From 2330c5f1a899e33ca896ee91713c9c7072a45f12 Mon Sep 17 00:00:00 2001 From: BlenderNeko <126974546+BlenderNeko@users.noreply.github.com> Date: Mon, 29 May 2023 00:13:11 +0200 Subject: [PATCH] first commit --- LICENSE | 674 ++++++++++++++++++++++++++++++++++++++++++++ README.md | 27 ++ __init__.py | 3 + nodes.py | 117 ++++++++ seecoder.py | 570 +++++++++++++++++++++++++++++++++++++ seecoder_utils.py | 108 +++++++ seet_tdecoder.py | 697 ++++++++++++++++++++++++++++++++++++++++++++++ swin.py | 657 +++++++++++++++++++++++++++++++++++++++++++ 8 files changed, 2853 insertions(+) create mode 100644 LICENSE create mode 100644 README.md create mode 100644 __init__.py create mode 100644 nodes.py create mode 100644 seecoder.py create mode 100644 seecoder_utils.py create mode 100644 seet_tdecoder.py create mode 100644 swin.py diff --git a/LICENSE b/LICENSE new file mode 100644 index 0000000..f288702 --- /dev/null +++ b/LICENSE @@ -0,0 +1,674 @@ + GNU GENERAL PUBLIC LICENSE + Version 3, 29 June 2007 + + Copyright (C) 2007 Free Software Foundation, Inc. + 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. + + + Copyright (C) + + 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 . + +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: + + Copyright (C) + 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 +. + + 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 +. diff --git a/README.md b/README.md new file mode 100644 index 0000000..14a2cd6 --- /dev/null +++ b/README.md @@ -0,0 +1,27 @@ +# ComfyUI SeeCoder nodes + +This repo contains 2 experimental WIP nodes for [ComfyUI](https://github.com/comfyanonymous/ComfyUI) that let's you use [SeeCoders](https://github.com/SHI-Labs/Prompt-Free-Diffusion). + +## getting SeeCoders +You can find the seecoders [here](https://huggingface.co/shi-labs/prompt-free-diffusion). They have to be placed at `models/seecoders` + +## nodes: + +### SEECoderImageEncode + +this node can be used to create an embedding from an image + +- **image**: the image to encode +- **seecoder_name**: the name of the seecoder + +### ConcatConditioning + +this node can be used to concat different embeddings together, so you can e.g. create both a text and a visual embedding and concat them together. + +- **conditioning_to**: a set of embeddings to concat something to +- **conditioning_from**: the embedding to concat behind those in **conditioning_to** + +## TODO: + + - [ ] support for non safetensor formats + - [ ] bring attention layers in line with ones used in comfy \ No newline at end of file diff --git a/__init__.py b/__init__.py new file mode 100644 index 0000000..e74c2f2 --- /dev/null +++ b/__init__.py @@ -0,0 +1,3 @@ +from .nodes import NODE_CLASS_MAPPINGS + +__all__ = ['NODE_CLASS_MAPPINGS'] \ No newline at end of file diff --git a/nodes.py b/nodes.py new file mode 100644 index 0000000..ac4814e --- /dev/null +++ b/nodes.py @@ -0,0 +1,117 @@ + +import folder_paths +import torch +from seecoder.seecoder import SemanticExtractionEncoder, QueryTransformer, Decoder +from seecoder.swin import SwinTransformer +import safetensors.torch +import os +import sys + +sys.path.insert(0, os.path.join(os.path.dirname(os.path.realpath(__file__)), "comfy")) + +import comfy.model_management + +folder_paths.folder_names_and_paths["seecoder"] = ([os.path.join(folder_paths.models_dir, "seecoders")], folder_paths.supported_ckpt_extensions) + +_swine_config = { + "embed_dim" : 192, + "depths" : [ 2, 2, 18, 2 ], + "num_heads" : [ 6, 12, 24, 48 ], + "window_size" : 12, + "ape" : False, + "drop_path_rate" : 0.3, + "patch_norm" : True, +} + +_decoder_config = { + "inchannels" : {'res3' : 384, 'res4' : 768, 'res5' : 1536}, + "trans_input_tags" : ['res3', 'res4', 'res5'], + "trans_dim" : 768, + "trans_dropout" : 0.1, + "trans_nheads" : 8, + "trans_feedforward_dim" : 1024, + "trans_num_layers" : 6, +} + +_qt_config = { + "in_channels":768, + "hidden_dim":768, + "num_queries":[4, 144], + "nheads":8, + "num_layers":9, + "feedforward_dim":2048, + "pre_norm":False, + "num_feature_levels":3, + "enforce_input_project":False, + "with_fea2d_pos":False +} + +class SEECoderImageEncode: + @classmethod + def INPUT_TYPES(s): + return {"required": { + "seecoder_name": (folder_paths.get_filename_list("seecoder"), ), + "image": ("IMAGE",), + }} + RETURN_TYPES = ("CONDITIONING",) + FUNCTION = "SEECoderEncode" + + CATEGORY = "conditioning" + + def SEECoderEncode(self, seecoder_name, image): + device = comfy.model_management.get_torch_device() + path = folder_paths.get_full_path("seecoder", seecoder_name) + sd = safetensors.torch.load_file(path, device="cpu") + sd = {k[10:] if k.startswith('ctx.image.') else k: v for k,v in sd.items()} + is_pa = any([x.startswith("qtransformer.pe_layer") for x in sd.keys()]) + + swine_config = _swine_config.copy() + decoder_config = _decoder_config.copy() + qt_config = _qt_config.copy() + if is_pa: + qt_config['with_fea2d_pos'] = True + + swine = SwinTransformer(**swine_config) + decoder = Decoder(**decoder_config) + queryTransformer = QueryTransformer(**qt_config) + + SEE_encoder = SemanticExtractionEncoder(swine, decoder, queryTransformer) + SEE_encoder.load_state_dict(sd) + SEE_encoder = SEE_encoder.to(device) + SEE_encoder.eval() + encoding = SEE_encoder(image.movedim(-1,1).to(device)).cpu() + + return ([[encoding, {}]], ) + +class ConcatConditioning: + @classmethod + def INPUT_TYPES(s): + return {"required": { + "conditioning_to": ("CONDITIONING",), + "conditioning_from": ("CONDITIONING",), + }} + RETURN_TYPES = ("CONDITIONING",) + FUNCTION = "SEECoderEncode" + + CATEGORY = "_for_testing" + + def SEECoderEncode(self, conditioning_to, conditioning_from): + out = [] + + if len(conditioning_from) > 1: + print("Warning: ConditioningAverage conditioning_from contains more than 1 cond, only the first one will actually be applied to conditioning_to.") + + cond_from = conditioning_from[0][0] + + for i in range(len(conditioning_to)): + t1 = conditioning_to[i][0] + tw = torch.cat((t1, cond_from),1) + n = [tw, conditioning_to[i][1].copy()] + out.append(n) + + return (out, ) + +NODE_CLASS_MAPPINGS = { + "SEECoderImageEncode": SEECoderImageEncode, + "ConcatConditioning": ConcatConditioning, +} \ No newline at end of file diff --git a/seecoder.py b/seecoder.py new file mode 100644 index 0000000..b4330b3 --- /dev/null +++ b/seecoder.py @@ -0,0 +1,570 @@ +import torch +import torch.nn as nn +import torch.nn.functional as F +import copy + +from .seecoder_utils import with_pos_embed + +########### +# helpers # +########### + +def _get_clones(module, N): + return nn.ModuleList([copy.deepcopy(module) for i in range(N)]) + +def _get_activation_fn(activation): + """Return an activation function given a string""" + if activation == "relu": + return F.relu + if activation == "gelu": + return F.gelu + if activation == "glu": + return F.glu + raise RuntimeError(f"activation should be relu/gelu, not {activation}.") + +def c2_xavier_fill(module): + # Caffe2 implementation of XavierFill in fact + nn.init.kaiming_uniform_(module.weight, a=1) + if module.bias is not None: + nn.init.constant_(module.bias, 0) + +def with_pos_embed(x, pos): + return x if pos is None else x + pos + +########### +# Modules # +########### + +class Conv2d_Convenience(nn.Conv2d): + def __init__(self, *args, **kwargs): + norm = kwargs.pop("norm", None) + activation = kwargs.pop("activation", None) + super().__init__(*args, **kwargs) + self.norm = norm + self.activation = activation + + def forward(self, x): + x = F.conv2d( + x, self.weight, self.bias, self.stride, self.padding, self.dilation, self.groups) + if self.norm is not None: + x = self.norm(x) + if self.activation is not None: + x = self.activation(x) + return x + +class DecoderLayer(nn.Module): + def __init__(self, + dim=256, + feedforward_dim=1024, + dropout=0.1, + activation="relu", + n_heads=8,): + + super().__init__() + + self.self_attn = nn.MultiheadAttention(dim, n_heads, dropout=dropout) + self.dropout1 = nn.Dropout(dropout) + self.norm1 = nn.LayerNorm(dim) + + self.linear1 = nn.Linear(dim, feedforward_dim) + self.activation = _get_activation_fn(activation) + self.dropout2 = nn.Dropout(dropout) + self.linear2 = nn.Linear(feedforward_dim, dim) + self.dropout3 = nn.Dropout(dropout) + self.norm2 = nn.LayerNorm(dim) + + def forward(self, x): + h = x + h1 = self.self_attn(x, x, x, attn_mask=None)[0] + h = h + self.dropout1(h1) + h = self.norm1(h) + + h2 = self.linear2(self.dropout2(self.activation(self.linear1(h)))) + h = h + self.dropout3(h2) + h = self.norm2(h) + return h + +class DecoderLayerStacked(nn.Module): + def __init__(self, layer, num_layers, norm=None): + super().__init__() + self.layers = _get_clones(layer, num_layers) + self.num_layers = num_layers + self.norm = norm + + def forward(self, x): + h = x + for _, layer in enumerate(self.layers): + h = layer(h) + if self.norm is not None: + h = self.norm(h) + return h + +class SelfAttentionLayer(nn.Module): + def __init__(self, channels, nhead, dropout=0.0, + activation="relu", normalize_before=False): + super().__init__() + self.self_attn = nn.MultiheadAttention(channels, nhead, dropout=dropout) + + self.norm = nn.LayerNorm(channels) + self.dropout = nn.Dropout(dropout) + + self.activation = _get_activation_fn(activation) + self.normalize_before = normalize_before + + self._reset_parameters() + + def _reset_parameters(self): + for p in self.parameters(): + if p.dim() > 1: + nn.init.xavier_uniform_(p) + + def forward_post(self, + qkv, + qk_pos = None, + mask = None,): + h = qkv + qk = with_pos_embed(qkv, qk_pos).transpose(0, 1) + v = qkv.transpose(0, 1) + h1 = self.self_attn(qk, qk, v, attn_mask=mask)[0] + h1 = h1.transpose(0, 1) + h = h + self.dropout(h1) + h = self.norm(h) + return h + + def forward_pre(self, tgt, + tgt_mask = None, + tgt_key_padding_mask = None, + query_pos = None): + # deprecated + assert False + tgt2 = self.norm(tgt) + q = k = self.with_pos_embed(tgt2, query_pos) + tgt2 = self.self_attn(q, k, value=tgt2, attn_mask=tgt_mask, + key_padding_mask=tgt_key_padding_mask)[0] + tgt = tgt + self.dropout(tgt2) + return tgt + + def forward(self, *args, **kwargs): + if self.normalize_before: + return self.forward_pre(*args, **kwargs) + return self.forward_post(*args, **kwargs) + +class CrossAttentionLayer(nn.Module): + def __init__(self, channels, nhead, dropout=0.0, + activation="relu", normalize_before=False): + super().__init__() + self.multihead_attn = nn.MultiheadAttention(channels, nhead, dropout=dropout) + + self.norm = nn.LayerNorm(channels) + self.dropout = nn.Dropout(dropout) + + self.activation = _get_activation_fn(activation) + self.normalize_before = normalize_before + + self._reset_parameters() + + def _reset_parameters(self): + for p in self.parameters(): + if p.dim() > 1: + nn.init.xavier_uniform_(p) + + def forward_post(self, + q, + kv, + q_pos = None, + k_pos = None, + mask = None,): + h = q + q = with_pos_embed(q, q_pos).transpose(0, 1) + k = with_pos_embed(kv, k_pos).transpose(0, 1) + v = kv.transpose(0, 1) + h1 = self.multihead_attn(q, k, v, attn_mask=mask)[0] + h1 = h1.transpose(0, 1) + h = h + self.dropout(h1) + h = self.norm(h) + return h + + def forward_pre(self, tgt, memory, + memory_mask = None, + memory_key_padding_mask = None, + pos = None, + query_pos = None): + # Deprecated + assert False + tgt2 = self.norm(tgt) + tgt2 = self.multihead_attn(query=self.with_pos_embed(tgt2, query_pos), + key=self.with_pos_embed(memory, pos), + value=memory, attn_mask=memory_mask, + key_padding_mask=memory_key_padding_mask)[0] + tgt = tgt + self.dropout(tgt2) + return tgt + + def forward(self, *args, **kwargs): + if self.normalize_before: + return self.forward_pre(*args, **kwargs) + return self.forward_post(*args, **kwargs) + +class FeedForwardLayer(nn.Module): + def __init__(self, channels, hidden_channels=2048, dropout=0.0, + activation="relu", normalize_before=False): + super().__init__() + self.linear1 = nn.Linear(channels, hidden_channels) + self.dropout = nn.Dropout(dropout) + self.linear2 = nn.Linear(hidden_channels, channels) + self.norm = nn.LayerNorm(channels) + self.activation = _get_activation_fn(activation) + self.normalize_before = normalize_before + self._reset_parameters() + + def _reset_parameters(self): + for p in self.parameters(): + if p.dim() > 1: + nn.init.xavier_uniform_(p) + + def forward_post(self, x): + h = x + h1 = self.linear2(self.dropout(self.activation(self.linear1(h)))) + h = h + self.dropout(h1) + h = self.norm(h) + return h + + def forward_pre(self, x): + xn = self.norm(x) + h = x + h1 = self.linear2(self.dropout(self.activation(self.linear1(xn)))) + h = h + self.dropout(h1) + return h + + def forward(self, *args, **kwargs): + if self.normalize_before: + return self.forward_pre(*args, **kwargs) + return self.forward_post(*args, **kwargs) + +class MLP(nn.Module): + def __init__(self, in_channels, channels, out_channels, num_layers): + super().__init__() + self.num_layers = num_layers + h = [channels] * (num_layers - 1) + self.layers = nn.ModuleList( + nn.Linear(n, k) + for n, k in zip([in_channels]+h, h+[out_channels])) + + def forward(self, x): + for i, layer in enumerate(self.layers): + x = F.relu(layer(x)) if i < self.num_layers - 1 else layer(x) + return x + +class PPE_MLP(nn.Module): + def __init__(self, freq_num=20, freq_max=None, out_channel=768, mlp_layer=3): + import math + super().__init__() + self.freq_num = freq_num + self.freq_max = freq_max + self.out_channel = out_channel + self.mlp_layer = mlp_layer + self.twopi = 2 * math.pi + + mlp = [] + in_channel = freq_num*4 + for idx in range(mlp_layer): + linear = nn.Linear(in_channel, out_channel, bias=True) + nn.init.xavier_normal_(linear.weight) + nn.init.constant_(linear.bias, 0) + mlp.append(linear) + if idx != mlp_layer-1: + mlp.append(nn.SiLU()) + in_channel = out_channel + self.mlp = nn.Sequential(*mlp) + nn.init.constant_(self.mlp[-1].weight, 0) + + def forward(self, x, mask=None): + assert mask is None, "Mask not implemented" + h, w = x.shape[-2:] + minlen = min(h, w) + + h_embed, w_embed = torch.meshgrid(torch.arange(h), torch.arange(w), indexing='ij') + if self.training: + import numpy.random as npr + pertube_h, pertube_w = npr.uniform(-0.5, 0.5), npr.uniform(-0.5, 0.5) + else: + pertube_h, pertube_w = 0, 0 + + h_embed = (h_embed+0.5 - h/2 + pertube_h) / (minlen) * self.twopi + w_embed = (w_embed+0.5 - w/2 + pertube_w) / (minlen) * self.twopi + h_embed, w_embed = h_embed.to(x.device).to(x.dtype), w_embed.to(x.device).to(x.dtype) + + dim_t = torch.linspace(0, 1, self.freq_num, dtype=torch.float32, device=x.device) + freq_max = self.freq_max if self.freq_max is not None else minlen/2 + dim_t = freq_max ** dim_t.to(x.dtype) + + pos_h = h_embed[:, :, None] * dim_t + pos_w = w_embed[:, :, None] * dim_t + pos = torch.cat((pos_h.sin(), pos_h.cos(), pos_w.sin(), pos_w.cos()), dim=-1) + pos = self.mlp(pos) + pos = pos.permute(2, 0, 1)[None] + return pos + + def __repr__(self, _repr_indent=4): + head = "Positional encoding " + self.__class__.__name__ + body = [ + "num_pos_feats: {}".format(self.num_pos_feats), + "temperature: {}".format(self.temperature), + "normalize: {}".format(self.normalize), + "scale: {}".format(self.scale), + ] + # _repr_indent = 4 + lines = [head] + [" " * _repr_indent + line for line in body] + return "\n".join(lines) + +########### +# Decoder # +########### + +class Decoder(nn.Module): + def __init__( + self, + inchannels, + trans_input_tags, + trans_num_layers, + trans_dim, + trans_nheads, + trans_dropout, + trans_feedforward_dim,): + + super().__init__() + trans_inchannels = { + k: v for k, v in inchannels.items() if k in trans_input_tags} + fpn_inchannels = { + k: v for k, v in inchannels.items() if k not in trans_input_tags} + + self.trans_tags = sorted(list(trans_inchannels.keys())) + self.fpn_tags = sorted(list(fpn_inchannels.keys())) + self.all_tags = sorted(list(inchannels.keys())) + + if len(self.trans_tags)==0: + assert False # Not allowed + + self.num_trans_lvls = len(self.trans_tags) + + self.inproj_layers = nn.ModuleDict() + for tagi in self.trans_tags: + layeri = nn.Sequential( + nn.Conv2d(trans_inchannels[tagi], trans_dim, kernel_size=1), + nn.GroupNorm(32, trans_dim),) + nn.init.xavier_uniform_(layeri[0].weight, gain=1) + nn.init.constant_(layeri[0].bias, 0) + self.inproj_layers[tagi] = layeri + + tlayer = DecoderLayer( + dim = trans_dim, + n_heads = trans_nheads, + dropout = trans_dropout, + feedforward_dim = trans_feedforward_dim, + activation = 'relu',) + + self.transformer = DecoderLayerStacked(tlayer, trans_num_layers) + for p in self.transformer.parameters(): + if p.dim() > 1: + nn.init.xavier_uniform_(p) + self.level_embed = nn.Parameter(torch.Tensor(len(self.trans_tags), trans_dim)) + nn.init.normal_(self.level_embed) + + self.lateral_layers = nn.ModuleDict() + self.output_layers = nn.ModuleDict() + for tagi in self.all_tags: + lateral_conv = Conv2d_Convenience( + inchannels[tagi], trans_dim, kernel_size=1, + bias=False, norm=nn.GroupNorm(32, trans_dim)) + c2_xavier_fill(lateral_conv) + self.lateral_layers[tagi] = lateral_conv + + for tagi in self.fpn_tags: + output_conv = Conv2d_Convenience( + trans_dim, trans_dim, kernel_size=3, stride=1, padding=1, + bias=False, norm=nn.GroupNorm(32, trans_dim), activation=F.relu,) + c2_xavier_fill(output_conv) + self.output_layers[tagi] = output_conv + + def forward(self, features): + x = [] + spatial_shapes = {} + for idx, tagi in enumerate(self.trans_tags[::-1]): + xi = features[tagi] + xi = self.inproj_layers[tagi](xi) + bs, _, h, w = xi.shape + spatial_shapes[tagi] = (h, w) + xi = xi.flatten(2).transpose(1, 2) + self.level_embed[idx].view(1, 1, -1) + x.append(xi) + + x_length = [xi.shape[1] for xi in x] + x_concat = torch.cat(x, 1) + y_concat = self.transformer(x_concat) + y = torch.split(y_concat, x_length, dim=1) + + out = {} + for idx, tagi in enumerate(self.trans_tags[::-1]): + h, w = spatial_shapes[tagi] + yi = y[idx].transpose(1, 2).view(bs, -1, h, w) + out[tagi] = yi + + for idx, tagi in enumerate(self.all_tags[::-1]): + lconv = self.lateral_layers[tagi] + if tagi in self.trans_tags: + out[tagi] = out[tagi] + lconv(features[tagi]) + tag_save = tagi + else: + oconv = self.output_layers[tagi] + h = lconv(features[tagi]) + oprev = out[tag_save] + h = h + F.interpolate(oconv(oprev), size=h.shape[-2:], mode="bilinear", align_corners=False) + out[tagi] = h + + return out + +##################### +# Query Transformer # +##################### + +class QueryTransformer(nn.Module): + def __init__(self, + in_channels, + hidden_dim, + num_queries = [8, 144], + nheads = 8, + num_layers = 9, + feedforward_dim = 2048, + mask_dim = 256, + pre_norm = False, + num_feature_levels = 3, + enforce_input_project = False, + with_fea2d_pos = True): + + super().__init__() + + if with_fea2d_pos: + self.pe_layer = PPE_MLP(freq_num=20, freq_max=None, out_channel=hidden_dim, mlp_layer=3) + else: + self.pe_layer = None + + if in_channels!=hidden_dim or enforce_input_project: + self.input_proj = nn.ModuleList() + for _ in range(num_feature_levels): + self.input_proj.append(nn.Conv2d(in_channels, hidden_dim, kernel_size=1)) + c2_xavier_fill(self.input_proj[-1]) + else: + self.input_proj = None + + self.num_heads = nheads + self.num_layers = num_layers + self.transformer_selfatt_layers = nn.ModuleList() + self.transformer_crossatt_layers = nn.ModuleList() + self.transformer_feedforward_layers = nn.ModuleList() + + for _ in range(self.num_layers): + self.transformer_selfatt_layers.append( + SelfAttentionLayer( + channels=hidden_dim, + nhead=nheads, + dropout=0.0, + normalize_before=pre_norm, )) + + self.transformer_crossatt_layers.append( + CrossAttentionLayer( + channels=hidden_dim, + nhead=nheads, + dropout=0.0, + normalize_before=pre_norm, )) + + self.transformer_feedforward_layers.append( + FeedForwardLayer( + channels=hidden_dim, + hidden_channels=feedforward_dim, + dropout=0.0, + normalize_before=pre_norm, )) + + self.num_queries = num_queries + num_gq, num_lq = self.num_queries + self.init_query = nn.Embedding(num_gq+num_lq, hidden_dim) + self.query_pos_embedding = nn.Embedding(num_gq+num_lq, hidden_dim) + + self.num_feature_levels = num_feature_levels + self.level_embed = nn.Embedding(num_feature_levels, hidden_dim) + + def forward(self, x): + # x is a list of multi-scale feature + assert len(x) == self.num_feature_levels + fea2d = [] + fea2d_pos = [] + size_list = [] + + for i in range(self.num_feature_levels): + size_list.append(x[i].shape[-2:]) + if self.pe_layer is not None: + pi = self.pe_layer(x[i], None).flatten(2) + pi = pi.transpose(1, 2) + else: + pi = None + xi = self.input_proj[i](x[i]) if self.input_proj is not None else x[i] + xi = xi.flatten(2) + self.level_embed.weight[i][None, :, None] + xi = xi.transpose(1, 2) + fea2d.append(xi) + fea2d_pos.append(pi) + + bs, _, _ = fea2d[0].shape + num_gq, num_lq = self.num_queries + gquery = self.init_query.weight[:num_gq].unsqueeze(0).repeat(bs, 1, 1) + lquery = self.init_query.weight[num_gq:].unsqueeze(0).repeat(bs, 1, 1) + + gquery_pos = self.query_pos_embedding.weight[:num_gq].unsqueeze(0).repeat(bs, 1, 1) + lquery_pos = self.query_pos_embedding.weight[num_gq:].unsqueeze(0).repeat(bs, 1, 1) + + for i in range(self.num_layers): + level_index = i % self.num_feature_levels + + qout = self.transformer_crossatt_layers[i]( + q = lquery, + kv = fea2d[level_index], + q_pos = lquery_pos, + k_pos = fea2d_pos[level_index], + mask = None,) + lquery = qout + + qout = self.transformer_selfatt_layers[i]( + qkv = torch.cat([gquery, lquery], dim=1), + qk_pos = torch.cat([gquery_pos, lquery_pos], dim=1),) + + qout = self.transformer_feedforward_layers[i](qout) + + gquery = qout[:, :num_gq] + lquery = qout[:, num_gq:] + + output = torch.cat([gquery, lquery], dim=1) + + return output + +################## +# Main structure # +################## + +class SemanticExtractionEncoder(nn.Module): + def __init__(self, + imencoder_cfg, + imdecoder_cfg, + qtransformer_cfg): + super().__init__() + self.imencoder = imencoder_cfg + self.imdecoder = imdecoder_cfg + self.qtransformer = qtransformer_cfg + + def forward(self, x): + fea = self.imencoder(x) + hs = {'res3' : fea['res3'], + 'res4' : fea['res4'], + 'res5' : fea['res5'], } + hs = self.imdecoder(hs) + hs = [hs['res3'], hs['res4'], hs['res5']] + q = self.qtransformer(hs) + return q + + def encode(self, x): + return self(x) diff --git a/seecoder_utils.py b/seecoder_utils.py new file mode 100644 index 0000000..40acf4d --- /dev/null +++ b/seecoder_utils.py @@ -0,0 +1,108 @@ +import torch +import torch.nn as nn +import torch.nn.functional as F +from torch.nn.init import xavier_uniform_, constant_, uniform_, normal_ +import math +import copy + +def _get_clones(module, N): + return nn.ModuleList([copy.deepcopy(module) for i in range(N)]) + +def _get_activation_fn(activation): + """Return an activation function given a string""" + if activation == "relu": + return F.relu + if activation == "gelu": + return F.gelu + if activation == "glu": + return F.glu + raise RuntimeError(f"activation should be relu/gelu, not {activation}.") + +def _is_power_of_2(n): + if (not isinstance(n, int)) or (n < 0): + raise ValueError("invalid input for _is_power_of_2: {} (type: {})".format(n, type(n))) + return (n & (n-1) == 0) and n != 0 + +def c2_xavier_fill(module): + # Caffe2 implementation of XavierFill in fact + nn.init.kaiming_uniform_(module.weight, a=1) + if module.bias is not None: + nn.init.constant_(module.bias, 0) + +def with_pos_embed(x, pos): + return x if pos is None else x + pos + +class PositionEmbeddingSine(nn.Module): + def __init__(self, num_pos_feats=64, temperature=256, normalize=False, scale=None): + super().__init__() + self.num_pos_feats = num_pos_feats + self.temperature = temperature + self.normalize = normalize + if scale is not None and normalize is False: + raise ValueError("normalize should be True if scale is passed") + if scale is None: + scale = 2 * math.pi + self.scale = scale + + def forward(self, x, mask=None): + if mask is None: + mask = torch.zeros((x.size(0), x.size(2), x.size(3)), device=x.device, dtype=torch.bool) + not_mask = ~mask + h, w = not_mask.shape[-2:] + minlen = min(h, w) + h_embed = not_mask.cumsum(1, dtype=torch.float32) + w_embed = not_mask.cumsum(2, dtype=torch.float32) + if self.normalize: + eps = 1e-6 + h_embed = (h_embed - h/2) / (minlen + eps) * self.scale + w_embed = (w_embed - w/2) / (minlen + eps) * self.scale + + dim_t = torch.arange(self.num_pos_feats, dtype=torch.float32, device=x.device) + dim_t = self.temperature ** (2 * (dim_t // 2) / self.num_pos_feats) + + pos_w = w_embed[:, :, :, None] / dim_t + pos_h = h_embed[:, :, :, None] / dim_t + pos_w = torch.stack( + (pos_w[:, :, :, 0::2].sin(), pos_w[:, :, :, 1::2].cos()), dim=4 + ).flatten(3) + pos_h = torch.stack( + (pos_h[:, :, :, 0::2].sin(), pos_h[:, :, :, 1::2].cos()), dim=4 + ).flatten(3) + pos = torch.cat((pos_h, pos_w), dim=3).permute(0, 3, 1, 2) + return pos + + def __repr__(self, _repr_indent=4): + head = "Positional encoding " + self.__class__.__name__ + body = [ + "num_pos_feats: {}".format(self.num_pos_feats), + "temperature: {}".format(self.temperature), + "normalize: {}".format(self.normalize), + "scale: {}".format(self.scale), + ] + # _repr_indent = 4 + lines = [head] + [" " * _repr_indent + line for line in body] + return "\n".join(lines) + +class Conv2d_Convenience(nn.Conv2d): + def __init__(self, *args, **kwargs): + norm = kwargs.pop("norm", None) + activation = kwargs.pop("activation", None) + super().__init__(*args, **kwargs) + self.norm = norm + self.activation = activation + + def forward(self, x): + if not torch.jit.is_scripting(): + if x.numel() == 0 and self.training: + assert not isinstance( + self.norm, torch.nn.SyncBatchNorm + ), "SyncBatchNorm does not support empty inputs!" + x = F.conv2d( + x, self.weight, self.bias, self.stride, self.padding, self.dilation, self.groups + ) + if self.norm is not None: + x = self.norm(x) + if self.activation is not None: + x = self.activation(x) + return x + diff --git a/seet_tdecoder.py b/seet_tdecoder.py new file mode 100644 index 0000000..3d5585a --- /dev/null +++ b/seet_tdecoder.py @@ -0,0 +1,697 @@ +#import fvcore.nn.weight_init as weight_init #dependency only needed for training? +from typing import Optional +import torch +import torch.nn as nn +import torch.nn.functional as F + +from .seecoder_utils import PositionEmbeddingSine, _get_clones, _get_activation_fn + +########## +# helper # +########## + +def with_pos_embed(x, pos): + return x if pos is None else x + pos + +############## +# One Former # +############## + +class Transformer(nn.Module): + def __init__(self, + d_model=512, + nhead=8, + num_encoder_layers=6, + num_decoder_layers=6, + dim_feedforward=2048, + dropout=0.1, + activation="relu", + normalize_before=False, + return_intermediate_dec=False,): + + super().__init__() + encoder_layer = TransformerEncoderLayer( + d_model, nhead, dim_feedforward, dropout, activation, normalize_before) + encoder_norm = nn.LayerNorm(d_model) if normalize_before else None + self.encoder = TransformerEncoder(encoder_layer, num_encoder_layers, encoder_norm) + + decoder_layer = TransformerDecoderLayer( + d_model, nhead, dim_feedforward, dropout, activation, normalize_before) + decoder_norm = nn.LayerNorm(d_model) + self.decoder = TransformerDecoder( + decoder_layer, + num_decoder_layers, + decoder_norm, + return_intermediate=return_intermediate_dec,) + + self._reset_parameters() + + self.d_model = d_model + self.nhead = nhead + + def _reset_parameters(self): + for p in self.parameters(): + if p.dim() > 1: + nn.init.xavier_uniform_(p) + + def forward(self, src, mask, query_embed, pos_embed, task_token=None): + # flatten NxCxHxW to HWxNxC + bs, c, h, w = src.shape + src = src.flatten(2).permute(2, 0, 1) + pos_embed = pos_embed.flatten(2).permute(2, 0, 1) + query_embed = query_embed.unsqueeze(1).repeat(1, bs, 1) + if mask is not None: + mask = mask.flatten(1) + + if task_token is None: + tgt = torch.zeros_like(query_embed) + else: + tgt = task_token.repeat(query_embed.shape[0], 1, 1) + + memory = self.encoder(src, src_key_padding_mask=mask, pos=pos_embed) # src = memory + hs = self.decoder( + tgt, memory, memory_key_padding_mask=mask, pos=pos_embed, query_pos=query_embed + ) + return hs.transpose(1, 2), memory.permute(1, 2, 0).view(bs, c, h, w) + +class TransformerEncoder(nn.Module): + def __init__(self, encoder_layer, num_layers, norm=None): + super().__init__() + self.layers = _get_clones(encoder_layer, num_layers) + self.num_layers = num_layers + self.norm = norm + + def forward(self, src, mask=None, src_key_padding_mask=None, pos=None,): + output = src + for layer in self.layers: + output = layer( + output, src_mask=mask, src_key_padding_mask=src_key_padding_mask, pos=pos + ) + if self.norm is not None: + output = self.norm(output) + return output + +class TransformerDecoder(nn.Module): + def __init__(self, decoder_layer, num_layers, norm=None, return_intermediate=False): + super().__init__() + self.layers = _get_clones(decoder_layer, num_layers) + self.num_layers = num_layers + self.norm = norm + self.return_intermediate = return_intermediate + + def forward( + self, + tgt, + memory, + tgt_mask=None, + memory_mask=None, + tgt_key_padding_mask=None, + memory_key_padding_mask=None, + pos=None, + query_pos=None,): + + output = tgt + intermediate = [] + for layer in self.layers: + output = layer( + output, + memory, + tgt_mask=tgt_mask, + memory_mask=memory_mask, + tgt_key_padding_mask=tgt_key_padding_mask, + memory_key_padding_mask=memory_key_padding_mask, + pos=pos, + query_pos=query_pos, + ) + if self.return_intermediate: + intermediate.append(self.norm(output)) + + if self.norm is not None: + output = self.norm(output) + if self.return_intermediate: + intermediate.pop() + intermediate.append(output) + + if self.return_intermediate: + return torch.stack(intermediate) + + return output.unsqueeze(0) + +class TransformerEncoderLayer(nn.Module): + def __init__( + self, + d_model, + nhead, + dim_feedforward=2048, + dropout=0.1, + activation="relu", + normalize_before=False, ): + + super().__init__() + self.self_attn = nn.MultiheadAttention(d_model, nhead, dropout=dropout) + # Implementation of Feedforward model + self.linear1 = nn.Linear(d_model, dim_feedforward) + self.dropout = nn.Dropout(dropout) + self.linear2 = nn.Linear(dim_feedforward, d_model) + + self.norm1 = nn.LayerNorm(d_model) + self.norm2 = nn.LayerNorm(d_model) + self.dropout1 = nn.Dropout(dropout) + self.dropout2 = nn.Dropout(dropout) + + self.activation = _get_activation_fn(activation) + self.normalize_before = normalize_before + + def with_pos_embed(self, x, pos): + return x if pos is None else x + pos + + def forward_post( + self, + src, + src_mask = None, + src_key_padding_mask = None, + pos = None,): + + q = k = self.with_pos_embed(src, pos) + src2 = self.self_attn( + q, k, value=src, attn_mask=src_mask, key_padding_mask=src_key_padding_mask + )[0] + src = src + self.dropout1(src2) + src = self.norm1(src) + src2 = self.linear2(self.dropout(self.activation(self.linear1(src)))) + src = src + self.dropout2(src2) + src = self.norm2(src) + return src + + def forward_pre( + self, + src, + src_mask = None, + src_key_padding_mask = None, + pos = None,): + + src2 = self.norm1(src) + q = k = self.with_pos_embed(src2, pos) + src2 = self.self_attn( + q, k, value=src2, attn_mask=src_mask, key_padding_mask=src_key_padding_mask + )[0] + src = src + self.dropout1(src2) + src2 = self.norm2(src) + src2 = self.linear2(self.dropout(self.activation(self.linear1(src2)))) + src = src + self.dropout2(src2) + return src + + def forward( + self, + src, + src_mask = None, + src_key_padding_mask = None, + pos = None,): + if self.normalize_before: + return self.forward_pre(src, src_mask, src_key_padding_mask, pos) + return self.forward_post(src, src_mask, src_key_padding_mask, pos) + +class TransformerDecoderLayer(nn.Module): + def __init__( + self, + d_model, + nhead, + dim_feedforward=2048, + dropout=0.1, + activation="relu", + normalize_before=False,): + + super().__init__() + self.self_attn = nn.MultiheadAttention(d_model, nhead, dropout=dropout) + self.multihead_attn = nn.MultiheadAttention(d_model, nhead, dropout=dropout) + # Implementation of Feedforward model + self.linear1 = nn.Linear(d_model, dim_feedforward) + self.dropout = nn.Dropout(dropout) + self.linear2 = nn.Linear(dim_feedforward, d_model) + + self.norm1 = nn.LayerNorm(d_model) + self.norm2 = nn.LayerNorm(d_model) + self.norm3 = nn.LayerNorm(d_model) + self.dropout1 = nn.Dropout(dropout) + self.dropout2 = nn.Dropout(dropout) + self.dropout3 = nn.Dropout(dropout) + + self.activation = _get_activation_fn(activation) + self.normalize_before = normalize_before + + def with_pos_embed(self, x, pos): + return x if pos is None else x + pos + + def forward_post( + self, + tgt, + memory, + tgt_mask = None, + memory_mask = None, + tgt_key_padding_mask = None, + memory_key_padding_mask = None, + pos = None, + query_pos = None,): + + q = k = self.with_pos_embed(tgt, query_pos) + tgt2 = self.self_attn( + q, k, value=tgt, attn_mask=tgt_mask, key_padding_mask=tgt_key_padding_mask)[0] + tgt = tgt + self.dropout1(tgt2) + tgt = self.norm1(tgt) + tgt2 = self.multihead_attn( + query=self.with_pos_embed(tgt, query_pos), + key=self.with_pos_embed(memory, pos), + value=memory, + attn_mask=memory_mask, + key_padding_mask=memory_key_padding_mask,)[0] + tgt = tgt + self.dropout2(tgt2) + tgt = self.norm2(tgt) + tgt2 = self.linear2(self.dropout(self.activation(self.linear1(tgt)))) + tgt = tgt + self.dropout3(tgt2) + tgt = self.norm3(tgt) + return tgt + + def forward_pre( + self, + tgt, + memory, + tgt_mask = None, + memory_mask = None, + tgt_key_padding_mask = None, + memory_key_padding_mask = None, + pos = None, + query_pos = None,): + + tgt2 = self.norm1(tgt) + q = k = self.with_pos_embed(tgt2, query_pos) + tgt2 = self.self_attn( + q, k, value=tgt2, attn_mask=tgt_mask, key_padding_mask=tgt_key_padding_mask + )[0] + tgt = tgt + self.dropout1(tgt2) + tgt2 = self.norm2(tgt) + tgt2 = self.multihead_attn( + query=self.with_pos_embed(tgt2, query_pos), + key=self.with_pos_embed(memory, pos), + value=memory, + attn_mask=memory_mask, + key_padding_mask=memory_key_padding_mask, + )[0] + tgt = tgt + self.dropout2(tgt2) + tgt2 = self.norm3(tgt) + tgt2 = self.linear2(self.dropout(self.activation(self.linear1(tgt2)))) + tgt = tgt + self.dropout3(tgt2) + return tgt + + def forward( + self, + tgt, + memory, + tgt_mask = None, + memory_mask = None, + tgt_key_padding_mask = None, + memory_key_padding_mask = None, + pos = None, + query_pos = None, ): + + if self.normalize_before: + return self.forward_pre( + tgt, + memory, + tgt_mask, + memory_mask, + tgt_key_padding_mask, + memory_key_padding_mask, + pos, + query_pos,) + return self.forward_post( + tgt, + memory, + tgt_mask, + memory_mask, + tgt_key_padding_mask, + memory_key_padding_mask, + pos, + query_pos,) + +class SelfAttentionLayer(nn.Module): + + def __init__(self, d_model, nhead, dropout=0.0, + activation="relu", normalize_before=False): + super().__init__() + self.self_attn = nn.MultiheadAttention(d_model, nhead, dropout=dropout) + + self.norm = nn.LayerNorm(d_model) + self.dropout = nn.Dropout(dropout) + + self.activation = _get_activation_fn(activation) + self.normalize_before = normalize_before + + self._reset_parameters() + + def _reset_parameters(self): + for p in self.parameters(): + if p.dim() > 1: + nn.init.xavier_uniform_(p) + + def with_pos_embed(self, tensor, pos): + return tensor if pos is None else tensor + pos + + def forward_post(self, tgt, + tgt_mask = None, + tgt_key_padding_mask = None, + query_pos = None): + q = k = self.with_pos_embed(tgt, query_pos).transpose(0 ,1) + tgt2 = self.self_attn(q, k, value=tgt.transpose(0 ,1), attn_mask=tgt_mask, + key_padding_mask=tgt_key_padding_mask)[0] + tgt = tgt + self.dropout(tgt2.transpose(0 ,1)) + tgt = self.norm(tgt) + + return tgt + + def forward_pre(self, tgt, + tgt_mask = None, + tgt_key_padding_mask = None, + query_pos = None): + tgt2 = self.norm(tgt) + q = k = self.with_pos_embed(tgt2, query_pos) + tgt2 = self.self_attn(q, k, value=tgt2, attn_mask=tgt_mask, + key_padding_mask=tgt_key_padding_mask)[0] + tgt = tgt + self.dropout(tgt2) + + return tgt + + def forward(self, tgt, + tgt_mask = None, + tgt_key_padding_mask = None, + query_pos = None): + if self.normalize_before: + return self.forward_pre(tgt, tgt_mask, + tgt_key_padding_mask, query_pos) + return self.forward_post(tgt, tgt_mask, + tgt_key_padding_mask, query_pos) + +class CrossAttentionLayer(nn.Module): + + def __init__(self, d_model, nhead, dropout=0.0, + activation="relu", normalize_before=False): + super().__init__() + self.multihead_attn = nn.MultiheadAttention(d_model, nhead, dropout=dropout) + + self.norm = nn.LayerNorm(d_model) + self.dropout = nn.Dropout(dropout) + + self.activation = _get_activation_fn(activation) + self.normalize_before = normalize_before + + self._reset_parameters() + + def _reset_parameters(self): + for p in self.parameters(): + if p.dim() > 1: + nn.init.xavier_uniform_(p) + + def with_pos_embed(self, tensor, pos): + return tensor if pos is None else tensor + pos + + def forward_post(self, tgt, memory, + memory_mask = None, + memory_key_padding_mask = None, + pos = None, + query_pos = None): + tgt2 = self.multihead_attn(query=self.with_pos_embed(tgt, query_pos).transpose(0, 1), + key=self.with_pos_embed(memory, pos).transpose(0, 1), + value=memory.transpose(0, 1), attn_mask=memory_mask, + key_padding_mask=memory_key_padding_mask)[0] + tgt = tgt + self.dropout(tgt2.transpose(0, 1)) + tgt = self.norm(tgt) + + return tgt + + def forward_pre(self, tgt, memory, + memory_mask = None, + memory_key_padding_mask = None, + pos = None, + query_pos = None): + tgt2 = self.norm(tgt) + tgt2 = self.multihead_attn(query=self.with_pos_embed(tgt2, query_pos), + key=self.with_pos_embed(memory, pos), + value=memory, attn_mask=memory_mask, + key_padding_mask=memory_key_padding_mask)[0] + tgt = tgt + self.dropout(tgt2) + + return tgt + + def forward(self, tgt, memory, + memory_mask = None, + memory_key_padding_mask = None, + pos = None, + query_pos = None): + if self.normalize_before: + return self.forward_pre(tgt, memory, memory_mask, + memory_key_padding_mask, pos, query_pos) + return self.forward_post(tgt, memory, memory_mask, + memory_key_padding_mask, pos, query_pos) + +class FFNLayer(nn.Module): + + def __init__(self, d_model, dim_feedforward=2048, dropout=0.0, + activation="relu", normalize_before=False): + super().__init__() + # Implementation of Feedforward model + self.linear1 = nn.Linear(d_model, dim_feedforward) + self.dropout = nn.Dropout(dropout) + self.linear2 = nn.Linear(dim_feedforward, d_model) + + self.norm = nn.LayerNorm(d_model) + + self.activation = _get_activation_fn(activation) + self.normalize_before = normalize_before + + self._reset_parameters() + + def _reset_parameters(self): + for p in self.parameters(): + if p.dim() > 1: + nn.init.xavier_uniform_(p) + + def with_pos_embed(self, tensor, pos): + return tensor if pos is None else tensor + pos + + def forward_post(self, tgt): + tgt2 = self.linear2(self.dropout(self.activation(self.linear1(tgt)))) + tgt = tgt + self.dropout(tgt2) + tgt = self.norm(tgt) + return tgt + + def forward_pre(self, tgt): + tgt2 = self.norm(tgt) + tgt2 = self.linear2(self.dropout(self.activation(self.linear1(tgt2)))) + tgt = tgt + self.dropout(tgt2) + return tgt + + def forward(self, tgt): + if self.normalize_before: + return self.forward_pre(tgt) + return self.forward_post(tgt) + +class MLP(nn.Module): + """ Very simple multi-layer perceptron (also called FFN)""" + def __init__(self, input_dim, hidden_dim, output_dim, num_layers): + super().__init__() + self.num_layers = num_layers + h = [hidden_dim] * (num_layers - 1) + self.layers = nn.ModuleList(nn.Linear(n, k) for n, k in zip([input_dim] + h, h + [output_dim])) + + def forward(self, x): + for i, layer in enumerate(self.layers): + x = F.relu(layer(x)) if i < self.num_layers - 1 else layer(x) + return x + +class Seet_OneFormer_TDecoder(nn.Module): + def __init__( + self, + in_channels, + mask_classification, + num_classes, + hidden_dim, + num_queries, + nheads, + dropout, + dim_feedforward, + enc_layers, + is_train, + dec_layers, + class_dec_layers, + pre_norm, + mask_dim, + enforce_input_project, + use_task_norm,): + + super().__init__() + + assert mask_classification, "Only support mask classification model" + self.mask_classification = mask_classification + self.is_train = is_train + self.use_task_norm = use_task_norm + + # positional encoding + N_steps = hidden_dim // 2 + self.pe_layer = PositionEmbeddingSine(N_steps, normalize=True) + + self.class_transformer = Transformer( + d_model=hidden_dim, + dropout=dropout, + nhead=nheads, + dim_feedforward=dim_feedforward, + num_encoder_layers=enc_layers, + num_decoder_layers=class_dec_layers, + normalize_before=pre_norm, + return_intermediate_dec=False, + ) + + # define Transformer decoder here + self.num_heads = nheads + self.num_layers = dec_layers + self.transformer_self_attention_layers = nn.ModuleList() + self.transformer_cross_attention_layers = nn.ModuleList() + self.transformer_ffn_layers = nn.ModuleList() + + for _ in range(self.num_layers): + self.transformer_self_attention_layers.append( + SelfAttentionLayer( + d_model=hidden_dim, + nhead=nheads, + dropout=0.0, + normalize_before=pre_norm, + ) + ) + + self.transformer_cross_attention_layers.append( + CrossAttentionLayer( + d_model=hidden_dim, + nhead=nheads, + dropout=0.0, + normalize_before=pre_norm, + ) + ) + + self.transformer_ffn_layers.append( + FFNLayer( + d_model=hidden_dim, + dim_feedforward=dim_feedforward, + dropout=0.0, + normalize_before=pre_norm, + ) + ) + + self.decoder_norm = nn.LayerNorm(hidden_dim) + + self.num_queries = num_queries + # learnable query p.e. + self.query_embed = nn.Embedding(num_queries, hidden_dim) + + # level embedding (we always use 3 scales) + self.num_feature_levels = 3 + self.level_embed = nn.Embedding(self.num_feature_levels, hidden_dim) + self.input_proj = nn.ModuleList() + for _ in range(self.num_feature_levels): + if in_channels != hidden_dim or enforce_input_project: + self.input_proj.append(nn.Conv2d(in_channels, hidden_dim, kernel_size=1)) + #weight_init.c2_xavier_fill(self.input_proj[-1]) + else: + self.input_proj.append(nn.Sequential()) + + self.class_input_proj = nn.Conv2d(in_channels, hidden_dim, kernel_size=1) + #weight_init.c2_xavier_fill(self.class_input_proj) + + # output FFNs + if self.mask_classification: + self.class_embed = nn.Linear(hidden_dim, num_classes + 1) + self.mask_embed = MLP(hidden_dim, hidden_dim, mask_dim, 3) + + def forward(self, x, mask_features, tasks): + # x is a list of multi-scale feature + assert len(x) == self.num_feature_levels + src = [] + pos = [] + size_list = [] + + for i in range(self.num_feature_levels): + size_list.append(x[i].shape[-2:]) + pos.append(self.pe_layer(x[i], None).flatten(2)) + src.append(self.input_proj[i](x[i]).flatten(2) + self.level_embed.weight[i][None, :, None]) + pos[-1] = pos[-1].transpose(1, 2) + src[-1] = src[-1].transpose(1, 2) + + bs, _, _ = src[0].shape + + query_embed = self.query_embed.weight.unsqueeze(0).repeat(bs, 1, 1) + + tasks = tasks.unsqueeze(0) + if self.use_task_norm: + tasks = self.decoder_norm(tasks) + + feats = self.pe_layer(mask_features, None) + + out_t, _ = self.class_transformer( + feats, None, + self.query_embed.weight[:-1], + self.class_input_proj(mask_features), + tasks if self.use_task_norm else None) + out_t = out_t[0] + + out = torch.cat([out_t, tasks], dim=1) + + output = out.clone() + + predictions_class = [] + predictions_mask = [] + + # prediction heads on learnable query features + outputs_class, outputs_mask, attn_mask = self.forward_prediction_heads( + output, mask_features, attn_mask_target_size=size_list[0]) + predictions_class.append(outputs_class) + predictions_mask.append(outputs_mask) + + for i in range(self.num_layers): + level_index = i % self.num_feature_levels + attn_mask[torch.where(attn_mask.sum(-1) == attn_mask.shape[-1])] = False + + output = self.transformer_cross_attention_layers[i]( + output, src[level_index], + memory_mask=attn_mask, + memory_key_padding_mask=None, + pos=pos[level_index], query_pos=query_embed, ) + + output = self.transformer_self_attention_layers[i]( + output, tgt_mask=None, + tgt_key_padding_mask=None, + query_pos=query_embed, ) + + # FFN + output = self.transformer_ffn_layers[i](output) + + outputs_class, outputs_mask, attn_mask = self.forward_prediction_heads( + output, mask_features, attn_mask_target_size=size_list[(i + 1) % self.num_feature_levels]) + predictions_class.append(outputs_class) + predictions_mask.append(outputs_mask) + + assert len(predictions_class) == self.num_layers + 1 + + out = { + 'pred_logits': predictions_class[-1], + 'pred_masks': predictions_mask[-1],} + + return out + + def forward_prediction_heads(self, output, mask_features, attn_mask_target_size): + decoder_output = self.decoder_norm(output) + outputs_class = self.class_embed(decoder_output) + mask_embed = self.mask_embed(decoder_output) + outputs_mask = torch.einsum("bqc,bchw->bqhw", mask_embed, mask_features) + + attn_mask = F.interpolate(outputs_mask, size=attn_mask_target_size, mode="bilinear", align_corners=False) + attn_mask = (attn_mask.sigmoid().flatten(2).unsqueeze(1).repeat(1, self.num_heads, 1, 1).flatten(0, 1) < 0.5).bool() + attn_mask = attn_mask.detach() + + return outputs_class, outputs_mask, attn_mask diff --git a/swin.py b/swin.py new file mode 100644 index 0000000..bbfb25c --- /dev/null +++ b/swin.py @@ -0,0 +1,657 @@ +import torch +import torch.nn as nn +import torch.nn.functional as F +import torch.utils.checkpoint as checkpoint +import numpy as np + + +############################## +# timm.models.layers helpers # +############################## + +def drop_path(x, drop_prob: float = 0., training: bool = False, scale_by_keep: bool = True): + if drop_prob == 0. or not training: + return x + keep_prob = 1 - drop_prob + shape = (x.shape[0],) + (1,) * (x.ndim - 1) # work with diff dim tensors, not just 2D ConvNets + random_tensor = x.new_empty(shape).bernoulli_(keep_prob) + if keep_prob > 0.0 and scale_by_keep: + random_tensor.div_(keep_prob) + return x * random_tensor + +class DropPath(nn.Module): + """Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks). + """ + def __init__(self, drop_prob: float = 0., scale_by_keep: bool = True): + super(DropPath, self).__init__() + self.drop_prob = drop_prob + self.scale_by_keep = scale_by_keep + + def forward(self, x): + return drop_path(x, self.drop_prob, self.training, self.scale_by_keep) + + def extra_repr(self): + return f'drop_prob={round(self.drop_prob,3):0.3f}' + +def _ntuple(n): + def parse(x): + from itertools import repeat + import collections.abc + if isinstance(x, collections.abc.Iterable) and not isinstance(x, str): + return tuple(x) + return tuple(repeat(x, n)) + return parse + +to_1tuple = _ntuple(1) +to_2tuple = _ntuple(2) +to_3tuple = _ntuple(3) +to_4tuple = _ntuple(4) +to_ntuple = _ntuple + +def _trunc_normal_(tensor, mean, std, a, b): + import warnings + import math + + def norm_cdf(x): + return (1. + math.erf(x / math.sqrt(2.))) / 2. + + if (mean < a - 2 * std) or (mean > b + 2 * std): + warnings.warn("mean is more than 2 std from [a, b] in nn.init.trunc_normal_. " + "The distribution of values may be incorrect.", + stacklevel=2) + + l = norm_cdf((a - mean) / std) + u = norm_cdf((b - mean) / std) + tensor.uniform_(2 * l - 1, 2 * u - 1) + tensor.erfinv_() + tensor.mul_(std * math.sqrt(2.)) + tensor.add_(mean) + tensor.clamp_(min=a, max=b) + return tensor + +def trunc_normal_(tensor, mean=0., std=1., a=-2., b=2.): + with torch.no_grad(): + return _trunc_normal_(tensor, mean, std, a, b) + +############# +# main swin # +############# + +class Mlp(nn.Module): + """ Multilayer perceptron.""" + + def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU, drop=0.): + super().__init__() + out_features = out_features or in_features + hidden_features = hidden_features or in_features + self.fc1 = nn.Linear(in_features, hidden_features) + self.act = act_layer() + self.fc2 = nn.Linear(hidden_features, out_features) + self.drop = nn.Dropout(drop) + + def forward(self, x): + x = self.fc1(x) + x = self.act(x) + x = self.drop(x) + x = self.fc2(x) + x = self.drop(x) + return x + + +def window_partition(x, window_size): + """ + Args: + x: (B, H, W, C) + window_size (int): window size + Returns: + windows: (num_windows*B, window_size, window_size, C) + """ + B, H, W, C = x.shape + x = x.view(B, H // window_size, window_size, W // window_size, window_size, C) + windows = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(-1, window_size, window_size, C) + return windows + + +def window_reverse(windows, window_size, H, W): + """ + Args: + windows: (num_windows*B, window_size, window_size, C) + window_size (int): Window size + H (int): Height of image + W (int): Width of image + Returns: + x: (B, H, W, C) + """ + B = int(windows.shape[0] / (H * W / window_size / window_size)) + x = windows.view(B, H // window_size, W // window_size, window_size, window_size, -1) + x = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(B, H, W, -1) + return x + + +class WindowAttention(nn.Module): + """ Window based multi-head self attention (W-MSA) module with relative position bias. + It supports both of shifted and non-shifted window. + Args: + dim (int): Number of input channels. + window_size (tuple[int]): The height and width of the window. + num_heads (int): Number of attention heads. + qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True + qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set + attn_drop (float, optional): Dropout ratio of attention weight. Default: 0.0 + proj_drop (float, optional): Dropout ratio of output. Default: 0.0 + """ + + def __init__(self, dim, window_size, num_heads, qkv_bias=True, qk_scale=None, attn_drop=0., proj_drop=0.): + + super().__init__() + self.dim = dim + self.window_size = window_size # Wh, Ww + self.num_heads = num_heads + head_dim = dim // num_heads + self.scale = qk_scale or head_dim ** -0.5 + + # define a parameter table of relative position bias + self.relative_position_bias_table = nn.Parameter( + torch.zeros((2 * window_size[0] - 1) * (2 * window_size[1] - 1), num_heads)) # 2*Wh-1 * 2*Ww-1, nH + + # get pair-wise relative position index for each token inside the window + coords_h = torch.arange(self.window_size[0]) + coords_w = torch.arange(self.window_size[1]) + coords = torch.stack(torch.meshgrid([coords_h, coords_w], indexing='ij')) # 2, Wh, Ww + coords_flatten = torch.flatten(coords, 1) # 2, Wh*Ww + relative_coords = coords_flatten[:, :, None] - coords_flatten[:, None, :] # 2, Wh*Ww, Wh*Ww + relative_coords = relative_coords.permute(1, 2, 0).contiguous() # Wh*Ww, Wh*Ww, 2 + relative_coords[:, :, 0] += self.window_size[0] - 1 # shift to start from 0 + relative_coords[:, :, 1] += self.window_size[1] - 1 + relative_coords[:, :, 0] *= 2 * self.window_size[1] - 1 + relative_position_index = relative_coords.sum(-1) # Wh*Ww, Wh*Ww + self.register_buffer("relative_position_index", relative_position_index) + + self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias) + self.attn_drop = nn.Dropout(attn_drop) + self.proj = nn.Linear(dim, dim) + self.proj_drop = nn.Dropout(proj_drop) + + trunc_normal_(self.relative_position_bias_table, std=.02) + self.softmax = nn.Softmax(dim=-1) + + def forward(self, x, mask=None): + """ Forward function. + Args: + x: input features with shape of (num_windows*B, N, C) + mask: (0/-inf) mask with shape of (num_windows, Wh*Ww, Wh*Ww) or None + """ + B_, N, C = x.shape + qkv = self.qkv(x).reshape(B_, N, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4) + q, k, v = qkv[0], qkv[1], qkv[2] # make torchscript happy (cannot use tensor as tuple) + + q = q * self.scale + attn = (q @ k.transpose(-2, -1)) + + relative_position_bias = self.relative_position_bias_table[self.relative_position_index.view(-1)].view( + self.window_size[0] * self.window_size[1], self.window_size[0] * self.window_size[1], -1) # Wh*Ww,Wh*Ww,nH + relative_position_bias = relative_position_bias.permute(2, 0, 1).contiguous() # nH, Wh*Ww, Wh*Ww + attn = attn + relative_position_bias.unsqueeze(0) + + if mask is not None: + nW = mask.shape[0] + attn = attn.view(B_ // nW, nW, self.num_heads, N, N) + mask.unsqueeze(1).unsqueeze(0) + attn = attn.view(-1, self.num_heads, N, N) + attn = self.softmax(attn) + else: + attn = self.softmax(attn) + + attn = self.attn_drop(attn) + + x = (attn @ v).transpose(1, 2).reshape(B_, N, C) + x = self.proj(x) + x = self.proj_drop(x) + return x + + +class SwinTransformerBlock(nn.Module): + """ Swin Transformer Block. + Args: + dim (int): Number of input channels. + num_heads (int): Number of attention heads. + window_size (int): Window size. + shift_size (int): Shift size for SW-MSA. + mlp_ratio (float): Ratio of mlp hidden dim to embedding dim. + qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True + qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set. + drop (float, optional): Dropout rate. Default: 0.0 + attn_drop (float, optional): Attention dropout rate. Default: 0.0 + drop_path (float, optional): Stochastic depth rate. Default: 0.0 + act_layer (nn.Module, optional): Activation layer. Default: nn.GELU + norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm + """ + + def __init__(self, dim, num_heads, window_size=7, shift_size=0, + mlp_ratio=4., qkv_bias=True, qk_scale=None, drop=0., attn_drop=0., drop_path=0., + act_layer=nn.GELU, norm_layer=nn.LayerNorm): + super().__init__() + self.dim = dim + self.num_heads = num_heads + self.window_size = window_size + self.shift_size = shift_size + self.mlp_ratio = mlp_ratio + assert 0 <= self.shift_size < self.window_size, "shift_size must in 0-window_size" + + self.norm1 = norm_layer(dim) + self.attn = WindowAttention( + dim, window_size=to_2tuple(self.window_size), num_heads=num_heads, + qkv_bias=qkv_bias, qk_scale=qk_scale, attn_drop=attn_drop, proj_drop=drop) + + self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity() + self.norm2 = norm_layer(dim) + mlp_hidden_dim = int(dim * mlp_ratio) + self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop) + + self.H = None + self.W = None + + def forward(self, x, mask_matrix): + """ Forward function. + Args: + x: Input feature, tensor size (B, H*W, C). + H, W: Spatial resolution of the input feature. + mask_matrix: Attention mask for cyclic shift. + """ + B, L, C = x.shape + H, W = self.H, self.W + assert L == H * W, "input feature has wrong size" + + shortcut = x + x = self.norm1(x) + x = x.view(B, H, W, C) + + # pad feature maps to multiples of window size + pad_l = pad_t = 0 + pad_r = (self.window_size - W % self.window_size) % self.window_size + pad_b = (self.window_size - H % self.window_size) % self.window_size + x = F.pad(x, (0, 0, pad_l, pad_r, pad_t, pad_b)) + _, Hp, Wp, _ = x.shape + + # cyclic shift + if self.shift_size > 0: + shifted_x = torch.roll(x, shifts=(-self.shift_size, -self.shift_size), dims=(1, 2)) + attn_mask = mask_matrix + else: + shifted_x = x + attn_mask = None + + # partition windows + x_windows = window_partition(shifted_x, self.window_size) # nW*B, window_size, window_size, C + x_windows = x_windows.view(-1, self.window_size * self.window_size, C) # nW*B, window_size*window_size, C + + # W-MSA/SW-MSA + attn_windows = self.attn(x_windows, mask=attn_mask) # nW*B, window_size*window_size, C + + # merge windows + attn_windows = attn_windows.view(-1, self.window_size, self.window_size, C) + shifted_x = window_reverse(attn_windows, self.window_size, Hp, Wp) # B H' W' C + + # reverse cyclic shift + if self.shift_size > 0: + x = torch.roll(shifted_x, shifts=(self.shift_size, self.shift_size), dims=(1, 2)) + else: + x = shifted_x + + if pad_r > 0 or pad_b > 0: + x = x[:, :H, :W, :].contiguous() + + x = x.view(B, H * W, C) + + # FFN + x = shortcut + self.drop_path(x) + x = x + self.drop_path(self.mlp(self.norm2(x))) + + return x + + +class PatchMerging(nn.Module): + """ Patch Merging Layer + Args: + dim (int): Number of input channels. + norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm + """ + def __init__(self, dim, norm_layer=nn.LayerNorm): + super().__init__() + self.dim = dim + self.reduction = nn.Linear(4 * dim, 2 * dim, bias=False) + self.norm = norm_layer(4 * dim) + + def forward(self, x, H, W): + """ Forward function. + Args: + x: Input feature, tensor size (B, H*W, C). + H, W: Spatial resolution of the input feature. + """ + B, L, C = x.shape + assert L == H * W, "input feature has wrong size" + + x = x.view(B, H, W, C) + + # padding + pad_input = (H % 2 == 1) or (W % 2 == 1) + if pad_input: + x = F.pad(x, (0, 0, 0, W % 2, 0, H % 2)) + + x0 = x[:, 0::2, 0::2, :] # B H/2 W/2 C + x1 = x[:, 1::2, 0::2, :] # B H/2 W/2 C + x2 = x[:, 0::2, 1::2, :] # B H/2 W/2 C + x3 = x[:, 1::2, 1::2, :] # B H/2 W/2 C + x = torch.cat([x0, x1, x2, x3], -1) # B H/2 W/2 4*C + x = x.view(B, -1, 4 * C) # B H/2*W/2 4*C + + x = self.norm(x) + x = self.reduction(x) + + return x + + +class BasicLayer(nn.Module): + """ A basic Swin Transformer layer for one stage. + Args: + dim (int): Number of feature channels + depth (int): Depths of this stage. + num_heads (int): Number of attention head. + window_size (int): Local window size. Default: 7. + mlp_ratio (float): Ratio of mlp hidden dim to embedding dim. Default: 4. + qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True + qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set. + drop (float, optional): Dropout rate. Default: 0.0 + attn_drop (float, optional): Attention dropout rate. Default: 0.0 + drop_path (float | tuple[float], optional): Stochastic depth rate. Default: 0.0 + norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm + downsample (nn.Module | None, optional): Downsample layer at the end of the layer. Default: None + use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False. + """ + + def __init__(self, + dim, + depth, + num_heads, + window_size=7, + mlp_ratio=4., + qkv_bias=True, + qk_scale=None, + drop=0., + attn_drop=0., + drop_path=0., + norm_layer=nn.LayerNorm, + downsample=None, + use_checkpoint=False): + super().__init__() + self.window_size = window_size + self.shift_size = window_size // 2 + self.depth = depth + self.use_checkpoint = use_checkpoint + + # build blocks + self.blocks = nn.ModuleList([ + SwinTransformerBlock( + dim=dim, + num_heads=num_heads, + window_size=window_size, + shift_size=0 if (i % 2 == 0) else window_size // 2, + mlp_ratio=mlp_ratio, + qkv_bias=qkv_bias, + qk_scale=qk_scale, + drop=drop, + attn_drop=attn_drop, + drop_path=drop_path[i] if isinstance(drop_path, list) else drop_path, + norm_layer=norm_layer) + for i in range(depth)]) + + # patch merging layer + if downsample is not None: + self.downsample = downsample(dim=dim, norm_layer=norm_layer) + else: + self.downsample = None + + def forward(self, x, H, W): + """ Forward function. + Args: + x: Input feature, tensor size (B, H*W, C). + H, W: Spatial resolution of the input feature. + """ + + # calculate attention mask for SW-MSA + Hp = int(np.ceil(H / self.window_size)) * self.window_size + Wp = int(np.ceil(W / self.window_size)) * self.window_size + img_mask = torch.zeros((1, Hp, Wp, 1), device=x.device, dtype=x.dtype) # 1 Hp Wp 1 + h_slices = (slice(0, -self.window_size), + slice(-self.window_size, -self.shift_size), + slice(-self.shift_size, None)) + w_slices = (slice(0, -self.window_size), + slice(-self.window_size, -self.shift_size), + slice(-self.shift_size, None)) + cnt = 0 + for h in h_slices: + for w in w_slices: + img_mask[:, h, w, :] = cnt + cnt += 1 + + mask_windows = window_partition(img_mask, self.window_size) # nW, window_size, window_size, 1 + mask_windows = mask_windows.view(-1, self.window_size * self.window_size) + attn_mask = mask_windows.unsqueeze(1) - mask_windows.unsqueeze(2) + attn_mask = attn_mask.masked_fill(attn_mask != 0, float(-100.0)).masked_fill(attn_mask == 0, float(0.0)) + + for blk in self.blocks: + blk.H, blk.W = H, W + if self.use_checkpoint: + x = checkpoint.checkpoint(blk, x, attn_mask) + else: + x = blk(x, attn_mask) + if self.downsample is not None: + x_down = self.downsample(x, H, W) + Wh, Ww = (H + 1) // 2, (W + 1) // 2 + return x, H, W, x_down, Wh, Ww + else: + return x, H, W, x, H, W + + +class PatchEmbed(nn.Module): + """ Image to Patch Embedding + Args: + patch_size (int): Patch token size. Default: 4. + in_chans (int): Number of input image channels. Default: 3. + embed_dim (int): Number of linear projection output channels. Default: 96. + norm_layer (nn.Module, optional): Normalization layer. Default: None + """ + + def __init__(self, patch_size=4, in_chans=3, embed_dim=96, norm_layer=None): + super().__init__() + patch_size = to_2tuple(patch_size) + self.patch_size = patch_size + + self.in_chans = in_chans + self.embed_dim = embed_dim + + self.proj = nn.Conv2d(in_chans, embed_dim, kernel_size=patch_size, stride=patch_size) + if norm_layer is not None: + self.norm = norm_layer(embed_dim) + else: + self.norm = None + + def forward(self, x): + """Forward function.""" + # padding + _, _, H, W = x.size() + if W % self.patch_size[1] != 0: + x = F.pad(x, (0, self.patch_size[1] - W % self.patch_size[1])) + if H % self.patch_size[0] != 0: + x = F.pad(x, (0, 0, 0, self.patch_size[0] - H % self.patch_size[0])) + + x = self.proj(x) # B C Wh Ww + if self.norm is not None: + Wh, Ww = x.size(2), x.size(3) + x = x.flatten(2).transpose(1, 2) + x = self.norm(x) + x = x.transpose(1, 2).view(-1, self.embed_dim, Wh, Ww) + + return x + + +class SwinTransformer(nn.Module): + """ Swin Transformer backbone. + A PyTorch impl of : `Swin Transformer: Hierarchical Vision Transformer using Shifted Windows` - + https://arxiv.org/pdf/2103.14030 + Args: + pretrain_img_size (int): Input image size for training the pretrained model, + used in absolute postion embedding. Default 224. + patch_size (int | tuple(int)): Patch size. Default: 4. + in_chans (int): Number of input image channels. Default: 3. + embed_dim (int): Number of linear projection output channels. Default: 96. + depths (tuple[int]): Depths of each Swin Transformer stage. + num_heads (tuple[int]): Number of attention head of each stage. + window_size (int): Window size. Default: 7. + mlp_ratio (float): Ratio of mlp hidden dim to embedding dim. Default: 4. + qkv_bias (bool): If True, add a learnable bias to query, key, value. Default: True + qk_scale (float): Override default qk scale of head_dim ** -0.5 if set. + drop_rate (float): Dropout rate. + attn_drop_rate (float): Attention dropout rate. Default: 0. + drop_path_rate (float): Stochastic depth rate. Default: 0.2. + norm_layer (nn.Module): Normalization layer. Default: nn.LayerNorm. + ape (bool): If True, add absolute position embedding to the patch embedding. Default: False. + patch_norm (bool): If True, add normalization after patch embedding. Default: True. + out_indices (Sequence[int]): Output from which stages. + frozen_stages (int): Stages to be frozen (stop grad and set eval mode). + -1 means not freezing any parameters. + use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False. + """ + + def __init__(self, + pretrain_img_size=224, + patch_size=4, + in_chans=3, + embed_dim=96, + depths=[2, 2, 6, 2], + num_heads=[3, 6, 12, 24], + window_size=7, + mlp_ratio=4., + qkv_bias=True, + qk_scale=None, + drop_rate=0., + attn_drop_rate=0., + drop_path_rate=0.2, + norm_layer=nn.LayerNorm, + ape=False, + patch_norm=True, + out_indices=(0, 1, 2, 3), + frozen_stages=-1, + use_checkpoint=False): + super().__init__() + + self.pretrain_img_size = pretrain_img_size + self.num_layers = len(depths) + self.embed_dim = embed_dim + self.ape = ape + self.patch_norm = patch_norm + self.out_indices = out_indices + self.frozen_stages = frozen_stages + + # split image into non-overlapping patches + self.patch_embed = PatchEmbed( + patch_size=patch_size, in_chans=in_chans, embed_dim=embed_dim, + norm_layer=norm_layer if self.patch_norm else None) + + # absolute position embedding + if self.ape: + pretrain_img_size = to_2tuple(pretrain_img_size) + patch_size = to_2tuple(patch_size) + patches_resolution = [pretrain_img_size[0] // patch_size[0], pretrain_img_size[1] // patch_size[1]] + + self.absolute_pos_embed = nn.Parameter(torch.zeros(1, embed_dim, patches_resolution[0], patches_resolution[1])) + trunc_normal_(self.absolute_pos_embed, std=.02) + + self.pos_drop = nn.Dropout(p=drop_rate) + + # stochastic depth + dpr = [x.item() for x in torch.linspace(0, drop_path_rate, sum(depths))] # stochastic depth decay rule + + # build layers + self.layers = nn.ModuleList() + for i_layer in range(self.num_layers): + layer = BasicLayer( + dim=int(embed_dim * 2 ** i_layer), + depth=depths[i_layer], + num_heads=num_heads[i_layer], + window_size=window_size, + mlp_ratio=mlp_ratio, + qkv_bias=qkv_bias, + qk_scale=qk_scale, + drop=drop_rate, + attn_drop=attn_drop_rate, + drop_path=dpr[sum(depths[:i_layer]):sum(depths[:i_layer + 1])], + norm_layer=norm_layer, + downsample=PatchMerging if (i_layer < self.num_layers - 1) else None, + use_checkpoint=use_checkpoint) + self.layers.append(layer) + + num_features = [int(embed_dim * 2 ** i) for i in range(self.num_layers)] + self.num_features = num_features + + # add a norm layer for each output + for i_layer in out_indices: + layer = norm_layer(num_features[i_layer]) + layer_name = f'norm{i_layer}' + self.add_module(layer_name, layer) + + self._freeze_stages() + + def _freeze_stages(self): + if self.frozen_stages >= 0: + self.patch_embed.eval() + for param in self.patch_embed.parameters(): + param.requires_grad = False + + if self.frozen_stages >= 1 and self.ape: + self.absolute_pos_embed.requires_grad = False + + if self.frozen_stages >= 2: + self.pos_drop.eval() + for i in range(0, self.frozen_stages - 1): + m = self.layers[i] + m.eval() + for param in m.parameters(): + param.requires_grad = False + + def forward(self, x): + """Forward function.""" + x = self.patch_embed(x) + + Wh, Ww = x.size(2), x.size(3) + if self.ape: + # interpolate the position embedding to the corresponding size + absolute_pos_embed = F.interpolate(self.absolute_pos_embed, size=(Wh, Ww), mode='bicubic') + x = (x + absolute_pos_embed).flatten(2).transpose(1, 2) # B Wh*Ww C + else: + x = x.flatten(2).transpose(1, 2) + x = self.pos_drop(x) + + outs = [] + for i in range(self.num_layers): + layer = self.layers[i] + x_out, H, W, x, Wh, Ww = layer(x, Wh, Ww) + + if i in self.out_indices: + norm_layer = getattr(self, f'norm{i}') + x_out = norm_layer(x_out) + + out = x_out.view(-1, H, W, self.num_features[i]).permute(0, 3, 1, 2).contiguous() + outs.append(out) + + outputs = { + 'res2' : outs[0], + 'res3' : outs[1], + 'res4' : outs[2], + 'res5' : outs[3],} + return outputs + + def train(self, mode=True): + """Convert the model into training mode while keep layers freezed.""" + super(SwinTransformer, self).train(mode) + self._freeze_stages() + return self