update lfs

.gitattributes

@@ -0,0 +1,36 @@
+*.7z filter=lfs diff=lfs merge=lfs -text
+*.arrow filter=lfs diff=lfs merge=lfs -text
+*.bin filter=lfs diff=lfs merge=lfs -text
+*.bz2 filter=lfs diff=lfs merge=lfs -text
+*.ckpt filter=lfs diff=lfs merge=lfs -text
+*.ftz filter=lfs diff=lfs merge=lfs -text
+*.gz filter=lfs diff=lfs merge=lfs -text
+*.h5 filter=lfs diff=lfs merge=lfs -text
+*.joblib filter=lfs diff=lfs merge=lfs -text
+*.lfs.* filter=lfs diff=lfs merge=lfs -text
+*.mlmodel filter=lfs diff=lfs merge=lfs -text
+*.model filter=lfs diff=lfs merge=lfs -text
+*.msgpack filter=lfs diff=lfs merge=lfs -text
+*.npy filter=lfs diff=lfs merge=lfs -text
+*.npz filter=lfs diff=lfs merge=lfs -text
+*.onnx filter=lfs diff=lfs merge=lfs -text
+*.ot filter=lfs diff=lfs merge=lfs -text
+*.parquet filter=lfs diff=lfs merge=lfs -text
+*.pb filter=lfs diff=lfs merge=lfs -text
+*.pickle filter=lfs diff=lfs merge=lfs -text
+*.pkl filter=lfs diff=lfs merge=lfs -text
+*.pt filter=lfs diff=lfs merge=lfs -text
+*.pth filter=lfs diff=lfs merge=lfs -text
+*.rar filter=lfs diff=lfs merge=lfs -text
+*.safetensors filter=lfs diff=lfs merge=lfs -text
+saved_model/**/* filter=lfs diff=lfs merge=lfs -text
+*.tar.* filter=lfs diff=lfs merge=lfs -text
+*.tar filter=lfs diff=lfs merge=lfs -text
+*.tflite filter=lfs diff=lfs merge=lfs -text
+*.tgz filter=lfs diff=lfs merge=lfs -text
+*.wasm filter=lfs diff=lfs merge=lfs -text
+*.xz filter=lfs diff=lfs merge=lfs -text
+*.zip filter=lfs diff=lfs merge=lfs -text
+*.zst filter=lfs diff=lfs merge=lfs -text
+*tfevents* filter=lfs diff=lfs merge=lfs -text
+weights/yolo_world_v2_l_image_prompt_adapter-719a7afb.pth filter=lfs diff=lfs merge=lfs -text

Dockerfile

@@ -0,0 +1,43 @@
+FROM nvidia/cuda:11.8.0-devel-ubuntu22.04
+
+ARG MODEL="yolo_world_l_dual_vlpan_l2norm_2e-3_100e_4x8gpus_obj365v1_goldg_train_lvis_minival.py"
+ARG WEIGHT="yolo_world_l_clip_base_dual_vlpan_2e-3adamw_32xb16_100e_o365_goldg_train_pretrained-0e566235.pth"
+
+ENV FORCE_CUDA="1"
+ENV MMCV_WITH_OPS=1
+
+RUN apt-get update && apt-get install -y --no-install-recommends \
+    python3-pip     \
+    libgl1-mesa-glx \
+    libsm6          \
+    libxext6        \
+    libxrender-dev  \
+    libglib2.0-0    \
+    git             \
+    python3-dev     \
+    python3-wheel
+
+RUN pip3 install --upgrade pip \
+    && pip3 install   \
+        gradio        \
+        opencv-python \
+        supervision   \
+        mmengine      \
+        setuptools    \
+        openmim       \
+    && mim install mmcv==2.0.0 \
+    && pip3 install --no-cache-dir --index-url https://download.pytorch.org/whl/cu118 \
+        wheel         \
+        torch         \
+        torchvision   \
+        torchaudio
+
+COPY . /yolo
+WORKDIR /yolo
+
+RUN pip3 install -e .
+
+RUN curl -o weights/$WEIGHT -L https://huggingface.co/wondervictor/YOLO-World/resolve/main/$WEIGHT
+
+ENTRYPOINT [ "python3", "demo.py" ]
+CMD ["configs/pretrain/$MODEL", "weights/$WEIGHT"]

LICENSE

@@ -0,0 +1,1347 @@
+                    GNU GENERAL PUBLIC LICENSE
+                       Version 3, 29 June 2007
+
+ Copyright (C) 2007 Free Software Foundation, Inc. <https://fsf.org/>
+ Everyone is permitted to copy and distribute verbatim copies
+ of this license document, but changing it is not allowed.
+
+                            Preamble
+
+  The GNU General Public License is a free, copyleft license for
+software and other kinds of works.
+
+  The licenses for most software and other practical works are designed
+to take away your freedom to share and change the works.  By contrast,
+the GNU General Public License is intended to guarantee your freedom to
+share and change all versions of a program--to make sure it remains free
+software for all its users.  We, the Free Software Foundation, use the
+GNU General Public License for most of our software; it applies also to
+any other work released this way by its authors.  You can apply it to
+your programs, too.
+
+  When we speak of free software, we are referring to freedom, not
+price.  Our General Public Licenses are designed to make sure that you
+have the freedom to distribute copies of free software (and charge for
+them if you wish), that you receive source code or can get it if you
+want it, that you can change the software or use pieces of it in new
+free programs, and that you know you can do these things.
+
+  To protect your rights, we need to prevent others from denying you
+these rights or asking you to surrender the rights.  Therefore, you have
+certain responsibilities if you distribute copies of the software, or if
+you modify it: responsibilities to respect the freedom of others.
+
+  For example, if you distribute copies of such a program, whether
+gratis or for a fee, you must pass on to the recipients the same
+freedoms that you received.  You must make sure that they, too, receive
+or can get the source code.  And you must show them these terms so they
+know their rights.
+
+  Developers that use the GNU GPL protect your rights with two steps:
+(1) assert copyright on the software, and (2) offer you this License
+giving you legal permission to copy, distribute and/or modify it.
+
+  For the developers' and authors' protection, the GPL clearly explains
+that there is no warranty for this free software.  For both users' and
+authors' sake, the GPL requires that modified versions be marked as
+changed, so that their problems will not be attributed erroneously to
+authors of previous versions.
+
+  Some devices are designed to deny users access to install or run
+modified versions of the software inside them, although the manufacturer
+can do so.  This is fundamentally incompatible with the aim of
+protecting users' freedom to change the software.  The systematic
+pattern of such abuse occurs in the area of products for individuals to
+use, which is precisely where it is most unacceptable.  Therefore, we
+have designed this version of the GPL to prohibit the practice for those
+products.  If such problems arise substantially in other domains, we
+stand ready to extend this provision to those domains in future versions
+of the GPL, as needed to protect the freedom of users.
+
+  Finally, every program is threatened constantly by software patents.
+States should not allow patents to restrict development and use of
+software on general-purpose computers, but in those that do, we wish to
+avoid the special danger that patents applied to a free program could
+make it effectively proprietary.  To prevent this, the GPL assures that
+patents cannot be used to render the program non-free.
+
+  The precise terms and conditions for copying, distribution and
+modification follow.
+
+                       TERMS AND CONDITIONS
+
+  0. Definitions.
+
+  "This License" refers to version 3 of the GNU General Public License.
+
+  "Copyright" also means copyright-like laws that apply to other kinds of
+works, such as semiconductor masks.
+
+  "The Program" refers to any copyrightable work licensed under this
+License.  Each licensee is addressed as "you".  "Licensees" and
+"recipients" may be individuals or organizations.
+
+  To "modify" a work means to copy from or adapt all or part of the work
+in a fashion requiring copyright permission, other than the making of an
+exact copy.  The resulting work is called a "modified version" of the
+earlier work or a work "based on" the earlier work.
+
+  A "covered work" means either the unmodified Program or a work based
+on the Program.
+
+  To "propagate" a work means to do anything with it that, without
+permission, would make you directly or secondarily liable for
+infringement under applicable copyright law, except executing it on a
+computer or modifying a private copy.  Propagation includes copying,
+distribution (with or without modification), making available to the
+public, and in some countries other activities as well.
+
+  To "convey" a work means any kind of propagation that enables other
+parties to make or receive copies.  Mere interaction with a user through
+a computer network, with no transfer of a copy, is not conveying.
+
+  An interactive user interface displays "Appropriate Legal Notices"
+to the extent that it includes a convenient and prominently visible
+feature that (1) displays an appropriate copyright notice, and (2)
+tells the user that there is no warranty for the work (except to the
+extent that warranties are provided), that licensees may convey the
+work under this License, and how to view a copy of this License.  If
+the interface presents a list of user commands or options, such as a
+menu, a prominent item in the list meets this criterion.
+
+  1. Source Code.
+
+  The "source code" for a work means the preferred form of the work
+for making modifications to it.  "Object code" means any non-source
+form of a work.
+
+  A "Standard Interface" means an interface that either is an official
+standard defined by a recognized standards body, or, in the case of
+interfaces specified for a particular programming language, one that
+is widely used among developers working in that language.
+
+  The "System Libraries" of an executable work include anything, other
+than the work as a whole, that (a) is included in the normal form of
+packaging a Major Component, but which is not part of that Major
+Component, and (b) serves only to enable use of the work with that
+Major Component, or to implement a Standard Interface for which an
+implementation is available to the public in source code form.  A
+"Major Component", in this context, means a major essential component
+(kernel, window system, and so on) of the specific operating system
+(if any) on which the executable work runs, or a compiler used to
+produce the work, or an object code interpreter used to run it.
+
+  The "Corresponding Source" for a work in object code form means all
+the source code needed to generate, install, and (for an executable
+work) run the object code and to modify the work, including scripts to
+control those activities.  However, it does not include the work's
+System Libraries, or general-purpose tools or generally available free
+programs which are used unmodified in performing those activities but
+which are not part of the work.  For example, Corresponding Source
+includes interface definition files associated with source files for
+the work, and the source code for shared libraries and dynamically
+linked subprograms that the work is specifically designed to require,
+such as by intimate data communication or control flow between those
+subprograms and other parts of the work.
+
+  The Corresponding Source need not include anything that users
+can regenerate automatically from other parts of the Corresponding
+Source.
+
+  The Corresponding Source for a work in source code form is that
+same work.
+
+  2. Basic Permissions.
+
+  All rights granted under this License are granted for the term of
+copyright on the Program, and are irrevocable provided the stated
+conditions are met.  This License explicitly affirms your unlimited
+permission to run the unmodified Program.  The output from running a
+covered work is covered by this License only if the output, given its
+content, constitutes a covered work.  This License acknowledges your
+rights of fair use or other equivalent, as provided by copyright law.
+
+  You may make, run and propagate covered works that you do not
+convey, without conditions so long as your license otherwise remains
+in force.  You may convey covered works to others for the sole purpose
+of having them make modifications exclusively for you, or provide you
+with facilities for running those works, provided that you comply with
+the terms of this License in conveying all material for which you do
+not control copyright.  Those thus making or running the covered works
+for you must do so exclusively on your behalf, under your direction
+and control, on terms that prohibit them from making any copies of
+your copyrighted material outside their relationship with you.
+
+  Conveying under any other circumstances is permitted solely under
+the conditions stated below.  Sublicensing is not allowed; section 10
+makes it unnecessary.
+
+  3. Protecting Users' Legal Rights From Anti-Circumvention Law.
+
+  No covered work shall be deemed part of an effective technological
+measure under any applicable law fulfilling obligations under article
+11 of the WIPO copyright treaty adopted on 20 December 1996, or
+similar laws prohibiting or restricting circumvention of such
+measures.
+
+  When you convey a covered work, you waive any legal power to forbid
+circumvention of technological measures to the extent such circumvention
+is effected by exercising rights under this License with respect to
+the covered work, and you disclaim any intention to limit operation or
+modification of the work as a means of enforcing, against the work's
+users, your or third parties' legal rights to forbid circumvention of
+technological measures.
+
+  4. Conveying Verbatim Copies.
+
+  You may convey verbatim copies of the Program's source code as you
+receive it, in any medium, provided that you conspicuously and
+appropriately publish on each copy an appropriate copyright notice;
+keep intact all notices stating that this License and any
+non-permissive terms added in accord with section 7 apply to the code;
+keep intact all notices of the absence of any warranty; and give all
+recipients a copy of this License along with the Program.
+
+  You may charge any price or no price for each copy that you convey,
+and you may offer support or warranty protection for a fee.
+
+  5. Conveying Modified Source Versions.
+
+  You may convey a work based on the Program, or the modifications to
+produce it from the Program, in the form of source code under the
+terms of section 4, provided that you also meet all of these conditions:
+
+    a) The work must carry prominent notices stating that you modified
+    it, and giving a relevant date.
+
+    b) The work must carry prominent notices stating that it is
+    released under this License and any conditions added under section
+    7.  This requirement modifies the requirement in section 4 to
+    "keep intact all notices".
+
+    c) You must license the entire work, as a whole, under this
+    License to anyone who comes into possession of a copy.  This
+    License will therefore apply, along with any applicable section 7
+    additional terms, to the whole of the work, and all its parts,
+    regardless of how they are packaged.  This License gives no
+    permission to license the work in any other way, but it does not
+    invalidate such permission if you have separately received it.
+
+    d) If the work has interactive user interfaces, each must display
+    Appropriate Legal Notices; however, if the Program has interactive
+    interfaces that do not display Appropriate Legal Notices, your
+    work need not make them do so.
+
+  A compilation of a covered work with other separate and independent
+works, which are not by their nature extensions of the covered work,
+and which are not combined with it such as to form a larger program,
+in or on a volume of a storage or distribution medium, is called an
+"aggregate" if the compilation and its resulting copyright are not
+used to limit the access or legal rights of the compilation's users
+beyond what the individual works permit.  Inclusion of a covered work
+in an aggregate does not cause this License to apply to the other
+parts of the aggregate.
+
+  6. Conveying Non-Source Forms.
+
+  You may convey a covered work in object code form under the terms
+of sections 4 and 5, provided that you also convey the
+machine-readable Corresponding Source under the terms of this License,
+in one of these ways:
+
+    a) Convey the object code in, or embodied in, a physical product
+    (including a physical distribution medium), accompanied by the
+    Corresponding Source fixed on a durable physical medium
+    customarily used for software interchange.
+
+    b) Convey the object code in, or embodied in, a physical product
+    (including a physical distribution medium), accompanied by a
+    written offer, valid for at least three years and valid for as
+    long as you offer spare parts or customer support for that product
+    model, to give anyone who possesses the object code either (1) a
+    copy of the Corresponding Source for all the software in the
+    product that is covered by this License, on a durable physical
+    medium customarily used for software interchange, for a price no
+    more than your reasonable cost of physically performing this
+    conveying of source, or (2) access to copy the
+    Corresponding Source from a network server at no charge.
+
+    c) Convey individual copies of the object code with a copy of the
+    written offer to provide the Corresponding Source.  This
+    alternative is allowed only occasionally and noncommercially, and
+    only if you received the object code with such an offer, in accord
+    with subsection 6b.
+
+    d) Convey the object code by offering access from a designated
+    place (gratis or for a charge), and offer equivalent access to the
+    Corresponding Source in the same way through the same place at no
+    further charge.  You need not require recipients to copy the
+    Corresponding Source along with the object code.  If the place to
+    copy the object code is a network server, the Corresponding Source
+    may be on a different server (operated by you or a third party)
+    that supports equivalent copying facilities, provided you maintain
+    clear directions next to the object code saying where to find the
+    Corresponding Source.  Regardless of what server hosts the
+    Corresponding Source, you remain obligated to ensure that it is
+    available for as long as needed to satisfy these requirements.
+
+    e) Convey the object code using peer-to-peer transmission, provided
+    you inform other peers where the object code and Corresponding
+    Source of the work are being offered to the general public at no
+    charge under subsection 6d.
+
+  A separable portion of the object code, whose source code is excluded
+from the Corresponding Source as a System Library, need not be
+included in conveying the object code work.
+
+  A "User Product" is either (1) a "consumer product", which means any
+tangible personal property which is normally used for personal, family,
+or household purposes, or (2) anything designed or sold for incorporation
+into a dwelling.  In determining whether a product is a consumer product,
+doubtful cases shall be resolved in favor of coverage.  For a particular
+product received by a particular user, "normally used" refers to a
+typical or common use of that class of product, regardless of the status
+of the particular user or of the way in which the particular user
+actually uses, or expects or is expected to use, the product.  A product
+is a consumer product regardless of whether the product has substantial
+commercial, industrial or non-consumer uses, unless such uses represent
+the only significant mode of use of the product.
+
+  "Installation Information" for a User Product means any methods,
+procedures, authorization keys, or other information required to install
+and execute modified versions of a covered work in that User Product from
+a modified version of its Corresponding Source.  The information must
+suffice to ensure that the continued functioning of the modified object
+code is in no case prevented or interfered with solely because
+modification has been made.
+
+  If you convey an object code work under this section in, or with, or
+specifically for use in, a User Product, and the conveying occurs as
+part of a transaction in which the right of possession and use of the
+User Product is transferred to the recipient in perpetuity or for a
+fixed term (regardless of how the transaction is characterized), the
+Corresponding Source conveyed under this section must be accompanied
+by the Installation Information.  But this requirement does not apply
+if neither you nor any third party retains the ability to install
+modified object code on the User Product (for example, the work has
+been installed in ROM).
+
+  The requirement to provide Installation Information does not include a
+requirement to continue to provide support service, warranty, or updates
+for a work that has been modified or installed by the recipient, or for
+the User Product in which it has been modified or installed.  Access to a
+network may be denied when the modification itself materially and
+adversely affects the operation of the network or violates the rules and
+protocols for communication across the network.
+
+  Corresponding Source conveyed, and Installation Information provided,
+in accord with this section must be in a format that is publicly
+documented (and with an implementation available to the public in
+source code form), and must require no special password or key for
+unpacking, reading or copying.
+
+  7. Additional Terms.
+
+  "Additional permissions" are terms that supplement the terms of this
+License by making exceptions from one or more of its conditions.
+Additional permissions that are applicable to the entire Program shall
+be treated as though they were included in this License, to the extent
+that they are valid under applicable law.  If additional permissions
+apply only to part of the Program, that part may be used separately
+under those permissions, but the entire Program remains governed by
+this License without regard to the additional permissions.
+
+  When you convey a copy of a covered work, you may at your option
+remove any additional permissions from that copy, or from any part of
+it.  (Additional permissions may be written to require their own
+removal in certain cases when you modify the work.)  You may place
+additional permissions on material, added by you to a covered work,
+for which you have or can give appropriate copyright permission.
+
+  Notwithstanding any other provision of this License, for material you
+add to a covered work, you may (if authorized by the copyright holders of
+that material) supplement the terms of this License with terms:
+
+    a) Disclaiming warranty or limiting liability differently from the
+    terms of sections 15 and 16 of this License; or
+
+    b) Requiring preservation of specified reasonable legal notices or
+    author attributions in that material or in the Appropriate Legal
+    Notices displayed by works containing it; or
+
+    c) Prohibiting misrepresentation of the origin of that material, or
+    requiring that modified versions of such material be marked in
+    reasonable ways as different from the original version; or
+
+    d) Limiting the use for publicity purposes of names of licensors or
+    authors of the material; or
+
+    e) Declining to grant rights under trademark law for use of some
+    trade names, trademarks, or service marks; or
+
+    f) Requiring indemnification of licensors and authors of that
+    material by anyone who conveys the material (or modified versions of
+    it) with contractual assumptions of liability to the recipient, for
+    any liability that these contractual assumptions directly impose on
+    those licensors and authors.
+
+  All other non-permissive additional terms are considered "further
+restrictions" within the meaning of section 10.  If the Program as you
+received it, or any part of it, contains a notice stating that it is
+governed by this License along with a term that is a further
+restriction, you may remove that term.  If a license document contains
+a further restriction but permits relicensing or conveying under this
+License, you may add to a covered work material governed by the terms
+of that license document, provided that the further restriction does
+not survive such relicensing or conveying.
+
+  If you add terms to a covered work in accord with this section, you
+must place, in the relevant source files, a statement of the
+additional terms that apply to those files, or a notice indicating
+where to find the applicable terms.
+
+  Additional terms, permissive or non-permissive, may be stated in the
+form of a separately written license, or stated as exceptions;
+the above requirements apply either way.
+
+  8. Termination.
+
+  You may not propagate or modify a covered work except as expressly
+provided under this License.  Any attempt otherwise to propagate or
+modify it is void, and will automatically terminate your rights under
+this License (including any patent licenses granted under the third
+paragraph of section 11).
+
+  However, if you cease all violation of this License, then your
+license from a particular copyright holder is reinstated (a)
+provisionally, unless and until the copyright holder explicitly and
+finally terminates your license, and (b) permanently, if the copyright
+holder fails to notify you of the violation by some reasonable means
+prior to 60 days after the cessation.
+
+  Moreover, your license from a particular copyright holder is
+reinstated permanently if the copyright holder notifies you of the
+violation by some reasonable means, this is the first time you have
+received notice of violation of this License (for any work) from that
+copyright holder, and you cure the violation prior to 30 days after
+your receipt of the notice.
+
+  Termination of your rights under this section does not terminate the
+licenses of parties who have received copies or rights from you under
+this License.  If your rights have been terminated and not permanently
+reinstated, you do not qualify to receive new licenses for the same
+material under section 10.
+
+  9. Acceptance Not Required for Having Copies.
+
+  You are not required to accept this License in order to receive or
+run a copy of the Program.  Ancillary propagation of a covered work
+occurring solely as a consequence of using peer-to-peer transmission
+to receive a copy likewise does not require acceptance.  However,
+nothing other than this License grants you permission to propagate or
+modify any covered work.  These actions infringe copyright if you do
+not accept this License.  Therefore, by modifying or propagating a
+covered work, you indicate your acceptance of this License to do so.
+
+  10. Automatic Licensing of Downstream Recipients.
+
+  Each time you convey a covered work, the recipient automatically
+receives a license from the original licensors, to run, modify and
+propagate that work, subject to this License.  You are not responsible
+for enforcing compliance by third parties with this License.
+
+  An "entity transaction" is a transaction transferring control of an
+organization, or substantially all assets of one, or subdividing an
+organization, or merging organizations.  If propagation of a covered
+work results from an entity transaction, each party to that
+transaction who receives a copy of the work also receives whatever
+licenses to the work the party's predecessor in interest had or could
+give under the previous paragraph, plus a right to possession of the
+Corresponding Source of the work from the predecessor in interest, if
+the predecessor has it or can get it with reasonable efforts.
+
+  You may not impose any further restrictions on the exercise of the
+rights granted or affirmed under this License.  For example, you may
+not impose a license fee, royalty, or other charge for exercise of
+rights granted under this License, and you may not initiate litigation
+(including a cross-claim or counterclaim in a lawsuit) alleging that
+any patent claim is infringed by making, using, selling, offering for
+sale, or importing the Program or any portion of it.
+
+  11. Patents.
+
+  A "contributor" is a copyright holder who authorizes use under this
+License of the Program or a work on which the Program is based.  The
+work thus licensed is called the contributor's "contributor version".
+
+  A contributor's "essential patent claims" are all patent claims
+owned or controlled by the contributor, whether already acquired or
+hereafter acquired, that would be infringed by some manner, permitted
+by this License, of making, using, or selling its contributor version,
+but do not include claims that would be infringed only as a
+consequence of further modification of the contributor version.  For
+purposes of this definition, "control" includes the right to grant
+patent sublicenses in a manner consistent with the requirements of
+this License.
+
+  Each contributor grants you a non-exclusive, worldwide, royalty-free
+patent license under the contributor's essential patent claims, to
+make, use, sell, offer for sale, import and otherwise run, modify and
+propagate the contents of its contributor version.
+
+  In the following three paragraphs, a "patent license" is any express
+agreement or commitment, however denominated, not to enforce a patent
+(such as an express permission to practice a patent or covenant not to
+sue for patent infringement).  To "grant" such a patent license to a
+party means to make such an agreement or commitment not to enforce a
+patent against the party.
+
+  If you convey a covered work, knowingly relying on a patent license,
+and the Corresponding Source of the work is not available for anyone
+to copy, free of charge and under the terms of this License, through a
+publicly available network server or other readily accessible means,
+then you must either (1) cause the Corresponding Source to be so
+available, or (2) arrange to deprive yourself of the benefit of the
+patent license for this particular work, or (3) arrange, in a manner
+consistent with the requirements of this License, to extend the patent
+license to downstream recipients.  "Knowingly relying" means you have
+actual knowledge that, but for the patent license, your conveying the
+covered work in a country, or your recipient's use of the covered work
+in a country, would infringe one or more identifiable patents in that
+country that you have reason to believe are valid.
+
+  If, pursuant to or in connection with a single transaction or
+arrangement, you convey, or propagate by procuring conveyance of, a
+covered work, and grant a patent license to some of the parties
+receiving the covered work authorizing them to use, propagate, modify
+or convey a specific copy of the covered work, then the patent license
+you grant is automatically extended to all recipients of the covered
+work and works based on it.
+
+  A patent license is "discriminatory" if it does not include within
+the scope of its coverage, prohibits the exercise of, or is
+conditioned on the non-exercise of one or more of the rights that are
+specifically granted under this License.  You may not convey a covered
+work if you are a party to an arrangement with a third party that is
+in the business of distributing software, under which you make payment
+to the third party based on the extent of your activity of conveying
+the work, and under which the third party grants, to any of the
+parties who would receive the covered work from you, a discriminatory
+patent license (a) in connection with copies of the covered work
+conveyed by you (or copies made from those copies), or (b) primarily
+for and in connection with specific products or compilations that
+contain the covered work, unless you entered into that arrangement,
+or that patent license was granted, prior to 28 March 2007.
+
+  Nothing in this License shall be construed as excluding or limiting
+any implied license or other defenses to infringement that may
+otherwise be available to you under applicable patent law.
+
+  12. No Surrender of Others' Freedom.
+
+  If conditions are imposed on you (whether by court order, agreement or
+otherwise) that contradict the conditions of this License, they do not
+excuse you from the conditions of this License.  If you cannot convey a
+covered work so as to satisfy simultaneously your obligations under this
+License and any other pertinent obligations, then as a consequence you may
+not convey it at all.  For example, if you agree to terms that obligate you
+to collect a royalty for further conveying from those to whom you convey
+the Program, the only way you could satisfy both those terms and this
+License would be to refrain entirely from conveying the Program.
+
+  13. Use with the GNU Affero General Public License.
+
+  Notwithstanding any other provision of this License, you have
+permission to link or combine any covered work with a work licensed
+under version 3 of the GNU Affero General Public License into a single
+combined work, and to convey the resulting work.  The terms of this
+License will continue to apply to the part which is the covered work,
+but the special requirements of the GNU Affero General Public License,
+section 13, concerning interaction through a network will apply to the
+combination as such.
+
+  14. Revised Versions of this License.
+
+  The Free Software Foundation may publish revised and/or new versions of
+the GNU General Public License from time to time.  Such new versions will
+be similar in spirit to the present version, but may differ in detail to
+address new problems or concerns.
+
+  Each version is given a distinguishing version number.  If the
+Program specifies that a certain numbered version of the GNU General
+Public License "or any later version" applies to it, you have the
+option of following the terms and conditions either of that numbered
+version or of any later version published by the Free Software
+Foundation.  If the Program does not specify a version number of the
+GNU General Public License, you may choose any version ever published
+by the Free Software Foundation.
+
+  If the Program specifies that a proxy can decide which future
+versions of the GNU General Public License can be used, that proxy's
+public statement of acceptance of a version permanently authorizes you
+to choose that version for the Program.
+
+  Later license versions may give you additional or different
+permissions.  However, no additional obligations are imposed on any
+author or copyright holder as a result of your choosing to follow a
+later version.
+
+  15. Disclaimer of Warranty.
+
+  THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY
+APPLICABLE LAW.  EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT
+HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY
+OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO,
+THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
+PURPOSE.  THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM
+IS WITH YOU.  SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF
+ALL NECESSARY SERVICING, REPAIR OR CORRECTION.
+
+  16. Limitation of Liability.
+
+  IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING
+WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS
+THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY
+GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE
+USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF
+DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD
+PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS),
+EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF
+SUCH DAMAGES.
+
+  17. Interpretation of Sections 15 and 16.
+
+  If the disclaimer of warranty and limitation of liability provided
+above cannot be given local legal effect according to their terms,
+reviewing courts shall apply local law that most closely approximates
+an absolute waiver of all civil liability in connection with the
+Program, unless a warranty or assumption of liability accompanies a
+copy of the Program in return for a fee.
+
+                     END OF TERMS AND CONDITIONS
+
+            How to Apply These Terms to Your New Programs
+
+  If you develop a new program, and you want it to be of the greatest
+possible use to the public, the best way to achieve this is to make it
+free software which everyone can redistribute and change under these terms.
+
+  To do so, attach the following notices to the program.  It is safest
+to attach them to the start of each source file to most effectively
+state the exclusion of warranty; and each file should have at least
+the "copyright" line and a pointer to where the full notice is found.
+
+    <one line to give the program's name and a brief idea of what it does.>
+    Copyright (C) <year>  <name of author>
+
+    This program is free software: you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation, either version 3 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License
+    along with this program.  If not, see <https://www.gnu.org/licenses/>.
+
+Also add information on how to contact you by electronic and paper mail.
+
+  If the program does terminal interaction, make it output a short
+notice like this when it starts in an interactive mode:
+
+    <program>  Copyright (C) <year>  <name of author>
+    This program comes with ABSOLUTELY NO WARRANTY; for details type `show w'.
+    This is free software, and you are welcome to redistribute it
+    under certain conditions; type `show c' for details.
+
+The hypothetical commands `show w' and `show c' should show the appropriate
+parts of the General Public License.  Of course, your program's commands
+might be different; for a GUI interface, you would use an "about box".
+
+  You should also get your employer (if you work as a programmer) or school,
+if any, to sign a "copyright disclaimer" for the program, if necessary.
+For more information on this, and how to apply and follow the GNU GPL, see
+<https://www.gnu.org/licenses/>.
+
+  The GNU General Public License does not permit incorporating your program
+into proprietary programs.  If your program is a subroutine library, you
+may consider it more useful to permit linking proprietary applications with
+the library.  If this is what you want to do, use the GNU Lesser General
+Public License instead of this License.  But first, please read
+<https://www.gnu.org/licenses/why-not-lgpl.html>.                    GNU GENERAL PUBLIC LICENSE
+                       Version 3, 29 June 2007
+
+ Copyright (C) 2007 Free Software Foundation, Inc. <https://fsf.org/>
+ Everyone is permitted to copy and distribute verbatim copies
+ of this license document, but changing it is not allowed.
+
+                            Preamble
+
+  The GNU General Public License is a free, copyleft license for
+software and other kinds of works.
+
+  The licenses for most software and other practical works are designed
+to take away your freedom to share and change the works.  By contrast,
+the GNU General Public License is intended to guarantee your freedom to
+share and change all versions of a program--to make sure it remains free
+software for all its users.  We, the Free Software Foundation, use the
+GNU General Public License for most of our software; it applies also to
+any other work released this way by its authors.  You can apply it to
+your programs, too.
+
+  When we speak of free software, we are referring to freedom, not
+price.  Our General Public Licenses are designed to make sure that you
+have the freedom to distribute copies of free software (and charge for
+them if you wish), that you receive source code or can get it if you
+want it, that you can change the software or use pieces of it in new
+free programs, and that you know you can do these things.
+
+  To protect your rights, we need to prevent others from denying you
+these rights or asking you to surrender the rights.  Therefore, you have
+certain responsibilities if you distribute copies of the software, or if
+you modify it: responsibilities to respect the freedom of others.
+
+  For example, if you distribute copies of such a program, whether
+gratis or for a fee, you must pass on to the recipients the same
+freedoms that you received.  You must make sure that they, too, receive
+or can get the source code.  And you must show them these terms so they
+know their rights.
+
+  Developers that use the GNU GPL protect your rights with two steps:
+(1) assert copyright on the software, and (2) offer you this License
+giving you legal permission to copy, distribute and/or modify it.
+
+  For the developers' and authors' protection, the GPL clearly explains
+that there is no warranty for this free software.  For both users' and
+authors' sake, the GPL requires that modified versions be marked as
+changed, so that their problems will not be attributed erroneously to
+authors of previous versions.
+
+  Some devices are designed to deny users access to install or run
+modified versions of the software inside them, although the manufacturer
+can do so.  This is fundamentally incompatible with the aim of
+protecting users' freedom to change the software.  The systematic
+pattern of such abuse occurs in the area of products for individuals to
+use, which is precisely where it is most unacceptable.  Therefore, we
+have designed this version of the GPL to prohibit the practice for those
+products.  If such problems arise substantially in other domains, we
+stand ready to extend this provision to those domains in future versions
+of the GPL, as needed to protect the freedom of users.
+
+  Finally, every program is threatened constantly by software patents.
+States should not allow patents to restrict development and use of
+software on general-purpose computers, but in those that do, we wish to
+avoid the special danger that patents applied to a free program could
+make it effectively proprietary.  To prevent this, the GPL assures that
+patents cannot be used to render the program non-free.
+
+  The precise terms and conditions for copying, distribution and
+modification follow.
+
+                       TERMS AND CONDITIONS
+
+  0. Definitions.
+
+  "This License" refers to version 3 of the GNU General Public License.
+
+  "Copyright" also means copyright-like laws that apply to other kinds of
+works, such as semiconductor masks.
+
+  "The Program" refers to any copyrightable work licensed under this
+License.  Each licensee is addressed as "you".  "Licensees" and
+"recipients" may be individuals or organizations.
+
+  To "modify" a work means to copy from or adapt all or part of the work
+in a fashion requiring copyright permission, other than the making of an
+exact copy.  The resulting work is called a "modified version" of the
+earlier work or a work "based on" the earlier work.
+
+  A "covered work" means either the unmodified Program or a work based
+on the Program.
+
+  To "propagate" a work means to do anything with it that, without
+permission, would make you directly or secondarily liable for
+infringement under applicable copyright law, except executing it on a
+computer or modifying a private copy.  Propagation includes copying,
+distribution (with or without modification), making available to the
+public, and in some countries other activities as well.
+
+  To "convey" a work means any kind of propagation that enables other
+parties to make or receive copies.  Mere interaction with a user through
+a computer network, with no transfer of a copy, is not conveying.
+
+  An interactive user interface displays "Appropriate Legal Notices"
+to the extent that it includes a convenient and prominently visible
+feature that (1) displays an appropriate copyright notice, and (2)
+tells the user that there is no warranty for the work (except to the
+extent that warranties are provided), that licensees may convey the
+work under this License, and how to view a copy of this License.  If
+the interface presents a list of user commands or options, such as a
+menu, a prominent item in the list meets this criterion.
+
+  1. Source Code.
+
+  The "source code" for a work means the preferred form of the work
+for making modifications to it.  "Object code" means any non-source
+form of a work.
+
+  A "Standard Interface" means an interface that either is an official
+standard defined by a recognized standards body, or, in the case of
+interfaces specified for a particular programming language, one that
+is widely used among developers working in that language.
+
+  The "System Libraries" of an executable work include anything, other
+than the work as a whole, that (a) is included in the normal form of
+packaging a Major Component, but which is not part of that Major
+Component, and (b) serves only to enable use of the work with that
+Major Component, or to implement a Standard Interface for which an
+implementation is available to the public in source code form.  A
+"Major Component", in this context, means a major essential component
+(kernel, window system, and so on) of the specific operating system
+(if any) on which the executable work runs, or a compiler used to
+produce the work, or an object code interpreter used to run it.
+
+  The "Corresponding Source" for a work in object code form means all
+the source code needed to generate, install, and (for an executable
+work) run the object code and to modify the work, including scripts to
+control those activities.  However, it does not include the work's
+System Libraries, or general-purpose tools or generally available free
+programs which are used unmodified in performing those activities but
+which are not part of the work.  For example, Corresponding Source
+includes interface definition files associated with source files for
+the work, and the source code for shared libraries and dynamically
+linked subprograms that the work is specifically designed to require,
+such as by intimate data communication or control flow between those
+subprograms and other parts of the work.
+
+  The Corresponding Source need not include anything that users
+can regenerate automatically from other parts of the Corresponding
+Source.
+
+  The Corresponding Source for a work in source code form is that
+same work.
+
+  2. Basic Permissions.
+
+  All rights granted under this License are granted for the term of
+copyright on the Program, and are irrevocable provided the stated
+conditions are met.  This License explicitly affirms your unlimited
+permission to run the unmodified Program.  The output from running a
+covered work is covered by this License only if the output, given its
+content, constitutes a covered work.  This License acknowledges your
+rights of fair use or other equivalent, as provided by copyright law.
+
+  You may make, run and propagate covered works that you do not
+convey, without conditions so long as your license otherwise remains
+in force.  You may convey covered works to others for the sole purpose
+of having them make modifications exclusively for you, or provide you
+with facilities for running those works, provided that you comply with
+the terms of this License in conveying all material for which you do
+not control copyright.  Those thus making or running the covered works
+for you must do so exclusively on your behalf, under your direction
+and control, on terms that prohibit them from making any copies of
+your copyrighted material outside their relationship with you.
+
+  Conveying under any other circumstances is permitted solely under
+the conditions stated below.  Sublicensing is not allowed; section 10
+makes it unnecessary.
+
+  3. Protecting Users' Legal Rights From Anti-Circumvention Law.
+
+  No covered work shall be deemed part of an effective technological
+measure under any applicable law fulfilling obligations under article
+11 of the WIPO copyright treaty adopted on 20 December 1996, or
+similar laws prohibiting or restricting circumvention of such
+measures.
+
+  When you convey a covered work, you waive any legal power to forbid
+circumvention of technological measures to the extent such circumvention
+is effected by exercising rights under this License with respect to
+the covered work, and you disclaim any intention to limit operation or
+modification of the work as a means of enforcing, against the work's
+users, your or third parties' legal rights to forbid circumvention of
+technological measures.
+
+  4. Conveying Verbatim Copies.
+
+  You may convey verbatim copies of the Program's source code as you
+receive it, in any medium, provided that you conspicuously and
+appropriately publish on each copy an appropriate copyright notice;
+keep intact all notices stating that this License and any
+non-permissive terms added in accord with section 7 apply to the code;
+keep intact all notices of the absence of any warranty; and give all
+recipients a copy of this License along with the Program.
+
+  You may charge any price or no price for each copy that you convey,
+and you may offer support or warranty protection for a fee.
+
+  5. Conveying Modified Source Versions.
+
+  You may convey a work based on the Program, or the modifications to
+produce it from the Program, in the form of source code under the
+terms of section 4, provided that you also meet all of these conditions:
+
+    a) The work must carry prominent notices stating that you modified
+    it, and giving a relevant date.
+
+    b) The work must carry prominent notices stating that it is
+    released under this License and any conditions added under section
+    7.  This requirement modifies the requirement in section 4 to
+    "keep intact all notices".
+
+    c) You must license the entire work, as a whole, under this
+    License to anyone who comes into possession of a copy.  This
+    License will therefore apply, along with any applicable section 7
+    additional terms, to the whole of the work, and all its parts,
+    regardless of how they are packaged.  This License gives no
+    permission to license the work in any other way, but it does not
+    invalidate such permission if you have separately received it.
+
+    d) If the work has interactive user interfaces, each must display
+    Appropriate Legal Notices; however, if the Program has interactive
+    interfaces that do not display Appropriate Legal Notices, your
+    work need not make them do so.
+
+  A compilation of a covered work with other separate and independent
+works, which are not by their nature extensions of the covered work,
+and which are not combined with it such as to form a larger program,
+in or on a volume of a storage or distribution medium, is called an
+"aggregate" if the compilation and its resulting copyright are not
+used to limit the access or legal rights of the compilation's users
+beyond what the individual works permit.  Inclusion of a covered work
+in an aggregate does not cause this License to apply to the other
+parts of the aggregate.
+
+  6. Conveying Non-Source Forms.
+
+  You may convey a covered work in object code form under the terms
+of sections 4 and 5, provided that you also convey the
+machine-readable Corresponding Source under the terms of this License,
+in one of these ways:
+
+    a) Convey the object code in, or embodied in, a physical product
+    (including a physical distribution medium), accompanied by the
+    Corresponding Source fixed on a durable physical medium
+    customarily used for software interchange.
+
+    b) Convey the object code in, or embodied in, a physical product
+    (including a physical distribution medium), accompanied by a
+    written offer, valid for at least three years and valid for as
+    long as you offer spare parts or customer support for that product
+    model, to give anyone who possesses the object code either (1) a
+    copy of the Corresponding Source for all the software in the
+    product that is covered by this License, on a durable physical
+    medium customarily used for software interchange, for a price no
+    more than your reasonable cost of physically performing this
+    conveying of source, or (2) access to copy the
+    Corresponding Source from a network server at no charge.
+
+    c) Convey individual copies of the object code with a copy of the
+    written offer to provide the Corresponding Source.  This
+    alternative is allowed only occasionally and noncommercially, and
+    only if you received the object code with such an offer, in accord
+    with subsection 6b.
+
+    d) Convey the object code by offering access from a designated
+    place (gratis or for a charge), and offer equivalent access to the
+    Corresponding Source in the same way through the same place at no
+    further charge.  You need not require recipients to copy the
+    Corresponding Source along with the object code.  If the place to
+    copy the object code is a network server, the Corresponding Source
+    may be on a different server (operated by you or a third party)
+    that supports equivalent copying facilities, provided you maintain
+    clear directions next to the object code saying where to find the
+    Corresponding Source.  Regardless of what server hosts the
+    Corresponding Source, you remain obligated to ensure that it is
+    available for as long as needed to satisfy these requirements.
+
+    e) Convey the object code using peer-to-peer transmission, provided
+    you inform other peers where the object code and Corresponding
+    Source of the work are being offered to the general public at no
+    charge under subsection 6d.
+
+  A separable portion of the object code, whose source code is excluded
+from the Corresponding Source as a System Library, need not be
+included in conveying the object code work.
+
+  A "User Product" is either (1) a "consumer product", which means any
+tangible personal property which is normally used for personal, family,
+or household purposes, or (2) anything designed or sold for incorporation
+into a dwelling.  In determining whether a product is a consumer product,
+doubtful cases shall be resolved in favor of coverage.  For a particular
+product received by a particular user, "normally used" refers to a
+typical or common use of that class of product, regardless of the status
+of the particular user or of the way in which the particular user
+actually uses, or expects or is expected to use, the product.  A product
+is a consumer product regardless of whether the product has substantial
+commercial, industrial or non-consumer uses, unless such uses represent
+the only significant mode of use of the product.
+
+  "Installation Information" for a User Product means any methods,
+procedures, authorization keys, or other information required to install
+and execute modified versions of a covered work in that User Product from
+a modified version of its Corresponding Source.  The information must
+suffice to ensure that the continued functioning of the modified object
+code is in no case prevented or interfered with solely because
+modification has been made.
+
+  If you convey an object code work under this section in, or with, or
+specifically for use in, a User Product, and the conveying occurs as
+part of a transaction in which the right of possession and use of the
+User Product is transferred to the recipient in perpetuity or for a
+fixed term (regardless of how the transaction is characterized), the
+Corresponding Source conveyed under this section must be accompanied
+by the Installation Information.  But this requirement does not apply
+if neither you nor any third party retains the ability to install
+modified object code on the User Product (for example, the work has
+been installed in ROM).
+
+  The requirement to provide Installation Information does not include a
+requirement to continue to provide support service, warranty, or updates
+for a work that has been modified or installed by the recipient, or for
+the User Product in which it has been modified or installed.  Access to a
+network may be denied when the modification itself materially and
+adversely affects the operation of the network or violates the rules and
+protocols for communication across the network.
+
+  Corresponding Source conveyed, and Installation Information provided,
+in accord with this section must be in a format that is publicly
+documented (and with an implementation available to the public in
+source code form), and must require no special password or key for
+unpacking, reading or copying.
+
+  7. Additional Terms.
+
+  "Additional permissions" are terms that supplement the terms of this
+License by making exceptions from one or more of its conditions.
+Additional permissions that are applicable to the entire Program shall
+be treated as though they were included in this License, to the extent
+that they are valid under applicable law.  If additional permissions
+apply only to part of the Program, that part may be used separately
+under those permissions, but the entire Program remains governed by
+this License without regard to the additional permissions.
+
+  When you convey a copy of a covered work, you may at your option
+remove any additional permissions from that copy, or from any part of
+it.  (Additional permissions may be written to require their own
+removal in certain cases when you modify the work.)  You may place
+additional permissions on material, added by you to a covered work,
+for which you have or can give appropriate copyright permission.
+
+  Notwithstanding any other provision of this License, for material you
+add to a covered work, you may (if authorized by the copyright holders of
+that material) supplement the terms of this License with terms:
+
+    a) Disclaiming warranty or limiting liability differently from the
+    terms of sections 15 and 16 of this License; or
+
+    b) Requiring preservation of specified reasonable legal notices or
+    author attributions in that material or in the Appropriate Legal
+    Notices displayed by works containing it; or
+
+    c) Prohibiting misrepresentation of the origin of that material, or
+    requiring that modified versions of such material be marked in
+    reasonable ways as different from the original version; or
+
+    d) Limiting the use for publicity purposes of names of licensors or
+    authors of the material; or
+
+    e) Declining to grant rights under trademark law for use of some
+    trade names, trademarks, or service marks; or
+
+    f) Requiring indemnification of licensors and authors of that
+    material by anyone who conveys the material (or modified versions of
+    it) with contractual assumptions of liability to the recipient, for
+    any liability that these contractual assumptions directly impose on
+    those licensors and authors.
+
+  All other non-permissive additional terms are considered "further
+restrictions" within the meaning of section 10.  If the Program as you
+received it, or any part of it, contains a notice stating that it is
+governed by this License along with a term that is a further
+restriction, you may remove that term.  If a license document contains
+a further restriction but permits relicensing or conveying under this
+License, you may add to a covered work material governed by the terms
+of that license document, provided that the further restriction does
+not survive such relicensing or conveying.
+
+  If you add terms to a covered work in accord with this section, you
+must place, in the relevant source files, a statement of the
+additional terms that apply to those files, or a notice indicating
+where to find the applicable terms.
+
+  Additional terms, permissive or non-permissive, may be stated in the
+form of a separately written license, or stated as exceptions;
+the above requirements apply either way.
+
+  8. Termination.
+
+  You may not propagate or modify a covered work except as expressly
+provided under this License.  Any attempt otherwise to propagate or
+modify it is void, and will automatically terminate your rights under
+this License (including any patent licenses granted under the third
+paragraph of section 11).
+
+  However, if you cease all violation of this License, then your
+license from a particular copyright holder is reinstated (a)
+provisionally, unless and until the copyright holder explicitly and
+finally terminates your license, and (b) permanently, if the copyright
+holder fails to notify you of the violation by some reasonable means
+prior to 60 days after the cessation.
+
+  Moreover, your license from a particular copyright holder is
+reinstated permanently if the copyright holder notifies you of the
+violation by some reasonable means, this is the first time you have
+received notice of violation of this License (for any work) from that
+copyright holder, and you cure the violation prior to 30 days after
+your receipt of the notice.
+
+  Termination of your rights under this section does not terminate the
+licenses of parties who have received copies or rights from you under
+this License.  If your rights have been terminated and not permanently
+reinstated, you do not qualify to receive new licenses for the same
+material under section 10.
+
+  9. Acceptance Not Required for Having Copies.
+
+  You are not required to accept this License in order to receive or
+run a copy of the Program.  Ancillary propagation of a covered work
+occurring solely as a consequence of using peer-to-peer transmission
+to receive a copy likewise does not require acceptance.  However,
+nothing other than this License grants you permission to propagate or
+modify any covered work.  These actions infringe copyright if you do
+not accept this License.  Therefore, by modifying or propagating a
+covered work, you indicate your acceptance of this License to do so.
+
+  10. Automatic Licensing of Downstream Recipients.
+
+  Each time you convey a covered work, the recipient automatically
+receives a license from the original licensors, to run, modify and
+propagate that work, subject to this License.  You are not responsible
+for enforcing compliance by third parties with this License.
+
+  An "entity transaction" is a transaction transferring control of an
+organization, or substantially all assets of one, or subdividing an
+organization, or merging organizations.  If propagation of a covered
+work results from an entity transaction, each party to that
+transaction who receives a copy of the work also receives whatever
+licenses to the work the party's predecessor in interest had or could
+give under the previous paragraph, plus a right to possession of the
+Corresponding Source of the work from the predecessor in interest, if
+the predecessor has it or can get it with reasonable efforts.
+
+  You may not impose any further restrictions on the exercise of the
+rights granted or affirmed under this License.  For example, you may
+not impose a license fee, royalty, or other charge for exercise of
+rights granted under this License, and you may not initiate litigation
+(including a cross-claim or counterclaim in a lawsuit) alleging that
+any patent claim is infringed by making, using, selling, offering for
+sale, or importing the Program or any portion of it.
+
+  11. Patents.
+
+  A "contributor" is a copyright holder who authorizes use under this
+License of the Program or a work on which the Program is based.  The
+work thus licensed is called the contributor's "contributor version".
+
+  A contributor's "essential patent claims" are all patent claims
+owned or controlled by the contributor, whether already acquired or
+hereafter acquired, that would be infringed by some manner, permitted
+by this License, of making, using, or selling its contributor version,
+but do not include claims that would be infringed only as a
+consequence of further modification of the contributor version.  For
+purposes of this definition, "control" includes the right to grant
+patent sublicenses in a manner consistent with the requirements of
+this License.
+
+  Each contributor grants you a non-exclusive, worldwide, royalty-free
+patent license under the contributor's essential patent claims, to
+make, use, sell, offer for sale, import and otherwise run, modify and
+propagate the contents of its contributor version.
+
+  In the following three paragraphs, a "patent license" is any express
+agreement or commitment, however denominated, not to enforce a patent
+(such as an express permission to practice a patent or covenant not to
+sue for patent infringement).  To "grant" such a patent license to a
+party means to make such an agreement or commitment not to enforce a
+patent against the party.
+
+  If you convey a covered work, knowingly relying on a patent license,
+and the Corresponding Source of the work is not available for anyone
+to copy, free of charge and under the terms of this License, through a
+publicly available network server or other readily accessible means,
+then you must either (1) cause the Corresponding Source to be so
+available, or (2) arrange to deprive yourself of the benefit of the
+patent license for this particular work, or (3) arrange, in a manner
+consistent with the requirements of this License, to extend the patent
+license to downstream recipients.  "Knowingly relying" means you have
+actual knowledge that, but for the patent license, your conveying the
+covered work in a country, or your recipient's use of the covered work
+in a country, would infringe one or more identifiable patents in that
+country that you have reason to believe are valid.
+
+  If, pursuant to or in connection with a single transaction or
+arrangement, you convey, or propagate by procuring conveyance of, a
+covered work, and grant a patent license to some of the parties
+receiving the covered work authorizing them to use, propagate, modify
+or convey a specific copy of the covered work, then the patent license
+you grant is automatically extended to all recipients of the covered
+work and works based on it.
+
+  A patent license is "discriminatory" if it does not include within
+the scope of its coverage, prohibits the exercise of, or is
+conditioned on the non-exercise of one or more of the rights that are
+specifically granted under this License.  You may not convey a covered
+work if you are a party to an arrangement with a third party that is
+in the business of distributing software, under which you make payment
+to the third party based on the extent of your activity of conveying
+the work, and under which the third party grants, to any of the
+parties who would receive the covered work from you, a discriminatory
+patent license (a) in connection with copies of the covered work
+conveyed by you (or copies made from those copies), or (b) primarily
+for and in connection with specific products or compilations that
+contain the covered work, unless you entered into that arrangement,
+or that patent license was granted, prior to 28 March 2007.
+
+  Nothing in this License shall be construed as excluding or limiting
+any implied license or other defenses to infringement that may
+otherwise be available to you under applicable patent law.
+
+  12. No Surrender of Others' Freedom.
+
+  If conditions are imposed on you (whether by court order, agreement or
+otherwise) that contradict the conditions of this License, they do not
+excuse you from the conditions of this License.  If you cannot convey a
+covered work so as to satisfy simultaneously your obligations under this
+License and any other pertinent obligations, then as a consequence you may
+not convey it at all.  For example, if you agree to terms that obligate you
+to collect a royalty for further conveying from those to whom you convey
+the Program, the only way you could satisfy both those terms and this
+License would be to refrain entirely from conveying the Program.
+
+  13. Use with the GNU Affero General Public License.
+
+  Notwithstanding any other provision of this License, you have
+permission to link or combine any covered work with a work licensed
+under version 3 of the GNU Affero General Public License into a single
+combined work, and to convey the resulting work.  The terms of this
+License will continue to apply to the part which is the covered work,
+but the special requirements of the GNU Affero General Public License,
+section 13, concerning interaction through a network will apply to the
+combination as such.
+
+  14. Revised Versions of this License.
+
+  The Free Software Foundation may publish revised and/or new versions of
+the GNU General Public License from time to time.  Such new versions will
+be similar in spirit to the present version, but may differ in detail to
+address new problems or concerns.
+
+  Each version is given a distinguishing version number.  If the
+Program specifies that a certain numbered version of the GNU General
+Public License "or any later version" applies to it, you have the
+option of following the terms and conditions either of that numbered
+version or of any later version published by the Free Software
+Foundation.  If the Program does not specify a version number of the
+GNU General Public License, you may choose any version ever published
+by the Free Software Foundation.
+
+  If the Program specifies that a proxy can decide which future
+versions of the GNU General Public License can be used, that proxy's
+public statement of acceptance of a version permanently authorizes you
+to choose that version for the Program.
+
+  Later license versions may give you additional or different
+permissions.  However, no additional obligations are imposed on any
+author or copyright holder as a result of your choosing to follow a
+later version.
+
+  15. Disclaimer of Warranty.
+
+  THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY
+APPLICABLE LAW.  EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT
+HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY
+OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO,
+THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
+PURPOSE.  THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM
+IS WITH YOU.  SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF
+ALL NECESSARY SERVICING, REPAIR OR CORRECTION.
+
+  16. Limitation of Liability.
+
+  IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING
+WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS
+THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY
+GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE
+USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF
+DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD
+PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS),
+EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF
+SUCH DAMAGES.
+
+  17. Interpretation of Sections 15 and 16.
+
+  If the disclaimer of warranty and limitation of liability provided
+above cannot be given local legal effect according to their terms,
+reviewing courts shall apply local law that most closely approximates
+an absolute waiver of all civil liability in connection with the
+Program, unless a warranty or assumption of liability accompanies a
+copy of the Program in return for a fee.
+
+                     END OF TERMS AND CONDITIONS
+
+            How to Apply These Terms to Your New Programs
+
+  If you develop a new program, and you want it to be of the greatest
+possible use to the public, the best way to achieve this is to make it
+free software which everyone can redistribute and change under these terms.
+
+  To do so, attach the following notices to the program.  It is safest
+to attach them to the start of each source file to most effectively
+state the exclusion of warranty; and each file should have at least
+the "copyright" line and a pointer to where the full notice is found.
+
+    <one line to give the program's name and a brief idea of what it does.>
+    Copyright (C) <year>  <name of author>
+
+    This program is free software: you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation, either version 3 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License
+    along with this program.  If not, see <https://www.gnu.org/licenses/>.
+
+Also add information on how to contact you by electronic and paper mail.
+
+  If the program does terminal interaction, make it output a short
+notice like this when it starts in an interactive mode:
+
+    <program>  Copyright (C) <year>  <name of author>
+    This program comes with ABSOLUTELY NO WARRANTY; for details type `show w'.
+    This is free software, and you are welcome to redistribute it
+    under certain conditions; type `show c' for details.
+
+The hypothetical commands `show w' and `show c' should show the appropriate
+parts of the General Public License.  Of course, your program's commands
+might be different; for a GUI interface, you would use an "about box".
+
+  You should also get your employer (if you work as a programmer) or school,
+if any, to sign a "copyright disclaimer" for the program, if necessary.
+For more information on this, and how to apply and follow the GNU GPL, see
+<https://www.gnu.org/licenses/>.
+
+  The GNU General Public License does not permit incorporating your program
+into proprietary programs.  If your program is a subroutine library, you
+may consider it more useful to permit linking proprietary applications with
+the library.  If this is what you want to do, use the GNU Lesser General
+Public License instead of this License.  But first, please read
+<https://www.gnu.org/licenses/why-not-lgpl.html>.

README.md

@@ -0,0 +1,231 @@
+<div align="center">
+<img src="./assets/yolo_logo.png" width=60%>
+<br>
+<a href="https://scholar.google.com/citations?hl=zh-CN&user=PH8rJHYAAAAJ">Tianheng Cheng</a><sup><span>2,3,*</span></sup>, 
+<a href="https://linsong.info/">Lin Song</a><sup><span>1,📧,*</span></sup>,
+<a href="https://yxgeee.github.io/">Yixiao Ge</a><sup><span>1,🌟,2</span></sup>,
+<a href="http://eic.hust.edu.cn/professor/liuwenyu/"> Wenyu Liu</a><sup><span>3</span></sup>,
+<a href="https://xwcv.github.io/">Xinggang Wang</a><sup><span>3,📧</span></sup>,
+<a href="https://scholar.google.com/citations?user=4oXBp9UAAAAJ&hl=en">Ying Shan</a><sup><span>1,2</span></sup>
+</br>
+
+\* Equal contribution 🌟 Project lead 📧 Corresponding author
+
+<sup>1</sup> Tencent AI Lab,  <sup>2</sup> ARC Lab, Tencent PCG
+<sup>3</sup> Huazhong University of Science and Technology
+<br>
+<div>
+
+[![arxiv paper](https://img.shields.io/badge/Project-Page-green)](https://wondervictor.github.io/)
+[![arxiv paper](https://img.shields.io/badge/arXiv-Paper-red)](https://arxiv.org/abs/2401.17270)
+<a href="https://colab.research.google.com/github/AILab-CVC/YOLO-World/blob/master/inference.ipynb"><img src="https://colab.research.google.com/assets/colab-badge.svg" alt="Open In Colab"></a>
+[![demo](https://img.shields.io/badge/🤗HugginngFace-Spaces-orange)](https://huggingface.co/spaces/stevengrove/YOLO-World)
+[![Replicate](https://replicate.com/zsxkib/yolo-world/badge)](https://replicate.com/zsxkib/yolo-world)
+[![hfpaper](https://img.shields.io/badge/🤗HugginngFace-Paper-yellow)](https://huggingface.co/papers/2401.17270)
+[![license](https://img.shields.io/badge/License-GPLv3.0-blue)](LICENSE)
+[![yoloworldseg](https://img.shields.io/badge/YOLOWorldxEfficientSAM-🤗Spaces-orange)](https://huggingface.co/spaces/SkalskiP/YOLO-World)
+[![yologuide](https://img.shields.io/badge/📖Notebook-roboflow-purple)](https://supervision.roboflow.com/develop/notebooks/zero-shot-object-detection-with-yolo-world)
+[![deploy](https://media.roboflow.com/deploy.svg)](https://inference.roboflow.com/foundation/yolo_world/)
+
+</div>
+</div>
+
+## Notice
+
+We recommend that everyone **use English to communicate on issues**, as this helps developers from around the world discuss, share experiences, and answer questions together.
+
+## 🔥 Updates 
+`[2024-5-18]:` YOLO-World models have been [integrated with the FiftyOne computer vision toolkit](https://docs.voxel51.com/integrations/ultralytics.html#open-vocabulary-detection) for streamlined open-vocabulary inference across image and video datasets.  
+`[2024-5-16]:` Hey guys! Long time no see! This update contains (1) [fine-tuning guide](https://github.com/AILab-CVC/YOLO-World?#highlights--introduction) and (2) [TFLite Export](./docs/tflite_deploy.md) with INT8 Quantization.  
+`[2024-5-9]:` This update contains the real [`reparameterization`](./docs/reparameterize.md) 🪄, and it's better for fine-tuning on custom datasets and improves the training/inference efficiency 🚀!  
+`[2024-4-28]:` Long time no see! This update contains bugfixs and improvements: (1) ONNX demo; (2) image demo (support tensor input); (2) new pre-trained models; (3) image prompts; (4) simple version for fine-tuning / deployment; (5) guide for installation (include a `requirements.txt`).  
+`[2024-3-28]:` We provide: (1) more high-resolution pre-trained models (e.g., S, M, X) ([#142](https://github.com/AILab-CVC/YOLO-World/issues/142)); (2) pre-trained models with CLIP-Large text encoders. Most importantly, we preliminarily fix the **fine-tuning without `mask-refine`** and explore a new fine-tuning setting ([#160](https://github.com/AILab-CVC/YOLO-World/issues/160),[#76](https://github.com/AILab-CVC/YOLO-World/issues/76)). In addition, fine-tuning YOLO-World with `mask-refine` also obtains significant improvements, check more details in [configs/finetune_coco](./configs/finetune_coco/).  
+`[2024-3-16]:` We fix the bugs about the demo ([#110](https://github.com/AILab-CVC/YOLO-World/issues/110),[#94](https://github.com/AILab-CVC/YOLO-World/issues/94),[#129](https://github.com/AILab-CVC/YOLO-World/issues/129), [#125](https://github.com/AILab-CVC/YOLO-World/issues/125)) with visualizations of segmentation masks, and release [**YOLO-World with Embeddings**](./docs/prompt_yolo_world.md), which supports prompt tuning, text prompts and image prompts.  
+`[2024-3-3]:` We add the **high-resolution YOLO-World**, which supports `1280x1280` resolution with higher accuracy and better performance for small objects!  
+`[2024-2-29]:` We release the newest version of [ **YOLO-World-v2**](./docs/updates.md) with higher accuracy and faster speed! We hope the community can join us to improve YOLO-World!  
+`[2024-2-28]:` Excited to announce that YOLO-World has been accepted by **CVPR 2024**! We're continuing to make YOLO-World faster and stronger, as well as making it better to use for all.  
+`[2024-2-22]:` We sincerely thank [RoboFlow](https://roboflow.com/) and [@Skalskip92](https://twitter.com/skalskip92) for the [**Video Guide**](https://www.youtube.com/watch?v=X7gKBGVz4vs) about YOLO-World, nice work!  
+`[2024-2-18]:` We thank [@Skalskip92](https://twitter.com/skalskip92) for developing the wonderful segmentation demo via connecting YOLO-World and EfficientSAM. You can try it now at the [🤗 HuggingFace Spaces](https://huggingface.co/spaces/SkalskiP/YOLO-World).   
+`[2024-2-17]:` The largest model **X** of YOLO-World is released, which achieves better zero-shot performance!   
+`[2024-2-17]:` We release the code & models for **YOLO-World-Seg** now! YOLO-World now supports open-vocabulary / zero-shot object segmentation!  
+`[2024-2-15]:` The pre-traind YOLO-World-L with CC3M-Lite is released!     
+`[2024-2-14]:` We provide the [`image_demo`](demo.py) for inference on images or directories.   
+`[2024-2-10]:` We provide the [fine-tuning](./docs/finetuning.md) and [data](./docs/data.md) details for fine-tuning YOLO-World on the COCO dataset or the custom datasets!  
+`[2024-2-3]:` We support the `Gradio` demo now in the repo and you can build the YOLO-World demo on your own device!  
+`[2024-2-1]:` We've released the code and weights of YOLO-World now!  
+`[2024-2-1]:` We deploy the YOLO-World demo on [HuggingFace 🤗](https://huggingface.co/spaces/stevengrove/YOLO-World), you can try it now!  
+`[2024-1-31]:` We are excited to launch **YOLO-World**, a cutting-edge real-time open-vocabulary object detector.  
+
+
+## TODO
+
+YOLO-World is under active development and please stay tuned ☕️! 
+If you have suggestions📃 or ideas💡,**we would love for you to bring them up in the [Roadmap](https://github.com/AILab-CVC/YOLO-World/issues/109)** ❤️!
+> YOLO-World 目前正在积极开发中📃，如果你有建议或者想法💡，**我们非常希望您在 [Roadmap](https://github.com/AILab-CVC/YOLO-World/issues/109) 中提出来** ❤️！
+
+## [FAQ (Frequently Asked Questions)](https://github.com/AILab-CVC/YOLO-World/discussions/149)
+
+We have set up an FAQ about YOLO-World in the discussion on GitHub. We hope everyone can raise issues or solutions during use here, and we also hope that everyone can quickly find solutions from it.
+
+> 我们在GitHub的discussion中建立了关于YOLO-World的常见问答，这里将收集一些常见问题，同时大家可以在此提出使用中的问题或者解决方案，也希望大家能够从中快速寻找到解决方案
+
+
+## Highlights & Introduction
+
+This repo contains the PyTorch implementation, pre-trained weights, and pre-training/fine-tuning code for YOLO-World.
+
+* YOLO-World is pre-trained on large-scale datasets, including detection, grounding, and image-text datasets.
+
+* YOLO-World is the next-generation YOLO detector, with a strong open-vocabulary detection capability and grounding ability.
+
+* YOLO-World presents a *prompt-then-detect* paradigm for efficient user-vocabulary inference, which re-parameterizes vocabulary embeddings as parameters into the model and achieve superior inference speed. You can try to export your own detection model without extra training or fine-tuning in our [online demo](https://huggingface.co/spaces/stevengrove/YOLO-World)!
+
+
+<div align="center">
+<img width=800px src="./assets/yolo_arch.png">
+</div>
+## Model Zoo
+
+We've pre-trained YOLO-World-S/M/L from scratch and evaluate on the `LVIS val-1.0` and `LVIS minival`. We provide the pre-trained model weights and training logs for applications/research or re-producing the results.
+
+### Zero-shot Inference on LVIS dataset
+
+<div><font size=2>
+
+| model                                                                                                                | Pre-train Data       | Size | AP<sup>mini</su> | AP<sub>r</sub> | AP<sub>c</sub> | AP<sub>f</sub> | AP<sup>val</su> | AP<sub>r</sub> | AP<sub>c</sub> | AP<sub>f</sub> |                                                                                        weights                                                                                         |
+| :------------------------------------------------------------------------------------------------------------------- | :------------------- | :----------------- | :--------------: | :------------: | :------------: | :------------: | :-------------: | :------------: | :------------: | :------------: | :------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------: |
+| [YOLO-Worldv2-S](./configs/pretrain/yolo_world_v2_s_vlpan_bn_2e-3_100e_4x8gpus_obj365v1_goldg_train_lvis_minival.py) | O365+GoldG | 640 | 22.7 | 16.3 | 20.8 | 25.5 |  17.3 | 11.3 | 14.9 | 22.7 |[HF Checkpoints 🤗](https://huggingface.co/wondervictor/YOLO-World/blob/main/yolo_world_v2_s_obj365v1_goldg_pretrain-55b943ea.pth)|
+| [YOLO-Worldv2-S](./configs/pretrain/yolo_world_v2_s_vlpan_bn_2e-3_100e_4x8gpus_obj365v1_goldg_train_1280ft_lvis_minival.py) | O365+GoldG | 1280&#x1F538; | 24.1 | 18.7 | 22.0 | 26.9 |  18.8 | 14.1 | 16.3 | 23.8 |[HF Checkpoints 🤗](https://huggingface.co/wondervictor/YOLO-World/blob/main/yolo_world_v2_s_obj365v1_goldg_pretrain_1280ft-fc4ff4f7.pth)|  
+| [YOLO-Worldv2-M](./configs/pretrain/yolo_world_v2_m_vlpan_bn_2e-3_100e_4x8gpus_obj365v1_goldg_train_lvis_minival.py)  | O365+GoldG | 640 | 30.0 | 25.0  | 27.2 | 33.4 | 23.5 | 17.1 | 20.0 | 30.1 | [HF Checkpoints 🤗](https://huggingface.co/wondervictor/YOLO-World/blob/main/yolo_world_v2_m_obj365v1_goldg_pretrain-c6237d5b.pth)| 
+| [YOLO-Worldv2-M](./configs/pretrain/yolo_world_v2_m_vlpan_bn_2e-3_100e_4x8gpus_obj365v1_goldg_train_1280ft_lvis_minival.py) | O365+GoldG | 1280&#x1F538; | 31.6 | 24.5  | 29.0 | 35.1 | 25.3 | 19.3 | 22.0 | 31.7 | [HF Checkpoints 🤗](https://huggingface.co/wondervictor/YOLO-World/blob/main/yolo_world_v2_m_obj365v1_goldg_pretrain_1280ft-77d0346d.pth)| 
+| [YOLO-Worldv2-L](./configs/pretrain/yolo_world_v2_l_vlpan_bn_2e-3_100e_4x8gpus_obj365v1_goldg_train_lvis_minival.py) | O365+GoldG | 640 | 33.0 | 22.6 | 32.0 | 35.8 | 26.0 | 18.6 | 23.0 | 32.6 | [HF Checkpoints 🤗](https://huggingface.co/wondervictor/YOLO-World/blob/main/yolo_world_v2_l_obj365v1_goldg_pretrain-a82b1fe3.pth)| 
+| [YOLO-Worldv2-L](./configs/pretrain/yolo_world_v2_l_vlpan_bn_2e-3_100e_4x8gpus_obj365v1_goldg_train_1280ft_lvis_minival.py) | O365+GoldG | 1280&#x1F538; | 34.6 | 29.2 | 32.8 | 37.2 | 27.6 | 21.9 | 24.2 | 34.0 | [HF Checkpoints 🤗](https://huggingface.co/wondervictor/YOLO-World/blob/main/yolo_world_v2_l_obj365v1_goldg_pretrain_1280ft-9babe3f6.pth)| 
+| [YOLO-Worldv2-L (CLIP-Large)](./configs/pretrain/yolo_world_v2_l_clip_large_vlpan_bn_2e-3_100e_4x8gpus_obj365v1_goldg_train_lvis_minival.py) 🔥  | O365+GoldG | 640 | 34.0 | 22.0 | 32.6 | 37.4 | 27.1 | 19.9 | 23.9 | 33.9 | [HF Checkpoints 🤗](https://huggingface.co/wondervictor/YOLO-World/blob/main/yolo_world_v2_l_clip_large_o365v1_goldg_pretrain-8ff2e744.pth)|
+| [YOLO-Worldv2-L (CLIP-Large)](./configs/pretrain/yolo_world_v2_l_clip_large_vlpan_bn_2e-3_100e_4x8gpus_obj365v1_goldg_train_800ft_lvis_minival.py) 🔥  | O365+GoldG | 800&#x1F538; | 35.5 | 28.3 | 33.2 | 38.8 | 28.6 | 22.0 | 25.1 | 35.4 | [HF Checkpoints 🤗](https://huggingface.co/wondervictor/YOLO-World/blob/main/yolo_world_v2_l_clip_large_o365v1_goldg_pretrain_800ft-9df82e55.pth)|
+| [YOLO-Worldv2-L](./configs/pretrain/yolo_world_v2_l_vlpan_bn_2e-3_100e_4x8gpus_obj365v1_goldg_train_lvis_minival.py) | O365+GoldG+CC3M-Lite | 640 | 32.9 | 25.3 | 31.1 | 35.8 | 26.1 | 20.6 | 22.6 | 32.3 | [HF Checkpoints 🤗](https://huggingface.co/wondervictor/YOLO-World/blob/main/yolo_world_v2_l_obj365v1_goldg_cc3mlite_pretrain-ca93cd1f.pth)|
+| [YOLO-Worldv2-X](./configs/pretrain/yolo_world_v2_x_vlpan_bn_2e-3_100e_4x8gpus_obj365v1_goldg_train_lvis_minival.py) | O365+GoldG+CC3M-Lite | 640 | 35.4 | 28.7 | 32.9 | 38.7 | 28.4 | 20.6 | 25.6 | 35.0 | [HF Checkpoints 🤗](https://huggingface.co/wondervictor/YOLO-World/blob/main/yolo_world_v2_x_obj365v1_goldg_cc3mlite_pretrain-8698fbfa.pth) |
+| 🔥 [YOLO-Worldv2-X]() |  O365+GoldG+CC3M-Lite | 1280&#x1F538; | 37.4 | 30.5 | 35.2 | 40.7  | 29.8 | 21.1 | 26.8 | 37.0 | [HF Checkpoints 🤗](https://huggingface.co/wondervictor/YOLO-World/blob/main/yolo_world_v2_x_obj365v1_goldg_cc3mlite_pretrain_1280ft-14996a36.pth) |
+| [YOLO-Worldv2-XL](./configs/pretrain/yolo_world_v2_xl_vlpan_bn_2e-3_100e_4x8gpus_obj365v1_goldg_train_lvis_minival.py) | O365+GoldG+CC3M-Lite | 640 | 36.0 | 25.8 | 34.1 | 39.5 | 29.1 | 21.1 | 26.3 | 35.8 | [HF Checkpoints 🤗](https://huggingface.co/wondervictor/YOLO-World/blob/main/yolo_world_v2_xl_obj365v1_goldg_cc3mlite_pretrain-5daf1395.pth) |
+
+</font>
+</div>
+
+**NOTE:**
+1. AP<sup>mini</sup>: evaluated on LVIS `minival`.
+3. AP<sup>val</sup>: evaluated on LVIS `val 1.0`.
+4. [HuggingFace Mirror](https://hf-mirror.com/) provides the mirror of HuggingFace, which is a choice for users who are unable to reach.
+5. &#x1F538;: fine-tuning models with the pre-trained data.
+
+**Pre-training Logs:**
+
+We provide the pre-training logs of `YOLO-World-v2`. Due to the unexpected errors of the local machines, the training might be interrupted several times.
+
+| Model | YOLO-World-v2-S | YOLO-World-v2-M  | YOLO-World-v2-L | YOLO-World-v2-X |
+| :---  | :-------------: | :--------------: | :-------------: | :-------------: |
+|Pre-training Log | [Part-1](https://drive.google.com/file/d/1oib7pKfA2h1U_5-85H_s0Nz8jWd0R-WP/view?usp=drive_link), [Part-2](https://drive.google.com/file/d/11cZ6OZy80VTvBlZy3kzLAHCxx5Iix5-n/view?usp=drive_link) | [Part-1](https://drive.google.com/file/d/1E6vYSS8kBipGc8oQnsjAfeUAx8I9yOX7/view?usp=drive_link), [Part-2](https://drive.google.com/file/d/1fbM7vt2tgSeB8o_7tUDofWvpPNSViNj5/view?usp=drive_link) | [Part-1](https://drive.google.com/file/d/1Tola1QGJZTL6nGy3SBxKuknfNfREDm8J/view?usp=drive_link), [Part-2](https://drive.google.com/file/d/1mTBXniioUb0CdctCG4ckIU6idGo0NnH8/view?usp=drive_link) |  [Final part](https://drive.google.com/file/d/1aEUA_EPQbXOrpxHTQYB6ieGXudb1PLpd/view?usp=drive_link)| 
+
+
+## Getting started
+
+### 1. Installation
+
+YOLO-World is developed based on `torch==1.11.0` `mmyolo==0.6.0` and `mmdetection==3.0.0`. Check more details about `requirements` and `mmcv` in [docs/installation](./docs/installation.md).
+
+#### Clone Project 
+
+```bash
+git clone --recursive https://github.com/AILab-CVC/YOLO-World.git
+```
+#### Install
+
+```bash
+pip install torch wheel -q
+pip install -e .
+```
+
+### 2. Preparing Data
+
+We provide the details about the pre-training data in [docs/data](./docs/data.md).
+
+
+## Training & Evaluation
+
+We adopt the default [training](./tools/train.py) or [evaluation](./tools/test.py) scripts of [mmyolo](https://github.com/open-mmlab/mmyolo).
+We provide the configs for pre-training and fine-tuning in `configs/pretrain` and `configs/finetune_coco`.
+Training YOLO-World is easy:
+
+```bash
+chmod +x tools/dist_train.sh
+# sample command for pre-training, use AMP for mixed-precision training
+./tools/dist_train.sh configs/pretrain/yolo_world_l_t2i_bn_2e-4_100e_4x8gpus_obj365v1_goldg_train_lvis_minival.py 8 --amp
+```
+**NOTE:** YOLO-World is pre-trained on 4 nodes with 8 GPUs per node (32 GPUs in total). For pre-training, the `node_rank` and `nnodes` for multi-node training should be specified. 
+
+Evaluating YOLO-World is also easy:
+
+```bash
+chmod +x tools/dist_test.sh
+./tools/dist_test.sh path/to/config path/to/weights 8
+```
+
+**NOTE:** We mainly evaluate the performance on LVIS-minival for pre-training.
+
+## Fine-tuning YOLO-World
+
+<div align="center">
+<img src="./assets/finetune_yoloworld.png" width=800px>
+</div>
+
+
+<div align="center">
+<b><p>Chose your pre-trained YOLO-World and Fine-tune it!</p></b> 
+</div>
+
+
+YOLO-World supports **zero-shot inference**, and three types of **fine-tuning recipes**: **(1) normal fine-tuning**, **(2) prompt tuning**, and **(3) reparameterized fine-tuning**.
+
+* Normal Fine-tuning: we provide the details about fine-tuning YOLO-World in [docs/fine-tuning](./docs/finetuning.md).
+
+* Prompt Tuning: we provide more details ahout prompt tuning in [docs/prompt_yolo_world](./docs/prompt_yolo_world.md).
+
+* Reparameterized Fine-tuning: the reparameterized YOLO-World is more suitable for specific domains far from generic scenes. You can find more details in [docs/reparameterize](./docs/reparameterize.md).
+
+## Deployment
+
+We provide the details about deployment for downstream applications in [docs/deployment](./docs/deploy.md).
+You can directly download the ONNX model through the online [demo](https://huggingface.co/spaces/stevengrove/YOLO-World) in Huggingface Spaces 🤗.
+
+- [x] ONNX export and demo: [docs/deploy](https://github.com/AILab-CVC/YOLO-World/blob/master/docs/deploy.md)
+- [x] TFLite and INT8 Quantization: [docs/tflite_deploy](https://github.com/AILab-CVC/YOLO-World/blob/master/docs/tflite_deploy.md)
+- [ ] TensorRT: coming soon.
+- [ ] C++: coming soon.
+
+## Demo
+
+See [`demo`](./demo) for more details
+
+- [x] `gradio_demo.py`: Gradio demo, ONNX export
+- [x] `image_demo.py`: inference with images or a directory of images
+- [x] `simple_demo.py`: a simple demo of YOLO-World, using `array` (instead of path as input).
+- [x] `video_demo.py`: inference YOLO-World on videos.
+- [x] `inference.ipynb`: jupyter notebook for YOLO-World.
+- [x] [Google Colab Notebook](https://colab.research.google.com/drive/1F_7S5lSaFM06irBCZqjhbN7MpUXo6WwO?usp=sharing): We sincerely thank [Onuralp](https://github.com/onuralpszr) for sharing the [Colab Demo](https://colab.research.google.com/drive/1F_7S5lSaFM06irBCZqjhbN7MpUXo6WwO?usp=sharing), you can have a try 😊！
+
+## Acknowledgement
+
+We sincerely thank [mmyolo](https://github.com/open-mmlab/mmyolo), [mmdetection](https://github.com/open-mmlab/mmdetection), [GLIP](https://github.com/microsoft/GLIP), and [transformers](https://github.com/huggingface/transformers) for providing their wonderful code to the community!
+
+## Citations
+If you find YOLO-World is useful in your research or applications, please consider giving us a star 🌟 and citing it.
+
+```bibtex
+@inproceedings{Cheng2024YOLOWorld,
+  title={YOLO-World: Real-Time Open-Vocabulary Object Detection},
+  author={Cheng, Tianheng and Song, Lin and Ge, Yixiao and Liu, Wenyu and Wang, Xinggang and Shan, Ying},
+  booktitle={Proc. IEEE Conf. Computer Vision and Pattern Recognition (CVPR)},
+  year={2024}
+}
+```
+
+## Licence
+YOLO-World is under the GPL-v3 Licence and is supported for commercial usage. If you need a commercial license for YOLO-World, please feel free to contact us.

app.py

@@ -0,0 +1,295 @@
+# Copyright (c) Tencent Inc. All rights reserved.
+import os
+import sys
+import argparse
+import os.path as osp
+from io import BytesIO
+from functools import partial
+
+import cv2
+# import onnx
+import torch
+# import onnxsim
+import numpy as np
+import gradio as gr
+from PIL import Image
+import supervision as sv
+from torchvision.ops import nms
+from mmengine.runner import Runner
+from mmengine.dataset import Compose
+from mmengine.runner.amp import autocast
+from mmengine.config import Config, DictAction, ConfigDict
+from mmdet.datasets import CocoDataset
+from mmyolo.registry import RUNNERS
+
+from transformers import (AutoTokenizer, CLIPTextModelWithProjection)
+from transformers import (AutoProcessor, CLIPVisionModelWithProjection)
+
+BOUNDING_BOX_ANNOTATOR = sv.BoundingBoxAnnotator(thickness=2)
+MASK_ANNOTATOR = sv.MaskAnnotator()
+
+
+class LabelAnnotator(sv.LabelAnnotator):
+
+    @staticmethod
+    def resolve_text_background_xyxy(
+        center_coordinates,
+        text_wh,
+        position,
+    ):
+        center_x, center_y = center_coordinates
+        text_w, text_h = text_wh
+        return center_x, center_y, center_x + text_w, center_y + text_h
+
+
+LABEL_ANNOTATOR = LabelAnnotator(text_padding=4,
+                                 text_scale=0.5,
+                                 text_thickness=1)
+
+def generate_image_embeddings(prompt_image,
+                              vision_encoder,
+                              vision_processor,
+                              projector,
+                              device='cuda:0'):
+    prompt_image = prompt_image.convert('RGB')
+    inputs = vision_processor(images=[prompt_image],
+                              return_tensors="pt",
+                              padding=True)
+    inputs = inputs.to(device)
+    image_outputs = vision_encoder(**inputs)
+    img_feats = image_outputs.image_embeds.view(1, -1)
+    img_feats = img_feats / img_feats.norm(p=2, dim=-1, keepdim=True)
+    if projector is not None:
+        img_feats = projector(img_feats)
+    return img_feats
+
+
+def run_image(runner,
+              vision_encoder,
+              vision_processor,
+              padding_token,
+              image,
+              text,
+              prompt_image,
+              add_padding,
+              max_num_boxes,
+              score_thr,
+              nms_thr,
+              image_path='./work_dirs/demo.png'):
+    image = image.convert('RGB')
+    if prompt_image is not None:
+        texts = [['object'], [' ']]
+        projector = None
+        if hasattr(runner.model, 'image_prompt_encoder'):
+            projector = runner.model.image_prompt_encoder.projector
+        print(projector)
+        prompt_embeddings = generate_image_embeddings(
+            prompt_image,
+            vision_encoder=vision_encoder,
+            vision_processor=vision_processor,
+            projector=projector)
+        if add_padding == 'padding':
+            prompt_embeddings = torch.cat([prompt_embeddings, padding_token],
+                                          dim=0)
+        prompt_embeddings = prompt_embeddings / prompt_embeddings.norm(
+            p=2, dim=-1, keepdim=True)
+        runner.model.num_test_classes = prompt_embeddings.shape[0]
+        runner.model.setembeddings(prompt_embeddings[None])
+    else:
+        runner.model.setembeddings(None)
+        texts = [[t.strip()] for t in text.split(',')]
+    data_info = dict(img_id=0, img=np.array(image), texts=texts)
+    data_info = runner.pipeline(data_info)
+    data_batch = dict(inputs=data_info['inputs'].unsqueeze(0),
+                      data_samples=[data_info['data_samples']])
+
+    with autocast(enabled=False), torch.no_grad():
+        if (prompt_image is not None) and ('texts' in data_batch['data_samples'][
+                0]):
+            del data_batch['data_samples'][0]['texts']
+        output = runner.model.test_step(data_batch)[0]
+        pred_instances = output.pred_instances
+
+    keep = nms(pred_instances.bboxes,
+               pred_instances.scores,
+               iou_threshold=nms_thr)
+    pred_instances = pred_instances[keep]
+    pred_instances = pred_instances[pred_instances.scores.float() > score_thr]
+
+    if len(pred_instances.scores) > max_num_boxes:
+        indices = pred_instances.scores.float().topk(max_num_boxes)[1]
+        pred_instances = pred_instances[indices]
+
+    pred_instances = pred_instances.cpu().numpy()
+    if 'masks' in pred_instances:
+        masks = pred_instances['masks']
+    else:
+        masks = None
+    detections = sv.Detections(xyxy=pred_instances['bboxes'],
+                               class_id=pred_instances['labels'],
+                               confidence=pred_instances['scores'],
+                               mask=masks)
+    labels = [
+        f"{texts[class_id][0]} {confidence:0.2f}" for class_id, confidence in
+        zip(detections.class_id, detections.confidence)
+    ]
+
+    image = np.array(image)
+    image = cv2.cvtColor(image, cv2.COLOR_RGB2BGR)  # Convert RGB to BGR
+    image = BOUNDING_BOX_ANNOTATOR.annotate(image, detections)
+    image = LABEL_ANNOTATOR.annotate(image, detections, labels=labels)
+    if masks is not None:
+        image = MASK_ANNOTATOR.annotate(image, detections)
+    image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)  # Convert BGR to RGB
+    image = Image.fromarray(image)
+    return image
+
+
+def demo(runner, vision_encoder, vision_processor, padding_embed):
+    with gr.Blocks(title="YOLO-World") as demo:
+        with gr.Row():
+            gr.Markdown('<h1><center>YOLO-World: Real-Time Open-Vocabulary '
+                        'Object Detector</center></h1>')
+        with gr.Row():
+            image = gr.Image(type='pil', label='input image')
+            output_image = gr.Image(type='pil', label='output image')
+        with gr.Row():
+            with gr.Column(scale=0.3):
+                with gr.Row():
+                    prompt_image = gr.Image(type='pil',
+                                            label='Image Prompts',
+                                            height=300)
+                with gr.Row():
+                    add_padding = gr.Radio(["padding", "none"],
+                                           label="Padding Prompt",
+                                           info="whether add padding prompt")
+            with gr.Column(scale=0.3):
+                with gr.Row():
+                    input_text = gr.Textbox(
+                        lines=7,
+                        label='Text Prompts:\nEnter the classes to be detected, '
+                        'separated by comma',
+                        value=', '.join(CocoDataset.METAINFO['classes']),
+                        elem_id='textbox')
+            with gr.Column(scale=0.4):
+                max_num_boxes = gr.Slider(minimum=1,
+                                          maximum=300,
+                                          value=100,
+                                          step=1,
+                                          interactive=True,
+                                          label='Maximum Number Boxes')
+                score_thr = gr.Slider(minimum=0,
+                                      maximum=1,
+                                      value=0.05,
+                                      step=0.001,
+                                      interactive=True,
+                                      label='Score Threshold')
+                nms_thr = gr.Slider(minimum=0,
+                                    maximum=1,
+                                    value=0.7,
+                                    step=0.001,
+                                    interactive=True,
+                                    label='NMS Threshold')
+
+                with gr.Row():
+                    submit = gr.Button('Submit')
+                    clear = gr.Button('Clear')
+
+        exp_image_dir = "./gradio_examples/image_prompts/images/"
+        exp_prompt_dir = "./gradio_examples/image_prompts/prompts/"
+        example = gr.Examples(
+            examples=[
+                [
+                    exp_image_dir + "0.jpeg", exp_prompt_dir + "0.png", "",
+                    "none", 0.3, 0.5, 100
+                ],
+                [
+                    exp_image_dir + "1.png", exp_prompt_dir + "1.png", "",
+                    "padding", 0.2, 0.1, 100
+                ],
+                [
+                    exp_image_dir + "2.png", exp_prompt_dir + "2.png", "",
+                    "padding", 0.0, 0.1, 200
+                ],
+                [
+                    exp_image_dir + "3.png", exp_prompt_dir + "3.png", "",
+                    "padding", 0.3, 0.5, 100
+                ],
+                [
+                    exp_image_dir + "4.png", exp_prompt_dir + "4.png", "",
+                    "padding", 0.01, 0.1, 200
+                ],
+                [
+                    exp_image_dir + "5.png", exp_prompt_dir + "5.png", "",
+                    "none", 0.3, 0.5, 100
+                ],
+            ],
+            inputs=[
+                image, prompt_image, input_text, add_padding, score_thr,
+                nms_thr, max_num_boxes
+            ],
+        )
+
+        submit.click(
+            partial(run_image, runner, vision_encoder, vision_processor,
+                    padding_embed), [
+                        image,
+                        input_text,
+                        prompt_image,
+                        add_padding,
+                        max_num_boxes,
+                        score_thr,
+                        nms_thr,
+                    ], [output_image])
+        clear.click(lambda: [None, None, '', None], None,
+                    [image, prompt_image, input_text, output_image])
+
+        demo.launch(server_name='0.0.0.0',
+                    server_port=38721)  # port 80 does not work for me
+
+
+if __name__ == '__main__':
+    # args = parse_args()
+
+    config = "configs/prompt_tuning_coco/yolo_world_v2_l_vlpan_bn_2e-4_80e_8gpus_image_prompt_demo.py"
+    checkpoint = "weights/yolo_world_v2_l_image_prompt_adapter-719a7afb.pth"
+    # load config
+    cfg = Config.fromfile(config)
+    if cfg.get('work_dir', None) is None:
+        cfg.work_dir = osp.join('./work_dirs',
+                                osp.splitext(osp.basename(config))[0])
+
+    cfg.load_from = checkpoint
+
+    if 'runner_type' not in cfg:
+        runner = Runner.from_cfg(cfg)
+    else:
+        runner = RUNNERS.build(cfg)
+
+    runner.call_hook('before_run')
+    runner.load_or_resume()
+    pipeline = cfg.test_dataloader.dataset.pipeline
+    pipeline[0].type = 'mmdet.LoadImageFromNDArray'
+    runner.pipeline = Compose(pipeline)
+    runner.model.eval()
+
+    # init vision encoder
+    clip_model = "openai/clip-vit-base-patch32"
+    vision_model = CLIPVisionModelWithProjection.from_pretrained(clip_model)
+    processor = AutoProcessor.from_pretrained(clip_model)
+    device = 'cuda:0'
+    vision_model.to(device)
+
+    texts = [' ']
+    tokenizer = AutoTokenizer.from_pretrained(clip_model)
+    text_model = CLIPTextModelWithProjection.from_pretrained(clip_model)
+    # device = 'cuda:0'
+    text_model.to(device)
+    texts = tokenizer(text=texts, return_tensors='pt', padding=True)
+    texts = texts.to(device)
+    text_outputs = text_model(**texts)
+    txt_feats = text_outputs.text_embeds
+    txt_feats = txt_feats / txt_feats.norm(p=2, dim=-1, keepdim=True)
+    txt_feats = txt_feats.reshape(-1, txt_feats.shape[-1])
+    txt_feats = txt_feats[0].unsqueeze(0)
+    demo(runner, vision_model, processor, txt_feats)

configs/finetune_coco/README.md

@@ -0,0 +1,29 @@
+## Fine-tune YOLO-World on MS-COCO
+
+
+### Updates
+
+1. [2024-3-27]: Considering that fine-tuning YOLO-World on COCO **without `mask-refine`** obtains bad results, e.g., YOLO-World-L obtains 48.6 AP without `mask-refine` compared to 53.3 AP with `mask-refine`, we rethink the training process and explore new training schemes for fine-tuning without `mask-refine`.
+BTW, the COCO fine-tuning results are updated with higher performance (with `mask-refine`)!
+
+
+### COCO Results and Checkpoints
+
+**NOTE:**
+1. AP<sup>ZS</sup>: AP evaluated in the zero-shot setting (w/o fine-tuning on COCO dataset).
+2. `mask-refine`: refine the box annotations with masks, and add `CopyPaste` augmentation during training.
+
+| model | Schedule | `mask-refine` | efficient neck | AP<sup>ZS</sup>|  AP | AP<sub>50</sub> | AP<sub>75</sub> | weights | log |
+| :---- | :-------: | :----------: |:-------------: | :------------: | :-: | :--------------:| :-------------: |:------: | :-: |
+| [YOLO-World-v2-S](./yolo_world_v2_s_vlpan_bn_2e-4_80e_8gpus_mask-refine_finetune_coco.py) | AdamW, 2e-4, 80e | ✔️  | ✖️ | 37.5 | 46.1 | 62.0 | 49.9 | [HF Checkpoints](https://huggingface.co/wondervictor/YOLO-World/blob/main/yolo_world_v2_s_vlpan_bn_2e-4_80e_8gpus_mask-refine_finetune_coco_ep80-492dc329.pth) | [log](https://huggingface.co/wondervictor/YOLO-World/blob/main/yolo_world_v2_s_vlpan_bn_2e-4_80e_8gpus_mask-refine_finetune_coco_20240327_110411.log) |
+| [YOLO-World-v2-M](./yolo_world_v2_m_vlpan_bn_2e-4_80e_8gpus_mask-refine_finetune_coco.py) | AdamW, 2e-4, 80e | ✔️  | ✖️ | 42.8 | 51.0 | 67.5 | 55.2 | [HF Checkpoints](https://huggingface.co/wondervictor/YOLO-World/blob/main/yolo_world_v2_m_vlpan_bn_2e-4_80e_8gpus_mask-refine_finetune_coco_ep80-69c27ac7.pth) | [log](https://huggingface.co/wondervictor/YOLO-World/blob/main/yolo_world_v2_m_vlpan_bn_2e-4_80e_8gpus_mask-refine_finetune_coco_20240327_110411.log) |
+| [YOLO-World-v2-L](./yolo_world_v2_l_vlpan_bn_2e-4_80e_8gpus_mask-refine_finetune_coco.py) | AdamW, 2e-4, 80e | ✔️  | ✖️ | 45.1 | 53.9 | 70.9 | 58.8 | [HF Checkpoints](https://huggingface.co/wondervictor/YOLO-World/blob/main/yolo_world_v2_l_vlpan_bn_2e-4_80e_8gpus_mask-refine_finetune_coco_ep80-81c701ee.pth) | [log](https://huggingface.co/wondervictor/YOLO-World/blob/main/yolo_world_v2_l_vlpan_bn_2e-4_80e_8gpus_mask-refine_finetune_coco_20240326_160313.log) |
+| [YOLO-World-v2-X](./yolo_world_v2_x_vlpan_bn_2e-4_80e_8gpus_mask-refine_finetune_coco.py) | AdamW, 2e-4, 80e | ✔️  | ✖️ | 46.8 | 54.7 | 71.6 | 59.6 | [HF Checkpoints](https://huggingface.co/wondervictor/YOLO-World/blob/main/yolo_world_v2_x_vlpan_bn_2e-4_80e_8gpus_mask-refine_finetune_coco_ep80-76bc0cbd.pth) | [log](https://huggingface.co/wondervictor/YOLO-World/blob/main/yolo_world_v2_x_vlpan_bn_2e-4_80e_8gpus_mask-refine_finetune_coco_20240322_181232.log) |
+| [YOLO-World-v2-L](./yolo_world_v2_l_vlpan_bn_sgd_1e-3_40e_8gpus_finetune_coco.py) 🔥 | SGD, 1e-3, 40e | ✖️  | ✖️ | 45.1 | 52.8 | 69.5 | 57.8 | [HF Checkpoints](https://huggingface.co/wondervictor/YOLO-World/blob/main/yolo_world_v2_l_vlpan_bn_sgd_1e-3_40e_8gpus_finetune_coco_ep80-e1288152.pth) | [log](https://huggingface.co/wondervictor/YOLO-World/blob/main/yolo_world_v2_l_vlpan_bn_sgd_1e-3_40e_8gpus_finetuning_coco_20240327_014902.log) |
+
+
+### Reparameterized Training
+
+| model | Schedule | `mask-refine` | efficient neck | AP<sup>ZS</sup>|  AP | AP<sub>50</sub> | AP<sub>75</sub> | weights | log |
+| :---- | :-------: | :----------: |:-------------: | :------------: | :-: | :--------------:| :-------------: |:------: | :-: |
+| [YOLO-World-v2-S](./yolo_world_v2_s_rep_vlpan_bn_2e-4_80e_8gpus_mask-refine_finetune_coco.py) | AdamW, 2e-4, 80e | ✔️  | ✖️ | 37.5 | 46.3 | 62.8 | 50.4 | [HF Checkpoints]() | [log]() | 

configs/finetune_coco/yolo_world_l_dual_vlpan_2e-4_80e_8gpus_finetune_coco.py

@@ -0,0 +1,179 @@
+_base_ = (
+    '../../third_party/mmyolo/configs/yolov8/'
+    'yolov8_l_syncbn_fast_8xb16-500e_coco.py')
+custom_imports = dict(
+    imports=['yolo_world'],
+    allow_failed_imports=False)
+
+# hyper-parameters
+num_classes = 80
+num_training_classes = 80
+max_epochs = 80  # Maximum training epochs
+close_mosaic_epochs = 10
+save_epoch_intervals = 5
+text_channels = 512
+neck_embed_channels = [128, 256, _base_.last_stage_out_channels // 2]
+neck_num_heads = [4, 8, _base_.last_stage_out_channels // 2 // 32]
+base_lr = 2e-4
+weight_decay = 0.05
+train_batch_size_per_gpu = 16
+load_from='pretrained_models/yolo_world_l_clip_base_dual_vlpan_2e-3adamw_32xb16_100e_o365_goldg_train_pretrained-0e566235.pth'
+persistent_workers = False
+
+# model settings
+model = dict(
+    type='YOLOWorldDetector',
+    mm_neck=True,
+    num_train_classes=num_training_classes,
+    num_test_classes=num_classes,
+    data_preprocessor=dict(type='YOLOWDetDataPreprocessor'),
+    backbone=dict(
+        _delete_=True,
+        type='MultiModalYOLOBackbone',
+        image_model={{_base_.model.backbone}},
+        text_model=dict(
+            type='HuggingCLIPLanguageBackbone',
+            model_name='openai/clip-vit-base-patch32',
+            frozen_modules=['all'])),
+    neck=dict(type='YOLOWorldDualPAFPN',
+              guide_channels=text_channels,
+              embed_channels=neck_embed_channels,
+              num_heads=neck_num_heads,
+              block_cfg=dict(type='MaxSigmoidCSPLayerWithTwoConv'),
+              text_enhancder=dict(type='ImagePoolingAttentionModule',
+                                  embed_channels=256,
+                                  num_heads=8)),
+    bbox_head=dict(type='YOLOWorldHead',
+                   head_module=dict(type='YOLOWorldHeadModule',
+                                    embed_dims=text_channels,
+                                    num_classes=num_training_classes)),
+    train_cfg=dict(assigner=dict(num_classes=num_training_classes)))
+
+# dataset settings
+text_transform = [
+    dict(type='RandomLoadText',
+         num_neg_samples=(num_classes, num_classes),
+         max_num_samples=num_training_classes,
+         padding_to_max=True,
+         padding_value=''),
+    dict(type='mmdet.PackDetInputs',
+         meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape', 'flip',
+                    'flip_direction', 'texts'))
+]
+mosaic_affine_transform = [
+    dict(
+        type='MultiModalMosaic',
+        img_scale=_base_.img_scale,
+        pad_val=114.0,
+        pre_transform=_base_.pre_transform),
+    dict(
+        type='YOLOv5RandomAffine',
+        max_rotate_degree=0.0,
+        max_shear_degree=0.0,
+        max_aspect_ratio=100.,
+        scaling_ratio_range=(1 - _base_.affine_scale,
+                             1 + _base_.affine_scale),
+        # img_scale is (width, height)
+        border=(-_base_.img_scale[0] // 2, -_base_.img_scale[1] // 2),
+        border_val=(114, 114, 114))
+]
+train_pipeline = [
+    *_base_.pre_transform,
+    *mosaic_affine_transform,
+    dict(
+        type='YOLOv5MultiModalMixUp',
+        prob=_base_.mixup_prob,
+        pre_transform=[*_base_.pre_transform,
+                       *mosaic_affine_transform]),
+    *_base_.last_transform[:-1],
+    *text_transform
+]
+train_pipeline_stage2 = [
+    *_base_.train_pipeline_stage2[:-1],
+    *text_transform
+]
+coco_train_dataset = dict(
+    _delete_=True,
+    type='MultiModalDataset',
+    dataset=dict(
+        type='YOLOv5CocoDataset',
+        data_root='data/coco',
+        ann_file='annotations/instances_train2017.json',
+        data_prefix=dict(img='train2017/'),
+        filter_cfg=dict(filter_empty_gt=False, min_size=32)),
+    class_text_path='data/texts/coco_class_texts.json',
+    pipeline=train_pipeline)
+
+train_dataloader = dict(
+    persistent_workers=persistent_workers,
+    batch_size=train_batch_size_per_gpu,
+    collate_fn=dict(type='yolow_collate'),
+    dataset=coco_train_dataset)
+test_pipeline = [
+    *_base_.test_pipeline[:-1],
+    dict(type='LoadText'),
+    dict(
+        type='mmdet.PackDetInputs',
+        meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape',
+                   'scale_factor', 'pad_param', 'texts'))
+]
+coco_val_dataset = dict(
+    _delete_=True,
+    type='MultiModalDataset',
+    dataset=dict(
+        type='YOLOv5CocoDataset',
+        data_root='data/coco',
+        ann_file='annotations/instances_val2017.json',
+        data_prefix=dict(img='val2017/'),
+        filter_cfg=dict(filter_empty_gt=False, min_size=32)),
+    class_text_path='data/texts/coco_class_texts.json',
+    pipeline=test_pipeline)
+val_dataloader = dict(dataset=coco_val_dataset)
+test_dataloader = val_dataloader
+# training settings
+default_hooks = dict(
+    param_scheduler=dict(
+        scheduler_type='linear',
+        lr_factor=0.01,
+        max_epochs=max_epochs),
+    checkpoint=dict(
+        max_keep_ckpts=-1,
+        save_best=None,
+        interval=save_epoch_intervals))
+custom_hooks = [
+    dict(
+        type='EMAHook',
+        ema_type='ExpMomentumEMA',
+        momentum=0.0001,
+        update_buffers=True,
+        strict_load=False,
+        priority=49),
+    dict(
+        type='mmdet.PipelineSwitchHook',
+        switch_epoch=max_epochs - close_mosaic_epochs,
+        switch_pipeline=train_pipeline_stage2)
+]
+train_cfg = dict(
+    max_epochs=max_epochs,
+    val_interval=5,
+    dynamic_intervals=[((max_epochs - close_mosaic_epochs),
+                        _base_.val_interval_stage2)])
+optim_wrapper = dict(
+    optimizer=dict(
+        _delete_=True,
+        type='AdamW',
+        lr=base_lr,
+        weight_decay=weight_decay,
+        batch_size_per_gpu=train_batch_size_per_gpu),
+    paramwise_cfg=dict(
+        custom_keys={'backbone.text_model': dict(lr_mult=0.01),
+                     'logit_scale': dict(weight_decay=0.0)}),
+    constructor='YOLOWv5OptimizerConstructor')
+
+# evaluation settings
+val_evaluator = dict(
+    _delete_=True,
+    type='mmdet.CocoMetric',
+    proposal_nums=(100, 1, 10),
+    ann_file='data/coco/annotations/instances_val2017.json',
+    metric='bbox')

configs/finetune_coco/yolo_world_l_dual_vlpan_2e-4_80e_8gpus_mask-refine_finetune_coco.py

@@ -0,0 +1,181 @@
+_base_ = (
+    '../../third_party/mmyolo/configs/yolov8/'
+    'yolov8_l_mask-refine_syncbn_fast_8xb16-500e_coco.py')
+custom_imports = dict(
+    imports=['yolo_world'],
+    allow_failed_imports=False)
+
+# hyper-parameters
+num_classes = 80
+num_training_classes = 80
+max_epochs = 80  # Maximum training epochs
+close_mosaic_epochs = 10
+save_epoch_intervals = 5
+text_channels = 512
+neck_embed_channels = [128, 256, _base_.last_stage_out_channels // 2]
+neck_num_heads = [4, 8, _base_.last_stage_out_channels // 2 // 32]
+base_lr = 2e-4
+weight_decay = 0.05
+train_batch_size_per_gpu = 16
+load_from='pretrained_models/yolo_world_l_clip_base_dual_vlpan_2e-3adamw_32xb16_100e_o365_goldg_train_pretrained-0e566235.pth'
+persistent_workers = False
+
+# model settings
+model = dict(
+    type='YOLOWorldDetector',
+    mm_neck=True,
+    num_train_classes=num_training_classes,
+    num_test_classes=num_classes,
+    data_preprocessor=dict(type='YOLOWDetDataPreprocessor'),
+    backbone=dict(
+        _delete_=True,
+        type='MultiModalYOLOBackbone',
+        image_model={{_base_.model.backbone}},
+        text_model=dict(
+            type='HuggingCLIPLanguageBackbone',
+            model_name='openai/clip-vit-base-patch32',
+            frozen_modules=['all'])),
+    neck=dict(type='YOLOWorldDualPAFPN',
+              guide_channels=text_channels,
+              embed_channels=neck_embed_channels,
+              num_heads=neck_num_heads,
+              block_cfg=dict(type='MaxSigmoidCSPLayerWithTwoConv'),
+              text_enhancder=dict(type='ImagePoolingAttentionModule',
+                                  embed_channels=256,
+                                  num_heads=8)),
+    bbox_head=dict(type='YOLOWorldHead',
+                   head_module=dict(type='YOLOWorldHeadModule',
+                                    embed_dims=text_channels,
+                                    num_classes=num_training_classes)),
+    train_cfg=dict(assigner=dict(num_classes=num_training_classes)))
+
+# dataset settings
+text_transform = [
+    dict(type='RandomLoadText',
+         num_neg_samples=(num_classes, num_classes),
+         max_num_samples=num_training_classes,
+         padding_to_max=True,
+         padding_value=''),
+    dict(type='mmdet.PackDetInputs',
+         meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape', 'flip',
+                    'flip_direction', 'texts'))
+]
+mosaic_affine_transform = [
+    dict(
+        type='MultiModalMosaic',
+        img_scale=_base_.img_scale,
+        pad_val=114.0,
+        pre_transform=_base_.pre_transform),
+    dict(type='YOLOv5CopyPaste', prob=_base_.copypaste_prob),
+    dict(
+        type='YOLOv5RandomAffine',
+        max_rotate_degree=0.0,
+        max_shear_degree=0.0,
+        max_aspect_ratio=100.,
+        scaling_ratio_range=(1 - _base_.affine_scale,
+                             1 + _base_.affine_scale),
+        # img_scale is (width, height)
+        border=(-_base_.img_scale[0] // 2, -_base_.img_scale[1] // 2),
+        border_val=(114, 114, 114),
+        min_area_ratio=_base_.min_area_ratio,
+        use_mask_refine=_base_.use_mask2refine)
+]
+train_pipeline = [
+    *_base_.pre_transform,
+    *mosaic_affine_transform,
+    dict(
+        type='YOLOv5MultiModalMixUp',
+        prob=_base_.mixup_prob,
+        pre_transform=[*_base_.pre_transform,
+                       *mosaic_affine_transform]),
+    *_base_.last_transform[:-1],
+    *text_transform
+]
+train_pipeline_stage2 = [
+    *_base_.train_pipeline_stage2[:-1],
+    *text_transform
+]
+coco_train_dataset = dict(
+    _delete_=True,
+    type='MultiModalDataset',
+    dataset=dict(
+        type='YOLOv5CocoDataset',
+        data_root='data/coco',
+        ann_file='annotations/instances_train2017.json',
+        data_prefix=dict(img='train2017/'),
+        filter_cfg=dict(filter_empty_gt=False, min_size=32)),
+    class_text_path='data/texts/coco_class_texts.json',
+    pipeline=train_pipeline)
+
+train_dataloader = dict(
+    persistent_workers=persistent_workers,
+    batch_size=train_batch_size_per_gpu,
+    collate_fn=dict(type='yolow_collate'),
+    dataset=coco_train_dataset)
+test_pipeline = [
+    *_base_.test_pipeline[:-1],
+    dict(type='LoadText'),
+    dict(
+        type='mmdet.PackDetInputs',
+        meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape',
+                   'scale_factor', 'pad_param', 'texts'))
+]
+coco_val_dataset = dict(
+    _delete_=True,
+    type='MultiModalDataset',
+    dataset=dict(
+        type='YOLOv5CocoDataset',
+        data_root='data/coco',
+        ann_file='annotations/instances_val2017.json',
+        data_prefix=dict(img='val2017/'),
+        filter_cfg=dict(filter_empty_gt=False, min_size=32)),
+    class_text_path='data/texts/coco_class_texts.json',
+    pipeline=test_pipeline)
+val_dataloader = dict(dataset=coco_val_dataset)
+test_dataloader = val_dataloader
+# training settings
+default_hooks = dict(
+    param_scheduler=dict(
+        scheduler_type='linear',
+        lr_factor=0.01,
+        max_epochs=max_epochs),
+    checkpoint=dict(
+        max_keep_ckpts=-1,
+        save_best=None,
+        interval=save_epoch_intervals))
+custom_hooks = [
+    dict(
+        type='EMAHook',
+        ema_type='ExpMomentumEMA',
+        momentum=0.0001,
+        update_buffers=True,
+        strict_load=False,
+        priority=49),
+    dict(
+        type='mmdet.PipelineSwitchHook',
+        switch_epoch=max_epochs - close_mosaic_epochs,
+        switch_pipeline=train_pipeline_stage2)
+]
+train_cfg = dict(
+    max_epochs=max_epochs,
+    val_interval=5,
+    dynamic_intervals=[((max_epochs - close_mosaic_epochs),
+                        _base_.val_interval_stage2)])
+optim_wrapper = dict(
+    optimizer=dict(
+        _delete_=True,
+        type='AdamW',
+        lr=base_lr,
+        weight_decay=weight_decay,
+        batch_size_per_gpu=train_batch_size_per_gpu),
+    paramwise_cfg=dict(
+        custom_keys={'backbone.text_model': dict(lr_mult=0.01),
+                     'logit_scale': dict(weight_decay=0.0)}),
+    constructor='YOLOWv5OptimizerConstructor')
+# evaluation settings
+val_evaluator = dict(
+    _delete_=True,
+    type='mmdet.CocoMetric',
+    proposal_nums=(100, 1, 10),
+    ann_file='data/coco/annotations/instances_val2017.json',
+    metric='bbox')

configs/finetune_coco/yolo_world_l_efficient_neck_2e-4_80e_8gpus_mask-refine_finetune_coco.py

@@ -0,0 +1,159 @@
+_base_ = ('../../third_party/mmyolo/configs/yolov8/'
+          'yolov8_l_mask-refine_syncbn_fast_8xb16-500e_coco.py')
+custom_imports = dict(imports=['yolo_world'], allow_failed_imports=False)
+
+# hyper-parameters
+num_classes = 80
+num_training_classes = 80
+max_epochs = 80  # Maximum training epochs
+close_mosaic_epochs = 10
+save_epoch_intervals = 5
+text_channels = 512
+neck_embed_channels = [128, 256, _base_.last_stage_out_channels // 2]
+neck_num_heads = [4, 8, _base_.last_stage_out_channels // 2 // 32]
+base_lr = 2e-4
+weight_decay = 0.05
+train_batch_size_per_gpu = 16
+load_from = 'pretrained_models/yolo_world_l_clip_base_dual_vlpan_2e-3adamw_32xb16_100e_o365_goldg_train_pretrained-0e566235.pth'
+# huggingface text model
+text_model_name = 'openai/clip-vit-base-patch32'
+persistent_workers = False
+
+# model settings
+model = dict(
+    type='YOLOWorldDetector',
+    mm_neck=True,
+    num_train_classes=num_training_classes,
+    num_test_classes=num_classes,
+    data_preprocessor=dict(type='YOLOWDetDataPreprocessor'),
+    backbone=dict(
+        _delete_=True,
+        type='MultiModalYOLOBackbone',
+        image_model={{_base_.model.backbone}},
+        text_model=dict(
+            type='HuggingCLIPLanguageBackbone',
+            model_name=text_model_name,
+            frozen_modules=['all'])),
+    neck=dict(type='YOLOWorldPAFPN',
+              guide_channels=text_channels,
+              embed_channels=neck_embed_channels,
+              num_heads=neck_num_heads,
+              block_cfg=dict(type='EfficientCSPLayerWithTwoConv')),
+    bbox_head=dict(type='YOLOWorldHead',
+                   head_module=dict(type='YOLOWorldHeadModule',
+                                    embed_dims=text_channels,
+                                    num_classes=num_training_classes)),
+    train_cfg=dict(assigner=dict(num_classes=num_training_classes)))
+
+# dataset settings
+text_transform = [
+    dict(type='RandomLoadText',
+         num_neg_samples=(num_classes, num_classes),
+         max_num_samples=num_training_classes,
+         padding_to_max=True,
+         padding_value=''),
+    dict(type='mmdet.PackDetInputs',
+         meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape', 'flip',
+                    'flip_direction', 'texts'))
+]
+mosaic_affine_transform = [
+    dict(type='MultiModalMosaic',
+         img_scale=_base_.img_scale,
+         pad_val=114.0,
+         pre_transform=_base_.pre_transform),
+    dict(type='YOLOv5CopyPaste', prob=_base_.copypaste_prob),
+    dict(
+        type='YOLOv5RandomAffine',
+        max_rotate_degree=0.0,
+        max_shear_degree=0.0,
+        max_aspect_ratio=100.,
+        scaling_ratio_range=(1 - _base_.affine_scale, 1 + _base_.affine_scale),
+        # img_scale is (width, height)
+        border=(-_base_.img_scale[0] // 2, -_base_.img_scale[1] // 2),
+        border_val=(114, 114, 114),
+        min_area_ratio=_base_.min_area_ratio,
+        use_mask_refine=_base_.use_mask2refine)
+]
+train_pipeline = [
+    *_base_.pre_transform, *mosaic_affine_transform,
+    dict(type='YOLOv5MultiModalMixUp',
+         prob=_base_.mixup_prob,
+         pre_transform=[*_base_.pre_transform, *mosaic_affine_transform]),
+    *_base_.last_transform[:-1], *text_transform
+]
+train_pipeline_stage2 = [*_base_.train_pipeline_stage2[:-1], *text_transform]
+coco_train_dataset = dict(_delete_=True,
+                          type='MultiModalDataset',
+                          dataset=dict(
+                              type='YOLOv5CocoDataset',
+                              data_root='data/coco',
+                              ann_file='annotations/instances_train2017.json',
+                              data_prefix=dict(img='train2017/'),
+                              filter_cfg=dict(filter_empty_gt=False,
+                                              min_size=32)),
+                          class_text_path='data/texts/coco_class_texts.json',
+                          pipeline=train_pipeline)
+
+train_dataloader = dict(persistent_workers=persistent_workers,
+                        batch_size=train_batch_size_per_gpu,
+                        collate_fn=dict(type='yolow_collate'),
+                        dataset=coco_train_dataset)
+test_pipeline = [
+    *_base_.test_pipeline[:-1],
+    dict(type='LoadText'),
+    dict(type='mmdet.PackDetInputs',
+         meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape',
+                    'scale_factor', 'pad_param', 'texts'))
+]
+coco_val_dataset = dict(
+    _delete_=True,
+    type='MultiModalDataset',
+    dataset=dict(type='YOLOv5CocoDataset',
+                 data_root='data/coco',
+                 ann_file='annotations/instances_val2017.json',
+                 data_prefix=dict(img='val2017/'),
+                 filter_cfg=dict(filter_empty_gt=False, min_size=32)),
+    class_text_path='data/texts/coco_class_texts.json',
+    pipeline=test_pipeline)
+val_dataloader = dict(dataset=coco_val_dataset)
+test_dataloader = val_dataloader
+# training settings
+default_hooks = dict(param_scheduler=dict(scheduler_type='linear',
+                                          lr_factor=0.01,
+                                          max_epochs=max_epochs),
+                     checkpoint=dict(max_keep_ckpts=-1,
+                                     save_best=None,
+                                     interval=save_epoch_intervals))
+custom_hooks = [
+    dict(type='EMAHook',
+         ema_type='ExpMomentumEMA',
+         momentum=0.0001,
+         update_buffers=True,
+         strict_load=False,
+         priority=49),
+    dict(type='mmdet.PipelineSwitchHook',
+         switch_epoch=max_epochs - close_mosaic_epochs,
+         switch_pipeline=train_pipeline_stage2)
+]
+train_cfg = dict(max_epochs=max_epochs,
+                 val_interval=5,
+                 dynamic_intervals=[((max_epochs - close_mosaic_epochs),
+                                     _base_.val_interval_stage2)])
+optim_wrapper = dict(
+    optimizer=dict(
+        _delete_=True,
+        type='AdamW',
+        lr=base_lr,
+        weight_decay=weight_decay,
+        batch_size_per_gpu=train_batch_size_per_gpu),
+    paramwise_cfg=dict(
+        custom_keys={'backbone.text_model': dict(lr_mult=0.01),
+                     'logit_scale': dict(weight_decay=0.0)}),
+    constructor='YOLOWv5OptimizerConstructor')
+
+# evaluation settings
+val_evaluator = dict(_delete_=True,
+                     type='mmdet.CocoMetric',
+                     proposal_nums=(100, 1, 10),
+                     ann_file='data/coco/annotations/instances_val2017.json',
+                     metric='bbox')

configs/finetune_coco/yolo_world_v2_l_efficient_neck_2e-4_80e_8gpus_mask-refine_finetune_coco.py

@@ -0,0 +1,182 @@
+_base_ = (
+    '../../third_party/mmyolo/configs/yolov8/'
+    'yolov8_l_mask-refine_syncbn_fast_8xb16-500e_coco.py')
+custom_imports = dict(
+    imports=['yolo_world'],
+    allow_failed_imports=False)
+
+# hyper-parameters
+num_classes = 80
+num_training_classes = 80
+max_epochs = 80  # Maximum training epochs
+close_mosaic_epochs = 10
+save_epoch_intervals = 5
+text_channels = 512
+neck_embed_channels = [128, 256, _base_.last_stage_out_channels // 2]
+neck_num_heads = [4, 8, _base_.last_stage_out_channels // 2 // 32]
+base_lr = 2e-4
+weight_decay = 0.05
+train_batch_size_per_gpu = 16
+load_from = 'pretrained_models/yolo_world_l_clip_t2i_bn_2e-3adamw_32xb16-100e_obj365v1_goldg_cc3mlite_train-ca93cd1f.pth'
+text_model_name = '../pretrained_models/clip-vit-base-patch32-projection'
+text_model_name = 'openai/clip-vit-base-patch32'
+persistent_workers = False
+
+# model settings
+model = dict(
+    type='YOLOWorldDetector',
+    mm_neck=True,
+    num_train_classes=num_training_classes,
+    num_test_classes=num_classes,
+    data_preprocessor=dict(type='YOLOWDetDataPreprocessor'),
+    backbone=dict(
+        _delete_=True,
+        type='MultiModalYOLOBackbone',
+        image_model={{_base_.model.backbone}},
+        text_model=dict(
+            type='HuggingCLIPLanguageBackbone',
+            model_name=text_model_name,
+            frozen_modules=['all'])),
+    neck=dict(type='YOLOWorldPAFPN',
+              guide_channels=text_channels,
+              embed_channels=neck_embed_channels,
+              num_heads=neck_num_heads,
+              block_cfg=dict(type='EfficientCSPLayerWithTwoConv')),
+    bbox_head=dict(type='YOLOWorldHead',
+                   head_module=dict(type='YOLOWorldHeadModule',
+                                    use_bn_head=True,
+                                    embed_dims=text_channels,
+                                    num_classes=num_training_classes)),
+    train_cfg=dict(assigner=dict(num_classes=num_training_classes)))
+
+# dataset settings
+text_transform = [
+    dict(type='RandomLoadText',
+         num_neg_samples=(num_classes, num_classes),
+         max_num_samples=num_training_classes,
+         padding_to_max=True,
+         padding_value=''),
+    dict(type='mmdet.PackDetInputs',
+         meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape', 'flip',
+                    'flip_direction', 'texts'))
+]
+mosaic_affine_transform = [
+    dict(
+        type='MultiModalMosaic',
+        img_scale=_base_.img_scale,
+        pad_val=114.0,
+        pre_transform=_base_.pre_transform),
+    dict(type='YOLOv5CopyPaste', prob=_base_.copypaste_prob),
+    dict(
+        type='YOLOv5RandomAffine',
+        max_rotate_degree=0.0,
+        max_shear_degree=0.0,
+        max_aspect_ratio=100.,
+        scaling_ratio_range=(1 - _base_.affine_scale,
+                             1 + _base_.affine_scale),
+        # img_scale is (width, height)
+        border=(-_base_.img_scale[0] // 2, -_base_.img_scale[1] // 2),
+        border_val=(114, 114, 114),
+        min_area_ratio=_base_.min_area_ratio,
+        use_mask_refine=_base_.use_mask2refine)
+]
+train_pipeline = [
+    *_base_.pre_transform,
+    *mosaic_affine_transform,
+    dict(
+        type='YOLOv5MultiModalMixUp',
+        prob=_base_.mixup_prob,
+        pre_transform=[*_base_.pre_transform,
+                       *mosaic_affine_transform]),
+    *_base_.last_transform[:-1],
+    *text_transform
+]
+train_pipeline_stage2 = [
+    *_base_.train_pipeline_stage2[:-1],
+    *text_transform
+]
+coco_train_dataset = dict(
+    _delete_=True,
+    type='MultiModalDataset',
+    dataset=dict(
+        type='YOLOv5CocoDataset',
+        data_root='data/coco',
+        ann_file='annotations/instances_train2017.json',
+        data_prefix=dict(img='train2017/'),
+        filter_cfg=dict(filter_empty_gt=False, min_size=32)),
+    class_text_path='data/texts/coco_class_texts.json',
+    pipeline=train_pipeline)
+
+train_dataloader = dict(
+    persistent_workers=persistent_workers,
+    batch_size=train_batch_size_per_gpu,
+    collate_fn=dict(type='yolow_collate'),
+    dataset=coco_train_dataset)
+test_pipeline = [
+    *_base_.test_pipeline[:-1],
+    dict(type='LoadText'),
+    dict(
+        type='mmdet.PackDetInputs',
+        meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape',
+                   'scale_factor', 'pad_param', 'texts'))
+]
+coco_val_dataset = dict(
+    _delete_=True,
+    type='MultiModalDataset',
+    dataset=dict(
+        type='YOLOv5CocoDataset',
+        data_root='data/coco',
+        ann_file='annotations/instances_val2017.json',
+        data_prefix=dict(img='val2017/'),
+        filter_cfg=dict(filter_empty_gt=False, min_size=32)),
+    class_text_path='data/texts/coco_class_texts.json',
+    pipeline=test_pipeline)
+val_dataloader = dict(dataset=coco_val_dataset)
+test_dataloader = val_dataloader
+# training settings
+default_hooks = dict(
+    param_scheduler=dict(
+        scheduler_type='linear',
+        lr_factor=0.01,
+        max_epochs=max_epochs),
+    checkpoint=dict(
+        max_keep_ckpts=-1,
+        save_best=None,
+        interval=save_epoch_intervals))
+custom_hooks = [
+    dict(
+        type='EMAHook',
+        ema_type='ExpMomentumEMA',
+        momentum=0.0001,
+        update_buffers=True,
+        strict_load=False,
+        priority=49),
+    dict(
+        type='mmdet.PipelineSwitchHook',
+        switch_epoch=max_epochs - close_mosaic_epochs,
+        switch_pipeline=train_pipeline_stage2)
+]
+train_cfg = dict(
+    max_epochs=max_epochs,
+    val_interval=5,
+    dynamic_intervals=[((max_epochs - close_mosaic_epochs),
+                        _base_.val_interval_stage2)])
+optim_wrapper = dict(
+    optimizer=dict(
+        _delete_=True,
+        type='AdamW',
+        lr=base_lr,
+        weight_decay=weight_decay,
+        batch_size_per_gpu=train_batch_size_per_gpu),
+    paramwise_cfg=dict(
+        custom_keys={'backbone.text_model': dict(lr_mult=0.01),
+                     'logit_scale': dict(weight_decay=0.0)}),
+    constructor='YOLOWv5OptimizerConstructor')
+
+# evaluation settings
+val_evaluator = dict(
+    _delete_=True,
+    type='mmdet.CocoMetric',
+    proposal_nums=(100, 1, 10),
+    ann_file='data/coco/annotations/instances_val2017.json',
+    metric='bbox')

configs/finetune_coco/yolo_world_v2_l_vlpan_bn_2e-4_80e_8gpus_mask-refine_finetune_coco.py

@@ -0,0 +1,181 @@
+_base_ = (
+    '../../third_party/mmyolo/configs/yolov8/'
+    'yolov8_l_mask-refine_syncbn_fast_8xb16-500e_coco.py')
+custom_imports = dict(
+    imports=['yolo_world'],
+    allow_failed_imports=False)
+
+# hyper-parameters
+num_classes = 80
+num_training_classes = 80
+max_epochs = 80  # Maximum training epochs
+close_mosaic_epochs = 10
+save_epoch_intervals = 5
+text_channels = 512
+neck_embed_channels = [128, 256, _base_.last_stage_out_channels // 2]
+neck_num_heads = [4, 8, _base_.last_stage_out_channels // 2 // 32]
+base_lr = 2e-4
+weight_decay = 0.05
+train_batch_size_per_gpu = 16
+load_from = 'pretrained_models/yolo_world_l_clip_t2i_bn_2e-3adamw_32xb16-100e_obj365v1_goldg_cc3mlite_train-ca93cd1f.pth'
+text_model_name = '../pretrained_models/clip-vit-base-patch32-projection'
+text_model_name = 'openai/clip-vit-base-patch32'
+persistent_workers = False
+
+# model settings
+model = dict(
+    type='YOLOWorldDetector',
+    mm_neck=True,
+    num_train_classes=num_training_classes,
+    num_test_classes=num_classes,
+    data_preprocessor=dict(type='YOLOWDetDataPreprocessor'),
+    backbone=dict(
+        _delete_=True,
+        type='MultiModalYOLOBackbone',
+        image_model={{_base_.model.backbone}},
+        text_model=dict(
+            type='HuggingCLIPLanguageBackbone',
+            model_name=text_model_name,
+            frozen_modules=['all'])),
+    neck=dict(type='YOLOWorldPAFPN',
+              guide_channels=text_channels,
+              embed_channels=neck_embed_channels,
+              num_heads=neck_num_heads,
+              block_cfg=dict(type='MaxSigmoidCSPLayerWithTwoConv')),
+    bbox_head=dict(type='YOLOWorldHead',
+                   head_module=dict(type='YOLOWorldHeadModule',
+                                    use_bn_head=True,
+                                    embed_dims=text_channels,
+                                    num_classes=num_training_classes)),
+    train_cfg=dict(assigner=dict(num_classes=num_training_classes)))
+
+# dataset settings
+text_transform = [
+    dict(type='RandomLoadText',
+         num_neg_samples=(num_classes, num_classes),
+         max_num_samples=num_training_classes,
+         padding_to_max=True,
+         padding_value=''),
+    dict(type='mmdet.PackDetInputs',
+         meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape', 'flip',
+                    'flip_direction', 'texts'))
+]
+mosaic_affine_transform = [
+    dict(
+        type='MultiModalMosaic',
+        img_scale=_base_.img_scale,
+        pad_val=114.0,
+        pre_transform=_base_.pre_transform),
+    dict(type='YOLOv5CopyPaste', prob=_base_.copypaste_prob),
+    dict(
+        type='YOLOv5RandomAffine',
+        max_rotate_degree=0.0,
+        max_shear_degree=0.0,
+        max_aspect_ratio=100.,
+        scaling_ratio_range=(1 - _base_.affine_scale,
+                             1 + _base_.affine_scale),
+        # img_scale is (width, height)
+        border=(-_base_.img_scale[0] // 2, -_base_.img_scale[1] // 2),
+        border_val=(114, 114, 114),
+        min_area_ratio=_base_.min_area_ratio,
+        use_mask_refine=_base_.use_mask2refine)
+]
+train_pipeline = [
+    *_base_.pre_transform,
+    *mosaic_affine_transform,
+    dict(
+        type='YOLOv5MultiModalMixUp',
+        prob=_base_.mixup_prob,
+        pre_transform=[*_base_.pre_transform,
+                       *mosaic_affine_transform]),
+    *_base_.last_transform[:-1],
+    *text_transform
+]
+train_pipeline_stage2 = [
+    *_base_.train_pipeline_stage2[:-1],
+    *text_transform
+]
+coco_train_dataset = dict(
+    _delete_=True,
+    type='MultiModalDataset',
+    dataset=dict(
+        type='YOLOv5CocoDataset',
+        data_root='data/coco',
+        ann_file='annotations/instances_train2017.json',
+        data_prefix=dict(img='train2017/'),
+        filter_cfg=dict(filter_empty_gt=False, min_size=32)),
+    class_text_path='data/texts/coco_class_texts.json',
+    pipeline=train_pipeline)
+
+train_dataloader = dict(
+    persistent_workers=persistent_workers,
+    batch_size=train_batch_size_per_gpu,
+    collate_fn=dict(type='yolow_collate'),
+    dataset=coco_train_dataset)
+test_pipeline = [
+    *_base_.test_pipeline[:-1],
+    dict(type='LoadText'),
+    dict(
+        type='mmdet.PackDetInputs',
+        meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape',
+                   'scale_factor', 'pad_param', 'texts'))
+]
+coco_val_dataset = dict(
+    _delete_=True,
+    type='MultiModalDataset',
+    dataset=dict(
+        type='YOLOv5CocoDataset',
+        data_root='data/coco',
+        ann_file='annotations/instances_val2017.json',
+        data_prefix=dict(img='val2017/'),
+        filter_cfg=dict(filter_empty_gt=False, min_size=32)),
+    class_text_path='data/texts/coco_class_texts.json',
+    pipeline=test_pipeline)
+val_dataloader = dict(dataset=coco_val_dataset)
+test_dataloader = val_dataloader
+# training settings
+default_hooks = dict(
+    param_scheduler=dict(
+        scheduler_type='linear',
+        lr_factor=0.01,
+        max_epochs=max_epochs),
+    checkpoint=dict(
+        max_keep_ckpts=-1,
+        save_best=None,
+        interval=save_epoch_intervals))
+custom_hooks = [
+    dict(
+        type='EMAHook',
+        ema_type='ExpMomentumEMA',
+        momentum=0.0001,
+        update_buffers=True,
+        strict_load=False,
+        priority=49),
+    dict(
+        type='mmdet.PipelineSwitchHook',
+        switch_epoch=max_epochs - close_mosaic_epochs,
+        switch_pipeline=train_pipeline_stage2)
+]
+train_cfg = dict(
+    max_epochs=max_epochs,
+    val_interval=5,
+    dynamic_intervals=[((max_epochs - close_mosaic_epochs),
+                        _base_.val_interval_stage2)])
+optim_wrapper = dict(
+    optimizer=dict(
+        _delete_=True,
+        type='AdamW',
+        lr=base_lr,
+        weight_decay=weight_decay,
+        batch_size_per_gpu=train_batch_size_per_gpu),
+    paramwise_cfg=dict(
+        custom_keys={'backbone.text_model': dict(lr_mult=0.01),
+                     'logit_scale': dict(weight_decay=0.0)}),
+    constructor='YOLOWv5OptimizerConstructor')
+# evaluation settings
+val_evaluator = dict(
+    _delete_=True,
+    type='mmdet.CocoMetric',
+    proposal_nums=(100, 1, 10),
+    ann_file='data/coco/annotations/instances_val2017.json',
+    metric='bbox')

configs/finetune_coco/yolo_world_v2_l_vlpan_bn_sgd_1e-3_40e_8gpus_finetune_coco.py

@@ -0,0 +1,160 @@
+_base_ = ('../../third_party/mmyolo/configs/yolov8/'
+          'yolov8_l_syncbn_fast_8xb16-500e_coco.py')
+custom_imports = dict(imports=['yolo_world'], allow_failed_imports=False)
+
+# hyper-parameters
+num_classes = 80
+num_training_classes = 80
+max_epochs = 40  # Maximum training epochs
+close_mosaic_epochs = 30
+save_epoch_intervals = 5
+text_channels = 512
+neck_embed_channels = [128, 256, _base_.last_stage_out_channels // 2]
+neck_num_heads = [4, 8, _base_.last_stage_out_channels // 2 // 32]
+base_lr = 1e-3
+weight_decay = 0.0005
+train_batch_size_per_gpu = 16
+load_from = 'pretrained_models/yolo_world_l_clip_t2i_bn_2e-3adamw_32xb16-100e_obj365v1_goldg_cc3mlite_train-ca93cd1f.pth'
+text_model_name = '../pretrained_models/clip-vit-base-patch32-projection'
+text_model_name = 'openai/clip-vit-base-patch32'
+persistent_workers = False
+
+# model settings
+model = dict(type='YOLOWorldDetector',
+             mm_neck=True,
+             num_train_classes=num_training_classes,
+             num_test_classes=num_classes,
+             data_preprocessor=dict(type='YOLOWDetDataPreprocessor'),
+             backbone=dict(_delete_=True,
+                           type='MultiModalYOLOBackbone',
+                           image_model={{_base_.model.backbone}},
+                           text_model=dict(type='HuggingCLIPLanguageBackbone',
+                                           model_name=text_model_name,
+                                           frozen_modules=['all'])),
+             neck=dict(type='YOLOWorldPAFPN',
+                       guide_channels=text_channels,
+                       embed_channels=neck_embed_channels,
+                       num_heads=neck_num_heads,
+                       block_cfg=dict(type='MaxSigmoidCSPLayerWithTwoConv')),
+             bbox_head=dict(type='YOLOWorldHead',
+                            head_module=dict(
+                                type='YOLOWorldHeadModule',
+                                use_bn_head=True,
+                                embed_dims=text_channels,
+                                num_classes=num_training_classes)),
+             train_cfg=dict(assigner=dict(num_classes=num_training_classes)))
+
+# dataset settings
+text_transform = [
+    dict(type='RandomLoadText',
+         num_neg_samples=(num_classes, num_classes),
+         max_num_samples=num_training_classes,
+         padding_to_max=True,
+         padding_value=''),
+    dict(type='mmdet.PackDetInputs',
+         meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape', 'flip',
+                    'flip_direction', 'texts'))
+]
+mosaic_affine_transform = [
+    dict(type='MultiModalMosaic',
+         img_scale=_base_.img_scale,
+         pad_val=114.0,
+         pre_transform=_base_.pre_transform),
+    dict(
+        type='YOLOv5RandomAffine',
+        max_rotate_degree=0.0,
+        max_shear_degree=0.0,
+        max_aspect_ratio=100.,
+        scaling_ratio_range=(1 - _base_.affine_scale, 1 + _base_.affine_scale),
+        # img_scale is (width, height)
+        border=(-_base_.img_scale[0] // 2, -_base_.img_scale[1] // 2),
+        border_val=(114, 114, 114))
+]
+
+train_pipeline = [
+    *_base_.pre_transform, *mosaic_affine_transform,
+    dict(type='YOLOv5MultiModalMixUp',
+         prob=_base_.mixup_prob,
+         pre_transform=[*_base_.pre_transform, *mosaic_affine_transform]),
+    *_base_.last_transform[:-1], *text_transform
+]
+train_pipeline_stage2 = [*_base_.train_pipeline_stage2[:-1], *text_transform]
+
+coco_train_dataset = dict(_delete_=True,
+                          type='MultiModalDataset',
+                          dataset=dict(
+                              type='YOLOv5CocoDataset',
+                              data_root='data/coco',
+                              ann_file='annotations/instances_train2017.json',
+                              data_prefix=dict(img='train2017/'),
+                              filter_cfg=dict(filter_empty_gt=False,
+                                              min_size=32)),
+                          class_text_path='data/texts/coco_class_texts.json',
+                          pipeline=train_pipeline)
+
+train_dataloader = dict(persistent_workers=persistent_workers,
+                        batch_size=train_batch_size_per_gpu,
+                        collate_fn=dict(type='yolow_collate'),
+                        dataset=coco_train_dataset)
+test_pipeline = [
+    *_base_.test_pipeline[:-1],
+    dict(type='LoadText'),
+    dict(type='mmdet.PackDetInputs',
+         meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape',
+                    'scale_factor', 'pad_param', 'texts'))
+]
+coco_val_dataset = dict(
+    _delete_=True,
+    type='MultiModalDataset',
+    dataset=dict(type='YOLOv5CocoDataset',
+                 data_root='data/coco',
+                 ann_file='annotations/instances_val2017.json',
+                 data_prefix=dict(img='val2017/'),
+                 filter_cfg=dict(filter_empty_gt=False, min_size=32)),
+    class_text_path='data/texts/coco_class_texts.json',
+    pipeline=test_pipeline)
+val_dataloader = dict(dataset=coco_val_dataset)
+test_dataloader = val_dataloader
+# training settings
+default_hooks = dict(param_scheduler=dict(scheduler_type='linear',
+                                          lr_factor=0.01,
+                                          max_epochs=max_epochs),
+                     checkpoint=dict(max_keep_ckpts=-1,
+                                     save_best=None,
+                                     interval=save_epoch_intervals))
+custom_hooks = [
+    dict(type='EMAHook',
+         ema_type='ExpMomentumEMA',
+         momentum=0.0001,
+         update_buffers=True,
+         strict_load=False,
+         priority=49),
+    dict(type='mmdet.PipelineSwitchHook',
+         switch_epoch=max_epochs - close_mosaic_epochs,
+         switch_pipeline=train_pipeline_stage2)
+]
+train_cfg = dict(max_epochs=max_epochs,
+                 val_interval=5,
+                 dynamic_intervals=[((max_epochs - close_mosaic_epochs),
+                                     _base_.val_interval_stage2)])
+optim_wrapper = dict(optimizer=dict(
+    _delete_=True,
+    type='SGD',
+    lr=base_lr,
+    momentum=0.937,
+    nesterov=True,
+    weight_decay=weight_decay,
+    batch_size_per_gpu=train_batch_size_per_gpu),
+                     paramwise_cfg=dict(
+                         custom_keys={
+                             'backbone.text_model': dict(lr_mult=0.01),
+                             'logit_scale': dict(weight_decay=0.0)
+                         }),
+                     constructor='YOLOWv5OptimizerConstructor')
+
+# evaluation settings
+val_evaluator = dict(_delete_=True,
+                     type='mmdet.CocoMetric',
+                     proposal_nums=(100, 1, 10),
+                     ann_file='data/coco/annotations/instances_val2017.json',
+                     metric='bbox')

configs/finetune_coco/yolo_world_v2_l_vlpan_bn_sgd_1e-3_80e_8gpus_mask-refine_finetune_coco.py

@@ -0,0 +1,161 @@
+_base_ = ('../../third_party/mmyolo/configs/yolov8/'
+          'yolov8_l_mask-refine_syncbn_fast_8xb16-500e_coco.py')
+custom_imports = dict(imports=['yolo_world'], allow_failed_imports=False)
+
+# hyper-parameters
+num_classes = 80
+num_training_classes = 80
+max_epochs = 80  # Maximum training epochs
+close_mosaic_epochs = 10
+save_epoch_intervals = 5
+text_channels = 512
+neck_embed_channels = [128, 256, _base_.last_stage_out_channels // 2]
+neck_num_heads = [4, 8, _base_.last_stage_out_channels // 2 // 32]
+base_lr = 1e-3
+weight_decay = 0.0005
+train_batch_size_per_gpu = 16
+load_from = 'pretrained_models/yolo_world_l_clip_t2i_bn_2e-3adamw_32xb16-100e_obj365v1_goldg_cc3mlite_train-ca93cd1f.pth'
+text_model_name = '../pretrained_models/clip-vit-base-patch32-projection'
+text_model_name = 'openai/clip-vit-base-patch32'
+persistent_workers = False
+
+# model settings
+model = dict(type='YOLOWorldDetector',
+             mm_neck=True,
+             num_train_classes=num_training_classes,
+             num_test_classes=num_classes,
+             data_preprocessor=dict(type='YOLOWDetDataPreprocessor'),
+             backbone=dict(_delete_=True,
+                           type='MultiModalYOLOBackbone',
+                           image_model={{_base_.model.backbone}},
+                           text_model=dict(type='HuggingCLIPLanguageBackbone',
+                                           model_name=text_model_name,
+                                           frozen_modules=['all'])),
+             neck=dict(type='YOLOWorldPAFPN',
+                       guide_channels=text_channels,
+                       embed_channels=neck_embed_channels,
+                       num_heads=neck_num_heads,
+                       block_cfg=dict(type='MaxSigmoidCSPLayerWithTwoConv')),
+             bbox_head=dict(type='YOLOWorldHead',
+                            head_module=dict(
+                                type='YOLOWorldHeadModule',
+                                use_bn_head=True,
+                                embed_dims=text_channels,
+                                num_classes=num_training_classes)),
+             train_cfg=dict(assigner=dict(num_classes=num_training_classes)))
+
+# dataset settings
+text_transform = [
+    dict(type='RandomLoadText',
+         num_neg_samples=(num_classes, num_classes),
+         max_num_samples=num_training_classes,
+         padding_to_max=True,
+         padding_value=''),
+    dict(type='mmdet.PackDetInputs',
+         meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape', 'flip',
+                    'flip_direction', 'texts'))
+]
+mosaic_affine_transform = [
+    dict(type='MultiModalMosaic',
+         img_scale=_base_.img_scale,
+         pad_val=114.0,
+         pre_transform=_base_.pre_transform),
+    dict(type='YOLOv5CopyPaste', prob=_base_.copypaste_prob),
+    dict(
+        type='YOLOv5RandomAffine',
+        max_rotate_degree=0.0,
+        max_shear_degree=0.0,
+        max_aspect_ratio=100.,
+        scaling_ratio_range=(1 - _base_.affine_scale, 1 + _base_.affine_scale),
+        # img_scale is (width, height)
+        border=(-_base_.img_scale[0] // 2, -_base_.img_scale[1] // 2),
+        border_val=(114, 114, 114),
+        min_area_ratio=_base_.min_area_ratio,
+        use_mask_refine=_base_.use_mask2refine)
+]
+train_pipeline = [
+    *_base_.pre_transform, *mosaic_affine_transform,
+    dict(type='YOLOv5MultiModalMixUp',
+         prob=_base_.mixup_prob,
+         pre_transform=[*_base_.pre_transform, *mosaic_affine_transform]),
+    *_base_.last_transform[:-1], *text_transform
+]
+train_pipeline_stage2 = [*_base_.train_pipeline_stage2[:-1], *text_transform]
+coco_train_dataset = dict(_delete_=True,
+                          type='MultiModalDataset',
+                          dataset=dict(
+                              type='YOLOv5CocoDataset',
+                              data_root='data/coco',
+                              ann_file='annotations/instances_train2017.json',
+                              data_prefix=dict(img='train2017/'),
+                              filter_cfg=dict(filter_empty_gt=False,
+                                              min_size=32)),
+                          class_text_path='data/texts/coco_class_texts.json',
+                          pipeline=train_pipeline)
+
+train_dataloader = dict(persistent_workers=persistent_workers,
+                        batch_size=train_batch_size_per_gpu,
+                        collate_fn=dict(type='yolow_collate'),
+                        dataset=coco_train_dataset)
+test_pipeline = [
+    *_base_.test_pipeline[:-1],
+    dict(type='LoadText'),
+    dict(type='mmdet.PackDetInputs',
+         meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape',
+                    'scale_factor', 'pad_param', 'texts'))
+]
+coco_val_dataset = dict(
+    _delete_=True,
+    type='MultiModalDataset',
+    dataset=dict(type='YOLOv5CocoDataset',
+                 data_root='data/coco',
+                 ann_file='annotations/instances_val2017.json',
+                 data_prefix=dict(img='val2017/'),
+                 filter_cfg=dict(filter_empty_gt=False, min_size=32)),
+    class_text_path='data/texts/coco_class_texts.json',
+    pipeline=test_pipeline)
+val_dataloader = dict(dataset=coco_val_dataset)
+test_dataloader = val_dataloader
+# training settings
+default_hooks = dict(param_scheduler=dict(scheduler_type='linear',
+                                          lr_factor=0.01,
+                                          max_epochs=max_epochs),
+                     checkpoint=dict(max_keep_ckpts=-1,
+                                     save_best=None,
+                                     interval=save_epoch_intervals))
+custom_hooks = [
+    dict(type='EMAHook',
+         ema_type='ExpMomentumEMA',
+         momentum=0.0001,
+         update_buffers=True,
+         strict_load=False,
+         priority=49),
+    dict(type='mmdet.PipelineSwitchHook',
+         switch_epoch=max_epochs - close_mosaic_epochs,
+         switch_pipeline=train_pipeline_stage2)
+]
+train_cfg = dict(max_epochs=max_epochs,
+                 val_interval=5,
+                 dynamic_intervals=[((max_epochs - close_mosaic_epochs),
+                                     _base_.val_interval_stage2)])
+optim_wrapper = dict(optimizer=dict(
+    _delete_=True,
+    type='SGD',
+    lr=base_lr,
+    momentum=0.937,
+    nesterov=True,
+    weight_decay=weight_decay,
+    batch_size_per_gpu=train_batch_size_per_gpu),
+                     paramwise_cfg=dict(
+                         custom_keys={
+                             'backbone.text_model': dict(lr_mult=0.01),
+                             'logit_scale': dict(weight_decay=0.0)
+                         }),
+                     constructor='YOLOWv5OptimizerConstructor')
+
+# evaluation settings
+val_evaluator = dict(_delete_=True,
+                     type='mmdet.CocoMetric',
+                     proposal_nums=(100, 1, 10),
+                     ann_file='data/coco/annotations/instances_val2017.json',
+                     metric='bbox')

configs/finetune_coco/yolo_world_v2_m_vlpan_bn_2e-4_80e_8gpus_mask-refine_finetune_coco.py

@@ -0,0 +1,182 @@
+_base_ = (
+    '../../third_party/mmyolo/configs/yolov8/'
+    'yolov8_m_mask-refine_syncbn_fast_8xb16-500e_coco.py')
+custom_imports = dict(
+    imports=['yolo_world'],
+    allow_failed_imports=False)
+
+# hyper-parameters
+num_classes = 80
+num_training_classes = 80
+max_epochs = 80  # Maximum training epochs
+close_mosaic_epochs = 10
+save_epoch_intervals = 5
+text_channels = 512
+neck_embed_channels = [128, 256, _base_.last_stage_out_channels // 2]
+neck_num_heads = [4, 8, _base_.last_stage_out_channels // 2 // 32]
+base_lr = 2e-4
+weight_decay = 0.05
+train_batch_size_per_gpu = 16
+load_from = 'pretrained_models/yolo_world_m_clip_t2i_bn_2e-3adamw_32xb16-100e_obj365v1_goldg_train-c6237d5b.pth'
+# text_model_name = '../pretrained_models/clip-vit-base-patch32-projection'
+text_model_name = 'openai/clip-vit-base-patch32'
+persistent_workers = False
+
+# model settings
+model = dict(
+    type='YOLOWorldDetector',
+    mm_neck=True,
+    num_train_classes=num_training_classes,
+    num_test_classes=num_classes,
+    data_preprocessor=dict(type='YOLOWDetDataPreprocessor'),
+    backbone=dict(
+        _delete_=True,
+        type='MultiModalYOLOBackbone',
+        image_model={{_base_.model.backbone}},
+        text_model=dict(
+            type='HuggingCLIPLanguageBackbone',
+            model_name=text_model_name,
+            frozen_modules=['all'])),
+    neck=dict(type='YOLOWorldPAFPN',
+              guide_channels=text_channels,
+              embed_channels=neck_embed_channels,
+              num_heads=neck_num_heads,
+              block_cfg=dict(type='MaxSigmoidCSPLayerWithTwoConv')),
+    bbox_head=dict(type='YOLOWorldHead',
+                   head_module=dict(type='YOLOWorldHeadModule',
+                                    use_bn_head=True,
+                                    embed_dims=text_channels,
+                                    num_classes=num_training_classes)),
+    train_cfg=dict(assigner=dict(num_classes=num_training_classes)))
+
+# dataset settings
+text_transform = [
+    dict(type='RandomLoadText',
+         num_neg_samples=(num_classes, num_classes),
+         max_num_samples=num_training_classes,
+         padding_to_max=True,
+         padding_value=''),
+    dict(type='mmdet.PackDetInputs',
+         meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape', 'flip',
+                    'flip_direction', 'texts'))
+]
+mosaic_affine_transform = [
+    dict(
+        type='MultiModalMosaic',
+        img_scale=_base_.img_scale,
+        pad_val=114.0,
+        pre_transform=_base_.pre_transform),
+    dict(type='YOLOv5CopyPaste', prob=_base_.copypaste_prob),
+    dict(
+        type='YOLOv5RandomAffine',
+        max_rotate_degree=0.0,
+        max_shear_degree=0.0,
+        max_aspect_ratio=100.,
+        scaling_ratio_range=(1 - _base_.affine_scale,
+                             1 + _base_.affine_scale),
+        # img_scale is (width, height)
+        border=(-_base_.img_scale[0] // 2, -_base_.img_scale[1] // 2),
+        border_val=(114, 114, 114),
+        min_area_ratio=_base_.min_area_ratio,
+        use_mask_refine=_base_.use_mask2refine)
+]
+train_pipeline = [
+    *_base_.pre_transform,
+    *mosaic_affine_transform,
+    dict(
+        type='YOLOv5MultiModalMixUp',
+        prob=_base_.mixup_prob,
+        pre_transform=[*_base_.pre_transform,
+                       *mosaic_affine_transform]),
+    *_base_.last_transform[:-1],
+    *text_transform
+]
+train_pipeline_stage2 = [
+    *_base_.train_pipeline_stage2[:-1],
+    *text_transform
+]
+coco_train_dataset = dict(
+    _delete_=True,
+    type='MultiModalDataset',
+    dataset=dict(
+        type='YOLOv5CocoDataset',
+        data_root='data/coco',
+        ann_file='annotations/instances_train2017.json',
+        data_prefix=dict(img='train2017/'),
+        filter_cfg=dict(filter_empty_gt=False, min_size=32)),
+    class_text_path='data/texts/coco_class_texts.json',
+    pipeline=train_pipeline)
+
+train_dataloader = dict(
+    persistent_workers=persistent_workers,
+    batch_size=train_batch_size_per_gpu,
+    collate_fn=dict(type='yolow_collate'),
+    dataset=coco_train_dataset)
+test_pipeline = [
+    *_base_.test_pipeline[:-1],
+    dict(type='LoadText'),
+    dict(
+        type='mmdet.PackDetInputs',
+        meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape',
+                   'scale_factor', 'pad_param', 'texts'))
+]
+coco_val_dataset = dict(
+    _delete_=True,
+    type='MultiModalDataset',
+    dataset=dict(
+        type='YOLOv5CocoDataset',
+        data_root='data/coco',
+        ann_file='annotations/instances_val2017.json',
+        data_prefix=dict(img='val2017/'),
+        filter_cfg=dict(filter_empty_gt=False, min_size=32)),
+    class_text_path='data/texts/coco_class_texts.json',
+    pipeline=test_pipeline)
+val_dataloader = dict(dataset=coco_val_dataset)
+test_dataloader = val_dataloader
+# training settings
+default_hooks = dict(
+    param_scheduler=dict(
+        scheduler_type='linear',
+        lr_factor=0.01,
+        max_epochs=max_epochs),
+    checkpoint=dict(
+        max_keep_ckpts=-1,
+        save_best=None,
+        interval=save_epoch_intervals))
+custom_hooks = [
+    dict(
+        type='EMAHook',
+        ema_type='ExpMomentumEMA',
+        momentum=0.0001,
+        update_buffers=True,
+        strict_load=False,
+        priority=49),
+    dict(
+        type='mmdet.PipelineSwitchHook',
+        switch_epoch=max_epochs - close_mosaic_epochs,
+        switch_pipeline=train_pipeline_stage2)
+]
+train_cfg = dict(
+    max_epochs=max_epochs,
+    val_interval=5,
+    dynamic_intervals=[((max_epochs - close_mosaic_epochs),
+                        _base_.val_interval_stage2)])
+optim_wrapper = dict(
+    optimizer=dict(
+        _delete_=True,
+        type='AdamW',
+        lr=base_lr,
+        weight_decay=weight_decay,
+        batch_size_per_gpu=train_batch_size_per_gpu),
+    paramwise_cfg=dict(
+        custom_keys={'backbone.text_model': dict(lr_mult=0.01),
+                     'logit_scale': dict(weight_decay=0.0)}),
+    constructor='YOLOWv5OptimizerConstructor')
+
+# evaluation settings
+val_evaluator = dict(
+    _delete_=True,
+    type='mmdet.CocoMetric',
+    proposal_nums=(100, 1, 10),
+    ann_file='data/coco/annotations/instances_val2017.json',
+    metric='bbox')

configs/finetune_coco/yolo_world_v2_s_bn_2e-4_80e_8gpus_mask-refine_finetune_coco.py

@@ -0,0 +1,145 @@
+_base_ = ('../../third_party/mmyolo/configs/yolov8/'
+          'yolov8_s_mask-refine_syncbn_fast_8xb16-500e_coco.py')
+custom_imports = dict(imports=['yolo_world'], allow_failed_imports=False)
+
+# hyper-parameters
+num_classes = 80
+num_training_classes = 80
+max_epochs = 80  # Maximum training epochs
+close_mosaic_epochs = 10
+save_epoch_intervals = 5
+text_channels = 512
+neck_embed_channels = [128, 256, _base_.last_stage_out_channels // 2]
+neck_num_heads = [4, 8, _base_.last_stage_out_channels // 2 // 32]
+base_lr = 2e-4
+weight_decay = 0.05
+train_batch_size_per_gpu = 16
+load_from = '../FastDet/output_models/pretrain_yolow-v8_s_clipv2_frozen_te_noprompt_t2i_bn_2e-3adamw_scale_lr_wd_32xb16-100e_obj365v1_goldg_cc3mram250k_train_lviseval-e3592307_rep_conv.pth'
+persistent_workers = False
+mixup_prob = 0.15
+copypaste_prob = 0.3
+
+# model settings
+model = dict(type='SimpleYOLOWorldDetector',
+             mm_neck=True,
+             num_train_classes=num_classes,
+             num_test_classes=num_classes,
+             reparameterized=True,
+             data_preprocessor=dict(type='YOLOv5DetDataPreprocessor'),
+             backbone=dict(_delete_=True,
+                           type='MultiModalYOLOBackbone',
+                           text_model=None,
+                           image_model={{_base_.model.backbone}},
+                           with_text_model=False),
+             neck=dict(type='YOLOWorldPAFPN',
+                       guide_channels=text_channels,
+                       embed_channels=neck_embed_channels,
+                       num_heads=neck_num_heads,
+                       block_cfg=dict(type='EfficientCSPLayerWithTwoConv')),
+             bbox_head=dict(head_module=dict(type='RepYOLOWorldHeadModule',
+                                             embed_dims=text_channels,
+                                             num_guide=num_classes,
+                                             num_classes=num_classes)),
+             train_cfg=dict(assigner=dict(num_classes=num_classes)))
+
+# dataset settings
+final_transform = [
+    dict(type='mmdet.PackDetInputs',
+         meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape', 'flip',
+                    'flip_direction'))
+]
+mosaic_affine_transform = [
+    dict(type='Mosaic',
+         img_scale=_base_.img_scale,
+         pad_val=114.0,
+         pre_transform=_base_.pre_transform),
+    dict(type='YOLOv5CopyPaste', prob=copypaste_prob),
+    dict(
+        type='YOLOv5RandomAffine',
+        max_rotate_degree=0.0,
+        max_shear_degree=0.0,
+        max_aspect_ratio=100.,
+        scaling_ratio_range=(1 - _base_.affine_scale, 1 + _base_.affine_scale),
+        # img_scale is (width, height)
+        border=(-_base_.img_scale[0] // 2, -_base_.img_scale[1] // 2),
+        border_val=(114, 114, 114),
+        min_area_ratio=_base_.min_area_ratio,
+        use_mask_refine=_base_.use_mask2refine)
+]
+train_pipeline = [
+    *_base_.pre_transform, *mosaic_affine_transform,
+    dict(type='YOLOv5MixUp',
+         prob=mixup_prob,
+         pre_transform=[*_base_.pre_transform, *mosaic_affine_transform]),
+    *_base_.last_transform[:-1], *final_transform
+]
+
+train_pipeline_stage2 = [*_base_.train_pipeline_stage2[:-1], *final_transform]
+
+coco_train_dataset = dict(type='YOLOv5CocoDataset',
+                          data_root='data/coco',
+                          ann_file='annotations/instances_train2017.json',
+                          data_prefix=dict(img='train2017/'),
+                          filter_cfg=dict(filter_empty_gt=False, min_size=32),
+                          pipeline=train_pipeline)
+
+train_dataloader = dict(persistent_workers=persistent_workers,
+                        batch_size=train_batch_size_per_gpu,
+                        collate_fn=dict(type='yolow_collate'),
+                        dataset=coco_train_dataset)
+
+train_dataloader = dict(persistent_workers=persistent_workers,
+                        batch_size=train_batch_size_per_gpu,
+                        collate_fn=dict(type='yolow_collate'),
+                        dataset=coco_train_dataset)
+test_pipeline = [
+    *_base_.test_pipeline[:-1],
+    dict(type='mmdet.PackDetInputs',
+         meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape',
+                    'scale_factor', 'pad_param'))
+]
+coco_val_dataset = dict(type='YOLOv5CocoDataset',
+                        data_root='data/coco',
+                        ann_file='annotations/instances_val2017.json',
+                        data_prefix=dict(img='val2017/'),
+                        filter_cfg=dict(filter_empty_gt=False, min_size=32),
+                        pipeline=test_pipeline)
+
+val_dataloader = dict(dataset=coco_val_dataset)
+test_dataloader = val_dataloader
+# training settings
+default_hooks = dict(param_scheduler=dict(scheduler_type='linear',
+                                          lr_factor=0.01,
+                                          max_epochs=max_epochs),
+                     checkpoint=dict(max_keep_ckpts=-1,
+                                     save_best=None,
+                                     interval=save_epoch_intervals))
+custom_hooks = [
+    dict(type='EMAHook',
+         ema_type='ExpMomentumEMA',
+         momentum=0.0001,
+         update_buffers=True,
+         strict_load=False,
+         priority=49),
+    dict(type='mmdet.PipelineSwitchHook',
+         switch_epoch=max_epochs - close_mosaic_epochs,
+         switch_pipeline=train_pipeline_stage2)
+]
+train_cfg = dict(max_epochs=max_epochs,
+                 val_interval=5,
+                 dynamic_intervals=[((max_epochs - close_mosaic_epochs),
+                                     _base_.val_interval_stage2)])
+optim_wrapper = dict(optimizer=dict(
+    _delete_=True,
+    type='AdamW',
+    lr=base_lr,
+    weight_decay=weight_decay,
+    batch_size_per_gpu=train_batch_size_per_gpu),
+                     constructor='YOLOWv5OptimizerConstructor')
+
+# evaluation settings
+val_evaluator = dict(_delete_=True,
+                     type='mmdet.CocoMetric',
+                     proposal_nums=(100, 1, 10),
+                     ann_file='data/coco/annotations/instances_val2017.json',
+                     metric='bbox')

configs/finetune_coco/yolo_world_v2_s_rep_vlpan_bn_2e-4_80e_8gpus_mask-refine_finetune_coco.py

@@ -0,0 +1,146 @@
+_base_ = ('../../third_party/mmyolo/configs/yolov8/'
+          'yolov8_s_mask-refine_syncbn_fast_8xb16-500e_coco.py')
+custom_imports = dict(imports=['yolo_world'], allow_failed_imports=False)
+
+# hyper-parameters
+num_classes = 80
+num_training_classes = 80
+max_epochs = 80  # Maximum training epochs
+close_mosaic_epochs = 10
+save_epoch_intervals = 5
+text_channels = 512
+neck_embed_channels = [128, 256, _base_.last_stage_out_channels // 2]
+neck_num_heads = [4, 8, _base_.last_stage_out_channels // 2 // 32]
+base_lr = 2e-4
+weight_decay = 0.05
+train_batch_size_per_gpu = 16
+load_from = '../FastDet/output_models/yolo_world_s_clip_t2i_bn_2e-3adamw_32xb16-100e_obj365v1_goldg_train-55b943ea_rep_conv.pth'
+persistent_workers = False
+mixup_prob = 0.15
+copypaste_prob = 0.3
+
+# model settings
+model = dict(type='SimpleYOLOWorldDetector',
+             mm_neck=True,
+             num_train_classes=num_classes,
+             num_test_classes=num_classes,
+             reparameterized=True,
+             data_preprocessor=dict(type='YOLOv5DetDataPreprocessor'),
+             backbone=dict(_delete_=True,
+                           type='MultiModalYOLOBackbone',
+                           text_model=None,
+                           image_model={{_base_.model.backbone}},
+                           with_text_model=False),
+             neck=dict(type='YOLOWorldPAFPN',
+                       guide_channels=num_classes,
+                       embed_channels=neck_embed_channels,
+                       num_heads=neck_num_heads,
+                       block_cfg=dict(type='RepConvMaxSigmoidCSPLayerWithTwoConv',
+                                      guide_channels=num_classes)),
+             bbox_head=dict(head_module=dict(type='RepYOLOWorldHeadModule',
+                                             embed_dims=text_channels,
+                                             num_guide=num_classes,
+                                             num_classes=num_classes)),
+             train_cfg=dict(assigner=dict(num_classes=num_classes)))
+
+# dataset settings
+final_transform = [
+    dict(type='mmdet.PackDetInputs',
+         meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape', 'flip',
+                    'flip_direction'))
+]
+mosaic_affine_transform = [
+    dict(type='Mosaic',
+         img_scale=_base_.img_scale,
+         pad_val=114.0,
+         pre_transform=_base_.pre_transform),
+    dict(type='YOLOv5CopyPaste', prob=copypaste_prob),
+    dict(
+        type='YOLOv5RandomAffine',
+        max_rotate_degree=0.0,
+        max_shear_degree=0.0,
+        max_aspect_ratio=100.,
+        scaling_ratio_range=(1 - _base_.affine_scale, 1 + _base_.affine_scale),
+        # img_scale is (width, height)
+        border=(-_base_.img_scale[0] // 2, -_base_.img_scale[1] // 2),
+        border_val=(114, 114, 114),
+        min_area_ratio=_base_.min_area_ratio,
+        use_mask_refine=_base_.use_mask2refine)
+]
+train_pipeline = [
+    *_base_.pre_transform, *mosaic_affine_transform,
+    dict(type='YOLOv5MixUp',
+         prob=mixup_prob,
+         pre_transform=[*_base_.pre_transform, *mosaic_affine_transform]),
+    *_base_.last_transform[:-1], *final_transform
+]
+
+train_pipeline_stage2 = [*_base_.train_pipeline_stage2[:-1], *final_transform]
+
+coco_train_dataset = dict(type='YOLOv5CocoDataset',
+                          data_root='data/coco',
+                          ann_file='annotations/instances_train2017.json',
+                          data_prefix=dict(img='train2017/'),
+                          filter_cfg=dict(filter_empty_gt=False, min_size=32),
+                          pipeline=train_pipeline)
+
+train_dataloader = dict(persistent_workers=persistent_workers,
+                        batch_size=train_batch_size_per_gpu,
+                        collate_fn=dict(type='yolow_collate'),
+                        dataset=coco_train_dataset)
+
+train_dataloader = dict(persistent_workers=persistent_workers,
+                        batch_size=train_batch_size_per_gpu,
+                        collate_fn=dict(type='yolow_collate'),
+                        dataset=coco_train_dataset)
+test_pipeline = [
+    *_base_.test_pipeline[:-1],
+    dict(type='mmdet.PackDetInputs',
+         meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape',
+                    'scale_factor', 'pad_param'))
+]
+coco_val_dataset = dict(type='YOLOv5CocoDataset',
+                        data_root='data/coco',
+                        ann_file='annotations/instances_val2017.json',
+                        data_prefix=dict(img='val2017/'),
+                        filter_cfg=dict(filter_empty_gt=False, min_size=32),
+                        pipeline=test_pipeline)
+
+val_dataloader = dict(dataset=coco_val_dataset)
+test_dataloader = val_dataloader
+# training settings
+default_hooks = dict(param_scheduler=dict(scheduler_type='linear',
+                                          lr_factor=0.01,
+                                          max_epochs=max_epochs),
+                     checkpoint=dict(max_keep_ckpts=-1,
+                                     save_best=None,
+                                     interval=save_epoch_intervals))
+custom_hooks = [
+    dict(type='EMAHook',
+         ema_type='ExpMomentumEMA',
+         momentum=0.0001,
+         update_buffers=True,
+         strict_load=False,
+         priority=49),
+    dict(type='mmdet.PipelineSwitchHook',
+         switch_epoch=max_epochs - close_mosaic_epochs,
+         switch_pipeline=train_pipeline_stage2)
+]
+train_cfg = dict(max_epochs=max_epochs,
+                 val_interval=5,
+                 dynamic_intervals=[((max_epochs - close_mosaic_epochs),
+                                     _base_.val_interval_stage2)])
+optim_wrapper = dict(optimizer=dict(
+    _delete_=True,
+    type='AdamW',
+    lr=base_lr,
+    weight_decay=weight_decay,
+    batch_size_per_gpu=train_batch_size_per_gpu),
+                     constructor='YOLOWv5OptimizerConstructor')
+
+# evaluation settings
+val_evaluator = dict(_delete_=True,
+                     type='mmdet.CocoMetric',
+                     proposal_nums=(100, 1, 10),
+                     ann_file='data/coco/annotations/instances_val2017.json',
+                     metric='bbox')

configs/finetune_coco/yolo_world_v2_s_vlpan_bn_2e-4_80e_8gpus_mask-refine_finetune_coco.py

@@ -0,0 +1,184 @@
+_base_ = (
+    '../../third_party/mmyolo/configs/yolov8/'
+    'yolov8_s_mask-refine_syncbn_fast_8xb16-500e_coco.py')
+custom_imports = dict(
+    imports=['yolo_world'],
+    allow_failed_imports=False)
+
+# hyper-parameters
+num_classes = 80
+num_training_classes = 80
+max_epochs = 80  # Maximum training epochs
+close_mosaic_epochs = 10
+save_epoch_intervals = 5
+text_channels = 512
+neck_embed_channels = [128, 256, _base_.last_stage_out_channels // 2]
+neck_num_heads = [4, 8, _base_.last_stage_out_channels // 2 // 32]
+base_lr = 2e-4
+weight_decay = 0.05
+train_batch_size_per_gpu = 16
+load_from = 'pretrained_models/yolo_world_s_clip_t2i_bn_2e-3adamw_32xb16-100e_obj365v1_goldg_train-55b943ea.pth'
+text_model_name = '../pretrained_models/clip-vit-base-patch32-projection'
+text_model_name = 'openai/clip-vit-base-patch32'
+persistent_workers = False
+mixup_prob = 0.15
+copypaste_prob = 0.3
+
+# model settings
+model = dict(
+    type='YOLOWorldDetector',
+    mm_neck=True,
+    num_train_classes=num_training_classes,
+    num_test_classes=num_classes,
+    data_preprocessor=dict(type='YOLOWDetDataPreprocessor'),
+    backbone=dict(
+        _delete_=True,
+        type='MultiModalYOLOBackbone',
+        image_model={{_base_.model.backbone}},
+        text_model=dict(
+            type='HuggingCLIPLanguageBackbone',
+            model_name=text_model_name,
+            frozen_modules=['all'])),
+    neck=dict(type='YOLOWorldPAFPN',
+              guide_channels=text_channels,
+              embed_channels=neck_embed_channels,
+              num_heads=neck_num_heads,
+              block_cfg=dict(type='MaxSigmoidCSPLayerWithTwoConv')),
+    bbox_head=dict(type='YOLOWorldHead',
+                   head_module=dict(type='YOLOWorldHeadModule',
+                                    use_bn_head=True,
+                                    embed_dims=text_channels,
+                                    num_classes=num_training_classes)),
+    train_cfg=dict(assigner=dict(num_classes=num_training_classes)))
+
+# dataset settings
+text_transform = [
+    dict(type='RandomLoadText',
+         num_neg_samples=(num_classes, num_classes),
+         max_num_samples=num_training_classes,
+         padding_to_max=True,
+         padding_value=''),
+    dict(type='mmdet.PackDetInputs',
+         meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape', 'flip',
+                    'flip_direction', 'texts'))
+]
+mosaic_affine_transform = [
+    dict(
+        type='MultiModalMosaic',
+        img_scale=_base_.img_scale,
+        pad_val=114.0,
+        pre_transform=_base_.pre_transform),
+    dict(type='YOLOv5CopyPaste', prob=copypaste_prob),
+    dict(
+        type='YOLOv5RandomAffine',
+        max_rotate_degree=0.0,
+        max_shear_degree=0.0,
+        max_aspect_ratio=100.,
+        scaling_ratio_range=(1 - _base_.affine_scale,
+                             1 + _base_.affine_scale),
+        # img_scale is (width, height)
+        border=(-_base_.img_scale[0] // 2, -_base_.img_scale[1] // 2),
+        border_val=(114, 114, 114),
+        min_area_ratio=_base_.min_area_ratio,
+        use_mask_refine=_base_.use_mask2refine)
+]
+train_pipeline = [
+    *_base_.pre_transform,
+    *mosaic_affine_transform,
+    dict(
+        type='YOLOv5MultiModalMixUp',
+        prob=mixup_prob,
+        pre_transform=[*_base_.pre_transform,
+                       *mosaic_affine_transform]),
+    *_base_.last_transform[:-1],
+    *text_transform
+]
+train_pipeline_stage2 = [
+    *_base_.train_pipeline_stage2[:-1],
+    *text_transform
+]
+coco_train_dataset = dict(
+    _delete_=True,
+    type='MultiModalDataset',
+    dataset=dict(
+        type='YOLOv5CocoDataset',
+        data_root='data/coco',
+        ann_file='annotations/instances_train2017.json',
+        data_prefix=dict(img='train2017/'),
+        filter_cfg=dict(filter_empty_gt=False, min_size=32)),
+    class_text_path='data/texts/coco_class_texts.json',
+    pipeline=train_pipeline)
+
+train_dataloader = dict(
+    persistent_workers=persistent_workers,
+    batch_size=train_batch_size_per_gpu,
+    collate_fn=dict(type='yolow_collate'),
+    dataset=coco_train_dataset)
+test_pipeline = [
+    *_base_.test_pipeline[:-1],
+    dict(type='LoadText'),
+    dict(
+        type='mmdet.PackDetInputs',
+        meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape',
+                   'scale_factor', 'pad_param', 'texts'))
+]
+coco_val_dataset = dict(
+    _delete_=True,
+    type='MultiModalDataset',
+    dataset=dict(
+        type='YOLOv5CocoDataset',
+        data_root='data/coco',
+        ann_file='annotations/instances_val2017.json',
+        data_prefix=dict(img='val2017/'),
+        filter_cfg=dict(filter_empty_gt=False, min_size=32)),
+    class_text_path='data/texts/coco_class_texts.json',
+    pipeline=test_pipeline)
+val_dataloader = dict(dataset=coco_val_dataset)
+test_dataloader = val_dataloader
+# training settings
+default_hooks = dict(
+    param_scheduler=dict(
+        scheduler_type='linear',
+        lr_factor=0.01,
+        max_epochs=max_epochs),
+    checkpoint=dict(
+        max_keep_ckpts=-1,
+        save_best=None,
+        interval=save_epoch_intervals))
+custom_hooks = [
+    dict(
+        type='EMAHook',
+        ema_type='ExpMomentumEMA',
+        momentum=0.0001,
+        update_buffers=True,
+        strict_load=False,
+        priority=49),
+    dict(
+        type='mmdet.PipelineSwitchHook',
+        switch_epoch=max_epochs - close_mosaic_epochs,
+        switch_pipeline=train_pipeline_stage2)
+]
+train_cfg = dict(
+    max_epochs=max_epochs,
+    val_interval=5,
+    dynamic_intervals=[((max_epochs - close_mosaic_epochs),
+                        _base_.val_interval_stage2)])
+optim_wrapper = dict(
+    optimizer=dict(
+        _delete_=True,
+        type='AdamW',
+        lr=base_lr,
+        weight_decay=weight_decay,
+        batch_size_per_gpu=train_batch_size_per_gpu),
+    paramwise_cfg=dict(
+        custom_keys={'backbone.text_model': dict(lr_mult=0.01),
+                     'logit_scale': dict(weight_decay=0.0)}),
+    constructor='YOLOWv5OptimizerConstructor')
+
+# evaluation settings
+val_evaluator = dict(
+    _delete_=True,
+    type='mmdet.CocoMetric',
+    proposal_nums=(100, 1, 10),
+    ann_file='data/coco/annotations/instances_val2017.json',
+    metric='bbox')

configs/finetune_coco/yolo_world_v2_x_vlpan_bn_2e-4_80e_8gpus_mask-refine_finetune_coco.py

@@ -0,0 +1,183 @@
+_base_ = (
+    '../../third_party/mmyolo/configs/yolov8/'
+    'yolov8_x_mask-refine_syncbn_fast_8xb16-500e_coco.py')
+custom_imports = dict(
+    imports=['yolo_world'],
+    allow_failed_imports=False)
+
+# hyper-parameters
+num_classes = 80
+num_training_classes = 80
+max_epochs = 80  # Maximum training epochs
+close_mosaic_epochs = 10
+save_epoch_intervals = 5
+text_channels = 512
+neck_embed_channels = [128, 256, _base_.last_stage_out_channels // 2]
+neck_num_heads = [4, 8, _base_.last_stage_out_channels // 2 // 32]
+base_lr = 2e-4
+weight_decay = 0.05
+train_batch_size_per_gpu = 16
+load_from = 'pretrained_models/yolo_world_x_clip_t2i_bn_2e-3adamw_32xb16-100e_obj365v1_goldg_cc250k_train_lviseval-8698fbfa.pth'
+text_model_name = '../pretrained_models/clip-vit-base-patch32-projection'
+text_model_name = 'openai/clip-vit-base-patch32'
+persistent_workers = False
+
+# model settings
+model = dict(
+    type='YOLOWorldDetector',
+    mm_neck=True,
+    num_train_classes=num_training_classes,
+    num_test_classes=num_classes,
+    data_preprocessor=dict(type='YOLOWDetDataPreprocessor'),
+    backbone=dict(
+        _delete_=True,
+        type='MultiModalYOLOBackbone',
+        image_model={{_base_.model.backbone}},
+        text_model=dict(
+            type='HuggingCLIPLanguageBackbone',
+            model_name=text_model_name,
+            frozen_modules=['all'])),
+    neck=dict(type='YOLOWorldPAFPN',
+              guide_channels=text_channels,
+              embed_channels=neck_embed_channels,
+              num_heads=neck_num_heads,
+              block_cfg=dict(type='MaxSigmoidCSPLayerWithTwoConv')),
+    bbox_head=dict(type='YOLOWorldHead',
+                   head_module=dict(type='YOLOWorldHeadModule',
+                                    use_bn_head=True,
+                                    embed_dims=text_channels,
+                                    num_classes=num_training_classes)),
+    train_cfg=dict(assigner=dict(num_classes=num_training_classes)))
+
+# dataset settings
+text_transform = [
+    dict(type='RandomLoadText',
+         num_neg_samples=(num_classes, num_classes),
+         max_num_samples=num_training_classes,
+         padding_to_max=True,
+         padding_value=''),
+    dict(type='mmdet.PackDetInputs',
+         meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape', 'flip',
+                    'flip_direction', 'texts'))
+]
+mosaic_affine_transform = [
+    dict(
+        type='MultiModalMosaic',
+        img_scale=_base_.img_scale,
+        pad_val=114.0,
+        pre_transform=_base_.pre_transform),
+    dict(type='YOLOv5CopyPaste', prob=_base_.copypaste_prob),
+    dict(
+        type='YOLOv5RandomAffine',
+        max_rotate_degree=0.0,
+        max_shear_degree=0.0,
+        max_aspect_ratio=100.,
+        scaling_ratio_range=(1 - _base_.affine_scale,
+                             1 + _base_.affine_scale),
+        # img_scale is (width, height)
+        border=(-_base_.img_scale[0] // 2, -_base_.img_scale[1] // 2),
+        border_val=(114, 114, 114),
+        min_area_ratio=_base_.min_area_ratio,
+        use_mask_refine=_base_.use_mask2refine)
+]
+train_pipeline = [
+    *_base_.pre_transform,
+    *mosaic_affine_transform,
+    dict(
+        type='YOLOv5MultiModalMixUp',
+        prob=_base_.mixup_prob,
+        pre_transform=[*_base_.pre_transform,
+                       *mosaic_affine_transform]),
+    *_base_.last_transform[:-1],
+    *text_transform
+]
+train_pipeline_stage2 = [
+    *_base_.train_pipeline_stage2[:-1],
+    *text_transform
+]
+coco_train_dataset = dict(
+    _delete_=True,
+    type='MultiModalDataset',
+    dataset=dict(
+        type='YOLOv5CocoDataset',
+        data_root='data/coco',
+        ann_file='annotations/instances_train2017.json',
+        data_prefix=dict(img='train2017/'),
+        filter_cfg=dict(filter_empty_gt=False, min_size=32)),
+    class_text_path='data/texts/coco_class_texts.json',
+    pipeline=train_pipeline)
+
+train_dataloader = dict(
+    persistent_workers=persistent_workers,
+    batch_size=train_batch_size_per_gpu,
+    collate_fn=dict(type='yolow_collate'),
+    dataset=coco_train_dataset)
+
+test_pipeline = [
+    *_base_.test_pipeline[:-1],
+    dict(type='LoadText'),
+    dict(
+        type='mmdet.PackDetInputs',
+        meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape',
+                   'scale_factor', 'pad_param', 'texts'))
+]
+
+coco_val_dataset = dict(
+    _delete_=True,
+    type='MultiModalDataset',
+    dataset=dict(
+        type='YOLOv5CocoDataset',
+        data_root='data/coco',
+        ann_file='annotations/instances_val2017.json',
+        data_prefix=dict(img='val2017/'),
+        filter_cfg=dict(filter_empty_gt=False, min_size=32)),
+    class_text_path='data/texts/coco_class_texts.json',
+    pipeline=test_pipeline)
+val_dataloader = dict(dataset=coco_val_dataset)
+test_dataloader = val_dataloader
+# training settings
+default_hooks = dict(
+    param_scheduler=dict(
+        scheduler_type='linear',
+        lr_factor=0.01,
+        max_epochs=max_epochs),
+    checkpoint=dict(
+        max_keep_ckpts=-1,
+        save_best=None,
+        interval=save_epoch_intervals))
+custom_hooks = [
+    dict(
+        type='EMAHook',
+        ema_type='ExpMomentumEMA',
+        momentum=0.0001,
+        update_buffers=True,
+        strict_load=False,
+        priority=49),
+    dict(
+        type='mmdet.PipelineSwitchHook',
+        switch_epoch=max_epochs - close_mosaic_epochs,
+        switch_pipeline=train_pipeline_stage2)
+]
+train_cfg = dict(
+    max_epochs=max_epochs,
+    val_interval=5,
+    dynamic_intervals=[((max_epochs - close_mosaic_epochs),
+                        _base_.val_interval_stage2)])
+optim_wrapper = dict(
+    optimizer=dict(
+        _delete_=True,
+        type='AdamW',
+        lr=base_lr,
+        weight_decay=weight_decay,
+        batch_size_per_gpu=train_batch_size_per_gpu),
+    paramwise_cfg=dict(
+        custom_keys={'backbone.text_model': dict(lr_mult=0.01),
+                     'logit_scale': dict(weight_decay=0.0)}),
+    constructor='YOLOWv5OptimizerConstructor')
+# evaluation settings
+val_evaluator = dict(
+    _delete_=True,
+    type='mmdet.CocoMetric',
+    proposal_nums=(100, 1, 10),
+    ann_file='data/coco/annotations/instances_val2017.json',
+    metric='bbox')

configs/finetune_coco/yolo_world_v2_xl_vlpan_bn_2e-4_80e_8gpus_mask-refine_finetune_coco.py

@@ -0,0 +1,173 @@
+_base_ = ('../../third_party/mmyolo/configs/yolov8/'
+          'yolov8_x_mask-refine_syncbn_fast_8xb16-500e_coco.py')
+custom_imports = dict(imports=['yolo_world'], allow_failed_imports=False)
+
+# hyper-parameters
+num_classes = 80
+num_training_classes = 80
+max_epochs = 80  # Maximum training epochs
+close_mosaic_epochs = 10
+save_epoch_intervals = 5
+text_channels = 512
+neck_embed_channels = [128, 256, _base_.last_stage_out_channels // 2]
+neck_num_heads = [4, 8, _base_.last_stage_out_channels // 2 // 32]
+base_lr = 2e-4
+weight_decay = 0.05
+train_batch_size_per_gpu = 16
+text_model_name = '../pretrained_models/clip-vit-base-patch32-projection'
+text_model_name = 'openai/clip-vit-base-patch32'
+persistent_workers = False
+
+# scaling model from X to XL
+deepen_factor = 1.0
+widen_factor = 1.5
+
+backbone = _base_.model.backbone
+backbone.update(deepen_factor=deepen_factor, widen_factor=widen_factor)
+
+# model settings
+model = dict(type='YOLOWorldDetector',
+             mm_neck=True,
+             num_train_classes=num_training_classes,
+             num_test_classes=num_classes,
+             data_preprocessor=dict(type='YOLOWDetDataPreprocessor'),
+             backbone=dict(_delete_=True,
+                           type='MultiModalYOLOBackbone',
+                           image_model=backbone,
+                           text_model=dict(type='HuggingCLIPLanguageBackbone',
+                                           model_name=text_model_name,
+                                           frozen_modules=['all'])),
+             neck=dict(type='YOLOWorldPAFPN',
+                       deepen_factor=deepen_factor,
+                       widen_factor=widen_factor,
+                       guide_channels=text_channels,
+                       embed_channels=neck_embed_channels,
+                       num_heads=neck_num_heads,
+                       block_cfg=dict(type='MaxSigmoidCSPLayerWithTwoConv')),
+             bbox_head=dict(type='YOLOWorldHead',
+                            head_module=dict(
+                                type='YOLOWorldHeadModule',
+                                widen_factor=widen_factor,
+                                use_bn_head=True,
+                                embed_dims=text_channels,
+                                num_classes=num_training_classes)),
+             train_cfg=dict(assigner=dict(num_classes=num_training_classes)))
+
+# dataset settings
+text_transform = [
+    dict(type='RandomLoadText',
+         num_neg_samples=(num_classes, num_classes),
+         max_num_samples=num_training_classes,
+         padding_to_max=True,
+         padding_value=''),
+    dict(type='mmdet.PackDetInputs',
+         meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape', 'flip',
+                    'flip_direction', 'texts'))
+]
+mosaic_affine_transform = [
+    dict(type='MultiModalMosaic',
+         img_scale=_base_.img_scale,
+         pad_val=114.0,
+         pre_transform=_base_.pre_transform),
+    dict(type='YOLOv5CopyPaste', prob=_base_.copypaste_prob),
+    dict(
+        type='YOLOv5RandomAffine',
+        max_rotate_degree=0.0,
+        max_shear_degree=0.0,
+        max_aspect_ratio=100.,
+        scaling_ratio_range=(1 - _base_.affine_scale, 1 + _base_.affine_scale),
+        # img_scale is (width, height)
+        border=(-_base_.img_scale[0] // 2, -_base_.img_scale[1] // 2),
+        border_val=(114, 114, 114),
+        min_area_ratio=_base_.min_area_ratio,
+        use_mask_refine=_base_.use_mask2refine)
+]
+train_pipeline = [
+    *_base_.pre_transform, *mosaic_affine_transform,
+    dict(type='YOLOv5MultiModalMixUp',
+         prob=_base_.mixup_prob,
+         pre_transform=[*_base_.pre_transform, *mosaic_affine_transform]),
+    *_base_.last_transform[:-1], *text_transform
+]
+train_pipeline_stage2 = [*_base_.train_pipeline_stage2[:-1], *text_transform]
+coco_train_dataset = dict(_delete_=True,
+                          type='MultiModalDataset',
+                          dataset=dict(
+                              type='YOLOv5CocoDataset',
+                              data_root='data/coco',
+                              ann_file='annotations/instances_train2017.json',
+                              data_prefix=dict(img='train2017/'),
+                              filter_cfg=dict(filter_empty_gt=False,
+                                              min_size=32)),
+                          class_text_path='data/texts/coco_class_texts.json',
+                          pipeline=train_pipeline)
+
+train_dataloader = dict(persistent_workers=persistent_workers,
+                        batch_size=train_batch_size_per_gpu,
+                        collate_fn=dict(type='yolow_collate'),
+                        dataset=coco_train_dataset)
+
+test_pipeline = [
+    *_base_.test_pipeline[:-1],
+    dict(type='LoadText'),
+    dict(type='mmdet.PackDetInputs',
+         meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape',
+                    'scale_factor', 'pad_param', 'texts'))
+]
+
+coco_val_dataset = dict(
+    _delete_=True,
+    type='MultiModalDataset',
+    dataset=dict(type='YOLOv5CocoDataset',
+                 data_root='data/coco',
+                 ann_file='annotations/instances_val2017.json',
+                 data_prefix=dict(img='val2017/'),
+                 filter_cfg=dict(filter_empty_gt=False, min_size=32)),
+    class_text_path='data/texts/coco_class_texts.json',
+    pipeline=test_pipeline)
+val_dataloader = dict(dataset=coco_val_dataset)
+test_dataloader = val_dataloader
+# training settings
+default_hooks = dict(param_scheduler=dict(scheduler_type='linear',
+                                          lr_factor=0.01,
+                                          max_epochs=max_epochs),
+                     checkpoint=dict(max_keep_ckpts=-1,
+                                     save_best=None,
+                                     interval=save_epoch_intervals))
+custom_hooks = [
+    dict(type='EMAHook',
+         ema_type='ExpMomentumEMA',
+         momentum=0.0001,
+         update_buffers=True,
+         strict_load=False,
+         priority=49),
+    dict(type='mmdet.PipelineSwitchHook',
+         switch_epoch=max_epochs - close_mosaic_epochs,
+         switch_pipeline=train_pipeline_stage2)
+]
+train_cfg = dict(max_epochs=max_epochs,
+                 val_interval=5,
+                 dynamic_intervals=[((max_epochs - close_mosaic_epochs),
+                                     _base_.val_interval_stage2)])
+optim_wrapper = dict(optimizer=dict(
+    _delete_=True,
+    type='AdamW',
+    lr=base_lr,
+    weight_decay=weight_decay,
+    batch_size_per_gpu=train_batch_size_per_gpu),
+                     paramwise_cfg=dict(bias_decay_mult=0.0,
+                                        norm_decay_mult=0.0,
+                                        custom_keys={
+                                            'backbone.text_model':
+                                            dict(lr_mult=0.01),
+                                            'logit_scale':
+                                            dict(weight_decay=0.0)
+                                        }),
+                     constructor='YOLOWv5OptimizerConstructor')
+
+# evaluation settings
+val_evaluator = dict(_delete_=True,
+                     type='mmdet.CocoMetric',
+                     proposal_nums=(100, 1, 10),
+                     ann_file='data/coco/annotations/instances_val2017.json',
+                     metric='bbox')

configs/pretrain/yolo_world_v2_l_clip_large_vlpan_bn_2e-3_100e_4x8gpus_obj365v1_goldg_train_800ft_lvis_minival.py

@@ -0,0 +1,200 @@
+_base_ = ('../../third_party/mmyolo/configs/yolov8/'
+          'yolov8_l_syncbn_fast_8xb16-500e_coco.py')
+custom_imports = dict(imports=['yolo_world'],
+                      allow_failed_imports=False)
+
+# hyper-parameters
+num_classes = 1203
+num_training_classes = 80
+max_epochs = 100  # Maximum training epochs
+close_mosaic_epochs = 2
+save_epoch_intervals = 2
+text_channels = 768
+neck_embed_channels = [128, 256, _base_.last_stage_out_channels // 2]
+neck_num_heads = [4, 8, _base_.last_stage_out_channels // 2 // 32]
+base_lr = 2e-3
+weight_decay = 0.0125
+train_batch_size_per_gpu = 16
+# text_model_name = '../pretrained_models/clip-vit-large-patch14-336'
+text_model_name = 'openai/clip-vit-large-patch14-336'
+img_scale = (800, 800)
+
+# model settings
+model = dict(
+    type='YOLOWorldDetector',
+    mm_neck=True,
+    num_train_classes=num_training_classes,
+    num_test_classes=num_classes,
+    data_preprocessor=dict(type='YOLOWDetDataPreprocessor'),
+    backbone=dict(
+        _delete_=True,
+        type='MultiModalYOLOBackbone',
+        image_model={{_base_.model.backbone}},
+        text_model=dict(
+            type='HuggingCLIPLanguageBackbone',
+            model_name=text_model_name,
+            frozen_modules=['all'])),
+    neck=dict(type='YOLOWorldPAFPN',
+              guide_channels=text_channels,
+              embed_channels=neck_embed_channels,
+              num_heads=neck_num_heads,
+              block_cfg=dict(type='MaxSigmoidCSPLayerWithTwoConv')),
+    bbox_head=dict(type='YOLOWorldHead',
+                   head_module=dict(type='YOLOWorldHeadModule',
+                                    use_bn_head=True,
+                                    embed_dims=text_channels,
+                                    num_classes=num_training_classes)),
+    train_cfg=dict(assigner=dict(num_classes=num_training_classes)))
+
+# dataset settings
+text_transform = [
+    dict(type='RandomLoadText',
+         num_neg_samples=(num_classes, num_classes),
+         max_num_samples=num_training_classes,
+         padding_to_max=True,
+         padding_value=''),
+    dict(type='mmdet.PackDetInputs',
+         meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape', 'flip',
+                    'flip_direction', 'texts'))
+]
+train_pipeline = [
+    *_base_.pre_transform,
+    dict(type='MultiModalMosaic',
+         img_scale=img_scale,
+         pad_val=114.0,
+         pre_transform=_base_.pre_transform),
+    dict(
+        type='YOLOv5RandomAffine',
+        max_rotate_degree=0.0,
+        max_shear_degree=0.0,
+        scaling_ratio_range=(1 - _base_.affine_scale, 1 + _base_.affine_scale),
+        max_aspect_ratio=_base_.max_aspect_ratio,
+        border=(-img_scale[0] // 2, -img_scale[1] // 2),
+        border_val=(114, 114, 114)),
+    *_base_.last_transform[:-1],
+    *text_transform,
+]
+
+train_pipeline_stage2 = [
+    *_base_.pre_transform,
+    dict(type='YOLOv5KeepRatioResize', scale=img_scale),
+    dict(
+        type='LetterResize',
+        scale=img_scale,
+        allow_scale_up=True,
+        pad_val=dict(img=114.0)),
+    dict(
+        type='YOLOv5RandomAffine',
+        max_rotate_degree=0.0,
+        max_shear_degree=0.0,
+        scaling_ratio_range=(1 - _base_.affine_scale, 1 + _base_.affine_scale),
+        max_aspect_ratio=_base_.max_aspect_ratio,
+        border_val=(114, 114, 114)),
+    *_base_.last_transform[:-1],
+    *text_transform
+]
+
+obj365v1_train_dataset = dict(
+    type='MultiModalDataset',
+    dataset=dict(
+        type='YOLOv5Objects365V1Dataset',
+        data_root='data/objects365v1/',
+        ann_file='annotations/objects365_train.json',
+        data_prefix=dict(img='train/'),
+        filter_cfg=dict(filter_empty_gt=False, min_size=32)),
+    class_text_path='data/texts/obj365v1_class_texts.json',
+    pipeline=train_pipeline)
+
+mg_train_dataset = dict(type='YOLOv5MixedGroundingDataset',
+                        data_root='data/mixed_grounding/',
+                        ann_file='annotations/final_mixed_train_no_coco.json',
+                        data_prefix=dict(img='gqa/images/'),
+                        filter_cfg=dict(filter_empty_gt=False, min_size=32),
+                        pipeline=train_pipeline)
+
+flickr_train_dataset = dict(
+    type='YOLOv5MixedGroundingDataset',
+    data_root='data/flickr/',
+    ann_file='annotations/final_flickr_separateGT_train.json',
+    data_prefix=dict(img='full_images/'),
+    filter_cfg=dict(filter_empty_gt=True, min_size=32),
+    pipeline=train_pipeline)
+
+train_dataloader = dict(batch_size=train_batch_size_per_gpu,
+                        collate_fn=dict(type='yolow_collate'),
+                        dataset=dict(_delete_=True,
+                                     type='ConcatDataset',
+                                     datasets=[
+                                         obj365v1_train_dataset,
+                                         flickr_train_dataset, mg_train_dataset
+                                     ],
+                                     ignore_keys=['classes', 'palette']))
+
+test_pipeline = [
+    dict(type='LoadImageFromFile'),
+    dict(type='YOLOv5KeepRatioResize', scale=img_scale),
+    dict(
+        type='LetterResize',
+        scale=img_scale,
+        allow_scale_up=False,
+        pad_val=dict(img=114)),
+    dict(type='LoadAnnotations', with_bbox=True, _scope_='mmdet'),
+    dict(type='LoadText'),
+    dict(type='mmdet.PackDetInputs',
+         meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape',
+                    'scale_factor', 'pad_param', 'texts'))
+]
+
+coco_val_dataset = dict(
+    _delete_=True,
+    type='MultiModalDataset',
+    dataset=dict(type='YOLOv5LVISV1Dataset',
+                 data_root='data/coco/',
+                 test_mode=True,
+                 ann_file='lvis/lvis_v1_minival_inserted_image_name.json',
+                 data_prefix=dict(img=''),
+                 batch_shapes_cfg=None),
+    class_text_path='data/texts/lvis_v1_class_texts.json',
+    pipeline=test_pipeline)
+val_dataloader = dict(dataset=coco_val_dataset)
+test_dataloader = val_dataloader
+
+val_evaluator = dict(type='mmdet.LVISMetric',
+                     ann_file='data/coco/lvis/lvis_v1_minival_inserted_image_name.json',
+                     metric='bbox')
+test_evaluator = val_evaluator
+
+# training settings
+default_hooks = dict(param_scheduler=dict(max_epochs=max_epochs),
+                     checkpoint=dict(interval=save_epoch_intervals,
+                                     rule='greater'))
+custom_hooks = [
+    dict(type='EMAHook',
+         ema_type='ExpMomentumEMA',
+         momentum=0.0001,
+         update_buffers=True,
+         strict_load=False,
+         priority=49),
+    dict(type='mmdet.PipelineSwitchHook',
+         switch_epoch=max_epochs - close_mosaic_epochs,
+         switch_pipeline=train_pipeline_stage2)
+]
+train_cfg = dict(max_epochs=max_epochs,
+                 val_interval=10,
+                 dynamic_intervals=[((max_epochs - close_mosaic_epochs),
+                                     _base_.val_interval_stage2)])
+optim_wrapper = dict(optimizer=dict(
+    _delete_=True,
+    type='AdamW',
+    lr=base_lr,
+    weight_decay=weight_decay,
+    batch_size_per_gpu=train_batch_size_per_gpu),
+                     paramwise_cfg=dict(bias_decay_mult=0.0,
+                                        norm_decay_mult=0.0,
+                                        custom_keys={
+                                            'backbone.text_model':
+                                            dict(lr_mult=0.01),
+                                            'logit_scale':
+                                            dict(weight_decay=0.0)
+                                        }),
+                     constructor='YOLOWv5OptimizerConstructor')

configs/pretrain/yolo_world_v2_l_clip_large_vlpan_bn_2e-3_100e_4x8gpus_obj365v1_goldg_train_lvis_minival.py

@@ -0,0 +1,171 @@
+_base_ = ('../../third_party/mmyolo/configs/yolov8/'
+          'yolov8_l_syncbn_fast_8xb16-500e_coco.py')
+custom_imports = dict(imports=['yolo_world'],
+                      allow_failed_imports=False)
+
+# hyper-parameters
+num_classes = 1203
+num_training_classes = 80
+max_epochs = 100  # Maximum training epochs
+close_mosaic_epochs = 2
+save_epoch_intervals = 2
+text_channels = 768
+neck_embed_channels = [128, 256, _base_.last_stage_out_channels // 2]
+neck_num_heads = [4, 8, _base_.last_stage_out_channels // 2 // 32]
+base_lr = 2e-3
+weight_decay = 0.0125
+train_batch_size_per_gpu = 16
+# text_model_name = '../pretrained_models/clip-vit-large-patch14-336'
+text_model_name = 'openai/clip-vit-large-patch14-336'
+# model settings
+model = dict(
+    type='YOLOWorldDetector',
+    mm_neck=True,
+    num_train_classes=num_training_classes,
+    num_test_classes=num_classes,
+    data_preprocessor=dict(type='YOLOWDetDataPreprocessor'),
+    backbone=dict(
+        _delete_=True,
+        type='MultiModalYOLOBackbone',
+        image_model={{_base_.model.backbone}},
+        text_model=dict(
+            type='HuggingCLIPLanguageBackbone',
+            model_name=text_model_name,
+            frozen_modules=['all'])),
+    neck=dict(type='YOLOWorldPAFPN',
+              guide_channels=text_channels,
+              embed_channels=neck_embed_channels,
+              num_heads=neck_num_heads,
+              block_cfg=dict(type='MaxSigmoidCSPLayerWithTwoConv')),
+    bbox_head=dict(type='YOLOWorldHead',
+                   head_module=dict(type='YOLOWorldHeadModule',
+                                    use_bn_head=True,
+                                    embed_dims=text_channels,
+                                    num_classes=num_training_classes)),
+    train_cfg=dict(assigner=dict(num_classes=num_training_classes)))
+
+# dataset settings
+text_transform = [
+    dict(type='RandomLoadText',
+         num_neg_samples=(num_classes, num_classes),
+         max_num_samples=num_training_classes,
+         padding_to_max=True,
+         padding_value=''),
+    dict(type='mmdet.PackDetInputs',
+         meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape', 'flip',
+                    'flip_direction', 'texts'))
+]
+train_pipeline = [
+    *_base_.pre_transform,
+    dict(type='MultiModalMosaic',
+         img_scale=_base_.img_scale,
+         pad_val=114.0,
+         pre_transform=_base_.pre_transform),
+    dict(
+        type='YOLOv5RandomAffine',
+        max_rotate_degree=0.0,
+        max_shear_degree=0.0,
+        scaling_ratio_range=(1 - _base_.affine_scale, 1 + _base_.affine_scale),
+        max_aspect_ratio=_base_.max_aspect_ratio,
+        border=(-_base_.img_scale[0] // 2, -_base_.img_scale[1] // 2),
+        border_val=(114, 114, 114)),
+    *_base_.last_transform[:-1],
+    *text_transform,
+]
+train_pipeline_stage2 = [*_base_.train_pipeline_stage2[:-1], *text_transform]
+obj365v1_train_dataset = dict(
+    type='MultiModalDataset',
+    dataset=dict(
+        type='YOLOv5Objects365V1Dataset',
+        data_root='data/objects365v1/',
+        ann_file='annotations/objects365_train.json',
+        data_prefix=dict(img='train/'),
+        filter_cfg=dict(filter_empty_gt=False, min_size=32)),
+    class_text_path='data/texts/obj365v1_class_texts.json',
+    pipeline=train_pipeline)
+
+mg_train_dataset = dict(type='YOLOv5MixedGroundingDataset',
+                        data_root='data/mixed_grounding/',
+                        ann_file='annotations/final_mixed_train_no_coco.json',
+                        data_prefix=dict(img='gqa/images/'),
+                        filter_cfg=dict(filter_empty_gt=False, min_size=32),
+                        pipeline=train_pipeline)
+
+flickr_train_dataset = dict(
+    type='YOLOv5MixedGroundingDataset',
+    data_root='data/flickr/',
+    ann_file='annotations/final_flickr_separateGT_train.json',
+    data_prefix=dict(img='full_images/'),
+    filter_cfg=dict(filter_empty_gt=True, min_size=32),
+    pipeline=train_pipeline)
+
+train_dataloader = dict(batch_size=train_batch_size_per_gpu,
+                        collate_fn=dict(type='yolow_collate'),
+                        dataset=dict(_delete_=True,
+                                     type='ConcatDataset',
+                                     datasets=[
+                                         obj365v1_train_dataset,
+                                         flickr_train_dataset, mg_train_dataset
+                                     ],
+                                     ignore_keys=['classes', 'palette']))
+
+test_pipeline = [
+    *_base_.test_pipeline[:-1],
+    dict(type='LoadText'),
+    dict(type='mmdet.PackDetInputs',
+         meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape',
+                    'scale_factor', 'pad_param', 'texts'))
+]
+coco_val_dataset = dict(
+    _delete_=True,
+    type='MultiModalDataset',
+    dataset=dict(type='YOLOv5LVISV1Dataset',
+                 data_root='data/coco/',
+                 test_mode=True,
+                 ann_file='lvis/lvis_v1_minival_inserted_image_name.json',
+                 data_prefix=dict(img=''),
+                 batch_shapes_cfg=None),
+    class_text_path='data/texts/lvis_v1_class_texts.json',
+    pipeline=test_pipeline)
+val_dataloader = dict(dataset=coco_val_dataset)
+test_dataloader = val_dataloader
+
+val_evaluator = dict(type='mmdet.LVISMetric',
+                     ann_file='data/coco/lvis/lvis_v1_minival_inserted_image_name.json',
+                     metric='bbox')
+test_evaluator = val_evaluator
+
+# training settings
+default_hooks = dict(param_scheduler=dict(max_epochs=max_epochs),
+                     checkpoint=dict(interval=save_epoch_intervals,
+                                     rule='greater'))
+custom_hooks = [
+    dict(type='EMAHook',
+         ema_type='ExpMomentumEMA',
+         momentum=0.0001,
+         update_buffers=True,
+         strict_load=False,
+         priority=49),
+    dict(type='mmdet.PipelineSwitchHook',
+         switch_epoch=max_epochs - close_mosaic_epochs,
+         switch_pipeline=train_pipeline_stage2)
+]
+train_cfg = dict(max_epochs=max_epochs,
+                 val_interval=10,
+                 dynamic_intervals=[((max_epochs - close_mosaic_epochs),
+                                     _base_.val_interval_stage2)])
+optim_wrapper = dict(optimizer=dict(
+    _delete_=True,
+    type='AdamW',
+    lr=base_lr,
+    weight_decay=weight_decay,
+    batch_size_per_gpu=train_batch_size_per_gpu),
+                     paramwise_cfg=dict(bias_decay_mult=0.0,
+                                        norm_decay_mult=0.0,
+                                        custom_keys={
+                                            'backbone.text_model':
+                                            dict(lr_mult=0.01),
+                                            'logit_scale':
+                                            dict(weight_decay=0.0)
+                                        }),
+                     constructor='YOLOWv5OptimizerConstructor')

configs/pretrain/yolo_world_v2_l_vlpan_bn_2e-3_100e_4x8gpus_obj365v1_goldg_train_1280ft_lvis_minival.py

@@ -0,0 +1,202 @@
+_base_ = ('../../third_party/mmyolo/configs/yolov8/'
+          'yolov8_l_syncbn_fast_8xb16-500e_coco.py')
+custom_imports = dict(imports=['yolo_world'],
+                      allow_failed_imports=False)
+
+# hyper-parameters
+num_classes = 1203
+num_training_classes = 80
+max_epochs = 20  # Maximum training epochs
+close_mosaic_epochs = 2
+save_epoch_intervals = 2
+text_channels = 512
+neck_embed_channels = [128, 256, _base_.last_stage_out_channels // 2]
+neck_num_heads = [4, 8, _base_.last_stage_out_channels // 2 // 32]
+base_lr = 2e-4
+weight_decay = 0.025
+train_batch_size_per_gpu = 4
+load_from = "pretrained_models/yolo_world_v2_l_obj365v1_goldg_pretrain-a82b1fe3.pth"
+# text_model_name = '../pretrained_models/clip-vit-base-patch32-projection'
+text_model_name = 'openai/clip-vit-base-patch32'
+img_scale = (1280, 1280)
+
+# model settings
+model = dict(
+    type='YOLOWorldDetector',
+    mm_neck=True,
+    num_train_classes=num_training_classes,
+    num_test_classes=num_classes,
+    data_preprocessor=dict(type='YOLOWDetDataPreprocessor'),
+    backbone=dict(
+        _delete_=True,
+        type='MultiModalYOLOBackbone',
+        image_model={{_base_.model.backbone}},
+        text_model=dict(
+            type='HuggingCLIPLanguageBackbone',
+            model_name=text_model_name,
+            frozen_modules=['all'])),
+    neck=dict(type='YOLOWorldPAFPN',
+              guide_channels=text_channels,
+              embed_channels=neck_embed_channels,
+              num_heads=neck_num_heads,
+              block_cfg=dict(type='MaxSigmoidCSPLayerWithTwoConv')),
+    bbox_head=dict(type='YOLOWorldHead',
+                   head_module=dict(type='YOLOWorldHeadModule',
+                                    use_bn_head=True,
+                                    embed_dims=text_channels,
+                                    num_classes=num_training_classes)),
+    train_cfg=dict(assigner=dict(num_classes=num_training_classes)))
+
+# dataset settings
+text_transform = [
+    dict(type='RandomLoadText',
+         num_neg_samples=(num_classes, num_classes),
+         max_num_samples=num_training_classes,
+         padding_to_max=True,
+         padding_value=''),
+    dict(type='mmdet.PackDetInputs',
+         meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape', 'flip',
+                    'flip_direction', 'texts'))
+]
+train_pipeline = [
+    *_base_.pre_transform,
+    dict(type='MultiModalMosaic',
+         img_scale=img_scale,
+         pad_val=114.0,
+         pre_transform=_base_.pre_transform),
+    dict(
+        type='YOLOv5RandomAffine',
+        max_rotate_degree=0.0,
+        max_shear_degree=0.0,
+        scaling_ratio_range=(1 - _base_.affine_scale, 1 + _base_.affine_scale),
+        max_aspect_ratio=_base_.max_aspect_ratio,
+        border=(-img_scale[0] // 2, -img_scale[1] // 2),
+        border_val=(114, 114, 114)),
+    *_base_.last_transform[:-1],
+    *text_transform,
+]
+
+train_pipeline_stage2 = [
+    *_base_.pre_transform,
+    dict(type='YOLOv5KeepRatioResize', scale=img_scale),
+    dict(
+        type='LetterResize',
+        scale=img_scale,
+        allow_scale_up=True,
+        pad_val=dict(img=114.0)),
+    dict(
+        type='YOLOv5RandomAffine',
+        max_rotate_degree=0.0,
+        max_shear_degree=0.0,
+        scaling_ratio_range=(1 - _base_.affine_scale, 1 + _base_.affine_scale),
+        max_aspect_ratio=_base_.max_aspect_ratio,
+        border_val=(114, 114, 114)),
+    *_base_.last_transform[:-1],
+    *text_transform
+]
+
+obj365v1_train_dataset = dict(
+    type='MultiModalDataset',
+    dataset=dict(
+        type='YOLOv5Objects365V1Dataset',
+        data_root='data/objects365v1/',
+        ann_file='annotations/objects365_train.json',
+        data_prefix=dict(img='train/'),
+        filter_cfg=dict(filter_empty_gt=False, min_size=32)),
+    class_text_path='data/texts/obj365v1_class_texts.json',
+    pipeline=train_pipeline)
+
+mg_train_dataset = dict(type='YOLOv5MixedGroundingDataset',
+                        data_root='data/mixed_grounding/',
+                        ann_file='annotations/final_mixed_train_no_coco.json',
+                        data_prefix=dict(img='gqa/images/'),
+                        filter_cfg=dict(filter_empty_gt=False, min_size=32),
+                        pipeline=train_pipeline)
+
+flickr_train_dataset = dict(
+    type='YOLOv5MixedGroundingDataset',
+    data_root='data/flickr/',
+    ann_file='annotations/final_flickr_separateGT_train.json',
+    data_prefix=dict(img='full_images/'),
+    filter_cfg=dict(filter_empty_gt=True, min_size=32),
+    pipeline=train_pipeline)
+
+train_dataloader = dict(batch_size=train_batch_size_per_gpu,
+                        collate_fn=dict(type='yolow_collate'),
+                        dataset=dict(_delete_=True,
+                                     type='ConcatDataset',
+                                     datasets=[
+                                         obj365v1_train_dataset,
+                                         flickr_train_dataset, mg_train_dataset
+                                     ],
+                                     ignore_keys=['classes', 'palette']))
+
+test_pipeline = [
+    dict(type='LoadImageFromFile'),
+    dict(type='YOLOv5KeepRatioResize', scale=img_scale),
+    dict(
+        type='LetterResize',
+        scale=img_scale,
+        allow_scale_up=False,
+        pad_val=dict(img=114)),
+    dict(type='LoadAnnotations', with_bbox=True, _scope_='mmdet'),
+    dict(type='LoadText'),
+    dict(type='mmdet.PackDetInputs',
+         meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape',
+                    'scale_factor', 'pad_param', 'texts'))
+]
+
+coco_val_dataset = dict(
+    _delete_=True,
+    type='MultiModalDataset',
+    dataset=dict(type='YOLOv5LVISV1Dataset',
+                 data_root='data/coco/',
+                 test_mode=True,
+                 ann_file='lvis/lvis_v1_minival_inserted_image_name.json',
+                 data_prefix=dict(img=''),
+                 batch_shapes_cfg=None),
+    class_text_path='data/texts/lvis_v1_class_texts.json',
+    pipeline=test_pipeline)
+val_dataloader = dict(dataset=coco_val_dataset)
+test_dataloader = val_dataloader
+
+val_evaluator = dict(type='mmdet.LVISMetric',
+                     ann_file='data/coco/lvis/lvis_v1_minival_inserted_image_name.json',
+                     metric='bbox')
+test_evaluator = val_evaluator
+
+# training settings
+default_hooks = dict(param_scheduler=dict(max_epochs=max_epochs),
+                     checkpoint=dict(interval=save_epoch_intervals,
+                                     rule='greater'))
+custom_hooks = [
+    dict(type='EMAHook',
+         ema_type='ExpMomentumEMA',
+         momentum=0.0001,
+         update_buffers=True,
+         strict_load=False,
+         priority=49),
+    dict(type='mmdet.PipelineSwitchHook',
+         switch_epoch=max_epochs - close_mosaic_epochs,
+         switch_pipeline=train_pipeline_stage2)
+]
+train_cfg = dict(max_epochs=max_epochs,
+                 val_interval=10,
+                 dynamic_intervals=[((max_epochs - close_mosaic_epochs),
+                                     _base_.val_interval_stage2)])
+
+optim_wrapper = dict(optimizer=dict(
+    _delete_=True,
+    type='AdamW',
+    lr=base_lr,
+    weight_decay=weight_decay,
+    batch_size_per_gpu=train_batch_size_per_gpu),
+                     paramwise_cfg=dict(bias_decay_mult=0.0,
+                                        norm_decay_mult=0.0,
+                                        custom_keys={
+                                            'backbone.text_model':
+                                            dict(lr_mult=0.01),
+                                            'logit_scale':
+                                            dict(weight_decay=0.0)
+                                        }),
+                     constructor='YOLOWv5OptimizerConstructor')

configs/pretrain/yolo_world_v2_l_vlpan_bn_2e-3_100e_4x8gpus_obj365v1_goldg_train_lvis_minival.py

@@ -0,0 +1,171 @@
+_base_ = ('../../third_party/mmyolo/configs/yolov8/'
+          'yolov8_l_syncbn_fast_8xb16-500e_coco.py')
+custom_imports = dict(imports=['yolo_world'],
+                      allow_failed_imports=False)
+
+# hyper-parameters
+num_classes = 1203
+num_training_classes = 80
+max_epochs = 100  # Maximum training epochs
+close_mosaic_epochs = 2
+save_epoch_intervals = 2
+text_channels = 512
+neck_embed_channels = [128, 256, _base_.last_stage_out_channels // 2]
+neck_num_heads = [4, 8, _base_.last_stage_out_channels // 2 // 32]
+base_lr = 2e-3
+weight_decay = 0.05 / 2
+train_batch_size_per_gpu = 16
+# text_model_name = '../pretrained_models/clip-vit-base-patch32-projection'
+text_model_name = 'openai/clip-vit-base-patch32'
+# model settings
+model = dict(
+    type='YOLOWorldDetector',
+    mm_neck=True,
+    num_train_classes=num_training_classes,
+    num_test_classes=num_classes,
+    data_preprocessor=dict(type='YOLOWDetDataPreprocessor'),
+    backbone=dict(
+        _delete_=True,
+        type='MultiModalYOLOBackbone',
+        image_model={{_base_.model.backbone}},
+        text_model=dict(
+            type='HuggingCLIPLanguageBackbone',
+            model_name=text_model_name,
+            frozen_modules=['all'])),
+    neck=dict(type='YOLOWorldPAFPN',
+              guide_channels=text_channels,
+              embed_channels=neck_embed_channels,
+              num_heads=neck_num_heads,
+              block_cfg=dict(type='MaxSigmoidCSPLayerWithTwoConv')),
+    bbox_head=dict(type='YOLOWorldHead',
+                   head_module=dict(type='YOLOWorldHeadModule',
+                                    use_bn_head=True,
+                                    embed_dims=text_channels,
+                                    num_classes=num_training_classes)),
+    train_cfg=dict(assigner=dict(num_classes=num_training_classes)))
+
+# dataset settings
+text_transform = [
+    dict(type='RandomLoadText',
+         num_neg_samples=(num_classes, num_classes),
+         max_num_samples=num_training_classes,
+         padding_to_max=True,
+         padding_value=''),
+    dict(type='mmdet.PackDetInputs',
+         meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape', 'flip',
+                    'flip_direction', 'texts'))
+]
+train_pipeline = [
+    *_base_.pre_transform,
+    dict(type='MultiModalMosaic',
+         img_scale=_base_.img_scale,
+         pad_val=114.0,
+         pre_transform=_base_.pre_transform),
+    dict(
+        type='YOLOv5RandomAffine',
+        max_rotate_degree=0.0,
+        max_shear_degree=0.0,
+        scaling_ratio_range=(1 - _base_.affine_scale, 1 + _base_.affine_scale),
+        max_aspect_ratio=_base_.max_aspect_ratio,
+        border=(-_base_.img_scale[0] // 2, -_base_.img_scale[1] // 2),
+        border_val=(114, 114, 114)),
+    *_base_.last_transform[:-1],
+    *text_transform,
+]
+train_pipeline_stage2 = [*_base_.train_pipeline_stage2[:-1], *text_transform]
+obj365v1_train_dataset = dict(
+    type='MultiModalDataset',
+    dataset=dict(
+        type='YOLOv5Objects365V1Dataset',
+        data_root='data/objects365v1/',
+        ann_file='annotations/objects365_train.json',
+        data_prefix=dict(img='train/'),
+        filter_cfg=dict(filter_empty_gt=False, min_size=32)),
+    class_text_path='data/texts/obj365v1_class_texts.json',
+    pipeline=train_pipeline)
+
+mg_train_dataset = dict(type='YOLOv5MixedGroundingDataset',
+                        data_root='data/mixed_grounding/',
+                        ann_file='annotations/final_mixed_train_no_coco.json',
+                        data_prefix=dict(img='gqa/images/'),
+                        filter_cfg=dict(filter_empty_gt=False, min_size=32),
+                        pipeline=train_pipeline)
+
+flickr_train_dataset = dict(
+    type='YOLOv5MixedGroundingDataset',
+    data_root='data/flickr/',
+    ann_file='annotations/final_flickr_separateGT_train.json',
+    data_prefix=dict(img='full_images/'),
+    filter_cfg=dict(filter_empty_gt=True, min_size=32),
+    pipeline=train_pipeline)
+
+train_dataloader = dict(batch_size=train_batch_size_per_gpu,
+                        collate_fn=dict(type='yolow_collate'),
+                        dataset=dict(_delete_=True,
+                                     type='ConcatDataset',
+                                     datasets=[
+                                         obj365v1_train_dataset,
+                                         flickr_train_dataset, mg_train_dataset
+                                     ],
+                                     ignore_keys=['classes', 'palette']))
+
+test_pipeline = [
+    *_base_.test_pipeline[:-1],
+    dict(type='LoadText'),
+    dict(type='mmdet.PackDetInputs',
+         meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape',
+                    'scale_factor', 'pad_param', 'texts'))
+]
+coco_val_dataset = dict(
+    _delete_=True,
+    type='MultiModalDataset',
+    dataset=dict(type='YOLOv5LVISV1Dataset',
+                 data_root='data/coco/',
+                 test_mode=True,
+                 ann_file='lvis/lvis_v1_minival_inserted_image_name.json',
+                 data_prefix=dict(img=''),
+                 batch_shapes_cfg=None),
+    class_text_path='data/texts/lvis_v1_class_texts.json',
+    pipeline=test_pipeline)
+val_dataloader = dict(dataset=coco_val_dataset)
+test_dataloader = val_dataloader
+
+val_evaluator = dict(type='mmdet.LVISMetric',
+                     ann_file='data/coco/lvis/lvis_v1_minival_inserted_image_name.json',
+                     metric='bbox')
+test_evaluator = val_evaluator
+
+# training settings
+default_hooks = dict(param_scheduler=dict(max_epochs=max_epochs),
+                     checkpoint=dict(interval=save_epoch_intervals,
+                                     rule='greater'))
+custom_hooks = [
+    dict(type='EMAHook',
+         ema_type='ExpMomentumEMA',
+         momentum=0.0001,
+         update_buffers=True,
+         strict_load=False,
+         priority=49),
+    dict(type='mmdet.PipelineSwitchHook',
+         switch_epoch=max_epochs - close_mosaic_epochs,
+         switch_pipeline=train_pipeline_stage2)
+]
+train_cfg = dict(max_epochs=max_epochs,
+                 val_interval=10,
+                 dynamic_intervals=[((max_epochs - close_mosaic_epochs),
+                                     _base_.val_interval_stage2)])
+optim_wrapper = dict(optimizer=dict(
+    _delete_=True,
+    type='AdamW',
+    lr=base_lr,
+    weight_decay=weight_decay,
+    batch_size_per_gpu=train_batch_size_per_gpu),
+                     paramwise_cfg=dict(bias_decay_mult=0.0,
+                                        norm_decay_mult=0.0,
+                                        custom_keys={
+                                            'backbone.text_model':
+                                            dict(lr_mult=0.01),
+                                            'logit_scale':
+                                            dict(weight_decay=0.0)
+                                        }),
+                     constructor='YOLOWv5OptimizerConstructor')

configs/pretrain/yolo_world_v2_l_vlpan_bn_2e-3_100e_4x8gpus_obj365v1_goldg_train_lvis_val.py

@@ -0,0 +1,171 @@
+_base_ = ('../../third_party/mmyolo/configs/yolov8/'
+          'yolov8_l_syncbn_fast_8xb16-500e_coco.py')
+custom_imports = dict(imports=['yolo_world'],
+                      allow_failed_imports=False)
+
+# hyper-parameters
+num_classes = 1203
+num_training_classes = 80
+max_epochs = 100  # Maximum training epochs
+close_mosaic_epochs = 2
+save_epoch_intervals = 2
+text_channels = 512
+neck_embed_channels = [128, 256, _base_.last_stage_out_channels // 2]
+neck_num_heads = [4, 8, _base_.last_stage_out_channels // 2 // 32]
+base_lr = 2e-3
+weight_decay = 0.05 / 2
+train_batch_size_per_gpu = 16
+# text_model_name = '../pretrained_models/clip-vit-base-patch32-projection'
+text_model_name = 'openai/clip-vit-base-patch32'
+# model settings
+model = dict(
+    type='YOLOWorldDetector',
+    mm_neck=True,
+    num_train_classes=num_training_classes,
+    num_test_classes=num_classes,
+    data_preprocessor=dict(type='YOLOWDetDataPreprocessor'),
+    backbone=dict(
+        _delete_=True,
+        type='MultiModalYOLOBackbone',
+        image_model={{_base_.model.backbone}},
+        text_model=dict(
+            type='HuggingCLIPLanguageBackbone',
+            model_name=text_model_name,
+            frozen_modules=['all'])),
+    neck=dict(type='YOLOWorldPAFPN',
+              guide_channels=text_channels,
+              embed_channels=neck_embed_channels,
+              num_heads=neck_num_heads,
+              block_cfg=dict(type='MaxSigmoidCSPLayerWithTwoConv')),
+    bbox_head=dict(type='YOLOWorldHead',
+                   head_module=dict(type='YOLOWorldHeadModule',
+                                    use_bn_head=True,
+                                    embed_dims=text_channels,
+                                    num_classes=num_training_classes)),
+    train_cfg=dict(assigner=dict(num_classes=num_training_classes)))
+
+# dataset settings
+text_transform = [
+    dict(type='RandomLoadText',
+         num_neg_samples=(num_classes, num_classes),
+         max_num_samples=num_training_classes,
+         padding_to_max=True,
+         padding_value=''),
+    dict(type='mmdet.PackDetInputs',
+         meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape', 'flip',
+                    'flip_direction', 'texts'))
+]
+train_pipeline = [
+    *_base_.pre_transform,
+    dict(type='MultiModalMosaic',
+         img_scale=_base_.img_scale,
+         pad_val=114.0,
+         pre_transform=_base_.pre_transform),
+    dict(
+        type='YOLOv5RandomAffine',
+        max_rotate_degree=0.0,
+        max_shear_degree=0.0,
+        scaling_ratio_range=(1 - _base_.affine_scale, 1 + _base_.affine_scale),
+        max_aspect_ratio=_base_.max_aspect_ratio,
+        border=(-_base_.img_scale[0] // 2, -_base_.img_scale[1] // 2),
+        border_val=(114, 114, 114)),
+    *_base_.last_transform[:-1],
+    *text_transform,
+]
+train_pipeline_stage2 = [*_base_.train_pipeline_stage2[:-1], *text_transform]
+obj365v1_train_dataset = dict(
+    type='MultiModalDataset',
+    dataset=dict(
+        type='YOLOv5Objects365V1Dataset',
+        data_root='data/objects365v1/',
+        ann_file='annotations/objects365_train.json',
+        data_prefix=dict(img='train/'),
+        filter_cfg=dict(filter_empty_gt=False, min_size=32)),
+    class_text_path='data/texts/obj365v1_class_texts.json',
+    pipeline=train_pipeline)
+
+mg_train_dataset = dict(type='YOLOv5MixedGroundingDataset',
+                        data_root='data/mixed_grounding/',
+                        ann_file='annotations/final_mixed_train_no_coco.json',
+                        data_prefix=dict(img='gqa/images/'),
+                        filter_cfg=dict(filter_empty_gt=False, min_size=32),
+                        pipeline=train_pipeline)
+
+flickr_train_dataset = dict(
+    type='YOLOv5MixedGroundingDataset',
+    data_root='data/flickr/',
+    ann_file='annotations/final_flickr_separateGT_train.json',
+    data_prefix=dict(img='full_images/'),
+    filter_cfg=dict(filter_empty_gt=True, min_size=32),
+    pipeline=train_pipeline)
+
+train_dataloader = dict(batch_size=train_batch_size_per_gpu,
+                        collate_fn=dict(type='yolow_collate'),
+                        dataset=dict(_delete_=True,
+                                     type='ConcatDataset',
+                                     datasets=[
+                                         obj365v1_train_dataset,
+                                         flickr_train_dataset, mg_train_dataset
+                                     ],
+                                     ignore_keys=['classes', 'palette']))
+
+test_pipeline = [
+    *_base_.test_pipeline[:-1],
+    dict(type='LoadText'),
+    dict(type='mmdet.PackDetInputs',
+         meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape',
+                    'scale_factor', 'pad_param', 'texts'))
+]
+coco_val_dataset = dict(
+    _delete_=True,
+    type='MultiModalDataset',
+    dataset=dict(type='YOLOv5LVISV1Dataset',
+                 data_root='data/coco/',
+                 test_mode=True,
+                 ann_file='lvis/lvis_v1_val.json',
+                 data_prefix=dict(img=''),
+                 batch_shapes_cfg=None),
+    class_text_path='data/texts/lvis_v1_class_texts.json',
+    pipeline=test_pipeline)
+val_dataloader = dict(dataset=coco_val_dataset)
+test_dataloader = val_dataloader
+
+val_evaluator = dict(type='mmdet.LVISMetric',
+                     ann_file='data/coco/lvis/lvis_v1_val.json',
+                     metric='bbox')
+test_evaluator = val_evaluator
+
+# training settings
+default_hooks = dict(param_scheduler=dict(max_epochs=max_epochs),
+                     checkpoint=dict(interval=save_epoch_intervals,
+                                     rule='greater'))
+custom_hooks = [
+    dict(type='EMAHook',
+         ema_type='ExpMomentumEMA',
+         momentum=0.0001,
+         update_buffers=True,
+         strict_load=False,
+         priority=49),
+    dict(type='mmdet.PipelineSwitchHook',
+         switch_epoch=max_epochs - close_mosaic_epochs,
+         switch_pipeline=train_pipeline_stage2)
+]
+train_cfg = dict(max_epochs=max_epochs,
+                 val_interval=10,
+                 dynamic_intervals=[((max_epochs - close_mosaic_epochs),
+                                     _base_.val_interval_stage2)])
+optim_wrapper = dict(optimizer=dict(
+    _delete_=True,
+    type='AdamW',
+    lr=base_lr,
+    weight_decay=weight_decay,
+    batch_size_per_gpu=train_batch_size_per_gpu),
+                     paramwise_cfg=dict(bias_decay_mult=0.0,
+                                        norm_decay_mult=0.0,
+                                        custom_keys={
+                                            'backbone.text_model':
+                                            dict(lr_mult=0.01),
+                                            'logit_scale':
+                                            dict(weight_decay=0.0)
+                                        }),
+                     constructor='YOLOWv5OptimizerConstructor')

configs/pretrain/yolo_world_v2_m_vlpan_bn_2e-3_100e_4x8gpus_obj365v1_goldg_train_1280ft_lvis_minival.py

@@ -0,0 +1,198 @@
+_base_ = ('../../third_party/mmyolo/configs/yolov8/'
+          'yolov8_m_syncbn_fast_8xb16-500e_coco.py')
+custom_imports = dict(imports=['yolo_world'],
+                      allow_failed_imports=False)
+
+# hyper-parameters
+num_classes = 1203
+num_training_classes = 80
+max_epochs = 100  # Maximum training epochs
+close_mosaic_epochs = 2
+save_epoch_intervals = 2
+text_channels = 512
+neck_embed_channels = [128, 256, _base_.last_stage_out_channels // 2]
+neck_num_heads = [4, 8, _base_.last_stage_out_channels // 2 // 32]
+base_lr = 2e-3
+weight_decay = 0.05 / 2
+train_batch_size_per_gpu = 16
+text_model_name = '../pretrained_models/clip-vit-base-patch32-projection'
+img_scale = (1280, 1280)
+
+text_model_name = 'openai/clip-vit-base-patch32'
+# model settings
+model = dict(
+    type='YOLOWorldDetector',
+    mm_neck=True,
+    num_train_classes=num_training_classes,
+    num_test_classes=num_classes,
+    data_preprocessor=dict(type='YOLOWDetDataPreprocessor'),
+    backbone=dict(
+        _delete_=True,
+        type='MultiModalYOLOBackbone',
+        image_model={{_base_.model.backbone}},
+        text_model=dict(
+            type='HuggingCLIPLanguageBackbone',
+            model_name=text_model_name,
+            frozen_modules=['all'])),
+    neck=dict(type='YOLOWorldPAFPN',
+              guide_channels=text_channels,
+              embed_channels=neck_embed_channels,
+              num_heads=neck_num_heads,
+              block_cfg=dict(type='MaxSigmoidCSPLayerWithTwoConv')),
+    bbox_head=dict(type='YOLOWorldHead',
+                   head_module=dict(type='YOLOWorldHeadModule',
+                                    use_bn_head=True,
+                                    embed_dims=text_channels,
+                                    num_classes=num_training_classes)),
+    train_cfg=dict(assigner=dict(num_classes=num_training_classes)))
+
+# dataset settings
+text_transform = [
+    dict(type='RandomLoadText',
+         num_neg_samples=(num_classes, num_classes),
+         max_num_samples=num_training_classes,
+         padding_to_max=True,
+         padding_value=''),
+    dict(type='mmdet.PackDetInputs',
+         meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape', 'flip',
+                    'flip_direction', 'texts'))
+]
+train_pipeline = [
+    *_base_.pre_transform,
+    dict(type='MultiModalMosaic',
+         img_scale=img_scale,
+         pad_val=114.0,
+         pre_transform=_base_.pre_transform),
+    dict(
+        type='YOLOv5RandomAffine',
+        max_rotate_degree=0.0,
+        max_shear_degree=0.0,
+        scaling_ratio_range=(1 - _base_.affine_scale, 1 + _base_.affine_scale),
+        max_aspect_ratio=_base_.max_aspect_ratio,
+        border=(-img_scale[0] // 2, -img_scale[1] // 2),
+        border_val=(114, 114, 114)),
+    *_base_.last_transform[:-1],
+    *text_transform,
+]
+
+train_pipeline_stage2 = [
+    *_base_.pre_transform,
+    dict(type='YOLOv5KeepRatioResize', scale=img_scale),
+    dict(
+        type='LetterResize',
+        scale=img_scale,
+        allow_scale_up=True,
+        pad_val=dict(img=114.0)),
+    dict(
+        type='YOLOv5RandomAffine',
+        max_rotate_degree=0.0,
+        max_shear_degree=0.0,
+        scaling_ratio_range=(1 - _base_.affine_scale, 1 + _base_.affine_scale),
+        max_aspect_ratio=_base_.max_aspect_ratio,
+        border_val=(114, 114, 114)),
+    *_base_.last_transform[:-1],
+    *text_transform
+]
+obj365v1_train_dataset = dict(
+    type='MultiModalDataset',
+    dataset=dict(
+        type='YOLOv5Objects365V1Dataset',
+        data_root='data/objects365v1/',
+        ann_file='annotations/objects365_train.json',
+        data_prefix=dict(img='train/'),
+        filter_cfg=dict(filter_empty_gt=False, min_size=32)),
+    class_text_path='data/texts/obj365v1_class_texts.json',
+    pipeline=train_pipeline)
+
+mg_train_dataset = dict(type='YOLOv5MixedGroundingDataset',
+                        data_root='data/mixed_grounding/',
+                        ann_file='annotations/final_mixed_train_no_coco.json',
+                        data_prefix=dict(img='gqa/images/'),
+                        filter_cfg=dict(filter_empty_gt=False, min_size=32),
+                        pipeline=train_pipeline)
+
+flickr_train_dataset = dict(
+    type='YOLOv5MixedGroundingDataset',
+    data_root='data/flickr/',
+    ann_file='annotations/final_flickr_separateGT_train.json',
+    data_prefix=dict(img='full_images/'),
+    filter_cfg=dict(filter_empty_gt=True, min_size=32),
+    pipeline=train_pipeline)
+
+train_dataloader = dict(batch_size=train_batch_size_per_gpu,
+                        collate_fn=dict(type='yolow_collate'),
+                        dataset=dict(_delete_=True,
+                                     type='ConcatDataset',
+                                     datasets=[
+                                         obj365v1_train_dataset,
+                                         flickr_train_dataset, mg_train_dataset
+                                     ],
+                                     ignore_keys=['classes', 'palette']))
+
+test_pipeline = [
+    dict(type='LoadImageFromFile'),
+    dict(type='YOLOv5KeepRatioResize', scale=img_scale),
+    dict(
+        type='LetterResize',
+        scale=img_scale,
+        allow_scale_up=False,
+        pad_val=dict(img=114)),
+    dict(type='LoadAnnotations', with_bbox=True, _scope_='mmdet'),
+    dict(type='LoadText'),
+    dict(type='mmdet.PackDetInputs',
+         meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape',
+                    'scale_factor', 'pad_param', 'texts'))
+]
+coco_val_dataset = dict(
+    _delete_=True,
+    type='MultiModalDataset',
+    dataset=dict(type='YOLOv5LVISV1Dataset',
+                 data_root='data/coco/',
+                 test_mode=True,
+                 ann_file='lvis/lvis_v1_minival_inserted_image_name.json',
+                 data_prefix=dict(img=''),
+                 batch_shapes_cfg=None),
+    class_text_path='data/texts/lvis_v1_class_texts.json',
+    pipeline=test_pipeline)
+val_dataloader = dict(dataset=coco_val_dataset)
+test_dataloader = val_dataloader
+
+val_evaluator = dict(type='mmdet.LVISMetric',
+                     ann_file='data/coco/lvis/lvis_v1_minival_inserted_image_name.json',
+                     metric='bbox')
+test_evaluator = val_evaluator
+
+# training settings
+default_hooks = dict(param_scheduler=dict(max_epochs=max_epochs),
+                     checkpoint=dict(interval=save_epoch_intervals,
+                                     rule='greater'))
+custom_hooks = [
+    dict(type='EMAHook',
+         ema_type='ExpMomentumEMA',
+         momentum=0.0001,
+         update_buffers=True,
+         strict_load=False,
+         priority=49),
+    dict(type='mmdet.PipelineSwitchHook',
+         switch_epoch=max_epochs - close_mosaic_epochs,
+         switch_pipeline=train_pipeline_stage2)
+]
+train_cfg = dict(max_epochs=max_epochs,
+                 val_interval=10,
+                 dynamic_intervals=[((max_epochs - close_mosaic_epochs),
+                                     _base_.val_interval_stage2)])
+optim_wrapper = dict(optimizer=dict(
+    _delete_=True,
+    type='AdamW',
+    lr=base_lr,
+    weight_decay=weight_decay,
+    batch_size_per_gpu=train_batch_size_per_gpu),
+                     paramwise_cfg=dict(bias_decay_mult=0.0,
+                                        norm_decay_mult=0.0,
+                                        custom_keys={
+                                            'backbone.text_model':
+                                            dict(lr_mult=0.01),
+                                            'logit_scale':
+                                            dict(weight_decay=0.0)
+                                        }),
+                     constructor='YOLOWv5OptimizerConstructor')

configs/pretrain/yolo_world_v2_m_vlpan_bn_2e-3_100e_4x8gpus_obj365v1_goldg_train_lvis_minival.py

@@ -0,0 +1,171 @@
+_base_ = ('../../third_party/mmyolo/configs/yolov8/'
+          'yolov8_m_syncbn_fast_8xb16-500e_coco.py')
+custom_imports = dict(imports=['yolo_world'],
+                      allow_failed_imports=False)
+
+# hyper-parameters
+num_classes = 1203
+num_training_classes = 80
+max_epochs = 100  # Maximum training epochs
+close_mosaic_epochs = 2
+save_epoch_intervals = 2
+text_channels = 512
+neck_embed_channels = [128, 256, _base_.last_stage_out_channels // 2]
+neck_num_heads = [4, 8, _base_.last_stage_out_channels // 2 // 32]
+base_lr = 2e-3
+weight_decay = 0.05 / 2
+train_batch_size_per_gpu = 16
+text_model_name = 'openai/clip-vit-large-patch14-336'
+text_model_name = 'openai/clip-vit-base-patch32'
+# model settings
+model = dict(
+    type='YOLOWorldDetector',
+    mm_neck=True,
+    num_train_classes=num_training_classes,
+    num_test_classes=num_classes,
+    data_preprocessor=dict(type='YOLOWDetDataPreprocessor'),
+    backbone=dict(
+        _delete_=True,
+        type='MultiModalYOLOBackbone',
+        image_model={{_base_.model.backbone}},
+        text_model=dict(
+            type='HuggingCLIPLanguageBackbone',
+            model_name=text_model_name,
+            frozen_modules=['all'])),
+    neck=dict(type='YOLOWorldPAFPN',
+              guide_channels=text_channels,
+              embed_channels=neck_embed_channels,
+              num_heads=neck_num_heads,
+              block_cfg=dict(type='MaxSigmoidCSPLayerWithTwoConv')),
+    bbox_head=dict(type='YOLOWorldHead',
+                   head_module=dict(type='YOLOWorldHeadModule',
+                                    use_bn_head=True,
+                                    embed_dims=text_channels,
+                                    num_classes=num_training_classes)),
+    train_cfg=dict(assigner=dict(num_classes=num_training_classes)))
+
+# dataset settings
+text_transform = [
+    dict(type='RandomLoadText',
+         num_neg_samples=(num_classes, num_classes),
+         max_num_samples=num_training_classes,
+         padding_to_max=True,
+         padding_value=''),
+    dict(type='mmdet.PackDetInputs',
+         meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape', 'flip',
+                    'flip_direction', 'texts'))
+]
+train_pipeline = [
+    *_base_.pre_transform,
+    dict(type='MultiModalMosaic',
+         img_scale=_base_.img_scale,
+         pad_val=114.0,
+         pre_transform=_base_.pre_transform),
+    dict(
+        type='YOLOv5RandomAffine',
+        max_rotate_degree=0.0,
+        max_shear_degree=0.0,
+        scaling_ratio_range=(1 - _base_.affine_scale, 1 + _base_.affine_scale),
+        max_aspect_ratio=_base_.max_aspect_ratio,
+        border=(-_base_.img_scale[0] // 2, -_base_.img_scale[1] // 2),
+        border_val=(114, 114, 114)),
+    *_base_.last_transform[:-1],
+    *text_transform,
+]
+train_pipeline_stage2 = [*_base_.train_pipeline_stage2[:-1], *text_transform]
+obj365v1_train_dataset = dict(
+    type='MultiModalDataset',
+    dataset=dict(
+        type='YOLOv5Objects365V1Dataset',
+        data_root='data/objects365v1/',
+        ann_file='annotations/objects365_train.json',
+        data_prefix=dict(img='train/'),
+        filter_cfg=dict(filter_empty_gt=False, min_size=32)),
+    class_text_path='data/texts/obj365v1_class_texts.json',
+    pipeline=train_pipeline)
+
+mg_train_dataset = dict(type='YOLOv5MixedGroundingDataset',
+                        data_root='data/mixed_grounding/',
+                        ann_file='annotations/final_mixed_train_no_coco.json',
+                        data_prefix=dict(img='gqa/images/'),
+                        filter_cfg=dict(filter_empty_gt=False, min_size=32),
+                        pipeline=train_pipeline)
+
+flickr_train_dataset = dict(
+    type='YOLOv5MixedGroundingDataset',
+    data_root='data/flickr/',
+    ann_file='annotations/final_flickr_separateGT_train.json',
+    data_prefix=dict(img='full_images/'),
+    filter_cfg=dict(filter_empty_gt=True, min_size=32),
+    pipeline=train_pipeline)
+
+train_dataloader = dict(batch_size=train_batch_size_per_gpu,
+                        collate_fn=dict(type='yolow_collate'),
+                        dataset=dict(_delete_=True,
+                                     type='ConcatDataset',
+                                     datasets=[
+                                         obj365v1_train_dataset,
+                                         flickr_train_dataset, mg_train_dataset
+                                     ],
+                                     ignore_keys=['classes', 'palette']))
+
+test_pipeline = [
+    *_base_.test_pipeline[:-1],
+    dict(type='LoadText'),
+    dict(type='mmdet.PackDetInputs',
+         meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape',
+                    'scale_factor', 'pad_param', 'texts'))
+]
+coco_val_dataset = dict(
+    _delete_=True,
+    type='MultiModalDataset',
+    dataset=dict(type='YOLOv5LVISV1Dataset',
+                 data_root='data/coco/',
+                 test_mode=True,
+                 ann_file='lvis/lvis_v1_minival_inserted_image_name.json',
+                 data_prefix=dict(img=''),
+                 batch_shapes_cfg=None),
+    class_text_path='data/texts/lvis_v1_class_texts.json',
+    pipeline=test_pipeline)
+val_dataloader = dict(dataset=coco_val_dataset)
+test_dataloader = val_dataloader
+
+val_evaluator = dict(type='mmdet.LVISMetric',
+                     ann_file='data/coco/lvis/lvis_v1_minival_inserted_image_name.json',
+                     metric='bbox')
+test_evaluator = val_evaluator
+
+# training settings
+default_hooks = dict(param_scheduler=dict(max_epochs=max_epochs),
+                     checkpoint=dict(interval=save_epoch_intervals,
+                                     rule='greater'))
+custom_hooks = [
+    dict(type='EMAHook',
+         ema_type='ExpMomentumEMA',
+         momentum=0.0001,
+         update_buffers=True,
+         strict_load=False,
+         priority=49),
+    dict(type='mmdet.PipelineSwitchHook',
+         switch_epoch=max_epochs - close_mosaic_epochs,
+         switch_pipeline=train_pipeline_stage2)
+]
+train_cfg = dict(max_epochs=max_epochs,
+                 val_interval=10,
+                 dynamic_intervals=[((max_epochs - close_mosaic_epochs),
+                                     _base_.val_interval_stage2)])
+optim_wrapper = dict(optimizer=dict(
+    _delete_=True,
+    type='AdamW',
+    lr=base_lr,
+    weight_decay=weight_decay,
+    batch_size_per_gpu=train_batch_size_per_gpu),
+                     paramwise_cfg=dict(bias_decay_mult=0.0,
+                                        norm_decay_mult=0.0,
+                                        custom_keys={
+                                            'backbone.text_model':
+                                            dict(lr_mult=0.01),
+                                            'logit_scale':
+                                            dict(weight_decay=0.0)
+                                        }),
+                     constructor='YOLOWv5OptimizerConstructor')

configs/pretrain/yolo_world_v2_m_vlpan_bn_noeinsum_2e-3_100e_4x8gpus_obj365v1_goldg_train_lvis_minival.py

@@ -0,0 +1,176 @@
+_base_ = ('../../third_party/mmyolo/configs/yolov8/'
+          'yolov8_m_syncbn_fast_8xb16-500e_coco.py')
+custom_imports = dict(imports=['yolo_world'],
+                      allow_failed_imports=False)
+
+# hyper-parameters
+num_classes = 1203
+num_training_classes = 80
+max_epochs = 100  # Maximum training epochs
+close_mosaic_epochs = 2
+save_epoch_intervals = 2
+text_channels = 512
+neck_embed_channels = [128, 256, _base_.last_stage_out_channels // 2]
+neck_num_heads = [4, 8, _base_.last_stage_out_channels // 2 // 32]
+base_lr = 2e-3
+weight_decay = 0.05 / 2
+train_batch_size_per_gpu = 16
+text_model_name = '../pretrained_models/clip-vit-base-patch32-projection'
+text_model_name = 'openai/clip-vit-base-patch32'
+# model settings
+model = dict(
+    type='YOLOWorldDetector',
+    mm_neck=True,
+    num_train_classes=num_training_classes,
+    num_test_classes=num_classes,
+    data_preprocessor=dict(type='YOLOWDetDataPreprocessor'),
+    backbone=dict(
+        _delete_=True,
+        type='MultiModalYOLOBackbone',
+        image_model={{_base_.model.backbone}},
+        text_model=dict(
+            type='HuggingCLIPLanguageBackbone',
+            model_name=text_model_name,
+            frozen_modules=['all'])),
+    neck=dict(type='YOLOWorldPAFPN',
+              guide_channels=text_channels,
+              embed_channels=neck_embed_channels,
+              num_heads=neck_num_heads,
+              block_cfg=dict(type='MaxSigmoidCSPLayerWithTwoConv',
+                             use_einsum=False)),
+    bbox_head=dict(type='YOLOWorldHead',
+                   head_module=dict(type='YOLOWorldHeadModule',
+                                    use_bn_head=True,
+                                    embed_dims=text_channels,
+                                    num_classes=num_training_classes,
+                                    use_einsum=False)),
+    train_cfg=dict(assigner=dict(num_classes=num_training_classes)))
+
+# dataset settings
+text_transform = [
+    dict(type='RandomLoadText',
+         num_neg_samples=(num_classes, num_classes),
+         max_num_samples=num_training_classes,
+         padding_to_max=True,
+         padding_value=''),
+    dict(type='mmdet.PackDetInputs',
+         meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape', 'flip',
+                    'flip_direction', 'texts'))
+]
+
+train_pipeline = [
+    *_base_.pre_transform,
+    dict(type='MultiModalMosaic',
+         img_scale=_base_.img_scale,
+         pad_val=114.0,
+         pre_transform=_base_.pre_transform),
+    dict(
+        type='YOLOv5RandomAffine',
+        max_rotate_degree=0.0,
+        max_shear_degree=0.0,
+        scaling_ratio_range=(1 - _base_.affine_scale, 1 + _base_.affine_scale),
+        max_aspect_ratio=_base_.max_aspect_ratio,
+        border=(-_base_.img_scale[0] // 2, -_base_.img_scale[1] // 2),
+        border_val=(114, 114, 114)),
+    *_base_.last_transform[:-1],
+    *text_transform,
+]
+
+train_pipeline_stage2 = [*_base_.train_pipeline_stage2[:-1], *text_transform]
+obj365v1_train_dataset = dict(
+    type='MultiModalDataset',
+    dataset=dict(
+        type='YOLOv5Objects365V1Dataset',
+        data_root='data/objects365v1/',
+        ann_file='annotations/objects365_train.json',
+        data_prefix=dict(img='train/'),
+        filter_cfg=dict(filter_empty_gt=False, min_size=32)),
+    class_text_path='data/texts/obj365v1_class_texts.json',
+    pipeline=train_pipeline)
+
+mg_train_dataset = dict(type='YOLOv5MixedGroundingDataset',
+                        data_root='data/mixed_grounding/',
+                        ann_file='annotations/final_mixed_train_no_coco.json',
+                        data_prefix=dict(img='gqa/images/'),
+                        filter_cfg=dict(filter_empty_gt=False, min_size=32),
+                        pipeline=train_pipeline)
+
+flickr_train_dataset = dict(
+    type='YOLOv5MixedGroundingDataset',
+    data_root='data/flickr/',
+    ann_file='annotations/final_flickr_separateGT_train.json',
+    data_prefix=dict(img='full_images/'),
+    filter_cfg=dict(filter_empty_gt=True, min_size=32),
+    pipeline=train_pipeline)
+
+train_dataloader = dict(batch_size=train_batch_size_per_gpu,
+                        collate_fn=dict(type='yolow_collate'),
+                        dataset=dict(_delete_=True,
+                                     type='ConcatDataset',
+                                     datasets=[
+                                         obj365v1_train_dataset,
+                                         flickr_train_dataset, mg_train_dataset
+                                     ],
+                                     ignore_keys=['classes', 'palette']))
+
+test_pipeline = [
+    *_base_.test_pipeline[:-1],
+    dict(type='LoadText'),
+    dict(type='mmdet.PackDetInputs',
+         meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape',
+                    'scale_factor', 'pad_param', 'texts'))
+]
+
+coco_val_dataset = dict(
+    _delete_=True,
+    type='MultiModalDataset',
+    dataset=dict(type='YOLOv5LVISV1Dataset',
+                 data_root='data/coco/',
+                 test_mode=True,
+                 ann_file='lvis/lvis_v1_minival_inserted_image_name.json',
+                 data_prefix=dict(img=''),
+                 batch_shapes_cfg=None),
+    class_text_path='data/texts/lvis_v1_class_texts.json',
+    pipeline=test_pipeline)
+val_dataloader = dict(dataset=coco_val_dataset)
+test_dataloader = val_dataloader
+
+val_evaluator = dict(type='mmdet.LVISMetric',
+                     ann_file='data/coco/lvis/lvis_v1_minival_inserted_image_name.json',
+                     metric='bbox')
+test_evaluator = val_evaluator
+
+# training settings
+default_hooks = dict(param_scheduler=dict(max_epochs=max_epochs),
+                     checkpoint=dict(interval=save_epoch_intervals,
+                                     rule='greater'))
+custom_hooks = [
+    dict(type='EMAHook',
+         ema_type='ExpMomentumEMA',
+         momentum=0.0001,
+         update_buffers=True,
+         strict_load=False,
+         priority=49),
+    dict(type='mmdet.PipelineSwitchHook',
+         switch_epoch=max_epochs - close_mosaic_epochs,
+         switch_pipeline=train_pipeline_stage2)
+]
+train_cfg = dict(max_epochs=max_epochs,
+                 val_interval=10,
+                 dynamic_intervals=[((max_epochs - close_mosaic_epochs),
+                                     _base_.val_interval_stage2)])
+optim_wrapper = dict(optimizer=dict(
+    _delete_=True,
+    type='AdamW',
+    lr=base_lr,
+    weight_decay=weight_decay,
+    batch_size_per_gpu=train_batch_size_per_gpu),
+                     paramwise_cfg=dict(bias_decay_mult=0.0,
+                                        norm_decay_mult=0.0,
+                                        custom_keys={
+                                            'backbone.text_model':
+                                            dict(lr_mult=0.01),
+                                            'logit_scale':
+                                            dict(weight_decay=0.0)
+                                        }),
+                     constructor='YOLOWv5OptimizerConstructor')

configs/pretrain/yolo_world_v2_s_vlpan_bn_2e-3_100e_4x8gpus_obj365v1_goldg_train_1280ft_lvis_minival.py

@@ -0,0 +1,195 @@
+_base_ = ('../../third_party/mmyolo/configs/yolov8/'
+          'yolov8_s_syncbn_fast_8xb16-500e_coco.py')
+custom_imports = dict(imports=['yolo_world'],
+                      allow_failed_imports=False)
+
+# hyper-parameters
+num_classes = 1203
+num_training_classes = 80
+max_epochs = 100  # Maximum training epochs
+close_mosaic_epochs = 2
+save_epoch_intervals = 2
+text_channels = 512
+neck_embed_channels = [128, 256, _base_.last_stage_out_channels // 2]
+neck_num_heads = [4, 8, _base_.last_stage_out_channels // 2 // 32]
+base_lr = 2e-4
+weight_decay = 0.025
+train_batch_size_per_gpu = 4
+img_scale = (1280, 1280)
+
+# model settings
+model = dict(
+    type='YOLOWorldDetector',
+    mm_neck=True,
+    num_train_classes=num_training_classes,
+    num_test_classes=num_classes,
+    data_preprocessor=dict(type='YOLOWDetDataPreprocessor'),
+    backbone=dict(
+        _delete_=True,
+        type='MultiModalYOLOBackbone',
+        image_model={{_base_.model.backbone}},
+        text_model=dict(
+            type='HuggingCLIPLanguageBackbone',
+            model_name='openai/clip-vit-base-patch32',
+            frozen_modules=['all'])),
+    neck=dict(type='YOLOWorldPAFPN',
+              guide_channels=text_channels,
+              embed_channels=neck_embed_channels,
+              num_heads=neck_num_heads,
+              block_cfg=dict(type='MaxSigmoidCSPLayerWithTwoConv')),
+    bbox_head=dict(type='YOLOWorldHead',
+                   head_module=dict(type='YOLOWorldHeadModule',
+                                    use_bn_head=True,
+                                    embed_dims=text_channels,
+                                    num_classes=num_training_classes)),
+    train_cfg=dict(assigner=dict(num_classes=num_training_classes)))
+
+# dataset settings
+text_transform = [
+    dict(type='RandomLoadText',
+         num_neg_samples=(num_classes, num_classes),
+         max_num_samples=num_training_classes,
+         padding_to_max=True,
+         padding_value=''),
+    dict(type='mmdet.PackDetInputs',
+         meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape', 'flip',
+                    'flip_direction', 'texts'))
+]
+train_pipeline = [
+    *_base_.pre_transform,
+    dict(type='MultiModalMosaic',
+         img_scale=img_scale,
+         pad_val=114.0,
+         pre_transform=_base_.pre_transform),
+    dict(
+        type='YOLOv5RandomAffine',
+        max_rotate_degree=0.0,
+        max_shear_degree=0.0,
+        scaling_ratio_range=(1 - _base_.affine_scale, 1 + _base_.affine_scale),
+        max_aspect_ratio=_base_.max_aspect_ratio,
+        border=(-img_scale[0] // 2, -img_scale[1] // 2),
+        border_val=(114, 114, 114)),
+    *_base_.last_transform[:-1],
+    *text_transform,
+]
+train_pipeline_stage2 = [
+    *_base_.pre_transform,
+    dict(type='YOLOv5KeepRatioResize', scale=img_scale),
+    dict(
+        type='LetterResize',
+        scale=img_scale,
+        allow_scale_up=True,
+        pad_val=dict(img=114.0)),
+    dict(
+        type='YOLOv5RandomAffine',
+        max_rotate_degree=0.0,
+        max_shear_degree=0.0,
+        scaling_ratio_range=(1 - _base_.affine_scale, 1 + _base_.affine_scale),
+        max_aspect_ratio=_base_.max_aspect_ratio,
+        border_val=(114, 114, 114)),
+    *_base_.last_transform[:-1],
+    *text_transform
+]
+obj365v1_train_dataset = dict(
+    type='MultiModalDataset',
+    dataset=dict(
+        type='YOLOv5Objects365V1Dataset',
+        data_root='data/objects365v1/',
+        ann_file='annotations/objects365_train.json',
+        data_prefix=dict(img='train/'),
+        filter_cfg=dict(filter_empty_gt=False, min_size=32)),
+    class_text_path='data/texts/obj365v1_class_texts.json',
+    pipeline=train_pipeline)
+
+mg_train_dataset = dict(type='YOLOv5MixedGroundingDataset',
+                        data_root='data/mixed_grounding/',
+                        ann_file='annotations/final_mixed_train_no_coco.json',
+                        data_prefix=dict(img='gqa/images/'),
+                        filter_cfg=dict(filter_empty_gt=False, min_size=32),
+                        pipeline=train_pipeline)
+
+flickr_train_dataset = dict(
+    type='YOLOv5MixedGroundingDataset',
+    data_root='data/flickr/',
+    ann_file='annotations/final_flickr_separateGT_train.json',
+    data_prefix=dict(img='full_images/'),
+    filter_cfg=dict(filter_empty_gt=True, min_size=32),
+    pipeline=train_pipeline)
+
+train_dataloader = dict(batch_size=train_batch_size_per_gpu,
+                        collate_fn=dict(type='yolow_collate'),
+                        dataset=dict(_delete_=True,
+                                     type='ConcatDataset',
+                                     datasets=[
+                                         obj365v1_train_dataset,
+                                         flickr_train_dataset, mg_train_dataset
+                                     ],
+                                     ignore_keys=['classes', 'palette']))
+test_pipeline = [
+    dict(type='LoadImageFromFile'),
+    dict(type='YOLOv5KeepRatioResize', scale=img_scale),
+    dict(
+        type='LetterResize',
+        scale=img_scale,
+        allow_scale_up=False,
+        pad_val=dict(img=114)),
+    dict(type='LoadAnnotations', with_bbox=True, _scope_='mmdet'),
+    dict(type='LoadText'),
+    dict(type='mmdet.PackDetInputs',
+         meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape',
+                    'scale_factor', 'pad_param', 'texts'))
+]
+
+coco_val_dataset = dict(
+    _delete_=True,
+    type='MultiModalDataset',
+    dataset=dict(type='YOLOv5LVISV1Dataset',
+                 data_root='data/coco/',
+                 test_mode=True,
+                 ann_file='lvis/lvis_v1_minival_inserted_image_name.json',
+                 data_prefix=dict(img=''),
+                 batch_shapes_cfg=None),
+    class_text_path='data/texts/lvis_v1_class_texts.json',
+    pipeline=test_pipeline)
+val_dataloader = dict(dataset=coco_val_dataset)
+test_dataloader = val_dataloader
+
+val_evaluator = dict(type='mmdet.LVISMetric',
+                     ann_file='data/coco/lvis/lvis_v1_minival_inserted_image_name.json',
+                     metric='bbox')
+test_evaluator = val_evaluator
+
+# training settings
+default_hooks = dict(param_scheduler=dict(max_epochs=max_epochs),
+                     checkpoint=dict(interval=save_epoch_intervals,
+                                     rule='greater'))
+custom_hooks = [
+    dict(type='EMAHook',
+         ema_type='ExpMomentumEMA',
+         momentum=0.0001,
+         update_buffers=True,
+         strict_load=False,
+         priority=49),
+    dict(type='mmdet.PipelineSwitchHook',
+         switch_epoch=max_epochs - close_mosaic_epochs,
+         switch_pipeline=train_pipeline_stage2)
+]
+train_cfg = dict(max_epochs=max_epochs,
+                 val_interval=10,
+                 dynamic_intervals=[((max_epochs - close_mosaic_epochs),
+                                     _base_.val_interval_stage2)])
+optim_wrapper = dict(optimizer=dict(
+    _delete_=True,
+    type='AdamW',
+    lr=base_lr,
+    weight_decay=weight_decay,
+    batch_size_per_gpu=train_batch_size_per_gpu),
+                     paramwise_cfg=dict(bias_decay_mult=0.0,
+                                        norm_decay_mult=0.0,
+                                        custom_keys={
+                                            'backbone.text_model':
+                                            dict(lr_mult=0.01),
+                                            'logit_scale':
+                                            dict(weight_decay=0.0)
+                                        }),
+                     constructor='YOLOWv5OptimizerConstructor')

configs/pretrain/yolo_world_v2_s_vlpan_bn_2e-3_100e_4x8gpus_obj365v1_goldg_train_lvis_minival.py

@@ -0,0 +1,170 @@
+_base_ = ('../../third_party/mmyolo/configs/yolov8/'
+          'yolov8_s_syncbn_fast_8xb16-500e_coco.py')
+custom_imports = dict(imports=['yolo_world'],
+                      allow_failed_imports=False)
+
+# hyper-parameters
+num_classes = 1203
+num_training_classes = 80
+max_epochs = 100  # Maximum training epochs
+close_mosaic_epochs = 2
+save_epoch_intervals = 2
+text_channels = 512
+neck_embed_channels = [128, 256, _base_.last_stage_out_channels // 2]
+neck_num_heads = [4, 8, _base_.last_stage_out_channels // 2 // 32]
+base_lr = 2e-3
+weight_decay = 0.05 / 2
+train_batch_size_per_gpu = 16
+
+# model settings
+model = dict(
+    type='YOLOWorldDetector',
+    mm_neck=True,
+    num_train_classes=num_training_classes,
+    num_test_classes=num_classes,
+    data_preprocessor=dict(type='YOLOWDetDataPreprocessor'),
+    backbone=dict(
+        _delete_=True,
+        type='MultiModalYOLOBackbone',
+        image_model={{_base_.model.backbone}},
+        text_model=dict(
+            type='HuggingCLIPLanguageBackbone',
+            model_name='openai/clip-vit-base-patch32',
+            frozen_modules=['all'])),
+    neck=dict(type='YOLOWorldPAFPN',
+              guide_channels=text_channels,
+              embed_channels=neck_embed_channels,
+              num_heads=neck_num_heads,
+              block_cfg=dict(type='MaxSigmoidCSPLayerWithTwoConv')),
+    bbox_head=dict(type='YOLOWorldHead',
+                   head_module=dict(type='YOLOWorldHeadModule',
+                                    use_bn_head=True,
+                                    embed_dims=text_channels,
+                                    num_classes=num_training_classes)),
+    train_cfg=dict(assigner=dict(num_classes=num_training_classes)))
+
+# dataset settings
+text_transform = [
+    dict(type='RandomLoadText',
+         num_neg_samples=(num_classes, num_classes),
+         max_num_samples=num_training_classes,
+         padding_to_max=True,
+         padding_value=''),
+    dict(type='mmdet.PackDetInputs',
+         meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape', 'flip',
+                    'flip_direction', 'texts'))
+]
+train_pipeline = [
+    *_base_.pre_transform,
+    dict(type='MultiModalMosaic',
+         img_scale=_base_.img_scale,
+         pad_val=114.0,
+         pre_transform=_base_.pre_transform),
+    dict(
+        type='YOLOv5RandomAffine',
+        max_rotate_degree=0.0,
+        max_shear_degree=0.0,
+        scaling_ratio_range=(1 - _base_.affine_scale, 1 + _base_.affine_scale),
+        max_aspect_ratio=_base_.max_aspect_ratio,
+        border=(-_base_.img_scale[0] // 2, -_base_.img_scale[1] // 2),
+        border_val=(114, 114, 114)),
+    *_base_.last_transform[:-1],
+    *text_transform,
+]
+train_pipeline_stage2 = [*_base_.train_pipeline_stage2[:-1], *text_transform]
+obj365v1_train_dataset = dict(
+    type='MultiModalDataset',
+    dataset=dict(
+        type='YOLOv5Objects365V1Dataset',
+        data_root='data/objects365v1/',
+        ann_file='annotations/objects365_train.json',
+        data_prefix=dict(img='train/'),
+        filter_cfg=dict(filter_empty_gt=False, min_size=32)),
+    class_text_path='data/texts/obj365v1_class_texts.json',
+    pipeline=train_pipeline)
+
+mg_train_dataset = dict(type='YOLOv5MixedGroundingDataset',
+                        data_root='data/mixed_grounding/',
+                        ann_file='annotations/final_mixed_train_no_coco.json',
+                        data_prefix=dict(img='gqa/images/'),
+                        filter_cfg=dict(filter_empty_gt=False, min_size=32),
+                        pipeline=train_pipeline)
+
+flickr_train_dataset = dict(
+    type='YOLOv5MixedGroundingDataset',
+    data_root='data/flickr/',
+    ann_file='annotations/final_flickr_separateGT_train.json',
+    data_prefix=dict(img='full_images/'),
+    filter_cfg=dict(filter_empty_gt=True, min_size=32),
+    pipeline=train_pipeline)
+
+train_dataloader = dict(batch_size=train_batch_size_per_gpu,
+                        collate_fn=dict(type='yolow_collate'),
+                        dataset=dict(_delete_=True,
+                                     type='ConcatDataset',
+                                     datasets=[
+                                         obj365v1_train_dataset,
+                                         flickr_train_dataset, mg_train_dataset
+                                     ],
+                                     ignore_keys=['classes', 'palette']))
+
+test_pipeline = [
+    *_base_.test_pipeline[:-1],
+    dict(type='LoadText'),
+    dict(type='mmdet.PackDetInputs',
+         meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape',
+                    'scale_factor', 'pad_param', 'texts'))
+]
+coco_val_dataset = dict(
+    _delete_=True,
+    type='MultiModalDataset',
+    dataset=dict(type='YOLOv5LVISV1Dataset',
+                 data_root='data/coco/',
+                 test_mode=True,
+                 ann_file='lvis/lvis_v1_minival_inserted_image_name.json',
+                 data_prefix=dict(img=''),
+                 batch_shapes_cfg=None),
+    class_text_path='data/texts/lvis_v1_class_texts.json',
+    pipeline=test_pipeline)
+val_dataloader = dict(dataset=coco_val_dataset)
+test_dataloader = val_dataloader
+
+val_evaluator = dict(type='mmdet.LVISMetric',
+                     ann_file='data/coco/lvis/lvis_v1_minival_inserted_image_name.json',
+                     metric='bbox')
+test_evaluator = val_evaluator
+
+# training settings
+default_hooks = dict(param_scheduler=dict(max_epochs=max_epochs),
+                     checkpoint=dict(interval=save_epoch_intervals,
+                                     rule='greater'))
+custom_hooks = [
+    dict(type='EMAHook',
+         ema_type='ExpMomentumEMA',
+         momentum=0.0001,
+         update_buffers=True,
+         strict_load=False,
+         priority=49),
+    dict(type='mmdet.PipelineSwitchHook',
+         switch_epoch=max_epochs - close_mosaic_epochs,
+         switch_pipeline=train_pipeline_stage2)
+]
+train_cfg = dict(max_epochs=max_epochs,
+                 val_interval=10,
+                 dynamic_intervals=[((max_epochs - close_mosaic_epochs),
+                                     _base_.val_interval_stage2)])
+optim_wrapper = dict(optimizer=dict(
+    _delete_=True,
+    type='AdamW',
+    lr=base_lr,
+    weight_decay=weight_decay,
+    batch_size_per_gpu=train_batch_size_per_gpu),
+                     paramwise_cfg=dict(bias_decay_mult=0.0,
+                                        norm_decay_mult=0.0,
+                                        custom_keys={
+                                            'backbone.text_model':
+                                            dict(lr_mult=0.01),
+                                            'logit_scale':
+                                            dict(weight_decay=0.0)
+                                        }),
+                     constructor='YOLOWv5OptimizerConstructor')

configs/pretrain/yolo_world_v2_x_vlpan_bn_2e-3_100e_4x8gpus_obj365v1_goldg_cc3mlite_train_lvis_minival.py

@@ -0,0 +1,183 @@
+_base_ = ('../../third_party/mmyolo/configs/yolov8/'
+          'yolov8_x_syncbn_fast_8xb16-500e_coco.py')
+custom_imports = dict(imports=['yolo_world'],
+                      allow_failed_imports=False)
+
+# hyper-parameters
+num_classes = 1203
+num_training_classes = 80
+max_epochs = 100  # Maximum training epochs
+close_mosaic_epochs = 2
+save_epoch_intervals = 2
+text_channels = 512
+neck_embed_channels = [128, 256, _base_.last_stage_out_channels // 2]
+neck_num_heads = [4, 8, _base_.last_stage_out_channels // 2 // 32]
+base_lr = 2e-3
+weight_decay = 0.05 / 2
+train_batch_size_per_gpu = 16
+# text_model_name = '../pretrained_models/clip-vit-base-patch32-projection'
+text_model_name = 'openai/clip-vit-base-patch32'
+# model settings
+model = dict(
+    type='YOLOWorldDetector',
+    mm_neck=True,
+    num_train_classes=num_training_classes,
+    num_test_classes=num_classes,
+    data_preprocessor=dict(type='YOLOWDetDataPreprocessor'),
+    backbone=dict(
+        _delete_=True,
+        type='MultiModalYOLOBackbone',
+        image_model={{_base_.model.backbone}},
+        text_model=dict(
+            type='HuggingCLIPLanguageBackbone',
+            model_name=text_model_name,
+            frozen_modules=['all'])),
+    neck=dict(type='YOLOWorldPAFPN',
+              guide_channels=text_channels,
+              embed_channels=neck_embed_channels,
+              num_heads=neck_num_heads,
+              block_cfg=dict(type='MaxSigmoidCSPLayerWithTwoConv')),
+    bbox_head=dict(type='YOLOWorldHead',
+                   head_module=dict(type='YOLOWorldHeadModule',
+                                    use_bn_head=True,
+                                    embed_dims=text_channels,
+                                    num_classes=num_training_classes)),
+    train_cfg=dict(assigner=dict(num_classes=num_training_classes)))
+
+# dataset settings
+text_transform = [
+    dict(type='RandomLoadText',
+         num_neg_samples=(num_classes, num_classes),
+         max_num_samples=num_training_classes,
+         padding_to_max=True,
+         padding_value=''),
+    dict(type='mmdet.PackDetInputs',
+         meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape', 'flip',
+                    'flip_direction', 'texts'))
+]
+train_pipeline = [
+    *_base_.pre_transform,
+    dict(type='MultiModalMosaic',
+         img_scale=_base_.img_scale,
+         pad_val=114.0,
+         pre_transform=_base_.pre_transform),
+    dict(
+        type='YOLOv5RandomAffine',
+        max_rotate_degree=0.0,
+        max_shear_degree=0.0,
+        scaling_ratio_range=(1 - _base_.affine_scale, 1 + _base_.affine_scale),
+        max_aspect_ratio=_base_.max_aspect_ratio,
+        border=(-_base_.img_scale[0] // 2, -_base_.img_scale[1] // 2),
+        border_val=(114, 114, 114)),
+    *_base_.last_transform[:-1],
+    *text_transform,
+]
+train_pipeline_stage2 = [*_base_.train_pipeline_stage2[:-1], *text_transform]
+obj365v1_train_dataset = dict(
+    type='MultiModalDataset',
+    dataset=dict(
+        type='YOLOv5Objects365V1Dataset',
+        data_root='data/objects365v1/',
+        ann_file='annotations/objects365_train.json',
+        data_prefix=dict(img='train/'),
+        filter_cfg=dict(filter_empty_gt=False, min_size=32)),
+    class_text_path='data/texts/obj365v1_class_texts.json',
+    pipeline=train_pipeline)
+
+mg_train_dataset = dict(type='YOLOv5MixedGroundingDataset',
+                        data_root='data/mixed_grounding/',
+                        ann_file='annotations/final_mixed_train_no_coco.json',
+                        data_prefix=dict(img='gqa/images/'),
+                        filter_cfg=dict(filter_empty_gt=False, min_size=32),
+                        pipeline=train_pipeline)
+
+flickr_train_dataset = dict(
+    type='YOLOv5MixedGroundingDataset',
+    data_root='data/flickr/',
+    ann_file='annotations/final_flickr_separateGT_train.json',
+    data_prefix=dict(img='full_images/'),
+    filter_cfg=dict(filter_empty_gt=True, min_size=32),
+    pipeline=train_pipeline)
+
+
+cc3m_train_dataset = dict(type='YOLOv5GeneralGroundingDataset',
+                          data_root='data/cc3m/',
+                          ann_file='annotations/cc3m_pseudo_annotations.json',
+                          data_prefix=dict(img='training'),
+                          filter_cfg=dict(filter_empty_gt=True, min_size=32),
+                          pipeline=train_pipeline)
+
+
+train_dataloader = dict(batch_size=train_batch_size_per_gpu,
+                        collate_fn=dict(type='yolow_collate'),
+                        dataset=dict(_delete_=True,
+                                     type='ConcatDataset',
+                                     datasets=[
+                                         obj365v1_train_dataset,
+                                         flickr_train_dataset,
+                                         mg_train_dataset,
+                                         cc3m_train_dataset
+                                     ],
+                                     ignore_keys=['classes', 'palette']))
+
+test_pipeline = [
+    *_base_.test_pipeline[:-1],
+    dict(type='LoadText'),
+    dict(type='mmdet.PackDetInputs',
+         meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape',
+                    'scale_factor', 'pad_param', 'texts'))
+]
+
+coco_val_dataset = dict(
+    _delete_=True,
+    type='MultiModalDataset',
+    dataset=dict(type='YOLOv5LVISV1Dataset',
+                 data_root='data/coco/',
+                 test_mode=True,
+                 ann_file='lvis/lvis_v1_minival_inserted_image_name.json',
+                 data_prefix=dict(img=''),
+                 batch_shapes_cfg=None),
+    class_text_path='data/texts/lvis_v1_class_texts.json',
+    pipeline=test_pipeline)
+val_dataloader = dict(dataset=coco_val_dataset)
+test_dataloader = val_dataloader
+
+val_evaluator = dict(type='mmdet.LVISMetric',
+                     ann_file='data/coco/lvis/lvis_v1_minival_inserted_image_name.json',
+                     metric='bbox')
+test_evaluator = val_evaluator
+
+# training settings
+default_hooks = dict(param_scheduler=dict(max_epochs=max_epochs),
+                     checkpoint=dict(interval=save_epoch_intervals,
+                                     rule='greater'))
+custom_hooks = [
+    dict(type='EMAHook',
+         ema_type='ExpMomentumEMA',
+         momentum=0.0001,
+         update_buffers=True,
+         strict_load=False,
+         priority=49),
+    dict(type='mmdet.PipelineSwitchHook',
+         switch_epoch=max_epochs - close_mosaic_epochs,
+         switch_pipeline=train_pipeline_stage2)
+]
+train_cfg = dict(max_epochs=max_epochs,
+                 val_interval=10,
+                 dynamic_intervals=[((max_epochs - close_mosaic_epochs),
+                                     _base_.val_interval_stage2)])
+optim_wrapper = dict(optimizer=dict(
+    _delete_=True,
+    type='AdamW',
+    lr=base_lr,
+    weight_decay=weight_decay,
+    batch_size_per_gpu=train_batch_size_per_gpu),
+                     paramwise_cfg=dict(bias_decay_mult=0.0,
+                                        norm_decay_mult=0.0,
+                                        custom_keys={
+                                            'backbone.text_model':
+                                            dict(lr_mult=0.01),
+                                            'logit_scale':
+                                            dict(weight_decay=0.0)
+                                        }),
+                     constructor='YOLOWv5OptimizerConstructor')

configs/pretrain/yolo_world_v2_x_vlpan_bn_2e-3_100e_4x8gpus_obj365v1_goldg_train_1280ft_lvis_minival.py

@@ -0,0 +1,199 @@
+_base_ = ('../../third_party/mmyolo/configs/yolov8/'
+          'yolov8_x_syncbn_fast_8xb16-500e_coco.py')
+custom_imports = dict(imports=['yolo_world'],
+                      allow_failed_imports=False)
+
+# hyper-parameters
+num_classes = 1203
+num_training_classes = 80
+max_epochs = 100  # Maximum training epochs
+close_mosaic_epochs = 2
+save_epoch_intervals = 2
+text_channels = 512
+neck_embed_channels = [128, 256, _base_.last_stage_out_channels // 2]
+neck_num_heads = [4, 8, _base_.last_stage_out_channels // 2 // 32]
+base_lr = 2e-3
+weight_decay = 0.05 / 2
+train_batch_size_per_gpu = 16
+text_model_name = '../pretrained_models/clip-vit-base-patch32-projection'
+# text_model_name = 'openai/clip-vit-base-patch32'
+img_scale = (1280, 1280)
+
+# model settings
+model = dict(
+    type='YOLOWorldDetector',
+    mm_neck=True,
+    num_train_classes=num_training_classes,
+    num_test_classes=num_classes,
+    data_preprocessor=dict(type='YOLOWDetDataPreprocessor'),
+    backbone=dict(
+        _delete_=True,
+        type='MultiModalYOLOBackbone',
+        image_model={{_base_.model.backbone}},
+        text_model=dict(
+            type='HuggingCLIPLanguageBackbone',
+            model_name=text_model_name,
+            frozen_modules=['all'])),
+    neck=dict(type='YOLOWorldPAFPN',
+              guide_channels=text_channels,
+              embed_channels=neck_embed_channels,
+              num_heads=neck_num_heads,
+              block_cfg=dict(type='MaxSigmoidCSPLayerWithTwoConv')),
+    bbox_head=dict(type='YOLOWorldHead',
+                   head_module=dict(type='YOLOWorldHeadModule',
+                                    use_bn_head=True,
+                                    embed_dims=text_channels,
+                                    num_classes=num_training_classes)),
+    train_cfg=dict(assigner=dict(num_classes=num_training_classes)))
+
+# dataset settings
+text_transform = [
+    dict(type='RandomLoadText',
+         num_neg_samples=(num_classes, num_classes),
+         max_num_samples=num_training_classes,
+         padding_to_max=True,
+         padding_value=''),
+    dict(type='mmdet.PackDetInputs',
+         meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape', 'flip',
+                    'flip_direction', 'texts'))
+]
+train_pipeline = [
+    *_base_.pre_transform,
+    dict(type='MultiModalMosaic',
+         img_scale=img_scale,
+         pad_val=114.0,
+         pre_transform=_base_.pre_transform),
+    dict(
+        type='YOLOv5RandomAffine',
+        max_rotate_degree=0.0,
+        max_shear_degree=0.0,
+        scaling_ratio_range=(1 - _base_.affine_scale, 1 + _base_.affine_scale),
+        max_aspect_ratio=_base_.max_aspect_ratio,
+        border=(-img_scale[0] // 2, -img_scale[1] // 2),
+        border_val=(114, 114, 114)),
+    *_base_.last_transform[:-1],
+    *text_transform,
+]
+train_pipeline_stage2 = [
+    *_base_.pre_transform,
+    dict(type='YOLOv5KeepRatioResize', scale=img_scale),
+    dict(
+        type='LetterResize',
+        scale=img_scale,
+        allow_scale_up=True,
+        pad_val=dict(img=114.0)),
+    dict(
+        type='YOLOv5RandomAffine',
+        max_rotate_degree=0.0,
+        max_shear_degree=0.0,
+        scaling_ratio_range=(1 - _base_.affine_scale, 1 + _base_.affine_scale),
+        max_aspect_ratio=_base_.max_aspect_ratio,
+        border_val=(114, 114, 114)),
+    *_base_.last_transform[:-1],
+    *text_transform
+]
+
+obj365v1_train_dataset = dict(
+    type='MultiModalDataset',
+    dataset=dict(
+        type='YOLOv5Objects365V1Dataset',
+        data_root='data/objects365v1/',
+        ann_file='annotations/objects365_train.json',
+        data_prefix=dict(img='train/'),
+        filter_cfg=dict(filter_empty_gt=False, min_size=32)),
+    class_text_path='data/texts/obj365v1_class_texts.json',
+    pipeline=train_pipeline)
+
+mg_train_dataset = dict(type='YOLOv5MixedGroundingDataset',
+                        data_root='data/mixed_grounding/',
+                        ann_file='annotations/final_mixed_train_no_coco.json',
+                        data_prefix=dict(img='gqa/images/'),
+                        filter_cfg=dict(filter_empty_gt=False, min_size=32),
+                        pipeline=train_pipeline)
+
+flickr_train_dataset = dict(
+    type='YOLOv5MixedGroundingDataset',
+    data_root='data/flickr/',
+    ann_file='annotations/final_flickr_separateGT_train.json',
+    data_prefix=dict(img='full_images/'),
+    filter_cfg=dict(filter_empty_gt=True, min_size=32),
+    pipeline=train_pipeline)
+
+train_dataloader = dict(batch_size=train_batch_size_per_gpu,
+                        collate_fn=dict(type='yolow_collate'),
+                        dataset=dict(_delete_=True,
+                                     type='ConcatDataset',
+                                     datasets=[
+                                         obj365v1_train_dataset,
+                                         flickr_train_dataset, mg_train_dataset
+                                     ],
+                                     ignore_keys=['classes', 'palette']))
+
+test_pipeline = [
+    dict(type='LoadImageFromFile'),
+    dict(type='YOLOv5KeepRatioResize', scale=img_scale),
+    dict(
+        type='LetterResize',
+        scale=img_scale,
+        allow_scale_up=False,
+        pad_val=dict(img=114)),
+    dict(type='LoadAnnotations', with_bbox=True, _scope_='mmdet'),
+    dict(type='LoadText'),
+    dict(type='mmdet.PackDetInputs',
+         meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape',
+                    'scale_factor', 'pad_param', 'texts'))
+]
+
+coco_val_dataset = dict(
+    _delete_=True,
+    type='MultiModalDataset',
+    dataset=dict(type='YOLOv5LVISV1Dataset',
+                 data_root='data/coco/',
+                 test_mode=True,
+                 ann_file='lvis/lvis_v1_minival_inserted_image_name.json',
+                 data_prefix=dict(img=''),
+                 batch_shapes_cfg=None),
+    class_text_path='data/texts/lvis_v1_class_texts.json',
+    pipeline=test_pipeline)
+val_dataloader = dict(dataset=coco_val_dataset)
+test_dataloader = val_dataloader
+
+val_evaluator = dict(type='mmdet.LVISMetric',
+                     ann_file='data/coco/lvis/lvis_v1_minival_inserted_image_name.json',
+                     metric='bbox')
+test_evaluator = val_evaluator
+
+# training settings
+default_hooks = dict(param_scheduler=dict(max_epochs=max_epochs),
+                     checkpoint=dict(interval=save_epoch_intervals,
+                                     rule='greater'))
+custom_hooks = [
+    dict(type='EMAHook',
+         ema_type='ExpMomentumEMA',
+         momentum=0.0001,
+         update_buffers=True,
+         strict_load=False,
+         priority=49),
+    dict(type='mmdet.PipelineSwitchHook',
+         switch_epoch=max_epochs - close_mosaic_epochs,
+         switch_pipeline=train_pipeline_stage2)
+]
+train_cfg = dict(max_epochs=max_epochs,
+                 val_interval=10,
+                 dynamic_intervals=[((max_epochs - close_mosaic_epochs),
+                                     _base_.val_interval_stage2)])
+optim_wrapper = dict(optimizer=dict(
+    _delete_=True,
+    type='AdamW',
+    lr=base_lr,
+    weight_decay=weight_decay,
+    batch_size_per_gpu=train_batch_size_per_gpu),
+                     paramwise_cfg=dict(bias_decay_mult=0.0,
+                                        norm_decay_mult=0.0,
+                                        custom_keys={
+                                            'backbone.text_model':
+                                            dict(lr_mult=0.01),
+                                            'logit_scale':
+                                            dict(weight_decay=0.0)
+                                        }),
+                     constructor='YOLOWv5OptimizerConstructor')

configs/pretrain/yolo_world_v2_x_vlpan_bn_2e-3_100e_4x8gpus_obj365v1_goldg_train_lvis_minival.py

@@ -0,0 +1,171 @@
+_base_ = ('../../third_party/mmyolo/configs/yolov8/'
+          'yolov8_x_syncbn_fast_8xb16-500e_coco.py')
+custom_imports = dict(imports=['yolo_world'],
+                      allow_failed_imports=False)
+
+# hyper-parameters
+num_classes = 1203
+num_training_classes = 80
+max_epochs = 100  # Maximum training epochs
+close_mosaic_epochs = 2
+save_epoch_intervals = 2
+text_channels = 512
+neck_embed_channels = [128, 256, _base_.last_stage_out_channels // 2]
+neck_num_heads = [4, 8, _base_.last_stage_out_channels // 2 // 32]
+base_lr = 2e-3
+weight_decay = 0.05 / 2
+train_batch_size_per_gpu = 16
+# text_model_name = '../pretrained_models/clip-vit-base-patch32-projection'
+text_model_name = 'openai/clip-vit-base-patch32'
+# model settings
+model = dict(
+    type='YOLOWorldDetector',
+    mm_neck=True,
+    num_train_classes=num_training_classes,
+    num_test_classes=num_classes,
+    data_preprocessor=dict(type='YOLOWDetDataPreprocessor'),
+    backbone=dict(
+        _delete_=True,
+        type='MultiModalYOLOBackbone',
+        image_model={{_base_.model.backbone}},
+        text_model=dict(
+            type='HuggingCLIPLanguageBackbone',
+            model_name=text_model_name,
+            frozen_modules=['all'])),
+    neck=dict(type='YOLOWorldPAFPN',
+              guide_channels=text_channels,
+              embed_channels=neck_embed_channels,
+              num_heads=neck_num_heads,
+              block_cfg=dict(type='MaxSigmoidCSPLayerWithTwoConv')),
+    bbox_head=dict(type='YOLOWorldHead',
+                   head_module=dict(type='YOLOWorldHeadModule',
+                                    use_bn_head=True,
+                                    embed_dims=text_channels,
+                                    num_classes=num_training_classes)),
+    train_cfg=dict(assigner=dict(num_classes=num_training_classes)))
+
+# dataset settings
+text_transform = [
+    dict(type='RandomLoadText',
+         num_neg_samples=(num_classes, num_classes),
+         max_num_samples=num_training_classes,
+         padding_to_max=True,
+         padding_value=''),
+    dict(type='mmdet.PackDetInputs',
+         meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape', 'flip',
+                    'flip_direction', 'texts'))
+]
+train_pipeline = [
+    *_base_.pre_transform,
+    dict(type='MultiModalMosaic',
+         img_scale=_base_.img_scale,
+         pad_val=114.0,
+         pre_transform=_base_.pre_transform),
+    dict(
+        type='YOLOv5RandomAffine',
+        max_rotate_degree=0.0,
+        max_shear_degree=0.0,
+        scaling_ratio_range=(1 - _base_.affine_scale, 1 + _base_.affine_scale),
+        max_aspect_ratio=_base_.max_aspect_ratio,
+        border=(-_base_.img_scale[0] // 2, -_base_.img_scale[1] // 2),
+        border_val=(114, 114, 114)),
+    *_base_.last_transform[:-1],
+    *text_transform,
+]
+train_pipeline_stage2 = [*_base_.train_pipeline_stage2[:-1], *text_transform]
+obj365v1_train_dataset = dict(
+    type='MultiModalDataset',
+    dataset=dict(
+        type='YOLOv5Objects365V1Dataset',
+        data_root='data/objects365v1/',
+        ann_file='annotations/objects365_train.json',
+        data_prefix=dict(img='train/'),
+        filter_cfg=dict(filter_empty_gt=False, min_size=32)),
+    class_text_path='data/texts/obj365v1_class_texts.json',
+    pipeline=train_pipeline)
+
+mg_train_dataset = dict(type='YOLOv5MixedGroundingDataset',
+                        data_root='data/mixed_grounding/',
+                        ann_file='annotations/final_mixed_train_no_coco.json',
+                        data_prefix=dict(img='gqa/images/'),
+                        filter_cfg=dict(filter_empty_gt=False, min_size=32),
+                        pipeline=train_pipeline)
+
+flickr_train_dataset = dict(
+    type='YOLOv5MixedGroundingDataset',
+    data_root='data/flickr/',
+    ann_file='annotations/final_flickr_separateGT_train.json',
+    data_prefix=dict(img='full_images/'),
+    filter_cfg=dict(filter_empty_gt=True, min_size=32),
+    pipeline=train_pipeline)
+
+train_dataloader = dict(batch_size=train_batch_size_per_gpu,
+                        collate_fn=dict(type='yolow_collate'),
+                        dataset=dict(_delete_=True,
+                                     type='ConcatDataset',
+                                     datasets=[
+                                         obj365v1_train_dataset,
+                                         flickr_train_dataset, mg_train_dataset
+                                     ],
+                                     ignore_keys=['classes', 'palette']))
+
+test_pipeline = [
+    *_base_.test_pipeline[:-1],
+    dict(type='LoadText'),
+    dict(type='mmdet.PackDetInputs',
+         meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape',
+                    'scale_factor', 'pad_param', 'texts'))
+]
+coco_val_dataset = dict(
+    _delete_=True,
+    type='MultiModalDataset',
+    dataset=dict(type='YOLOv5LVISV1Dataset',
+                 data_root='data/coco/',
+                 test_mode=True,
+                 ann_file='lvis/lvis_v1_minival_inserted_image_name.json',
+                 data_prefix=dict(img=''),
+                 batch_shapes_cfg=None),
+    class_text_path='data/texts/lvis_v1_class_texts.json',
+    pipeline=test_pipeline)
+val_dataloader = dict(dataset=coco_val_dataset)
+test_dataloader = val_dataloader
+
+val_evaluator = dict(type='mmdet.LVISMetric',
+                     ann_file='data/coco/lvis/lvis_v1_minival_inserted_image_name.json',
+                     metric='bbox')
+test_evaluator = val_evaluator
+
+# training settings
+default_hooks = dict(param_scheduler=dict(max_epochs=max_epochs),
+                     checkpoint=dict(interval=save_epoch_intervals,
+                                     rule='greater'))
+custom_hooks = [
+    dict(type='EMAHook',
+         ema_type='ExpMomentumEMA',
+         momentum=0.0001,
+         update_buffers=True,
+         strict_load=False,
+         priority=49),
+    dict(type='mmdet.PipelineSwitchHook',
+         switch_epoch=max_epochs - close_mosaic_epochs,
+         switch_pipeline=train_pipeline_stage2)
+]
+train_cfg = dict(max_epochs=max_epochs,
+                 val_interval=10,
+                 dynamic_intervals=[((max_epochs - close_mosaic_epochs),
+                                     _base_.val_interval_stage2)])
+optim_wrapper = dict(optimizer=dict(
+    _delete_=True,
+    type='AdamW',
+    lr=base_lr,
+    weight_decay=weight_decay,
+    batch_size_per_gpu=train_batch_size_per_gpu),
+                     paramwise_cfg=dict(bias_decay_mult=0.0,
+                                        norm_decay_mult=0.0,
+                                        custom_keys={
+                                            'backbone.text_model':
+                                            dict(lr_mult=0.01),
+                                            'logit_scale':
+                                            dict(weight_decay=0.0)
+                                        }),
+                     constructor='YOLOWv5OptimizerConstructor')

configs/pretrain/yolo_world_v2_xl_vlpan_bn_2e-3_100e_4x8gpus_obj365v1_goldg_train_lvis_minival.py

@@ -0,0 +1,185 @@
+_base_ = ('../../third_party/mmyolo/configs/yolov8/'
+          'yolov8_x_syncbn_fast_8xb16-500e_coco.py')
+custom_imports = dict(imports=['yolo_world'],
+                      allow_failed_imports=False)
+
+# hyper-parameters
+num_classes = 1203
+num_training_classes = 80
+max_epochs = 100  # Maximum training epochs
+close_mosaic_epochs = 2
+save_epoch_intervals = 2
+text_channels = 512
+neck_embed_channels = [128, 256, _base_.last_stage_out_channels // 2]
+neck_num_heads = [4, 8, _base_.last_stage_out_channels // 2 // 32]
+base_lr = 2e-3
+weight_decay = 0.05 / 2
+train_batch_size_per_gpu = 16
+text_model_name = '../pretrained_models/clip-vit-base-patch32-projection'
+text_model_name = 'openai/clip-vit-base-patch32'
+
+# scaling model from X to XL
+deepen_factor = 1.0
+widen_factor = 1.5
+
+backbone = _base_.model.backbone
+backbone.update(
+    deepen_factor=deepen_factor,
+    widen_factor=widen_factor
+)
+
+# model settings
+model = dict(
+    type='YOLOWorldDetector',
+    mm_neck=True,
+    num_train_classes=num_training_classes,
+    num_test_classes=num_classes,
+    data_preprocessor=dict(type='YOLOWDetDataPreprocessor'),
+    backbone=dict(
+        _delete_=True,
+        type='MultiModalYOLOBackbone',
+        image_model=backbone,
+        text_model=dict(
+            type='HuggingCLIPLanguageBackbone',
+            model_name=text_model_name,
+            frozen_modules=['all'])),
+    neck=dict(type='YOLOWorldPAFPN',
+              deepen_factor=deepen_factor,
+              widen_factor=widen_factor,
+              guide_channels=text_channels,
+              embed_channels=neck_embed_channels,
+              num_heads=neck_num_heads,
+              block_cfg=dict(type='MaxSigmoidCSPLayerWithTwoConv')),
+    bbox_head=dict(type='YOLOWorldHead',
+                   head_module=dict(type='YOLOWorldHeadModule',
+                                    widen_factor=widen_factor,
+                                    use_bn_head=True,
+                                    embed_dims=text_channels,
+                                    num_classes=num_training_classes)),
+    train_cfg=dict(assigner=dict(num_classes=num_training_classes)))
+
+# dataset settings
+text_transform = [
+    dict(type='RandomLoadText',
+         num_neg_samples=(num_classes, num_classes),
+         max_num_samples=num_training_classes,
+         padding_to_max=True,
+         padding_value=''),
+    dict(type='mmdet.PackDetInputs',
+         meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape', 'flip',
+                    'flip_direction', 'texts'))
+]
+train_pipeline = [
+    *_base_.pre_transform,
+    dict(type='MultiModalMosaic',
+         img_scale=_base_.img_scale,
+         pad_val=114.0,
+         pre_transform=_base_.pre_transform),
+    dict(
+        type='YOLOv5RandomAffine',
+        max_rotate_degree=0.0,
+        max_shear_degree=0.0,
+        scaling_ratio_range=(1 - _base_.affine_scale, 1 + _base_.affine_scale),
+        max_aspect_ratio=_base_.max_aspect_ratio,
+        border=(-_base_.img_scale[0] // 2, -_base_.img_scale[1] // 2),
+        border_val=(114, 114, 114)),
+    *_base_.last_transform[:-1],
+    *text_transform,
+]
+train_pipeline_stage2 = [*_base_.train_pipeline_stage2[:-1], *text_transform]
+obj365v1_train_dataset = dict(
+    type='MultiModalDataset',
+    dataset=dict(
+        type='YOLOv5Objects365V1Dataset',
+        data_root='data/objects365v1/',
+        ann_file='annotations/objects365_train.json',
+        data_prefix=dict(img='train/'),
+        filter_cfg=dict(filter_empty_gt=False, min_size=32)),
+    class_text_path='data/texts/obj365v1_class_texts.json',
+    pipeline=train_pipeline)
+
+mg_train_dataset = dict(type='YOLOv5MixedGroundingDataset',
+                        data_root='data/mixed_grounding/',
+                        ann_file='annotations/final_mixed_train_no_coco.json',
+                        data_prefix=dict(img='gqa/images/'),
+                        filter_cfg=dict(filter_empty_gt=False, min_size=32),
+                        pipeline=train_pipeline)
+
+flickr_train_dataset = dict(
+    type='YOLOv5MixedGroundingDataset',
+    data_root='data/flickr/',
+    ann_file='annotations/final_flickr_separateGT_train.json',
+    data_prefix=dict(img='full_images/'),
+    filter_cfg=dict(filter_empty_gt=True, min_size=32),
+    pipeline=train_pipeline)
+
+train_dataloader = dict(batch_size=train_batch_size_per_gpu,
+                        collate_fn=dict(type='yolow_collate'),
+                        dataset=dict(_delete_=True,
+                                     type='ConcatDataset',
+                                     datasets=[
+                                         obj365v1_train_dataset,
+                                         flickr_train_dataset, mg_train_dataset
+                                     ],
+                                     ignore_keys=['classes', 'palette']))
+
+test_pipeline = [
+    *_base_.test_pipeline[:-1],
+    dict(type='LoadText'),
+    dict(type='mmdet.PackDetInputs',
+         meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape',
+                    'scale_factor', 'pad_param', 'texts'))
+]
+coco_val_dataset = dict(
+    _delete_=True,
+    type='MultiModalDataset',
+    dataset=dict(type='YOLOv5LVISV1Dataset',
+                 data_root='data/coco/',
+                 test_mode=True,
+                 ann_file='lvis/lvis_v1_minival_inserted_image_name.json',
+                 data_prefix=dict(img=''),
+                 batch_shapes_cfg=None),
+    class_text_path='data/texts/lvis_v1_class_texts.json',
+    pipeline=test_pipeline)
+val_dataloader = dict(dataset=coco_val_dataset)
+test_dataloader = val_dataloader
+
+val_evaluator = dict(type='mmdet.LVISMetric',
+                     ann_file='data/coco/lvis/lvis_v1_minival_inserted_image_name.json',
+                     metric='bbox')
+test_evaluator = val_evaluator
+
+# training settings
+default_hooks = dict(param_scheduler=dict(max_epochs=max_epochs),
+                     checkpoint=dict(interval=save_epoch_intervals,
+                                     rule='greater'))
+custom_hooks = [
+    dict(type='EMAHook',
+         ema_type='ExpMomentumEMA',
+         momentum=0.0001,
+         update_buffers=True,
+         strict_load=False,
+         priority=49),
+    dict(type='mmdet.PipelineSwitchHook',
+         switch_epoch=max_epochs - close_mosaic_epochs,
+         switch_pipeline=train_pipeline_stage2)
+]
+train_cfg = dict(max_epochs=max_epochs,
+                 val_interval=10,
+                 dynamic_intervals=[((max_epochs - close_mosaic_epochs),
+                                     _base_.val_interval_stage2)])
+optim_wrapper = dict(optimizer=dict(
+    _delete_=True,
+    type='AdamW',
+    lr=base_lr,
+    weight_decay=weight_decay,
+    batch_size_per_gpu=train_batch_size_per_gpu),
+                     paramwise_cfg=dict(bias_decay_mult=0.0,
+                                        norm_decay_mult=0.0,
+                                        custom_keys={
+                                            'backbone.text_model':
+                                            dict(lr_mult=0.01),
+                                            'logit_scale':
+                                            dict(weight_decay=0.0)
+                                        }),
+                     constructor='YOLOWv5OptimizerConstructor')

configs/pretrain_v1/README.md

@@ -0,0 +1,21 @@
+## Pre-training YOLO-World-v1
+
+> The YOLO-World-v1 is an initial version, and now is nearly deprecated! We strongly suggest you use the [latest version](../pretrain/).
+
+
+
+### Zero-shot Inference on LVIS dataset
+
+| model                                                                                                                | Pre-train Data       | Size | AP<sup>mini</su> | AP<sub>r</sub> | AP<sub>c</sub> | AP<sub>f</sub> | AP<sup>val</su> | AP<sub>r</sub> | AP<sub>c</sub> | AP<sub>f</sub> |                                                                                        weights                                                                                         |
+| :------------------------------------------------------------------------------------------------------------------- | :------------------- | :----------------- | :--------------: | :------------: | :------------: | :------------: | :-------------: | :------------: | :------------: | :------------: | :------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------: |
+| [YOLO-World-S](./yolo_world_s_dual_vlpan_l2norm_2e-3_100e_4x8gpus_obj365v1_goldg_train_lvis_minival.py)   | O365+GoldG           |        640        |       24.3       |      16.6      |      22.1      |      27.7      |      17.8       |      11.0      |      14.8      |      24.0      |    [HF Checkpoints 🤗](https://huggingface.co/wondervictor/YOLO-World/resolve/main/yolo_world_s_clip_base_dual_vlpan_2e-3adamw_32xb16_100e_o365_goldg_train_pretrained-18bea4d2.pth)    |
+| [YOLO-World-M](./yolo_world_m_dual_l2norm_2e-4_100e_4x8gpus_obj365v1_goldg_train_lvis_minival.py)   | O365+GoldG           |        640        |       28.6       |      19.7      |      26.6      |      31.9      |      22.3       |      16.2      |      19.0      |      28.7      |    [HF Checkpoints 🤗](https://huggingface.co/wondervictor/YOLO-World/resolve/main/yolo_world_m_clip_base_dual_vlpan_2e-3adamw_32xb16_100e_o365_goldg_train_pretrained-2b7bd1be.pth)    |
+| [YOLO-World-L](./yolo_world_l_dual_vlpan_l2norm_2e-3_100e_4x8gpus_obj365v1_goldg_train_lvis_minival.py)   | O365+GoldG           |        640        |       32.5       |      22.3      |      30.6      |      36.1      |      24.8       |      17.8      |      22.4      |      32.5      |    [HF Checkpoints 🤗](https://huggingface.co/wondervictor/YOLO-World/resolve/main/yolo_world_l_clip_base_dual_vlpan_2e-3adamw_32xb16_100e_o365_goldg_train_pretrained-0e566235.pth)    |
+| [YOLO-World-L](./yolo_world_l_dual_vlpan_l2norm_2e-3_100e_4x8gpus_obj365v1_goldg_train_lvis_minival.py) | O365+GoldG+CC3M-Lite |        640        |       33.0       |      23.6      |      32.0      |      35.5      |      25.3       |      18.0      |      22.1      |      32.1      | [HF Checkpoints 🤗](https://huggingface.co/wondervictor/YOLO-World/blob/main/yolo_world_l_clip_base_dual_vlpan_2e-3adamw_32xb16_100e_o365_goldg_cc3mlite_train_pretrained-7a5eea3b.pth) |
+| [YOLO-World-X](./yolo_world_x_dual_vlpan_l2norm_2e-3_100e_4x8gpus_obj365v1_goldg_train_lvis_minival.py) | O365+GoldG+CC3M-Lite | 640 | 33.4 | 24.4 | 31.6 | 36.6 | 26.6 | 19.2 | 23.5 | 33.2 | [HF Checkpoints 🤗](https://huggingface.co/wondervictor/YOLO-World/blob/main/yolo_world_x_clip_base_dual_vlpan_2e-3adamw_32xb16_100e_o365_goldg_cc3mlite_train_pretrained-8cf6b025.pth) |
+
+
+**NOTE:**
+1. AP<sup>mini</sup>: evaluated on LVIS `minival`.
+3. AP<sup>val</sup>: evaluated on LVIS `val 1.0`.
+4. [HuggingFace Mirror](https://hf-mirror.com/) provides the mirror of HuggingFace, which is a choice for users who are unable to reach.

configs/pretrain_v1/yolo_world_l_dual_vlpan_l2norm_2e-3_100e_4x8gpus_obj365v1_goldg_train_lvis_minival.py

@@ -0,0 +1,172 @@
+_base_ = ('../../third_party/mmyolo/configs/yolov8/'
+          'yolov8_l_syncbn_fast_8xb16-500e_coco.py')
+custom_imports = dict(imports=['yolo_world'],
+                      allow_failed_imports=False)
+
+# hyper-parameters
+num_classes = 1203
+num_training_classes = 80
+max_epochs = 100  # Maximum training epochs
+close_mosaic_epochs = 2
+save_epoch_intervals = 2
+text_channels = 512
+neck_embed_channels = [128, 256, _base_.last_stage_out_channels // 2]
+neck_num_heads = [4, 8, _base_.last_stage_out_channels // 2 // 32]
+base_lr = 2e-3
+weight_decay = 0.05 / 2
+train_batch_size_per_gpu = 16
+
+# model settings
+model = dict(
+    type='YOLOWorldDetector',
+    mm_neck=True,
+    num_train_classes=num_training_classes,
+    num_test_classes=num_classes,
+    data_preprocessor=dict(type='YOLOWDetDataPreprocessor'),
+    backbone=dict(
+        _delete_=True,
+        type='MultiModalYOLOBackbone',
+        image_model={{_base_.model.backbone}},
+        text_model=dict(
+            type='HuggingCLIPLanguageBackbone',
+            model_name='openai/clip-vit-base-patch32',
+            frozen_modules=['all'])),
+    neck=dict(type='YOLOWorldDualPAFPN',
+              guide_channels=text_channels,
+              embed_channels=neck_embed_channels,
+              num_heads=neck_num_heads,
+              block_cfg=dict(type='MaxSigmoidCSPLayerWithTwoConv'),
+              text_enhancder=dict(type='ImagePoolingAttentionModule',
+                                  embed_channels=256,
+                                  num_heads=8)),
+    bbox_head=dict(type='YOLOWorldHead',
+                   head_module=dict(type='YOLOWorldHeadModule',
+                                    embed_dims=text_channels,
+                                    num_classes=num_training_classes)),
+    train_cfg=dict(assigner=dict(num_classes=num_training_classes)))
+
+# dataset settings
+text_transform = [
+    dict(type='RandomLoadText',
+         num_neg_samples=(num_classes, num_classes),
+         max_num_samples=num_training_classes,
+         padding_to_max=True,
+         padding_value=''),
+    dict(type='mmdet.PackDetInputs',
+         meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape', 'flip',
+                    'flip_direction', 'texts'))
+]
+train_pipeline = [
+    *_base_.pre_transform,
+    dict(type='MultiModalMosaic',
+         img_scale=_base_.img_scale,
+         pad_val=114.0,
+         pre_transform=_base_.pre_transform),
+    dict(
+        type='YOLOv5RandomAffine',
+        max_rotate_degree=0.0,
+        max_shear_degree=0.0,
+        scaling_ratio_range=(1 - _base_.affine_scale, 1 + _base_.affine_scale),
+        max_aspect_ratio=_base_.max_aspect_ratio,
+        border=(-_base_.img_scale[0] // 2, -_base_.img_scale[1] // 2),
+        border_val=(114, 114, 114)),
+    *_base_.last_transform[:-1],
+    *text_transform,
+]
+train_pipeline_stage2 = [*_base_.train_pipeline_stage2[:-1], *text_transform]
+obj365v1_train_dataset = dict(
+    type='MultiModalDataset',
+    dataset=dict(
+        type='YOLOv5Objects365V1Dataset',
+        data_root='data/objects365v1/',
+        ann_file='annotations/objects365_train.json',
+        data_prefix=dict(img='train/'),
+        filter_cfg=dict(filter_empty_gt=False, min_size=32)),
+    class_text_path='data/texts/obj365v1_class_texts.json',
+    pipeline=train_pipeline)
+
+mg_train_dataset = dict(type='YOLOv5MixedGroundingDataset',
+                        data_root='data/mixed_grounding/',
+                        ann_file='annotations/final_mixed_train_no_coco.json',
+                        data_prefix=dict(img='gqa/images/'),
+                        filter_cfg=dict(filter_empty_gt=False, min_size=32),
+                        pipeline=train_pipeline)
+
+flickr_train_dataset = dict(
+    type='YOLOv5MixedGroundingDataset',
+    data_root='data/flickr/',
+    ann_file='annotations/final_flickr_separateGT_train.json',
+    data_prefix=dict(img='full_images/'),
+    filter_cfg=dict(filter_empty_gt=True, min_size=32),
+    pipeline=train_pipeline)
+
+train_dataloader = dict(batch_size=train_batch_size_per_gpu,
+                        collate_fn=dict(type='yolow_collate'),
+                        dataset=dict(_delete_=True,
+                                     type='ConcatDataset',
+                                     datasets=[
+                                         obj365v1_train_dataset,
+                                         flickr_train_dataset, mg_train_dataset
+                                     ],
+                                     ignore_keys=['classes', 'palette']))
+
+test_pipeline = [
+    *_base_.test_pipeline[:-1],
+    dict(type='LoadText'),
+    dict(type='mmdet.PackDetInputs',
+         meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape',
+                    'scale_factor', 'pad_param', 'texts'))
+]
+coco_val_dataset = dict(
+    _delete_=True,
+    type='MultiModalDataset',
+    dataset=dict(type='YOLOv5LVISV1Dataset',
+                 data_root='data/coco/',
+                 test_mode=True,
+                 ann_file='lvis/lvis_v1_minival_inserted_image_name.json',
+                 data_prefix=dict(img=''),
+                 batch_shapes_cfg=None),
+    class_text_path='data/texts/lvis_v1_class_texts.json',
+    pipeline=test_pipeline)
+val_dataloader = dict(dataset=coco_val_dataset)
+test_dataloader = val_dataloader
+
+val_evaluator = dict(type='mmdet.LVISMetric',
+                     ann_file='data/coco/lvis/lvis_v1_minival_inserted_image_name.json',
+                     metric='bbox')
+test_evaluator = val_evaluator
+
+# training settings
+default_hooks = dict(param_scheduler=dict(max_epochs=max_epochs),
+                     checkpoint=dict(interval=save_epoch_intervals,
+                                     rule='greater'))
+custom_hooks = [
+    dict(type='EMAHook',
+         ema_type='ExpMomentumEMA',
+         momentum=0.0001,
+         update_buffers=True,
+         strict_load=False,
+         priority=49),
+    dict(type='mmdet.PipelineSwitchHook',
+         switch_epoch=max_epochs - close_mosaic_epochs,
+         switch_pipeline=train_pipeline_stage2)
+]
+train_cfg = dict(max_epochs=max_epochs,
+                 val_interval=10,
+                 dynamic_intervals=[((max_epochs - close_mosaic_epochs),
+                                     _base_.val_interval_stage2)])
+optim_wrapper = dict(optimizer=dict(
+    _delete_=True,
+    type='AdamW',
+    lr=base_lr,
+    weight_decay=weight_decay,
+    batch_size_per_gpu=train_batch_size_per_gpu),
+                     paramwise_cfg=dict(bias_decay_mult=0.0,
+                                        norm_decay_mult=0.0,
+                                        custom_keys={
+                                            'backbone.text_model':
+                                            dict(lr_mult=0.01),
+                                            'logit_scale':
+                                            dict(weight_decay=0.0)
+                                        }),
+                     constructor='YOLOWv5OptimizerConstructor')

configs/pretrain_v1/yolo_world_l_dual_vlpan_l2norm_2e-3_100e_4x8gpus_obj365v1_goldg_train_lvis_val.py

@@ -0,0 +1,172 @@
+_base_ = ('../../third_party/mmyolo/configs/yolov8/'
+          'yolov8_l_syncbn_fast_8xb16-500e_coco.py')
+custom_imports = dict(imports=['yolo_world'],
+                      allow_failed_imports=False)
+
+# hyper-parameters
+num_classes = 1203
+num_training_classes = 80
+max_epochs = 100  # Maximum training epochs
+close_mosaic_epochs = 2
+save_epoch_intervals = 2
+text_channels = 512
+neck_embed_channels = [128, 256, _base_.last_stage_out_channels // 2]
+neck_num_heads = [4, 8, _base_.last_stage_out_channels // 2 // 32]
+base_lr = 2e-3
+weight_decay = 0.05 / 2
+train_batch_size_per_gpu = 16
+
+# model settings
+model = dict(
+    type='YOLOWorldDetector',
+    mm_neck=True,
+    num_train_classes=num_training_classes,
+    num_test_classes=num_classes,
+    data_preprocessor=dict(type='YOLOWDetDataPreprocessor'),
+    backbone=dict(
+        _delete_=True,
+        type='MultiModalYOLOBackbone',
+        image_model={{_base_.model.backbone}},
+        text_model=dict(
+            type='HuggingCLIPLanguageBackbone',
+            model_name='openai/clip-vit-base-patch32',
+            frozen_modules=['all'])),
+    neck=dict(type='YOLOWorldDualPAFPN',
+              guide_channels=text_channels,
+              embed_channels=neck_embed_channels,
+              num_heads=neck_num_heads,
+              block_cfg=dict(type='MaxSigmoidCSPLayerWithTwoConv'),
+              text_enhancder=dict(type='ImagePoolingAttentionModule',
+                                  embed_channels=256,
+                                  num_heads=8)),
+    bbox_head=dict(type='YOLOWorldHead',
+                   head_module=dict(type='YOLOWorldHeadModule',
+                                    embed_dims=text_channels,
+                                    num_classes=num_training_classes)),
+    train_cfg=dict(assigner=dict(num_classes=num_training_classes)))
+
+# dataset settings
+text_transform = [
+    dict(type='RandomLoadText',
+         num_neg_samples=(num_classes, num_classes),
+         max_num_samples=num_training_classes,
+         padding_to_max=True,
+         padding_value=''),
+    dict(type='mmdet.PackDetInputs',
+         meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape', 'flip',
+                    'flip_direction', 'texts'))
+]
+train_pipeline = [
+    *_base_.pre_transform,
+    dict(type='MultiModalMosaic',
+         img_scale=_base_.img_scale,
+         pad_val=114.0,
+         pre_transform=_base_.pre_transform),
+    dict(
+        type='YOLOv5RandomAffine',
+        max_rotate_degree=0.0,
+        max_shear_degree=0.0,
+        scaling_ratio_range=(1 - _base_.affine_scale, 1 + _base_.affine_scale),
+        max_aspect_ratio=_base_.max_aspect_ratio,
+        border=(-_base_.img_scale[0] // 2, -_base_.img_scale[1] // 2),
+        border_val=(114, 114, 114)),
+    *_base_.last_transform[:-1],
+    *text_transform,
+]
+train_pipeline_stage2 = [*_base_.train_pipeline_stage2[:-1], *text_transform]
+obj365v1_train_dataset = dict(
+    type='MultiModalDataset',
+    dataset=dict(
+        type='YOLOv5Objects365V1Dataset',
+        data_root='data/objects365v1/',
+        ann_file='annotations/objects365_train.json',
+        data_prefix=dict(img='train/'),
+        filter_cfg=dict(filter_empty_gt=False, min_size=32)),
+    class_text_path='data/texts/obj365v1_class_texts.json',
+    pipeline=train_pipeline)
+
+mg_train_dataset = dict(type='YOLOv5MixedGroundingDataset',
+                        data_root='data/mixed_grounding/',
+                        ann_file='annotations/final_mixed_train_no_coco.json',
+                        data_prefix=dict(img='gqa/images/'),
+                        filter_cfg=dict(filter_empty_gt=False, min_size=32),
+                        pipeline=train_pipeline)
+
+flickr_train_dataset = dict(
+    type='YOLOv5MixedGroundingDataset',
+    data_root='data/flickr/',
+    ann_file='annotations/final_flickr_separateGT_train.json',
+    data_prefix=dict(img='full_images/'),
+    filter_cfg=dict(filter_empty_gt=True, min_size=32),
+    pipeline=train_pipeline)
+
+train_dataloader = dict(batch_size=train_batch_size_per_gpu,
+                        collate_fn=dict(type='yolow_collate'),
+                        dataset=dict(_delete_=True,
+                                     type='ConcatDataset',
+                                     datasets=[
+                                         obj365v1_train_dataset,
+                                         flickr_train_dataset, mg_train_dataset
+                                     ],
+                                     ignore_keys=['classes', 'palette']))
+
+test_pipeline = [
+    *_base_.test_pipeline[:-1],
+    dict(type='LoadText'),
+    dict(type='mmdet.PackDetInputs',
+         meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape',
+                    'scale_factor', 'pad_param', 'texts'))
+]
+coco_val_dataset = dict(
+    _delete_=True,
+    type='MultiModalDataset',
+    dataset=dict(type='YOLOv5LVISV1Dataset',
+                 data_root='data/coco/',
+                 test_mode=True,
+                 ann_file='lvis/lvis_v1_val.json',
+                 data_prefix=dict(img=''),
+                 batch_shapes_cfg=None),
+    class_text_path='data/texts/lvis_v1_class_texts.json',
+    pipeline=test_pipeline)
+val_dataloader = dict(dataset=coco_val_dataset)
+test_dataloader = val_dataloader
+
+val_evaluator = dict(type='mmdet.LVISMetric',
+                     ann_file='data/coco/lvis/lvis_v1_val.json',
+                     metric='bbox')
+test_evaluator = val_evaluator
+
+# training settings
+default_hooks = dict(param_scheduler=dict(max_epochs=max_epochs),
+                     checkpoint=dict(interval=save_epoch_intervals,
+                                     rule='greater'))
+custom_hooks = [
+    dict(type='EMAHook',
+         ema_type='ExpMomentumEMA',
+         momentum=0.0001,
+         update_buffers=True,
+         strict_load=False,
+         priority=49),
+    dict(type='mmdet.PipelineSwitchHook',
+         switch_epoch=max_epochs - close_mosaic_epochs,
+         switch_pipeline=train_pipeline_stage2)
+]
+train_cfg = dict(max_epochs=max_epochs,
+                 val_interval=10,
+                 dynamic_intervals=[((max_epochs - close_mosaic_epochs),
+                                     _base_.val_interval_stage2)])
+optim_wrapper = dict(optimizer=dict(
+    _delete_=True,
+    type='AdamW',
+    lr=base_lr,
+    weight_decay=weight_decay,
+    batch_size_per_gpu=train_batch_size_per_gpu),
+                     paramwise_cfg=dict(bias_decay_mult=0.0,
+                                        norm_decay_mult=0.0,
+                                        custom_keys={
+                                            'backbone.text_model':
+                                            dict(lr_mult=0.01),
+                                            'logit_scale':
+                                            dict(weight_decay=0.0)
+                                        }),
+                     constructor='YOLOWv5OptimizerConstructor')

configs/pretrain_v1/yolo_world_m_dual_vlpan_l2norm_2e-3_100e_4x8gpus_obj365v1_goldg_train_lvis_minival.py

@@ -0,0 +1,172 @@
+_base_ = ('../../third_party/mmyolo/configs/yolov8/'
+          'yolov8_m_syncbn_fast_8xb16-500e_coco.py')
+custom_imports = dict(imports=['yolo_world'],
+                      allow_failed_imports=False)
+
+# hyper-parameters
+num_classes = 1203
+num_training_classes = 80
+max_epochs = 100  # Maximum training epochs
+close_mosaic_epochs = 2
+save_epoch_intervals = 2
+text_channels = 512
+neck_embed_channels = [128, 256, _base_.last_stage_out_channels // 2]
+neck_num_heads = [4, 8, _base_.last_stage_out_channels // 2 // 32]
+base_lr = 2e-3
+weight_decay = 0.05 / 2
+train_batch_size_per_gpu = 16
+
+# model settings
+model = dict(
+    type='YOLOWorldDetector',
+    mm_neck=True,
+    num_train_classes=num_training_classes,
+    num_test_classes=num_classes,
+    data_preprocessor=dict(type='YOLOWDetDataPreprocessor'),
+    backbone=dict(
+        _delete_=True,
+        type='MultiModalYOLOBackbone',
+        image_model={{_base_.model.backbone}},
+        text_model=dict(
+            type='HuggingCLIPLanguageBackbone',
+            model_name='openai/clip-vit-base-patch32',
+            frozen_modules=['all'])),
+    neck=dict(type='YOLOWorldDualPAFPN',
+              guide_channels=text_channels,
+              embed_channels=neck_embed_channels,
+              num_heads=neck_num_heads,
+              block_cfg=dict(type='MaxSigmoidCSPLayerWithTwoConv'),
+              text_enhancder=dict(type='ImagePoolingAttentionModule',
+                                  embed_channels=256,
+                                  num_heads=8)),
+    bbox_head=dict(type='YOLOWorldHead',
+                   head_module=dict(type='YOLOWorldHeadModule',
+                                    embed_dims=text_channels,
+                                    num_classes=num_training_classes)),
+    train_cfg=dict(assigner=dict(num_classes=num_training_classes)))
+
+# dataset settings
+text_transform = [
+    dict(type='RandomLoadText',
+         num_neg_samples=(num_classes, num_classes),
+         max_num_samples=num_training_classes,
+         padding_to_max=True,
+         padding_value=''),
+    dict(type='mmdet.PackDetInputs',
+         meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape', 'flip',
+                    'flip_direction', 'texts'))
+]
+train_pipeline = [
+    *_base_.pre_transform,
+    dict(type='MultiModalMosaic',
+         img_scale=_base_.img_scale,
+         pad_val=114.0,
+         pre_transform=_base_.pre_transform),
+    dict(
+        type='YOLOv5RandomAffine',
+        max_rotate_degree=0.0,
+        max_shear_degree=0.0,
+        scaling_ratio_range=(1 - _base_.affine_scale, 1 + _base_.affine_scale),
+        max_aspect_ratio=_base_.max_aspect_ratio,
+        border=(-_base_.img_scale[0] // 2, -_base_.img_scale[1] // 2),
+        border_val=(114, 114, 114)),
+    *_base_.last_transform[:-1],
+    *text_transform,
+]
+train_pipeline_stage2 = [*_base_.train_pipeline_stage2[:-1], *text_transform]
+obj365v1_train_dataset = dict(
+    type='MultiModalDataset',
+    dataset=dict(
+        type='YOLOv5Objects365V1Dataset',
+        data_root='data/objects365v1/',
+        ann_file='annotations/objects365_train.json',
+        data_prefix=dict(img='train/'),
+        filter_cfg=dict(filter_empty_gt=False, min_size=32)),
+    class_text_path='data/texts/obj365v1_class_texts.json',
+    pipeline=train_pipeline)
+
+mg_train_dataset = dict(type='YOLOv5MixedGroundingDataset',
+                        data_root='data/mixed_grounding/',
+                        ann_file='annotations/final_mixed_train_no_coco.json',
+                        data_prefix=dict(img='gqa/images/'),
+                        filter_cfg=dict(filter_empty_gt=False, min_size=32),
+                        pipeline=train_pipeline)
+
+flickr_train_dataset = dict(
+    type='YOLOv5MixedGroundingDataset',
+    data_root='data/flickr/',
+    ann_file='annotations/final_flickr_separateGT_train.json',
+    data_prefix=dict(img='full_images/'),
+    filter_cfg=dict(filter_empty_gt=True, min_size=32),
+    pipeline=train_pipeline)
+
+train_dataloader = dict(batch_size=train_batch_size_per_gpu,
+                        collate_fn=dict(type='yolow_collate'),
+                        dataset=dict(_delete_=True,
+                                     type='ConcatDataset',
+                                     datasets=[
+                                         obj365v1_train_dataset,
+                                         flickr_train_dataset, mg_train_dataset
+                                     ],
+                                     ignore_keys=['classes', 'palette']))
+
+test_pipeline = [
+    *_base_.test_pipeline[:-1],
+    dict(type='LoadText'),
+    dict(type='mmdet.PackDetInputs',
+         meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape',
+                    'scale_factor', 'pad_param', 'texts'))
+]
+coco_val_dataset = dict(
+    _delete_=True,
+    type='MultiModalDataset',
+    dataset=dict(type='YOLOv5LVISV1Dataset',
+                 data_root='data/coco/',
+                 test_mode=True,
+                 ann_file='lvis/lvis_v1_minival_inserted_image_name.json',
+                 data_prefix=dict(img=''),
+                 batch_shapes_cfg=None),
+    class_text_path='data/texts/lvis_v1_class_texts.json',
+    pipeline=test_pipeline)
+val_dataloader = dict(dataset=coco_val_dataset)
+test_dataloader = val_dataloader
+
+val_evaluator = dict(type='mmdet.LVISMetric',
+                     ann_file='data/coco/lvis/lvis_v1_minival_inserted_image_name.json',
+                     metric='bbox')
+test_evaluator = val_evaluator
+
+# training settings
+default_hooks = dict(param_scheduler=dict(max_epochs=max_epochs),
+                     checkpoint=dict(interval=save_epoch_intervals,
+                                     rule='greater'))
+custom_hooks = [
+    dict(type='EMAHook',
+         ema_type='ExpMomentumEMA',
+         momentum=0.0001,
+         update_buffers=True,
+         strict_load=False,
+         priority=49),
+    dict(type='mmdet.PipelineSwitchHook',
+         switch_epoch=max_epochs - close_mosaic_epochs,
+         switch_pipeline=train_pipeline_stage2)
+]
+train_cfg = dict(max_epochs=max_epochs,
+                 val_interval=10,
+                 dynamic_intervals=[((max_epochs - close_mosaic_epochs),
+                                     _base_.val_interval_stage2)])
+optim_wrapper = dict(optimizer=dict(
+    _delete_=True,
+    type='AdamW',
+    lr=base_lr,
+    weight_decay=weight_decay,
+    batch_size_per_gpu=train_batch_size_per_gpu),
+                     paramwise_cfg=dict(bias_decay_mult=0.0,
+                                        norm_decay_mult=0.0,
+                                        custom_keys={
+                                            'backbone.text_model':
+                                            dict(lr_mult=0.01),
+                                            'logit_scale':
+                                            dict(weight_decay=0.0)
+                                        }),
+                     constructor='YOLOWv5OptimizerConstructor')

configs/pretrain_v1/yolo_world_s_dual_vlpan_l2norm_2e-3_100e_4x8gpus_obj365v1_goldg_train_lvis_minival.py

@@ -0,0 +1,172 @@
+_base_ = ('../../third_party/mmyolo/configs/yolov8/'
+          'yolov8_s_syncbn_fast_8xb16-500e_coco.py')
+custom_imports = dict(imports=['yolo_world'],
+                      allow_failed_imports=False)
+
+# hyper-parameters
+num_classes = 1203
+num_training_classes = 80
+max_epochs = 100  # Maximum training epochs
+close_mosaic_epochs = 2
+save_epoch_intervals = 2
+text_channels = 512
+neck_embed_channels = [128, 256, _base_.last_stage_out_channels // 2]
+neck_num_heads = [4, 8, _base_.last_stage_out_channels // 2 // 32]
+base_lr = 2e-3
+weight_decay = 0.05 / 2
+train_batch_size_per_gpu = 16
+
+# model settings
+model = dict(
+    type='YOLOWorldDetector',
+    mm_neck=True,
+    num_train_classes=num_training_classes,
+    num_test_classes=num_classes,
+    data_preprocessor=dict(type='YOLOWDetDataPreprocessor'),
+    backbone=dict(
+        _delete_=True,
+        type='MultiModalYOLOBackbone',
+        image_model={{_base_.model.backbone}},
+        text_model=dict(
+            type='HuggingCLIPLanguageBackbone',
+            model_name='openai/clip-vit-base-patch32',
+            frozen_modules=['all'])),
+    neck=dict(type='YOLOWorldDualPAFPN',
+              guide_channels=text_channels,
+              embed_channels=neck_embed_channels,
+              num_heads=neck_num_heads,
+              block_cfg=dict(type='MaxSigmoidCSPLayerWithTwoConv'),
+              text_enhancder=dict(type='ImagePoolingAttentionModule',
+                                  embed_channels=256,
+                                  num_heads=8)),
+    bbox_head=dict(type='YOLOWorldHead',
+                   head_module=dict(type='YOLOWorldHeadModule',
+                                    embed_dims=text_channels,
+                                    num_classes=num_training_classes)),
+    train_cfg=dict(assigner=dict(num_classes=num_training_classes)))
+
+# dataset settings
+text_transform = [
+    dict(type='RandomLoadText',
+         num_neg_samples=(num_classes, num_classes),
+         max_num_samples=num_training_classes,
+         padding_to_max=True,
+         padding_value=''),
+    dict(type='mmdet.PackDetInputs',
+         meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape', 'flip',
+                    'flip_direction', 'texts'))
+]
+train_pipeline = [
+    *_base_.pre_transform,
+    dict(type='MultiModalMosaic',
+         img_scale=_base_.img_scale,
+         pad_val=114.0,
+         pre_transform=_base_.pre_transform),
+    dict(
+        type='YOLOv5RandomAffine',
+        max_rotate_degree=0.0,
+        max_shear_degree=0.0,
+        scaling_ratio_range=(1 - _base_.affine_scale, 1 + _base_.affine_scale),
+        max_aspect_ratio=_base_.max_aspect_ratio,
+        border=(-_base_.img_scale[0] // 2, -_base_.img_scale[1] // 2),
+        border_val=(114, 114, 114)),
+    *_base_.last_transform[:-1],
+    *text_transform,
+]
+train_pipeline_stage2 = [*_base_.train_pipeline_stage2[:-1], *text_transform]
+obj365v1_train_dataset = dict(
+    type='MultiModalDataset',
+    dataset=dict(
+        type='YOLOv5Objects365V1Dataset',
+        data_root='data/objects365v1/',
+        ann_file='annotations/objects365_train.json',
+        data_prefix=dict(img='train/'),
+        filter_cfg=dict(filter_empty_gt=False, min_size=32)),
+    class_text_path='data/texts/obj365v1_class_texts.json',
+    pipeline=train_pipeline)
+
+mg_train_dataset = dict(type='YOLOv5MixedGroundingDataset',
+                        data_root='data/mixed_grounding/',
+                        ann_file='annotations/final_mixed_train_no_coco.json',
+                        data_prefix=dict(img='gqa/images/'),
+                        filter_cfg=dict(filter_empty_gt=False, min_size=32),
+                        pipeline=train_pipeline)
+
+flickr_train_dataset = dict(
+    type='YOLOv5MixedGroundingDataset',
+    data_root='data/flickr/',
+    ann_file='annotations/final_flickr_separateGT_train.json',
+    data_prefix=dict(img='full_images/'),
+    filter_cfg=dict(filter_empty_gt=True, min_size=32),
+    pipeline=train_pipeline)
+
+train_dataloader = dict(batch_size=train_batch_size_per_gpu,
+                        collate_fn=dict(type='yolow_collate'),
+                        dataset=dict(_delete_=True,
+                                     type='ConcatDataset',
+                                     datasets=[
+                                         obj365v1_train_dataset,
+                                         flickr_train_dataset, mg_train_dataset
+                                     ],
+                                     ignore_keys=['classes', 'palette']))
+
+test_pipeline = [
+    *_base_.test_pipeline[:-1],
+    dict(type='LoadText'),
+    dict(type='mmdet.PackDetInputs',
+         meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape',
+                    'scale_factor', 'pad_param', 'texts'))
+]
+coco_val_dataset = dict(
+    _delete_=True,
+    type='MultiModalDataset',
+    dataset=dict(type='YOLOv5LVISV1Dataset',
+                 data_root='data/coco/',
+                 test_mode=True,
+                 ann_file='lvis/lvis_v1_minival_inserted_image_name.json',
+                 data_prefix=dict(img=''),
+                 batch_shapes_cfg=None),
+    class_text_path='data/texts/lvis_v1_class_texts.json',
+    pipeline=test_pipeline)
+val_dataloader = dict(dataset=coco_val_dataset)
+test_dataloader = val_dataloader
+
+val_evaluator = dict(type='mmdet.LVISMetric',
+                     ann_file='data/coco/lvis/lvis_v1_minival_inserted_image_name.json',
+                     metric='bbox')
+test_evaluator = val_evaluator
+
+# training settings
+default_hooks = dict(param_scheduler=dict(max_epochs=max_epochs),
+                     checkpoint=dict(interval=save_epoch_intervals,
+                                     rule='greater'))
+custom_hooks = [
+    dict(type='EMAHook',
+         ema_type='ExpMomentumEMA',
+         momentum=0.0001,
+         update_buffers=True,
+         strict_load=False,
+         priority=49),
+    dict(type='mmdet.PipelineSwitchHook',
+         switch_epoch=max_epochs - close_mosaic_epochs,
+         switch_pipeline=train_pipeline_stage2)
+]
+train_cfg = dict(max_epochs=max_epochs,
+                 val_interval=10,
+                 dynamic_intervals=[((max_epochs - close_mosaic_epochs),
+                                     _base_.val_interval_stage2)])
+optim_wrapper = dict(optimizer=dict(
+    _delete_=True,
+    type='AdamW',
+    lr=base_lr,
+    weight_decay=weight_decay,
+    batch_size_per_gpu=train_batch_size_per_gpu),
+                     paramwise_cfg=dict(bias_decay_mult=0.0,
+                                        norm_decay_mult=0.0,
+                                        custom_keys={
+                                            'backbone.text_model':
+                                            dict(lr_mult=0.01),
+                                            'logit_scale':
+                                            dict(weight_decay=0.0)
+                                        }),
+                     constructor='YOLOWv5OptimizerConstructor')

configs/pretrain_v1/yolo_world_x_dual_vlpan_l2norm_2e-3_100e_4x8gpus_obj365v1_goldg_train_lvis_minival.py

@@ -0,0 +1,172 @@
+_base_ = ('../../third_party/mmyolo/configs/yolov8/'
+          'yolov8_x_syncbn_fast_8xb16-500e_coco.py')
+custom_imports = dict(imports=['yolo_world'],
+                      allow_failed_imports=False)
+
+# hyper-parameters
+num_classes = 1203
+num_training_classes = 80
+max_epochs = 100  # Maximum training epochs
+close_mosaic_epochs = 2
+save_epoch_intervals = 2
+text_channels = 512
+neck_embed_channels = [128, 256, _base_.last_stage_out_channels // 2]
+neck_num_heads = [4, 8, _base_.last_stage_out_channels // 2 // 32]
+base_lr = 2e-3
+weight_decay = 0.05 / 2
+train_batch_size_per_gpu = 16
+
+# model settings
+model = dict(
+    type='YOLOWorldDetector',
+    mm_neck=True,
+    num_train_classes=num_training_classes,
+    num_test_classes=num_classes,
+    data_preprocessor=dict(type='YOLOWDetDataPreprocessor'),
+    backbone=dict(
+        _delete_=True,
+        type='MultiModalYOLOBackbone',
+        image_model={{_base_.model.backbone}},
+        text_model=dict(
+            type='HuggingCLIPLanguageBackbone',
+            model_name='openai/clip-vit-base-patch32',
+            frozen_modules=['all'])),
+    neck=dict(type='YOLOWorldDualPAFPN',
+              guide_channels=text_channels,
+              embed_channels=neck_embed_channels,
+              num_heads=neck_num_heads,
+              block_cfg=dict(type='MaxSigmoidCSPLayerWithTwoConv'),
+              text_enhancder=dict(type='ImagePoolingAttentionModule',
+                                  embed_channels=256,
+                                  num_heads=8)),
+    bbox_head=dict(type='YOLOWorldHead',
+                   head_module=dict(type='YOLOWorldHeadModule',
+                                    embed_dims=text_channels,
+                                    num_classes=num_training_classes)),
+    train_cfg=dict(assigner=dict(num_classes=num_training_classes)))
+
+# dataset settings
+text_transform = [
+    dict(type='RandomLoadText',
+         num_neg_samples=(num_classes, num_classes),
+         max_num_samples=num_training_classes,
+         padding_to_max=True,
+         padding_value=''),
+    dict(type='mmdet.PackDetInputs',
+         meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape', 'flip',
+                    'flip_direction', 'texts'))
+]
+train_pipeline = [
+    *_base_.pre_transform,
+    dict(type='MultiModalMosaic',
+         img_scale=_base_.img_scale,
+         pad_val=114.0,
+         pre_transform=_base_.pre_transform),
+    dict(
+        type='YOLOv5RandomAffine',
+        max_rotate_degree=0.0,
+        max_shear_degree=0.0,
+        scaling_ratio_range=(1 - _base_.affine_scale, 1 + _base_.affine_scale),
+        max_aspect_ratio=_base_.max_aspect_ratio,
+        border=(-_base_.img_scale[0] // 2, -_base_.img_scale[1] // 2),
+        border_val=(114, 114, 114)),
+    *_base_.last_transform[:-1],
+    *text_transform,
+]
+train_pipeline_stage2 = [*_base_.train_pipeline_stage2[:-1], *text_transform]
+obj365v1_train_dataset = dict(
+    type='MultiModalDataset',
+    dataset=dict(
+        type='YOLOv5Objects365V1Dataset',
+        data_root='data/objects365v1/',
+        ann_file='annotations/objects365_train.json',
+        data_prefix=dict(img='train/'),
+        filter_cfg=dict(filter_empty_gt=False, min_size=32)),
+    class_text_path='data/texts/obj365v1_class_texts.json',
+    pipeline=train_pipeline)
+
+mg_train_dataset = dict(type='YOLOv5MixedGroundingDataset',
+                        data_root='data/mixed_grounding/',
+                        ann_file='annotations/final_mixed_train_no_coco.json',
+                        data_prefix=dict(img='gqa/images/'),
+                        filter_cfg=dict(filter_empty_gt=False, min_size=32),
+                        pipeline=train_pipeline)
+
+flickr_train_dataset = dict(
+    type='YOLOv5MixedGroundingDataset',
+    data_root='data/flickr/',
+    ann_file='annotations/final_flickr_separateGT_train.json',
+    data_prefix=dict(img='full_images/'),
+    filter_cfg=dict(filter_empty_gt=True, min_size=32),
+    pipeline=train_pipeline)
+
+train_dataloader = dict(batch_size=train_batch_size_per_gpu,
+                        collate_fn=dict(type='yolow_collate'),
+                        dataset=dict(_delete_=True,
+                                     type='ConcatDataset',
+                                     datasets=[
+                                         obj365v1_train_dataset,
+                                         flickr_train_dataset, mg_train_dataset
+                                     ],
+                                     ignore_keys=['classes', 'palette']))
+
+test_pipeline = [
+    *_base_.test_pipeline[:-1],
+    dict(type='LoadText'),
+    dict(type='mmdet.PackDetInputs',
+         meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape',
+                    'scale_factor', 'pad_param', 'texts'))
+]
+coco_val_dataset = dict(
+    _delete_=True,
+    type='MultiModalDataset',
+    dataset=dict(type='YOLOv5LVISV1Dataset',
+                 data_root='data/coco/',
+                 test_mode=True,
+                 ann_file='lvis/lvis_v1_minival_inserted_image_name.json',
+                 data_prefix=dict(img=''),
+                 batch_shapes_cfg=None),
+    class_text_path='data/texts/lvis_v1_class_texts.json',
+    pipeline=test_pipeline)
+val_dataloader = dict(dataset=coco_val_dataset)
+test_dataloader = val_dataloader
+
+val_evaluator = dict(type='mmdet.LVISMetric',
+                     ann_file='data/coco/lvis/lvis_v1_minival_inserted_image_name.json',
+                     metric='bbox')
+test_evaluator = val_evaluator
+
+# training settings
+default_hooks = dict(param_scheduler=dict(max_epochs=max_epochs),
+                     checkpoint=dict(interval=save_epoch_intervals,
+                                     rule='greater'))
+custom_hooks = [
+    dict(type='EMAHook',
+         ema_type='ExpMomentumEMA',
+         momentum=0.0001,
+         update_buffers=True,
+         strict_load=False,
+         priority=49),
+    dict(type='mmdet.PipelineSwitchHook',
+         switch_epoch=max_epochs - close_mosaic_epochs,
+         switch_pipeline=train_pipeline_stage2)
+]
+train_cfg = dict(max_epochs=max_epochs,
+                 val_interval=10,
+                 dynamic_intervals=[((max_epochs - close_mosaic_epochs),
+                                     _base_.val_interval_stage2)])
+optim_wrapper = dict(optimizer=dict(
+    _delete_=True,
+    type='AdamW',
+    lr=base_lr,
+    weight_decay=weight_decay,
+    batch_size_per_gpu=train_batch_size_per_gpu),
+                     paramwise_cfg=dict(bias_decay_mult=0.0,
+                                        norm_decay_mult=0.0,
+                                        custom_keys={
+                                            'backbone.text_model':
+                                            dict(lr_mult=0.01),
+                                            'logit_scale':
+                                            dict(weight_decay=0.0)
+                                        }),
+                     constructor='YOLOWv5OptimizerConstructor')

configs/prompt_tuning_coco/READEME.md

@@ -0,0 +1,12 @@
+## Prompt Tuning for YOLO-World
+
+### NOTE:
+
+This folder contains many experimental config files, which will be removed later!!
+
+### Experimental Results
+
+| Model | Config |  AP  | AP50 | AP75  | APS | APM | APL |
+| :---- | :----: | :--: | :--: | :---: | :-: | :-: | :-: |
+| YOLO-World-v2-L | Zero-shot | 45.7 | 61.6 | 49.8 | 29.9 | 50.0 | 60.8 |
+| [YOLO-World-v2-L](./../configs/prompt_tuning_coco/yolo_world_v2_l_vlpan_bn_2e-4_80e_8gpus_mask-refine_prompt_tuning_coco.py) | Prompt tuning | 47.9 | 64.3 | 52.5 | 31.9 | 52.6 | 61.3 | 

configs/prompt_tuning_coco/yolo_world_v2_l_vlpan_bn_2e-4_80e_8gpus_image_prompt_demo.py

@@ -0,0 +1,126 @@
+_base_ = ('../../third_party/mmyolo/configs/yolov8/'
+          'yolov8_l_syncbn_fast_8xb16-500e_coco.py')
+custom_imports = dict(imports=['yolo_world'], allow_failed_imports=False)
+
+# hyper-parameters
+num_classes = 80
+num_training_classes = 80
+max_epochs = 80  # Maximum training epochs
+close_mosaic_epochs = 10
+save_epoch_intervals = 5
+text_channels = 512
+neck_embed_channels = [128, 256, _base_.last_stage_out_channels // 2]
+neck_num_heads = [4, 8, _base_.last_stage_out_channels // 2 // 32]
+base_lr = 2e-4
+weight_decay = 0.05
+train_batch_size_per_gpu = 16
+persistent_workers = False
+text_model_name = 'openai/clip-vit-base-patch32'
+img_scale = (800, 800)
+
+# model settings
+model = dict(type='YOLOWorldImageDetector',
+             mm_neck=True,
+             num_train_classes=num_training_classes,
+             num_test_classes=num_classes,
+             vision_model=text_model_name,
+             prompt_dim=text_channels,
+             data_preprocessor=dict(type='YOLOv5DetDataPreprocessor'),
+             backbone=dict(_delete_=True,
+                           type='MultiModalYOLOBackbone',
+                           image_model={{_base_.model.backbone}},
+                           frozen_stages=4,
+                           text_model=dict(type='HuggingCLIPLanguageBackbone',
+                                           model_name=text_model_name,
+                                           frozen_modules=['all'])),
+             neck=dict(type='YOLOWorldPAFPN',
+                       freeze_all=True,
+                       guide_channels=text_channels,
+                       embed_channels=neck_embed_channels,
+                       num_heads=neck_num_heads,
+                       block_cfg=dict(type='MaxSigmoidCSPLayerWithTwoConv')),
+             bbox_head=dict(type='YOLOWorldHead',
+                            head_module=dict(
+                                type='YOLOWorldHeadModule',
+                                freeze_all=True,
+                                use_bn_head=True,
+                                embed_dims=text_channels,
+                                num_classes=num_training_classes)),
+             train_cfg=dict(assigner=dict(num_classes=num_training_classes)))
+
+# dataset settings
+coco_train_dataset = dict(type='YOLOv5CocoDataset',
+                          data_root='data/coco',
+                          ann_file='annotations/instances_train2017.json',
+                          data_prefix=dict(img='train2017/'),
+                          filter_cfg=dict(filter_empty_gt=False, min_size=32),
+                          pipeline=_base_.train_pipeline)
+
+train_dataloader = dict(persistent_workers=persistent_workers,
+                        batch_size=train_batch_size_per_gpu,
+                        collate_fn=dict(type='yolow_collate'),
+                        dataset=coco_train_dataset)
+test_pipeline = [
+    *_base_.test_pipeline[:-1],
+    dict(type='LoadText'),
+    dict(
+        type='mmdet.PackDetInputs',
+        meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape',
+                   'scale_factor', 'pad_param', 'texts'))
+]
+coco_val_dataset = dict(
+    _delete_=True,
+    type='MultiModalDataset',
+    dataset=dict(
+        type='YOLOv5CocoDataset',
+        data_root='data/coco',
+        ann_file='annotations/instances_val2017.json',
+        data_prefix=dict(img='val2017/'),
+        filter_cfg=dict(filter_empty_gt=False, min_size=32)),
+    class_text_path='data/texts/coco_class_texts.json',
+    pipeline=test_pipeline)
+val_dataloader = dict(dataset=coco_val_dataset)
+test_dataloader = val_dataloader
+# training settings
+default_hooks = dict(param_scheduler=dict(scheduler_type='linear',
+                                          lr_factor=0.01,
+                                          max_epochs=max_epochs),
+                     checkpoint=dict(max_keep_ckpts=-1,
+                                     save_best=None,
+                                     interval=save_epoch_intervals))
+custom_hooks = [
+    dict(type='EMAHook',
+         ema_type='ExpMomentumEMA',
+         momentum=0.0001,
+         update_buffers=True,
+         strict_load=False,
+         priority=49),
+    dict(type='mmdet.PipelineSwitchHook',
+         switch_epoch=max_epochs - close_mosaic_epochs,
+         switch_pipeline=_base_.train_pipeline_stage2)
+]
+train_cfg = dict(max_epochs=max_epochs,
+                 val_interval=5,
+                 dynamic_intervals=[((max_epochs - close_mosaic_epochs),
+                                     _base_.val_interval_stage2)])
+
+optim_wrapper = dict(optimizer=dict(
+    _delete_=True,
+    type='AdamW',
+    lr=base_lr,
+    weight_decay=weight_decay,
+    batch_size_per_gpu=train_batch_size_per_gpu),
+                     paramwise_cfg=dict(
+                         custom_keys={
+                             'backbone.text_model': dict(lr_mult=0.01),
+                             'logit_scale': dict(weight_decay=0.0),
+                             'embeddings': dict(weight_decay=0.0)
+                         }),
+                     constructor='YOLOWv5OptimizerConstructor')
+
+# evaluation settings
+val_evaluator = dict(_delete_=True,
+                     type='mmdet.CocoMetric',
+                     proposal_nums=(100, 1, 10),
+                     ann_file='data/coco/annotations/instances_val2017.json',
+                     metric='bbox')

configs/prompt_tuning_coco/yolo_world_v2_l_vlpan_bn_2e-4_80e_8gpus_mask-refine_prompt_tuning_coco.py

@@ -0,0 +1,161 @@
+_base_ = ('../../third_party/mmyolo/configs/yolov8/'
+          'yolov8_l_mask-refine_syncbn_fast_8xb16-500e_coco.py')
+custom_imports = dict(imports=['yolo_world'], allow_failed_imports=False)
+
+# hyper-parameters
+num_classes = 80
+num_training_classes = 80
+max_epochs = 80  # Maximum training epochs
+close_mosaic_epochs = 10
+save_epoch_intervals = 5
+text_channels = 512
+neck_embed_channels = [128, 256, _base_.last_stage_out_channels // 2]
+neck_num_heads = [4, 8, _base_.last_stage_out_channels // 2 // 32]
+base_lr = 2e-3
+weight_decay = 0.05
+train_batch_size_per_gpu = 16
+load_from = 'pretrained_models/yolo_world_l_clip_t2i_bn_2e-3adamw_32xb16-100e_obj365v1_goldg_cc3mlite_train-ca93cd1f.pth'
+persistent_workers = False
+
+# model settings
+model = dict(type='SimpleYOLOWorldDetector',
+             mm_neck=True,
+             num_train_classes=num_training_classes,
+             num_test_classes=num_classes,
+             embedding_path='embeddings/clip_vit_b32_coco_80_embeddings.npy',
+             prompt_dim=text_channels,
+             num_prompts=80,
+             data_preprocessor=dict(type='YOLOv5DetDataPreprocessor'),
+             backbone=dict(_delete_=True,
+                           type='MultiModalYOLOBackbone',
+                           text_model=None,
+                           image_model={{_base_.model.backbone}},
+                           frozen_stages=4,
+                           with_text_model=False),
+             neck=dict(type='YOLOWorldPAFPN',
+                       freeze_all=True,
+                       guide_channels=text_channels,
+                       embed_channels=neck_embed_channels,
+                       num_heads=neck_num_heads,
+                       block_cfg=dict(type='MaxSigmoidCSPLayerWithTwoConv')),
+             bbox_head=dict(type='YOLOWorldHead',
+                            head_module=dict(
+                                type='YOLOWorldHeadModule',
+                                freeze_all=True,
+                                use_bn_head=True,
+                                embed_dims=text_channels,
+                                num_classes=num_training_classes)),
+             train_cfg=dict(assigner=dict(num_classes=num_training_classes)))
+
+# dataset settings
+final_transform = [
+    dict(type='mmdet.PackDetInputs',
+         meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape', 'flip',
+                    'flip_direction'))
+]
+mosaic_affine_transform = [
+    dict(type='Mosaic',
+         img_scale=_base_.img_scale,
+         pad_val=114.0,
+         pre_transform=_base_.pre_transform),
+    dict(type='YOLOv5CopyPaste', prob=_base_.copypaste_prob),
+    dict(
+        type='YOLOv5RandomAffine',
+        max_rotate_degree=0.0,
+        max_shear_degree=0.0,
+        max_aspect_ratio=100.,
+        scaling_ratio_range=(1 - _base_.affine_scale, 1 + _base_.affine_scale),
+        # img_scale is (width, height)
+        border=(-_base_.img_scale[0] // 2, -_base_.img_scale[1] // 2),
+        border_val=(114, 114, 114),
+        min_area_ratio=_base_.min_area_ratio,
+        use_mask_refine=_base_.use_mask2refine)
+]
+train_pipeline = [
+    *_base_.pre_transform, *mosaic_affine_transform,
+    dict(type='YOLOv5MixUp',
+         prob=_base_.mixup_prob,
+         pre_transform=[*_base_.pre_transform, *mosaic_affine_transform]),
+    *_base_.last_transform[:-1], *final_transform
+]
+
+train_pipeline_stage2 = [*_base_.train_pipeline_stage2[:-1], *final_transform]
+
+coco_train_dataset = dict(type='YOLOv5CocoDataset',
+                          data_root='data/coco',
+                          ann_file='annotations/instances_train2017.json',
+                          data_prefix=dict(img='train2017/'),
+                          filter_cfg=dict(filter_empty_gt=False, min_size=32),
+                          pipeline=train_pipeline)
+
+train_dataloader = dict(persistent_workers=persistent_workers,
+                        batch_size=train_batch_size_per_gpu,
+                        collate_fn=dict(type='yolow_collate'),
+                        dataset=coco_train_dataset)
+
+train_dataloader = dict(persistent_workers=persistent_workers,
+                        batch_size=train_batch_size_per_gpu,
+                        collate_fn=dict(type='yolow_collate'),
+                        dataset=coco_train_dataset)
+test_pipeline = [
+    *_base_.test_pipeline[:-1],
+    dict(type='mmdet.PackDetInputs',
+         meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape',
+                    'scale_factor', 'pad_param'))
+]
+coco_val_dataset = dict(type='YOLOv5CocoDataset',
+                        data_root='data/coco',
+                        ann_file='annotations/instances_val2017.json',
+                        data_prefix=dict(img='val2017/'),
+                        filter_cfg=dict(filter_empty_gt=False, min_size=32),
+                        pipeline=test_pipeline)
+
+val_dataloader = dict(dataset=coco_val_dataset)
+test_dataloader = val_dataloader
+# training settings
+default_hooks = dict(param_scheduler=dict(scheduler_type='linear',
+                                          lr_factor=0.01,
+                                          max_epochs=max_epochs),
+                     checkpoint=dict(max_keep_ckpts=-1,
+                                     save_best=None,
+                                     interval=save_epoch_intervals))
+custom_hooks = [
+    dict(type='EMAHook',
+         ema_type='ExpMomentumEMA',
+         momentum=0.0001,
+         update_buffers=True,
+         strict_load=False,
+         priority=49),
+    dict(type='mmdet.PipelineSwitchHook',
+         switch_epoch=max_epochs - close_mosaic_epochs,
+         switch_pipeline=train_pipeline_stage2)
+]
+train_cfg = dict(max_epochs=max_epochs,
+                 val_interval=5,
+                 dynamic_intervals=[((max_epochs - close_mosaic_epochs),
+                                     _base_.val_interval_stage2)])
+optim_wrapper = dict(optimizer=dict(
+    _delete_=True,
+    type='AdamW',
+    lr=base_lr,
+    weight_decay=weight_decay,
+    batch_size_per_gpu=train_batch_size_per_gpu),
+                     paramwise_cfg=dict(bias_decay_mult=0.0,
+                                        norm_decay_mult=0.0,
+                                        custom_keys={
+                                            'backbone.text_model':
+                                            dict(lr_mult=0.01),
+                                            'logit_scale':
+                                            dict(weight_decay=0.0),
+                                            'embeddings':
+                                            dict(weight_decay=0.0)
+                                        }),
+                     constructor='YOLOWv5OptimizerConstructor')
+
+# evaluation settings
+val_evaluator = dict(_delete_=True,
+                     type='mmdet.CocoMetric',
+                     proposal_nums=(100, 1, 10),
+                     ann_file='data/coco/annotations/instances_val2017.json',
+                     metric='bbox')
+find_unused_parameters = True

configs/prompt_tuning_coco/yolo_world_v2_l_vlpan_bn_2e-4_80e_8gpus_prompt_tuning_coco.py

@@ -0,0 +1,117 @@
+_base_ = ('../../third_party/mmyolo/configs/yolov8/'
+          'yolov8_l_syncbn_fast_8xb16-500e_coco.py')
+custom_imports = dict(imports=['yolo_world'], allow_failed_imports=False)
+
+# hyper-parameters
+num_classes = 80
+num_training_classes = 80
+max_epochs = 80  # Maximum training epochs
+close_mosaic_epochs = 10
+save_epoch_intervals = 5
+text_channels = 512
+neck_embed_channels = [128, 256, _base_.last_stage_out_channels // 2]
+neck_num_heads = [4, 8, _base_.last_stage_out_channels // 2 // 32]
+base_lr = 2e-4
+weight_decay = 0.05
+train_batch_size_per_gpu = 16
+load_from = 'pretrained_models/yolo_world_l_clip_t2i_bn_2e-3adamw_32xb16-100e_obj365v1_goldg_cc3mlite_train-ca93cd1f.pth'
+persistent_workers = False
+
+# model settings
+model = dict(type='SimpleYOLOWorldDetector',
+             mm_neck=True,
+             num_train_classes=num_training_classes,
+             num_test_classes=num_classes,
+             embedding_path='embeddings/clip_vit_b32_coco_80_embeddings.npy',
+             prompt_dim=text_channels,
+             num_prompts=80,
+             freeze_prompt=False,
+             data_preprocessor=dict(type='YOLOv5DetDataPreprocessor'),
+             backbone=dict(_delete_=True,
+                           type='MultiModalYOLOBackbone',
+                           text_model=None,
+                           image_model={{_base_.model.backbone}},
+                           frozen_stages=4,
+                           with_text_model=False),
+             neck=dict(type='YOLOWorldPAFPN',
+                       freeze_all=True,
+                       guide_channels=text_channels,
+                       embed_channels=neck_embed_channels,
+                       num_heads=neck_num_heads,
+                       block_cfg=dict(type='MaxSigmoidCSPLayerWithTwoConv')),
+             bbox_head=dict(type='YOLOWorldHead',
+                            head_module=dict(
+                                type='YOLOWorldHeadModule',
+                                freeze_all=True,
+                                use_bn_head=True,
+                                embed_dims=text_channels,
+                                num_classes=num_training_classes)),
+             train_cfg=dict(assigner=dict(num_classes=num_training_classes)))
+
+# dataset settings
+coco_train_dataset = dict(type='YOLOv5CocoDataset',
+                          data_root='data/coco',
+                          ann_file='annotations/instances_train2017.json',
+                          data_prefix=dict(img='train2017/'),
+                          filter_cfg=dict(filter_empty_gt=False, min_size=32),
+                          pipeline=_base_.train_pipeline)
+
+train_dataloader = dict(persistent_workers=persistent_workers,
+                        batch_size=train_batch_size_per_gpu,
+                        collate_fn=dict(type='yolow_collate'),
+                        dataset=coco_train_dataset)
+
+coco_val_dataset = dict(type='YOLOv5CocoDataset',
+                        data_root='data/coco',
+                        ann_file='annotations/instances_val2017.json',
+                        data_prefix=dict(img='val2017/'),
+                        filter_cfg=dict(filter_empty_gt=False, min_size=32),
+                        pipeline=_base_.test_pipeline)
+
+val_dataloader = dict(dataset=coco_val_dataset)
+test_dataloader = val_dataloader
+# training settings
+default_hooks = dict(param_scheduler=dict(scheduler_type='linear',
+                                          lr_factor=0.01,
+                                          max_epochs=max_epochs),
+                     checkpoint=dict(max_keep_ckpts=-1,
+                                     save_best=None,
+                                     interval=save_epoch_intervals))
+custom_hooks = [
+    dict(type='EMAHook',
+         ema_type='ExpMomentumEMA',
+         momentum=0.0001,
+         update_buffers=True,
+         strict_load=False,
+         priority=49),
+    dict(type='mmdet.PipelineSwitchHook',
+         switch_epoch=max_epochs - close_mosaic_epochs,
+         switch_pipeline=_base_.train_pipeline_stage2)
+]
+train_cfg = dict(max_epochs=max_epochs,
+                 val_interval=5,
+                 dynamic_intervals=[((max_epochs - close_mosaic_epochs),
+                                     _base_.val_interval_stage2)])
+
+optim_wrapper = dict(optimizer=dict(
+    _delete_=True,
+    type='AdamW',
+    lr=base_lr,
+    weight_decay=weight_decay,
+    batch_size_per_gpu=train_batch_size_per_gpu),
+                     paramwise_cfg=dict(custom_keys={
+                                            'backbone.text_model':
+                                            dict(lr_mult=0.01),
+                                            'logit_scale':
+                                            dict(weight_decay=0.0),
+                                            'embeddings':
+                                            dict(weight_decay=0.0)
+                                        }),
+                     constructor='YOLOWv5OptimizerConstructor')
+
+# evaluation settings
+val_evaluator = dict(_delete_=True,
+                     type='mmdet.CocoMetric',
+                     proposal_nums=(100, 1, 10),
+                     ann_file='data/coco/annotations/instances_val2017.json',
+                     metric='bbox')

configs/prompt_tuning_coco/yolo_world_v2_l_vlpan_bn_sgd_1e-3_80e_8gpus_all_finetuning_coco.py

@@ -0,0 +1,109 @@
+_base_ = ('../../third_party/mmyolo/configs/yolov8/'
+          'yolov8_l_syncbn_fast_8xb16-500e_coco.py')
+custom_imports = dict(imports=['yolo_world'], allow_failed_imports=False)
+
+# hyper-parameters
+num_classes = 80
+num_training_classes = 80
+max_epochs = 40  # Maximum training epochs
+close_mosaic_epochs = 10
+save_epoch_intervals = 5
+text_channels = 512
+neck_embed_channels = [128, 256, _base_.last_stage_out_channels // 2]
+neck_num_heads = [4, 8, _base_.last_stage_out_channels // 2 // 32]
+base_lr = 1e-3
+weight_decay = 0.0005
+train_batch_size_per_gpu = 16
+load_from = 'pretrained_models/yolo_world_l_clip_t2i_bn_2e-3adamw_32xb16-100e_obj365v1_goldg_cc3mlite_train-ca93cd1f.pth'
+persistent_workers = False
+
+# model settings
+model = dict(type='SimpleYOLOWorldDetector',
+             mm_neck=True,
+             num_train_classes=num_training_classes,
+             num_test_classes=num_classes,
+             embedding_path='embeddings/clip_vit_b32_coco_80_embeddings.npy',
+             prompt_dim=text_channels,
+             num_prompts=80,
+             freeze_prompt=True,
+             data_preprocessor=dict(type='YOLOv5DetDataPreprocessor'),
+             backbone=dict(_delete_=True,
+                           type='MultiModalYOLOBackbone',
+                           text_model=None,
+                           image_model={{_base_.model.backbone}},
+                           with_text_model=False),
+             neck=dict(type='YOLOWorldPAFPN',
+                       freeze_all=False,
+                       guide_channels=text_channels,
+                       embed_channels=neck_embed_channels,
+                       num_heads=neck_num_heads,
+                       block_cfg=dict(type='MaxSigmoidCSPLayerWithTwoConv')),
+             bbox_head=dict(type='YOLOWorldHead',
+                            head_module=dict(
+                                type='YOLOWorldHeadModule',
+                                freeze_all=False,
+                                use_bn_head=True,
+                                embed_dims=text_channels,
+                                num_classes=num_training_classes)),
+             train_cfg=dict(assigner=dict(num_classes=num_training_classes)))
+
+# dataset settings
+coco_train_dataset = dict(type='YOLOv5CocoDataset',
+                          data_root='data/coco',
+                          ann_file='annotations/instances_train2017.json',
+                          data_prefix=dict(img='train2017/'),
+                          filter_cfg=dict(filter_empty_gt=False, min_size=32),
+                          pipeline=_base_.train_pipeline)
+
+train_dataloader = dict(persistent_workers=persistent_workers,
+                        batch_size=train_batch_size_per_gpu,
+                        collate_fn=dict(type='yolow_collate'),
+                        dataset=coco_train_dataset)
+
+coco_val_dataset = dict(type='YOLOv5CocoDataset',
+                        data_root='data/coco',
+                        ann_file='annotations/instances_val2017.json',
+                        data_prefix=dict(img='val2017/'),
+                        filter_cfg=dict(filter_empty_gt=False, min_size=32),
+                        pipeline=_base_.test_pipeline)
+
+val_dataloader = dict(dataset=coco_val_dataset)
+test_dataloader = val_dataloader
+# training settings
+default_hooks = dict(param_scheduler=dict(scheduler_type='linear',
+                                          lr_factor=0.01,
+                                          max_epochs=max_epochs),
+                     checkpoint=dict(max_keep_ckpts=-1,
+                                     save_best=None,
+                                     interval=save_epoch_intervals))
+custom_hooks = [
+    dict(type='EMAHook',
+         ema_type='ExpMomentumEMA',
+         momentum=0.0001,
+         update_buffers=True,
+         strict_load=False,
+         priority=49),
+    dict(type='mmdet.PipelineSwitchHook',
+         switch_epoch=max_epochs - close_mosaic_epochs,
+         switch_pipeline=_base_.train_pipeline_stage2)
+]
+train_cfg = dict(max_epochs=max_epochs,
+                 val_interval=5,
+                 dynamic_intervals=[((max_epochs - close_mosaic_epochs),
+                                     _base_.val_interval_stage2)])
+
+optim_wrapper = dict(optimizer=dict(
+    _delete_=True,
+    type='SGD',
+    lr=base_lr,
+    momentum=0.937,
+    nesterov=True,
+    weight_decay=weight_decay,
+    batch_size_per_gpu=train_batch_size_per_gpu))
+
+# evaluation settings
+val_evaluator = dict(_delete_=True,
+                     type='mmdet.CocoMetric',
+                     proposal_nums=(100, 1, 10),
+                     ann_file='data/coco/annotations/instances_val2017.json',
+                     metric='bbox')