Spaces:
Running
on
Zero
Running
on
Zero
Create app.py
Browse files
app.py
ADDED
|
@@ -0,0 +1,205 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import gradio as gr
|
| 2 |
+
|
| 3 |
+
import spaces
|
| 4 |
+
import torch
|
| 5 |
+
from gradio_rerun import Rerun
|
| 6 |
+
import rerun as rr
|
| 7 |
+
import rerun.blueprint as rrb
|
| 8 |
+
from pathlib import Path
|
| 9 |
+
import uuid
|
| 10 |
+
|
| 11 |
+
from mini_dust3r.api import OptimizedResult, inferece_dust3r, log_optimized_result
|
| 12 |
+
from mini_dust3r.model import AsymmetricCroCo3DStereo
|
| 13 |
+
from mini_dust3r.utils.misc import (
|
| 14 |
+
fill_default_args,
|
| 15 |
+
freeze_all_params,
|
| 16 |
+
is_symmetrized,
|
| 17 |
+
interleave,
|
| 18 |
+
transpose_to_landscape,
|
| 19 |
+
)
|
| 20 |
+
|
| 21 |
+
from .head import Cat_MLP_LocalFeatures_DPT_Pts3d
|
| 22 |
+
|
| 23 |
+
DEVICE = "cuda" if torch.cuda.is_available() else "CPU"
|
| 24 |
+
|
| 25 |
+
# model = AsymmetricCroCo3DStereo.from_pretrained(
|
| 26 |
+
# "naver/DUSt3R_ViTLarge_BaseDecoder_512_dpt"
|
| 27 |
+
# ).to(DEVICE)
|
| 28 |
+
|
| 29 |
+
|
| 30 |
+
|
| 31 |
+
from .linear_head import LinearPts3d
|
| 32 |
+
from .dpt_head import create_dpt_head
|
| 33 |
+
|
| 34 |
+
def head_factory(head_type, output_mode, net, has_conf=False):
|
| 35 |
+
"""" build a prediction head for the decoder
|
| 36 |
+
"""
|
| 37 |
+
if head_type == 'linear' and output_mode == 'pts3d':
|
| 38 |
+
return LinearPts3d(net, has_conf)
|
| 39 |
+
elif head_type == 'dpt' and output_mode == 'pts3d':
|
| 40 |
+
return create_dpt_head(net, has_conf=has_conf)
|
| 41 |
+
if head_type == 'catmlp+dpt' and output_mode.startswith('pts3d+desc'):
|
| 42 |
+
local_feat_dim = int(output_mode[10:])
|
| 43 |
+
assert net.dec_depth > 9
|
| 44 |
+
l2 = net.dec_depth
|
| 45 |
+
feature_dim = 256
|
| 46 |
+
last_dim = feature_dim // 2
|
| 47 |
+
out_nchan = 3
|
| 48 |
+
ed = net.enc_embed_dim
|
| 49 |
+
dd = net.dec_embed_dim
|
| 50 |
+
return Cat_MLP_LocalFeatures_DPT_Pts3d(net, local_feat_dim=local_feat_dim, has_conf=has_conf,
|
| 51 |
+
num_channels=out_nchan + has_conf,
|
| 52 |
+
feature_dim=feature_dim,
|
| 53 |
+
last_dim=last_dim,
|
| 54 |
+
hooks_idx=[0, l2 * 2 // 4, l2 * 3 // 4, l2],
|
| 55 |
+
dim_tokens=[ed, dd, dd, dd],
|
| 56 |
+
postprocess=postprocess,
|
| 57 |
+
depth_mode=net.depth_mode,
|
| 58 |
+
conf_mode=net.conf_mode,
|
| 59 |
+
head_type='regression')
|
| 60 |
+
else:
|
| 61 |
+
raise NotImplementedError(f"unexpected {head_type=} and {output_mode=}")
|
| 62 |
+
|
| 63 |
+
|
| 64 |
+
class AsymmetricMASt3R(AsymmetricCroCo3DStereo):
|
| 65 |
+
def __init__(self, desc_mode=('norm'), two_confs=False, desc_conf_mode=None, **kwargs):
|
| 66 |
+
self.desc_mode = desc_mode
|
| 67 |
+
self.two_confs = two_confs
|
| 68 |
+
self.desc_conf_mode = desc_conf_mode
|
| 69 |
+
super().__init__(**kwargs)
|
| 70 |
+
|
| 71 |
+
@classmethod
|
| 72 |
+
def from_pretrained(cls, pretrained_model_name_or_path, **kw):
|
| 73 |
+
if os.path.isfile(pretrained_model_name_or_path):
|
| 74 |
+
return load_model(pretrained_model_name_or_path, device='cpu')
|
| 75 |
+
else:
|
| 76 |
+
return super(AsymmetricMASt3R, cls).from_pretrained(pretrained_model_name_or_path, **kw)
|
| 77 |
+
|
| 78 |
+
def set_downstream_head(self, output_mode, head_type, landscape_only, depth_mode, conf_mode, patch_size, img_size, **kw):
|
| 79 |
+
assert img_size[0] % patch_size == 0 and img_size[
|
| 80 |
+
1] % patch_size == 0, f'{img_size=} must be multiple of {patch_size=}'
|
| 81 |
+
self.output_mode = output_mode
|
| 82 |
+
self.head_type = head_type
|
| 83 |
+
self.depth_mode = depth_mode
|
| 84 |
+
self.conf_mode = conf_mode
|
| 85 |
+
if self.desc_conf_mode is None:
|
| 86 |
+
self.desc_conf_mode = conf_mode
|
| 87 |
+
# allocate heads
|
| 88 |
+
self.downstream_head1 = head_factory(head_type, output_mode, self, has_conf=bool(conf_mode))
|
| 89 |
+
self.downstream_head2 = head_factory(head_type, output_mode, self, has_conf=bool(conf_mode))
|
| 90 |
+
# magic wrapper
|
| 91 |
+
self.head1 = transpose_to_landscape(self.downstream_head1, activate=landscape_only)
|
| 92 |
+
self.head2 = transpose_to_landscape(self.downstream_head2, activate=landscape_only)
|
| 93 |
+
|
| 94 |
+
|
| 95 |
+
|
| 96 |
+
model = AsymmetricMASt3R.from_pretrained(
|
| 97 |
+
"naver/MASt3R_ViTLarge_BaseDecoder_512_catmlpdpt_metric").to(DEVICE)
|
| 98 |
+
|
| 99 |
+
|
| 100 |
+
def create_blueprint(image_name_list: list[str], log_path: Path) -> rrb.Blueprint:
|
| 101 |
+
# dont show 2d views if there are more than 4 images as to not clutter the view
|
| 102 |
+
if len(image_name_list) > 4:
|
| 103 |
+
blueprint = rrb.Blueprint(
|
| 104 |
+
rrb.Horizontal(
|
| 105 |
+
rrb.Spatial3DView(origin=f"{log_path}"),
|
| 106 |
+
),
|
| 107 |
+
collapse_panels=True,
|
| 108 |
+
)
|
| 109 |
+
else:
|
| 110 |
+
blueprint = rrb.Blueprint(
|
| 111 |
+
rrb.Horizontal(
|
| 112 |
+
contents=[
|
| 113 |
+
rrb.Spatial3DView(origin=f"{log_path}"),
|
| 114 |
+
rrb.Vertical(
|
| 115 |
+
contents=[
|
| 116 |
+
rrb.Spatial2DView(
|
| 117 |
+
origin=f"{log_path}/camera_{i}/pinhole/",
|
| 118 |
+
contents=[
|
| 119 |
+
"+ $origin/**",
|
| 120 |
+
],
|
| 121 |
+
)
|
| 122 |
+
for i in range(len(image_name_list))
|
| 123 |
+
]
|
| 124 |
+
),
|
| 125 |
+
],
|
| 126 |
+
column_shares=[3, 1],
|
| 127 |
+
),
|
| 128 |
+
collapse_panels=True,
|
| 129 |
+
)
|
| 130 |
+
return blueprint
|
| 131 |
+
|
| 132 |
+
|
| 133 |
+
@spaces.GPU
|
| 134 |
+
def predict(image_name_list: list[str] | str):
|
| 135 |
+
# check if is list or string and if not raise error
|
| 136 |
+
if not isinstance(image_name_list, list) and not isinstance(image_name_list, str):
|
| 137 |
+
raise gr.Error(
|
| 138 |
+
f"Input must be a list of strings or a string, got: {type(image_name_list)}"
|
| 139 |
+
)
|
| 140 |
+
uuid_str = str(uuid.uuid4())
|
| 141 |
+
filename = Path(f"/tmp/gradio/{uuid_str}.rrd")
|
| 142 |
+
rr.init(f"{uuid_str}")
|
| 143 |
+
log_path = Path("world")
|
| 144 |
+
|
| 145 |
+
if isinstance(image_name_list, str):
|
| 146 |
+
image_name_list = [image_name_list]
|
| 147 |
+
|
| 148 |
+
optimized_results: OptimizedResult = inferece_dust3r(
|
| 149 |
+
image_dir_or_list=image_name_list,
|
| 150 |
+
model=model,
|
| 151 |
+
device=DEVICE,
|
| 152 |
+
batch_size=1,
|
| 153 |
+
)
|
| 154 |
+
|
| 155 |
+
blueprint: rrb.Blueprint = create_blueprint(image_name_list, log_path)
|
| 156 |
+
rr.send_blueprint(blueprint)
|
| 157 |
+
|
| 158 |
+
rr.set_time_sequence("sequence", 0)
|
| 159 |
+
log_optimized_result(optimized_results, log_path)
|
| 160 |
+
rr.save(filename.as_posix())
|
| 161 |
+
return filename.as_posix()
|
| 162 |
+
|
| 163 |
+
|
| 164 |
+
with gr.Blocks(
|
| 165 |
+
css=""".gradio-container {margin: 0 !important; min-width: 100%};""",
|
| 166 |
+
title="Mini-DUSt3R Demo",
|
| 167 |
+
) as demo:
|
| 168 |
+
# scene state is save so that you can change conf_thr, cam_size... without rerunning the inference
|
| 169 |
+
gr.HTML('<h2 style="text-align: center;">Mini-DUSt3R Demo</h2>')
|
| 170 |
+
gr.HTML(
|
| 171 |
+
'<p style="text-align: center;">Unofficial DUSt3R demo using the mini-dust3r pip package</p>'
|
| 172 |
+
)
|
| 173 |
+
gr.HTML(
|
| 174 |
+
'<p style="text-align: center;">More info <a href="https://github.com/pablovela5620/mini-dust3r">here</a></p>'
|
| 175 |
+
)
|
| 176 |
+
with gr.Tab(label="Single Image"):
|
| 177 |
+
with gr.Column():
|
| 178 |
+
single_image = gr.Image(type="filepath", height=300)
|
| 179 |
+
run_btn_single = gr.Button("Run")
|
| 180 |
+
rerun_viewer_single = Rerun(height=900)
|
| 181 |
+
run_btn_single.click(
|
| 182 |
+
fn=predict, inputs=[single_image], outputs=[rerun_viewer_single]
|
| 183 |
+
)
|
| 184 |
+
|
| 185 |
+
example_single_dir = Path("examples/single_image")
|
| 186 |
+
example_single_files = sorted(example_single_dir.glob("*.png"))
|
| 187 |
+
|
| 188 |
+
examples_single = gr.Examples(
|
| 189 |
+
examples=example_single_files,
|
| 190 |
+
inputs=[single_image],
|
| 191 |
+
outputs=[rerun_viewer_single],
|
| 192 |
+
fn=predict,
|
| 193 |
+
cache_examples="lazy",
|
| 194 |
+
)
|
| 195 |
+
with gr.Tab(label="Multi Image"):
|
| 196 |
+
with gr.Column():
|
| 197 |
+
multi_files = gr.File(file_count="multiple")
|
| 198 |
+
run_btn_multi = gr.Button("Run")
|
| 199 |
+
rerun_viewer_multi = Rerun(height=900)
|
| 200 |
+
run_btn_multi.click(
|
| 201 |
+
fn=predict, inputs=[multi_files], outputs=[rerun_viewer_multi]
|
| 202 |
+
)
|
| 203 |
+
|
| 204 |
+
|
| 205 |
+
demo.launch()
|