| 377 |
+
num_layers=1,
|
| 378 |
+
cross_attention_dim=cross_attention_dim,
|
| 379 |
+
norm_num_groups=resnet_groups,
|
| 380 |
+
use_linear_projection=use_linear_projection,
|
| 381 |
+
only_cross_attention=only_cross_attention,
|
| 382 |
+
upcast_attention=upcast_attention,
|
| 383 |
+
unet_use_cross_frame_attention=unet_use_cross_frame_attention,
|
| 384 |
+
unet_use_temporal_attention=unet_use_temporal_attention,
|
| 385 |
+
use_zero_convs=use_zero_convs,
|
| 386 |
+
)
|
| 387 |
+
)
|
| 388 |
+
motion_modules.append(
|
| 389 |
+
get_motion_module(
|
| 390 |
+
in_channels=out_channels,
|
| 391 |
+
motion_module_type=motion_module_type,
|
| 392 |
+
motion_module_kwargs=motion_module_kwargs,
|
| 393 |
+
)
|
| 394 |
+
if use_motion_module
|
| 395 |
+
else None
|
| 396 |
+
)
|
| 397 |
+
|
| 398 |
+
self.attentions = nn.ModuleList(attentions)
|
| 399 |
+
self.resnets = nn.ModuleList(resnets)
|
| 400 |
+
self.motion_modules = nn.ModuleList(motion_modules)
|
| 401 |
+
|
| 402 |
+
if add_downsample:
|
| 403 |
+
self.downsamplers = nn.ModuleList(
|
| 404 |
+
[
|
| 405 |
+
Downsample3D(
|
| 406 |
+
out_channels,
|
| 407 |
+
use_conv=True,
|
| 408 |
+
out_channels=out_channels,
|
| 409 |
+
padding=downsample_padding,
|
| 410 |
+
name="op",
|
| 411 |
+
)
|
| 412 |
+
]
|
| 413 |
+
)
|
| 414 |
+
else:
|
| 415 |
+
self.downsamplers = None
|
| 416 |
+
|
| 417 |
+
self.gradient_checkpointing = False
|
| 418 |
+
|
| 419 |
+
def forward(
|
| 420 |
+
self,
|
| 421 |
+
hidden_states,
|
| 422 |
+
temb=None,
|
| 423 |
+
encoder_hidden_states=None,
|
| 424 |
+
attention_mask=None,
|
| 425 |
+
):
|
| 426 |
+
output_states = ()
|
| 427 |
+
|
| 428 |
+
for i, (resnet, attn, motion_module) in enumerate(
|
| 429 |
+
zip(self.resnets, self.attentions, self.motion_modules)
|
| 430 |
+
):
|
| 431 |
+
# self.gradient_checkpointing = False
|
| 432 |
+
if self.training and self.gradient_checkpointing:
|
| 433 |
+
|
| 434 |
+
def create_custom_forward(module, return_dict=None):
|
| 435 |
+
def custom_forward(*inputs):
|
| 436 |
+
if return_dict is not None:
|
| 437 |
+
return module(*inputs, return_dict=return_dict)
|
| 438 |
+
else:
|
| 439 |
+
return module(*inputs)
|
| 440 |
+
|
| 441 |
+
return custom_forward
|
| 442 |
+
|
| 443 |
+
hidden_states = torch.utils.checkpoint.checkpoint(
|
| 444 |
+
create_custom_forward(resnet), hidden_states, temb
|
| 445 |
+
)
|
| 446 |
+
hidden_states = torch.utils.checkpoint.checkpoint(
|
| 447 |
+
create_custom_forward(attn, return_dict=False),
|
| 448 |
+
hidden_states,
|
| 449 |
+
encoder_hidden_states,
|
| 450 |
+
)[0]
|
| 451 |
+
|
| 452 |
+
# add motion module
|
| 453 |
+
hidden_states = (
|
| 454 |
+
motion_module(
|
| 455 |
+
hidden_states, temb, encoder_hidden_states=encoder_hidden_states
|
| 456 |
+
)
|
| 457 |
+
if motion_module is not None
|
| 458 |
+
else hidden_states
|
| 459 |
+
)
|
| 460 |
+
|
| 461 |
+
else:
|
| 462 |
+
hidden_states = resnet(hidden_states, temb)
|
| 463 |
+
hidden_states = attn(
|
| 464 |
+
hidden_states,
|
| 465 |
+
encoder_hidden_states=encoder_hidden_states,
|
| 466 |
+
).sample
|
| 467 |
+
|
| 468 |
+
# add motion module
|
| 469 |
+
hidden_states = (
|
| 470 |
+
motion_module(
|
| 471 |
+
hidden_states, temb, encoder_hidden_states=encoder_hidden_states
|
| 472 |
+
)
|
| 473 |
+
if motion_module is not None
|
| 474 |
+
else hidden_states
|
| 475 |
+
)
|
| 476 |
+
|
| 477 |
+
output_states += (hidden_states,)
|
| 478 |
+
|
| 479 |
+
if self.downsamplers is not None:
|
| 480 |
+
for downsampler in self.downsamplers:
|
| 481 |
+
hidden_states = downsampler(hidden_states)
|
| 482 |
+
|
| 483 |
+
output_states += (hidden_states,)
|
| 484 |
+
|
| 485 |
+
return hidden_states, output_states
|
| 486 |
+
|
| 487 |
+
|
| 488 |
+
class DownBlock3D(nn.Module):
|
| 489 |
+
def __init__(
|
| 490 |
+
self,
|
| 491 |
+
in_channels: int,
|
| 492 |
+
out_channels: int,
|
| 493 |
+
temb_channels: int,
|
| 494 |
+
dropout: float = 0.0,
|
| 495 |
+
num_layers: int = 1,
|
| 496 |
+
resnet_eps: float = 1e-6,
|
| 497 |
+
resnet_time_scale_shift: str = "default",
|
| 498 |
+
resnet_act_fn: str = "swish",
|
| 499 |
+
resnet_groups: int = 32,
|
| 500 |
+
resnet_pre_norm: bool = True,
|
| 501 |
+
output_scale_factor=1.0,
|
| 502 |
+
add_downsample=True,
|
| 503 |
+
downsample_padding=1,
|
| 504 |
+
use_inflated_groupnorm=None,
|
| 505 |
+
use_motion_module=None,
|
| 506 |
+
motion_module_type=None,
|
| 507 |
+
motion_module_kwargs=None,
|
| 508 |
+
):
|
| 509 |
+
super().__init__()
|
| 510 |
+
resnets = []
|
| 511 |
+
motion_modules = []
|
| 512 |
+
|
| 513 |
+
# use_motion_module = False
|
| 514 |
+
for i in range(num_layers):
|
| 515 |
+
in_channels = in_channels if i == 0 else out_channels
|
| 516 |
+
resnets.append(
|
| 517 |
+
ResnetBlock3D(
|
| 518 |
+
in_channels=in_channels,
|
| 519 |
+
out_channels=out_channels,
|
| 520 |
+
temb_channels=temb_channels,
|
| 521 |
+
eps=resnet_eps,
|
| 522 |
+
groups=resnet_groups,
|
| 523 |
+
dropout=dropout,
|
| 524 |
+
time_embedding_norm=resnet_time_scale_shift,
|
| 525 |
+
non_linearity=resnet_act_fn,
|
| 526 |
+
output_scale_factor=output_scale_factor,
|
| 527 |
+
pre_norm=resnet_pre_norm,
|
| 528 |
+
use_inflated_groupnorm=use_inflated_groupnorm,
|
| 529 |
+
)
|
| 530 |
+
)
|
| 531 |
+
motion_modules.append(
|
| 532 |
+
get_motion_module(
|
| 533 |
+
in_channels=out_channels,
|
| 534 |
+
motion_module_type=motion_module_type,
|
| 535 |
+
motion_module_kwargs=motion_module_kwargs,
|
| 536 |
+
)
|
| 537 |
+
if use_motion_module
|
| 538 |
+
else None
|
| 539 |
+
)
|
| 540 |
+
|
| 541 |
+
self.resnets = nn.ModuleList(resnets)
|
| 542 |
+
self.motion_modules = nn.ModuleList(motion_modules)
|
| 543 |
+
|
| 544 |
+
if add_downsample:
|
| 545 |
+
self.downsamplers = nn.ModuleList(
|
| 546 |
+
[
|
| 547 |
+
Downsample3D(
|
| 548 |
+
out_channels,
|
| 549 |
+
use_conv=True,
|
| 550 |
+
out_channels=out_channels,
|
| 551 |
+
padding=downsample_padding,
|
| 552 |
+
name="op",
|
| 553 |
+
)
|
| 554 |
+
]
|
| 555 |
+
)
|
| 556 |
+
else:
|
| 557 |
+
self.downsamplers = None
|
| 558 |
+
|
| 559 |
+
self.gradient_checkpointing = False
|
| 560 |
+
|
| 561 |
+
def forward(self, hidden_states, temb=None, encoder_hidden_states=None):
|
| 562 |
+
output_states = ()
|
| 563 |
+
|
| 564 |
+
for resnet, motion_module in zip(self.resnets, self.motion_modules):
|
| 565 |
+
# print(f"DownBlock3D {self.gradient_checkpointing = }")
|
| 566 |
+
if self.training and self.gradient_checkpointing:
|
| 567 |
+
|
| 568 |
+
def create_custom_forward(module):
|
| 569 |
+
def custom_forward(*inputs):
|
| 570 |
+
return module(*inputs)
|
| 571 |
+
|
| 572 |
+
return custom_forward
|
| 573 |
+
|
| 574 |
+
hidden_states = torch.utils.checkpoint.checkpoint(
|
| 575 |
+
create_custom_forward(resnet), hidden_states, temb
|
| 576 |
+
)
|
| 577 |
+
if motion_module is not None:
|
| 578 |
+
hidden_states = torch.utils.checkpoint.checkpoint(
|
| 579 |
+
create_custom_forward(motion_module),
|
| 580 |
+
hidden_states.requires_grad_(),
|
| 581 |
+
temb,
|
| 582 |
+
encoder_hidden_states,
|
| 583 |
+
)
|
| 584 |
+
else:
|
| 585 |
+
hidden_states = resnet(hidden_states, temb)
|
| 586 |
+
|
| 587 |
+
# add motion module
|
| 588 |
+
hidden_states = (
|
| 589 |
+
motion_module(
|
| 590 |
+
hidden_states, temb, encoder_hidden_states=encoder_hidden_states
|
| 591 |
+
)
|
| 592 |
+
if motion_module is not None
|
| 593 |
+
else hidden_states
|
| 594 |
+
)
|
| 595 |
+
|
| 596 |
+
output_states += (hidden_states,)
|
| 597 |
+
|
| 598 |
+
if self.downsamplers is not None:
|
| 599 |
+
for downsampler in self.downsamplers:
|
| 600 |
+
hidden_states = downsampler(hidden_states)
|
| 601 |
+
|
| 602 |
+
output_states += (hidden_states,)
|
| 603 |
+
|
| 604 |
+
return hidden_states, output_states
|
| 605 |
+
|
| 606 |
+
|
| 607 |
+
class CrossAttnUpBlock3D(nn.Module):
|
| 608 |
+
def __init__(
|
| 609 |
+
self,
|
| 610 |
+
in_channels: int,
|
| 611 |
+
out_channels: int,
|
| 612 |
+
prev_output_channel: int,
|
| 613 |
+
temb_channels: int,
|
| 614 |
+
dropout: float = 0.0,
|
| 615 |
+
num_layers: int = 1,
|
| 616 |
+
resnet_eps: float = 1e-6,
|
| 617 |
+
resnet_time_scale_shift: str = "default",
|
| 618 |
+
resnet_act_fn: str = "swish",
|
| 619 |
+
resnet_groups: int = 32,
|
| 620 |
+
resnet_pre_norm: bool = True,
|
| 621 |
+
attn_num_head_channels=1,
|
| 622 |
+
cross_attention_dim=1280,
|
| 623 |
+
output_scale_factor=1.0,
|
| 624 |
+
add_upsample=True,
|
| 625 |
+
dual_cross_attention=False,
|
| 626 |
+
use_linear_projection=False,
|
| 627 |
+
only_cross_attention=False,
|
| 628 |
+
upcast_attention=False,
|
| 629 |
+
unet_use_cross_frame_attention=None,
|
| 630 |
+
unet_use_temporal_attention=None,
|
| 631 |
+
use_motion_module=None,
|
| 632 |
+
use_inflated_groupnorm=None,
|
| 633 |
+
motion_module_type=None,
|
| 634 |
+
motion_module_kwargs=None,
|
| 635 |
+
use_zero_convs=False,
|
| 636 |
+
):
|
| 637 |
+
super().__init__()
|
| 638 |
+
resnets = []
|
| 639 |
+
attentions = []
|
| 640 |
+
motion_modules = []
|
| 641 |
+
|
| 642 |
+
self.has_cross_attention = True
|
| 643 |
+
self.attn_num_head_channels = attn_num_head_channels
|
| 644 |
+
|
| 645 |
+
for i in range(num_layers):
|
| 646 |
+
res_skip_channels = in_channels if (i == num_layers - 1) else out_channels
|
| 647 |
+
resnet_in_channels = prev_output_channel if i == 0 else out_channels
|
| 648 |
+
|
| 649 |
+
resnets.append(
|
| 650 |
+
ResnetBlock3D(
|
| 651 |
+
in_channels=resnet_in_channels + res_skip_channels,
|
| 652 |
+
out_channels=out_channels,
|
| 653 |
+
temb_channels=temb_channels,
|
| 654 |
+
eps=resnet_eps,
|
| 655 |
+
groups=resnet_groups,
|
| 656 |
+
dropout=dropout,
|
| 657 |
+
time_embedding_norm=resnet_time_scale_shift,
|
| 658 |
+
non_linearity=resnet_act_fn,
|
| 659 |
+
output_scale_factor=output_scale_factor,
|
| 660 |
+
pre_norm=resnet_pre_norm,
|
| 661 |
+
use_inflated_groupnorm=use_inflated_groupnorm,
|
| 662 |
+
)
|
| 663 |
+
)
|
| 664 |
+
if dual_cross_attention:
|
| 665 |
+
raise NotImplementedError
|
| 666 |
+
attentions.append(
|
| 667 |
+
Transformer3DModel(
|
| 668 |
+
attn_num_head_channels,
|
| 669 |
+
out_channels // attn_num_head_channels,
|
| 670 |
+
in_channels=out_channels,
|
| 671 |
+
num_layers=1,
|
| 672 |
+
cross_attention_dim=cross_attention_dim,
|
| 673 |
+
norm_num_groups=resnet_groups,
|
| 674 |
+
use_linear_projection=use_linear_projection,
|
| 675 |
+
only_cross_attention=only_cross_attention,
|
| 676 |
+
upcast_attention=upcast_attention,
|
| 677 |
+
unet_use_cross_frame_attention=unet_use_cross_frame_attention,
|
| 678 |
+
unet_use_temporal_attention=unet_use_temporal_attention,
|
| 679 |
+
use_zero_convs=use_zero_convs,
|
| 680 |
+
)
|
| 681 |
+
)
|
| 682 |
+
motion_modules.append(
|
| 683 |
+
get_motion_module(
|
| 684 |
+
in_channels=out_channels,
|
| 685 |
+
motion_module_type=motion_module_type,
|
| 686 |
+
motion_module_kwargs=motion_module_kwargs,
|
| 687 |
+
)
|
| 688 |
+
if use_motion_module
|
| 689 |
+
else None
|
| 690 |
+
)
|
| 691 |
+
|
| 692 |
+
self.attentions = nn.ModuleList(attentions)
|
| 693 |
+
self.resnets = nn.ModuleList(resnets)
|
| 694 |
+
self.motion_modules = nn.ModuleList(motion_modules)
|
| 695 |
+
|
| 696 |
+
if add_upsample:
|
| 697 |
+
self.upsamplers = nn.ModuleList(
|
| 698 |
+
[Upsample3D(out_channels, use_conv=True, out_channels=out_channels)]
|
| 699 |
+
)
|
| 700 |
+
else:
|
| 701 |
+
self.upsamplers = None
|
| 702 |
+
|
| 703 |
+
self.gradient_checkpointing = False
|
| 704 |
+
|
| 705 |
+
def forward(
|
| 706 |
+
self,
|
| 707 |
+
hidden_states,
|
| 708 |
+
res_hidden_states_tuple,
|
| 709 |
+
temb=None,
|
| 710 |
+
encoder_hidden_states=None,
|
| 711 |
+
upsample_size=None,
|
| 712 |
+
attention_mask=None,
|
| 713 |
+
):
|
| 714 |
+
for i, (resnet, attn, motion_module) in enumerate(
|
| 715 |
+
zip(self.resnets, self.attentions, self.motion_modules)
|
| 716 |
+
):
|
| 717 |
+
# pop res hidden states
|
| 718 |
+
res_hidden_states = res_hidden_states_tuple[-1]
|
| 719 |
+
res_hidden_states_tuple = res_hidden_states_tuple[:-1]
|
| 720 |
+
hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)
|
| 721 |
+
|
| 722 |
+
if self.training and self.gradient_checkpointing:
|
| 723 |
+
|
| 724 |
+
def create_custom_forward(module, return_dict=None):
|
| 725 |
+
def custom_forward(*inputs):
|
| 726 |
+
if return_dict is not None:
|
| 727 |
+
return module(*inputs, return_dict=return_dict)
|
| 728 |
+
else:
|
| 729 |
+
return module(*inputs)
|
| 730 |
+
|
| 731 |
+
return custom_forward
|
| 732 |
+
|
| 733 |
+
hidden_states = torch.utils.checkpoint.checkpoint(
|
| 734 |
+
create_custom_forward(resnet), hidden_states, temb
|
| 735 |
+
)
|
| 736 |
+
hidden_states = attn(
|
| 737 |
+
hidden_states,
|
| 738 |
+
encoder_hidden_states=encoder_hidden_states,
|
| 739 |
+
).sample
|
| 740 |
+
if motion_module is not None:
|
| 741 |
+
hidden_states = torch.utils.checkpoint.checkpoint(
|
| 742 |
+
create_custom_forward(motion_module),
|
| 743 |
+
hidden_states.requires_grad_(),
|
| 744 |
+
temb,
|
| 745 |
+
encoder_hidden_states,
|
| 746 |
+
)
|
| 747 |
+
|
| 748 |
+
else:
|
| 749 |
+
hidden_states = resnet(hidden_states, temb)
|
| 750 |
+
hidden_states = attn(
|
| 751 |
+
hidden_states,
|
| 752 |
+
encoder_hidden_states=encoder_hidden_states,
|
| 753 |
+
).sample
|
| 754 |
+
|
| 755 |
+
# add motion module
|
| 756 |
+
hidden_states = (
|
| 757 |
+
motion_module(
|
| 758 |
+
hidden_states, temb, encoder_hidden_states=encoder_hidden_states
|
| 759 |
+
)
|
| 760 |
+
if motion_module is not None
|
| 761 |
+
else hidden_states
|
| 762 |
+
)
|
| 763 |
+
|
| 764 |
+
if self.upsamplers is not None:
|
| 765 |
+
for upsampler in self.upsamplers:
|
| 766 |
+
hidden_states = upsampler(hidden_states, upsample_size)
|
| 767 |
+
|
| 768 |
+
return hidden_states
|
| 769 |
+
|
| 770 |
+
|
| 771 |
+
class UpBlock3D(nn.Module):
|
| 772 |
+
def __init__(
|
| 773 |
+
self,
|
| 774 |
+
in_channels: int,
|
| 775 |
+
prev_output_channel: int,
|
| 776 |
+
out_channels: int,
|
| 777 |
+
temb_channels: int,
|
| 778 |
+
dropout: float = 0.0,
|
| 779 |
+
num_layers: int = 1,
|
| 780 |
+
resnet_eps: float = 1e-6,
|
| 781 |
+
resnet_time_scale_shift: str = "default",
|
| 782 |
+
resnet_act_fn: str = "swish",
|
| 783 |
+
resnet_groups: int = 32,
|
| 784 |
+
resnet_pre_norm: bool = True,
|
| 785 |
+
output_scale_factor=1.0,
|
| 786 |
+
add_upsample=True,
|
| 787 |
+
use_inflated_groupnorm=None,
|
| 788 |
+
use_motion_module=None,
|
| 789 |
+
motion_module_type=None,
|
| 790 |
+
motion_module_kwargs=None,
|
| 791 |
+
):
|
| 792 |
+
super().__init__()
|
| 793 |
+
resnets = []
|
| 794 |
+
motion_modules = []
|
| 795 |
+
|
| 796 |
+
# use_motion_module = False
|
| 797 |
+
for i in range(num_layers):
|
| 798 |
+
res_skip_channels = in_channels if (i == num_layers - 1) else out_channels
|
| 799 |
+
resnet_in_channels = prev_output_channel if i == 0 else out_channels
|
| 800 |
+
|
| 801 |
+
resnets.append(
|
| 802 |
+
ResnetBlock3D(
|
| 803 |
+
in_channels=resnet_in_channels + res_skip_channels,
|
| 804 |
+
out_channels=out_channels,
|
| 805 |
+
temb_channels=temb_channels,
|
| 806 |
+
eps=resnet_eps,
|
| 807 |
+
groups=resnet_groups,
|
| 808 |
+
dropout=dropout,
|
| 809 |
+
time_embedding_norm=resnet_time_scale_shift,
|
| 810 |
+
non_linearity=resnet_act_fn,
|
| 811 |
+
output_scale_factor=output_scale_factor,
|
| 812 |
+
pre_norm=resnet_pre_norm,
|
| 813 |
+
use_inflated_groupnorm=use_inflated_groupnorm,
|
| 814 |
+
)
|
| 815 |
+
)
|
| 816 |
+
motion_modules.append(
|
| 817 |
+
get_motion_module(
|
| 818 |
+
in_channels=out_channels,
|
| 819 |
+
motion_module_type=motion_module_type,
|
| 820 |
+
motion_module_kwargs=motion_module_kwargs,
|
| 821 |
+
)
|
| 822 |
+
if use_motion_module
|
| 823 |
+
else None
|
| 824 |
+
)
|
| 825 |
+
|
| 826 |
+
self.resnets = nn.ModuleList(resnets)
|
| 827 |
+
self.motion_modules = nn.ModuleList(motion_modules)
|
| 828 |
+
|
| 829 |
+
if add_upsample:
|
| 830 |
+
self.upsamplers = nn.ModuleList(
|
| 831 |
+
[Upsample3D(out_channels, use_conv=True, out_channels=out_channels)]
|
| 832 |
+
)
|
| 833 |
+
else:
|
| 834 |
+
self.upsamplers = None
|
| 835 |
+
|
| 836 |
+
self.gradient_checkpointing = False
|
| 837 |
+
|
| 838 |
+
def forward(
|
| 839 |
+
self,
|
| 840 |
+
hidden_states,
|
| 841 |
+
res_hidden_states_tuple,
|
| 842 |
+
temb=None,
|
| 843 |
+
upsample_size=None,
|
| 844 |
+
encoder_hidden_states=None,
|
| 845 |
+
):
|
| 846 |
+
for resnet, motion_module in zip(self.resnets, self.motion_modules):
|
| 847 |
+
# pop res hidden states
|
| 848 |
+
res_hidden_states = res_hidden_states_tuple[-1]
|
| 849 |
+
res_hidden_states_tuple = res_hidden_states_tuple[:-1]
|
| 850 |
+
hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)
|
| 851 |
+
|
| 852 |
+
# print(f"UpBlock3D {self.gradient_checkpointing = }")
|
| 853 |
+
if self.training and self.gradient_checkpointing:
|
| 854 |
+
|
| 855 |
+
def create_custom_forward(module):
|
| 856 |
+
def custom_forward(*inputs):
|
| 857 |
+
return module(*inputs)
|
| 858 |
+
|
| 859 |
+
return custom_forward
|
| 860 |
+
|
| 861 |
+
hidden_states = torch.utils.checkpoint.checkpoint(
|
| 862 |
+
create_custom_forward(resnet), hidden_states, temb
|
| 863 |
+
)
|
| 864 |
+
if motion_module is not None:
|
| 865 |
+
hidden_states = torch.utils.checkpoint.checkpoint(
|
| 866 |
+
create_custom_forward(motion_module),
|
| 867 |
+
hidden_states.requires_grad_(),
|
| 868 |
+
temb,
|
| 869 |
+
encoder_hidden_states,
|
| 870 |
+
)
|
| 871 |
+
else:
|
| 872 |
+
hidden_states = resnet(hidden_states, temb)
|
| 873 |
+
hidden_states = (
|
| 874 |
+
motion_module(
|
| 875 |
+
hidden_states, temb, encoder_hidden_states=encoder_hidden_states
|
| 876 |
+
)
|
| 877 |
+
if motion_module is not None
|
| 878 |
+
else hidden_states
|
| 879 |
+
)
|
| 880 |
+
|
| 881 |
+
if self.upsamplers is not None:
|
| 882 |
+
for upsampler in self.upsamplers:
|
| 883 |
+
hidden_states = upsampler(hidden_states, upsample_size)
|
| 884 |
+
|
| 885 |
+
return hidden_states
|
genwarp/ops.py
ADDED
|
@@ -0,0 +1,130 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
from typing import Dict
|
| 2 |
+
from jaxtyping import Float
|
| 3 |
+
|
| 4 |
+
import numpy as np
|
| 5 |
+
import torch
|
| 6 |
+
from torch import Tensor
|
| 7 |
+
import torch.nn.functional as F
|
| 8 |
+
|
| 9 |
+
from einops import rearrange
|
| 10 |
+
from splatting import splatting_function
|
| 11 |
+
|
| 12 |
+
def sph2cart(
|
| 13 |
+
azi: Float[Tensor, 'B'],
|
| 14 |
+
ele: Float[Tensor, 'B'],
|
| 15 |
+
r: Float[Tensor, 'B']
|
| 16 |
+
) -> Float[Tensor, 'B 3']:
|
| 17 |
+
# z-up, y-right, x-back
|
| 18 |
+
rcos = r * torch.cos(ele)
|
| 19 |
+
pos_cart = torch.stack([
|
| 20 |
+
rcos * torch.cos(azi),
|
| 21 |
+
rcos * torch.sin(azi),
|
| 22 |
+
r * torch.sin(ele)
|
| 23 |
+
], dim=1)
|
| 24 |
+
|
| 25 |
+
return pos_cart
|
| 26 |
+
|
| 27 |
+
def get_viewport_matrix(
|
| 28 |
+
width: int,
|
| 29 |
+
height: int,
|
| 30 |
+
batch_size: int=1,
|
| 31 |
+
device: torch.device=None,
|
| 32 |
+
) -> Float[Tensor, 'B 4 4']:
|
| 33 |
+
N = torch.tensor(
|
| 34 |
+
[[width/2, 0, 0, width/2],
|
| 35 |
+
[0, height/2, 0, height/2],
|
| 36 |
+
[0, 0, 1/2, 1/2],
|
| 37 |
+
[0, 0, 0, 1]],
|
| 38 |
+
dtype=torch.float32,
|
| 39 |
+
device=device
|
| 40 |
+
)[None].repeat(batch_size, 1, 1)
|
| 41 |
+
return N
|
| 42 |
+
|
| 43 |
+
def get_projection_matrix(
|
| 44 |
+
fovy: Float[Tensor, 'B'],
|
| 45 |
+
aspect_wh: float,
|
| 46 |
+
near: float,
|
| 47 |
+
far: float
|
| 48 |
+
) -> Float[Tensor, 'B 4 4']:
|
| 49 |
+
batch_size = fovy.shape[0]
|
| 50 |
+
proj_mtx = torch.zeros(batch_size, 4, 4, dtype=torch.float32)
|
| 51 |
+
proj_mtx[:, 0, 0] = 1.0 / (torch.tan(fovy / 2.0) * aspect_wh)
|
| 52 |
+
proj_mtx[:, 1, 1] = -1.0 / torch.tan(
|
| 53 |
+
fovy / 2.0
|
| 54 |
+
) # add a negative sign here as the y axis is flipped in nvdiffrast output
|
| 55 |
+
proj_mtx[:, 2, 2] = -(far + near) / (far - near)
|
| 56 |
+
proj_mtx[:, 2, 3] = -2.0 * far * near / (far - near)
|
| 57 |
+
proj_mtx[:, 3, 2] = -1.0
|
| 58 |
+
return proj_mtx
|
| 59 |
+
|
| 60 |
+
def camera_lookat(
|
| 61 |
+
eye: Float[Tensor, 'B 3'],
|
| 62 |
+
target: Float[Tensor, 'B 3'],
|
| 63 |
+
up: Float[Tensor, 'B 3']
|
| 64 |
+
) -> Float[Tensor, 'B 4 4']:
|
| 65 |
+
B = eye.shape[0]
|
| 66 |
+
f = F.normalize(eye - target)
|
| 67 |
+
l = F.normalize(torch.linalg.cross(up, f))
|
| 68 |
+
u = F.normalize(torch.linalg.cross(f, l))
|
| 69 |
+
|
| 70 |
+
R = torch.stack((l, u, f), dim=1) # B 3 3
|
| 71 |
+
M_R = torch.eye(4, dtype=torch.float32)[None].repeat((B, 1, 1))
|
| 72 |
+
M_R[..., :3, :3] = R
|
| 73 |
+
|
| 74 |
+
T = - eye
|
| 75 |
+
M_T = torch.eye(4, dtype=torch.float32)[None].repeat((B, 1, 1))
|
| 76 |
+
M_T[..., :3, 3] = T
|
| 77 |
+
|
| 78 |
+
return (M_R @ M_T).to(dtype=torch.float32)
|
| 79 |
+
|
| 80 |
+
def focal_length_to_fov(
|
| 81 |
+
focal_length: float,
|
| 82 |
+
censor_length: float = 24.
|
| 83 |
+
) -> float:
|
| 84 |
+
return 2 * np.arctan(censor_length / focal_length / 2.)
|
| 85 |
+
|
| 86 |
+
def forward_warper(
|
| 87 |
+
image: Float[Tensor, 'B C H W'],
|
| 88 |
+
screen: Float[Tensor, 'B (H W) 2'],
|
| 89 |
+
pcd: Float[Tensor, 'B (H W) 4'],
|
| 90 |
+
mvp_mtx: Float[Tensor, 'B 4 4'],
|
| 91 |
+
viewport_mtx: Float[Tensor, 'B 4 4'],
|
| 92 |
+
alpha: float = 0.5
|
| 93 |
+
) -> Dict[str, Tensor]:
|
| 94 |
+
H, W = image.shape[2:4]
|
| 95 |
+
|
| 96 |
+
# Projection.
|
| 97 |
+
points_c = pcd @ mvp_mtx.mT
|
| 98 |
+
points_ndc = points_c / points_c[..., 3:4]
|
| 99 |
+
# To screen.
|
| 100 |
+
coords_new = points_ndc @ viewport_mtx.mT
|
| 101 |
+
|
| 102 |
+
# Masking invalid pixels.
|
| 103 |
+
invalid = coords_new[..., 2] <= 0
|
| 104 |
+
coords_new[invalid] = -1000000 if coords_new.dtype == torch.float32 else -1e+4
|
| 105 |
+
|
| 106 |
+
# Calculate flow and importance for splatting.
|
| 107 |
+
new_z = points_c[..., 2:3]
|
| 108 |
+
flow = coords_new[..., :2] - screen[..., :2]
|
| 109 |
+
## Importance.
|
| 110 |
+
importance = alpha / new_z
|
| 111 |
+
importance -= importance.amin((1, 2), keepdim=True)
|
| 112 |
+
importance /= importance.amax((1, 2), keepdim=True) + 1e-6
|
| 113 |
+
importance = importance * 10 - 10
|
| 114 |
+
## Rearrange.
|
| 115 |
+
importance = rearrange(importance, 'b (h w) c -> b c h w', h=H, w=W)
|
| 116 |
+
flow = rearrange(flow, 'b (h w) c -> b c h w', h=H, w=W)
|
| 117 |
+
|
| 118 |
+
# Splatting.
|
| 119 |
+
warped = splatting_function('softmax', image, flow, importance, eps=1e-6)
|
| 120 |
+
## mask is 1 where there is no splat
|
| 121 |
+
mask = (warped == 0.).all(dim=1, keepdim=True).to(image.dtype)
|
| 122 |
+
flow2 = rearrange(coords_new[..., :2], 'b (h w) c -> b c h w', h=H, w=W)
|
| 123 |
+
|
| 124 |
+
output = dict(
|
| 125 |
+
warped=warped,
|
| 126 |
+
mask=mask,
|
| 127 |
+
correspondence=flow2
|
| 128 |
+
)
|
| 129 |
+
|
| 130 |
+
return output
|
requirements.txt
ADDED
|
@@ -0,0 +1,14 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
torch==2.0.1
|
| 2 |
+
torchvision==0.15.2
|
| 3 |
+
diffusers
|
| 4 |
+
accelerate
|
| 5 |
+
transformers
|
| 6 |
+
scipy
|
| 7 |
+
opencv-python
|
| 8 |
+
omegaconf
|
| 9 |
+
einops
|
| 10 |
+
roma
|
| 11 |
+
jaxtyping
|
| 12 |
+
timm==0.6.7
|
| 13 |
+
matplotlib==3.6.2
|
| 14 |
+
gradio_model3dgscamera
|