nodes/kijai-ComfyUI-APISR-KJ
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Initial commit

Browse Source
This commit is contained in:
Kijai
2024-03-19 15:59:43 +02:00
parent c718e8e8dd
commit c31c774a0f
21 changed files with 6744 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files
.gitignore
+7
View File
@@ -0,0 +1,7 @@
__pycache__
/venv
.vscode
*.ckpt
*.pth
types
models
LICENSE
+674
View File
@@ -0,0 +1,674 @@
GNU GENERAL PUBLIC LICENSE
Version 3, 29 June 2007
Copyright (C) 2007 Free Software Foundation, Inc. <https://fsf.org/>
Everyone is permitted to copy and distribute verbatim copies
of this license document, but changing it is not allowed.
Preamble
The GNU General Public License is a free, copyleft license for
software and other kinds of works.
The licenses for most software and other practical works are designed
to take away your freedom to share and change the works. By contrast,
the GNU General Public License is intended to guarantee your freedom to
share and change all versions of a program--to make sure it remains free
software for all its users. We, the Free Software Foundation, use the
GNU General Public License for most of our software; it applies also to
any other work released this way by its authors. You can apply it to
your programs, too.
When we speak of free software, we are referring to freedom, not
price. Our General Public Licenses are designed to make sure that you
have the freedom to distribute copies of free software (and charge for
them if you wish), that you receive source code or can get it if you
want it, that you can change the software or use pieces of it in new
free programs, and that you know you can do these things.
To protect your rights, we need to prevent others from denying you
these rights or asking you to surrender the rights. Therefore, you have
certain responsibilities if you distribute copies of the software, or if
you modify it: responsibilities to respect the freedom of others.
For example, if you distribute copies of such a program, whether
gratis or for a fee, you must pass on to the recipients the same
freedoms that you received. You must make sure that they, too, receive
or can get the source code. And you must show them these terms so they
know their rights.
Developers that use the GNU GPL protect your rights with two steps:
(1) assert copyright on the software, and (2) offer you this License
giving you legal permission to copy, distribute and/or modify it.
For the developers' and authors' protection, the GPL clearly explains
that there is no warranty for this free software. For both users' and
authors' sake, the GPL requires that modified versions be marked as
changed, so that their problems will not be attributed erroneously to
authors of previous versions.
Some devices are designed to deny users access to install or run
modified versions of the software inside them, although the manufacturer
can do so. This is fundamentally incompatible with the aim of
protecting users' freedom to change the software. The systematic
pattern of such abuse occurs in the area of products for individuals to
use, which is precisely where it is most unacceptable. Therefore, we
have designed this version of the GPL to prohibit the practice for those
products. If such problems arise substantially in other domains, we
stand ready to extend this provision to those domains in future versions
of the GPL, as needed to protect the freedom of users.
Finally, every program is threatened constantly by software patents.
States should not allow patents to restrict development and use of
software on general-purpose computers, but in those that do, we wish to
avoid the special danger that patents applied to a free program could
make it effectively proprietary. To prevent this, the GPL assures that
patents cannot be used to render the program non-free.
The precise terms and conditions for copying, distribution and
modification follow.
TERMS AND CONDITIONS
0. Definitions.
"This License" refers to version 3 of the GNU General Public License.
"Copyright" also means copyright-like laws that apply to other kinds of
works, such as semiconductor masks.
"The Program" refers to any copyrightable work licensed under this
License. Each licensee is addressed as "you". "Licensees" and
"recipients" may be individuals or organizations.
To "modify" a work means to copy from or adapt all or part of the work
in a fashion requiring copyright permission, other than the making of an
exact copy. The resulting work is called a "modified version" of the
earlier work or a work "based on" the earlier work.
A "covered work" means either the unmodified Program or a work based
on the Program.
To "propagate" a work means to do anything with it that, without
permission, would make you directly or secondarily liable for
infringement under applicable copyright law, except executing it on a
computer or modifying a private copy. Propagation includes copying,
distribution (with or without modification), making available to the
public, and in some countries other activities as well.
To "convey" a work means any kind of propagation that enables other
parties to make or receive copies. Mere interaction with a user through
a computer network, with no transfer of a copy, is not conveying.
An interactive user interface displays "Appropriate Legal Notices"
to the extent that it includes a convenient and prominently visible
feature that (1) displays an appropriate copyright notice, and (2)
tells the user that there is no warranty for the work (except to the
extent that warranties are provided), that licensees may convey the
work under this License, and how to view a copy of this License. If
the interface presents a list of user commands or options, such as a
menu, a prominent item in the list meets this criterion.
1. Source Code.
The "source code" for a work means the preferred form of the work
for making modifications to it. "Object code" means any non-source
form of a work.
A "Standard Interface" means an interface that either is an official
standard defined by a recognized standards body, or, in the case of
interfaces specified for a particular programming language, one that
is widely used among developers working in that language.
The "System Libraries" of an executable work include anything, other
than the work as a whole, that (a) is included in the normal form of
packaging a Major Component, but which is not part of that Major
Component, and (b) serves only to enable use of the work with that
Major Component, or to implement a Standard Interface for which an
implementation is available to the public in source code form. A
"Major Component", in this context, means a major essential component
(kernel, window system, and so on) of the specific operating system
(if any) on which the executable work runs, or a compiler used to
produce the work, or an object code interpreter used to run it.
The "Corresponding Source" for a work in object code form means all
the source code needed to generate, install, and (for an executable
work) run the object code and to modify the work, including scripts to
control those activities. However, it does not include the work's
System Libraries, or general-purpose tools or generally available free
programs which are used unmodified in performing those activities but
which are not part of the work. For example, Corresponding Source
includes interface definition files associated with source files for
the work, and the source code for shared libraries and dynamically
linked subprograms that the work is specifically designed to require,
such as by intimate data communication or control flow between those
subprograms and other parts of the work.
The Corresponding Source need not include anything that users
can regenerate automatically from other parts of the Corresponding
Source.
The Corresponding Source for a work in source code form is that
same work.
2. Basic Permissions.
All rights granted under this License are granted for the term of
copyright on the Program, and are irrevocable provided the stated
conditions are met. This License explicitly affirms your unlimited
permission to run the unmodified Program. The output from running a
covered work is covered by this License only if the output, given its
content, constitutes a covered work. This License acknowledges your
rights of fair use or other equivalent, as provided by copyright law.
You may make, run and propagate covered works that you do not
convey, without conditions so long as your license otherwise remains
in force. You may convey covered works to others for the sole purpose
of having them make modifications exclusively for you, or provide you
with facilities for running those works, provided that you comply with
the terms of this License in conveying all material for which you do
not control copyright. Those thus making or running the covered works
for you must do so exclusively on your behalf, under your direction
and control, on terms that prohibit them from making any copies of
your copyrighted material outside their relationship with you.
Conveying under any other circumstances is permitted solely under
the conditions stated below. Sublicensing is not allowed; section 10
makes it unnecessary.
3. Protecting Users' Legal Rights From Anti-Circumvention Law.
No covered work shall be deemed part of an effective technological
measure under any applicable law fulfilling obligations under article
11 of the WIPO copyright treaty adopted on 20 December 1996, or
similar laws prohibiting or restricting circumvention of such
measures.
When you convey a covered work, you waive any legal power to forbid
circumvention of technological measures to the extent such circumvention
is effected by exercising rights under this License with respect to
the covered work, and you disclaim any intention to limit operation or
modification of the work as a means of enforcing, against the work's
users, your or third parties' legal rights to forbid circumvention of
technological measures.
4. Conveying Verbatim Copies.
You may convey verbatim copies of the Program's source code as you
receive it, in any medium, provided that you conspicuously and
appropriately publish on each copy an appropriate copyright notice;
keep intact all notices stating that this License and any
non-permissive terms added in accord with section 7 apply to the code;
keep intact all notices of the absence of any warranty; and give all
recipients a copy of this License along with the Program.
You may charge any price or no price for each copy that you convey,
and you may offer support or warranty protection for a fee.
5. Conveying Modified Source Versions.
You may convey a work based on the Program, or the modifications to
produce it from the Program, in the form of source code under the
terms of section 4, provided that you also meet all of these conditions:
a) The work must carry prominent notices stating that you modified
it, and giving a relevant date.
b) The work must carry prominent notices stating that it is
released under this License and any conditions added under section
7. This requirement modifies the requirement in section 4 to
"keep intact all notices".
c) You must license the entire work, as a whole, under this
License to anyone who comes into possession of a copy. This
License will therefore apply, along with any applicable section 7
additional terms, to the whole of the work, and all its parts,
regardless of how they are packaged. This License gives no
permission to license the work in any other way, but it does not
invalidate such permission if you have separately received it.
d) If the work has interactive user interfaces, each must display
Appropriate Legal Notices; however, if the Program has interactive
interfaces that do not display Appropriate Legal Notices, your
work need not make them do so.
A compilation of a covered work with other separate and independent
works, which are not by their nature extensions of the covered work,
and which are not combined with it such as to form a larger program,
in or on a volume of a storage or distribution medium, is called an
"aggregate" if the compilation and its resulting copyright are not
used to limit the access or legal rights of the compilation's users
beyond what the individual works permit. Inclusion of a covered work
in an aggregate does not cause this License to apply to the other
parts of the aggregate.
6. Conveying Non-Source Forms.
You may convey a covered work in object code form under the terms
of sections 4 and 5, provided that you also convey the
machine-readable Corresponding Source under the terms of this License,
in one of these ways:
a) Convey the object code in, or embodied in, a physical product
(including a physical distribution medium), accompanied by the
Corresponding Source fixed on a durable physical medium
customarily used for software interchange.
b) Convey the object code in, or embodied in, a physical product
(including a physical distribution medium), accompanied by a
written offer, valid for at least three years and valid for as
long as you offer spare parts or customer support for that product
model, to give anyone who possesses the object code either (1) a
copy of the Corresponding Source for all the software in the
product that is covered by this License, on a durable physical
medium customarily used for software interchange, for a price no
more than your reasonable cost of physically performing this
conveying of source, or (2) access to copy the
Corresponding Source from a network server at no charge.
c) Convey individual copies of the object code with a copy of the
written offer to provide the Corresponding Source. This
alternative is allowed only occasionally and noncommercially, and
only if you received the object code with such an offer, in accord
with subsection 6b.
d) Convey the object code by offering access from a designated
place (gratis or for a charge), and offer equivalent access to the
Corresponding Source in the same way through the same place at no
further charge. You need not require recipients to copy the
Corresponding Source along with the object code. If the place to
copy the object code is a network server, the Corresponding Source
may be on a different server (operated by you or a third party)
that supports equivalent copying facilities, provided you maintain
clear directions next to the object code saying where to find the
Corresponding Source. Regardless of what server hosts the
Corresponding Source, you remain obligated to ensure that it is
available for as long as needed to satisfy these requirements.
e) Convey the object code using peer-to-peer transmission, provided
you inform other peers where the object code and Corresponding
Source of the work are being offered to the general public at no
charge under subsection 6d.
A separable portion of the object code, whose source code is excluded
from the Corresponding Source as a System Library, need not be
included in conveying the object code work.
A "User Product" is either (1) a "consumer product", which means any
tangible personal property which is normally used for personal, family,
or household purposes, or (2) anything designed or sold for incorporation
into a dwelling. In determining whether a product is a consumer product,
doubtful cases shall be resolved in favor of coverage. For a particular
product received by a particular user, "normally used" refers to a
typical or common use of that class of product, regardless of the status
of the particular user or of the way in which the particular user
actually uses, or expects or is expected to use, the product. A product
is a consumer product regardless of whether the product has substantial
commercial, industrial or non-consumer uses, unless such uses represent
the only significant mode of use of the product.
"Installation Information" for a User Product means any methods,
procedures, authorization keys, or other information required to install
and execute modified versions of a covered work in that User Product from
a modified version of its Corresponding Source. The information must
suffice to ensure that the continued functioning of the modified object
code is in no case prevented or interfered with solely because
modification has been made.
If you convey an object code work under this section in, or with, or
specifically for use in, a User Product, and the conveying occurs as
part of a transaction in which the right of possession and use of the
User Product is transferred to the recipient in perpetuity or for a
fixed term (regardless of how the transaction is characterized), the
Corresponding Source conveyed under this section must be accompanied
by the Installation Information. But this requirement does not apply
if neither you nor any third party retains the ability to install
modified object code on the User Product (for example, the work has
been installed in ROM).
The requirement to provide Installation Information does not include a
requirement to continue to provide support service, warranty, or updates
for a work that has been modified or installed by the recipient, or for
the User Product in which it has been modified or installed. Access to a
network may be denied when the modification itself materially and
adversely affects the operation of the network or violates the rules and
protocols for communication across the network.
Corresponding Source conveyed, and Installation Information provided,
in accord with this section must be in a format that is publicly
documented (and with an implementation available to the public in
source code form), and must require no special password or key for
unpacking, reading or copying.
7. Additional Terms.
"Additional permissions" are terms that supplement the terms of this
License by making exceptions from one or more of its conditions.
Additional permissions that are applicable to the entire Program shall
be treated as though they were included in this License, to the extent
that they are valid under applicable law. If additional permissions
apply only to part of the Program, that part may be used separately
under those permissions, but the entire Program remains governed by
this License without regard to the additional permissions.
When you convey a copy of a covered work, you may at your option
remove any additional permissions from that copy, or from any part of
it. (Additional permissions may be written to require their own
removal in certain cases when you modify the work.) You may place
additional permissions on material, added by you to a covered work,
for which you have or can give appropriate copyright permission.
Notwithstanding any other provision of this License, for material you
add to a covered work, you may (if authorized by the copyright holders of
that material) supplement the terms of this License with terms:
a) Disclaiming warranty or limiting liability differently from the
terms of sections 15 and 16 of this License; or
b) Requiring preservation of specified reasonable legal notices or
author attributions in that material or in the Appropriate Legal
Notices displayed by works containing it; or
c) Prohibiting misrepresentation of the origin of that material, or
requiring that modified versions of such material be marked in
reasonable ways as different from the original version; or
d) Limiting the use for publicity purposes of names of licensors or
authors of the material; or
e) Declining to grant rights under trademark law for use of some
trade names, trademarks, or service marks; or
f) Requiring indemnification of licensors and authors of that
material by anyone who conveys the material (or modified versions of
it) with contractual assumptions of liability to the recipient, for
any liability that these contractual assumptions directly impose on
those licensors and authors.
All other non-permissive additional terms are considered "further
restrictions" within the meaning of section 10. If the Program as you
received it, or any part of it, contains a notice stating that it is
governed by this License along with a term that is a further
restriction, you may remove that term. If a license document contains
a further restriction but permits relicensing or conveying under this
License, you may add to a covered work material governed by the terms
of that license document, provided that the further restriction does
not survive such relicensing or conveying.
If you add terms to a covered work in accord with this section, you
must place, in the relevant source files, a statement of the
additional terms that apply to those files, or a notice indicating
where to find the applicable terms.
Additional terms, permissive or non-permissive, may be stated in the
form of a separately written license, or stated as exceptions;
the above requirements apply either way.
8. Termination.
You may not propagate or modify a covered work except as expressly
provided under this License. Any attempt otherwise to propagate or
modify it is void, and will automatically terminate your rights under
this License (including any patent licenses granted under the third
paragraph of section 11).
However, if you cease all violation of this License, then your
license from a particular copyright holder is reinstated (a)
provisionally, unless and until the copyright holder explicitly and
finally terminates your license, and (b) permanently, if the copyright
holder fails to notify you of the violation by some reasonable means
prior to 60 days after the cessation.
Moreover, your license from a particular copyright holder is
reinstated permanently if the copyright holder notifies you of the
violation by some reasonable means, this is the first time you have
received notice of violation of this License (for any work) from that
copyright holder, and you cure the violation prior to 30 days after
your receipt of the notice.
Termination of your rights under this section does not terminate the
licenses of parties who have received copies or rights from you under
this License. If your rights have been terminated and not permanently
reinstated, you do not qualify to receive new licenses for the same
material under section 10.
9. Acceptance Not Required for Having Copies.
You are not required to accept this License in order to receive or
run a copy of the Program. Ancillary propagation of a covered work
occurring solely as a consequence of using peer-to-peer transmission
to receive a copy likewise does not require acceptance. However,
nothing other than this License grants you permission to propagate or
modify any covered work. These actions infringe copyright if you do
not accept this License. Therefore, by modifying or propagating a
covered work, you indicate your acceptance of this License to do so.
10. Automatic Licensing of Downstream Recipients.
Each time you convey a covered work, the recipient automatically
receives a license from the original licensors, to run, modify and
propagate that work, subject to this License. You are not responsible
for enforcing compliance by third parties with this License.
An "entity transaction" is a transaction transferring control of an
organization, or substantially all assets of one, or subdividing an
organization, or merging organizations. If propagation of a covered
work results from an entity transaction, each party to that
transaction who receives a copy of the work also receives whatever
licenses to the work the party's predecessor in interest had or could
give under the previous paragraph, plus a right to possession of the
Corresponding Source of the work from the predecessor in interest, if
the predecessor has it or can get it with reasonable efforts.
You may not impose any further restrictions on the exercise of the
rights granted or affirmed under this License. For example, you may
not impose a license fee, royalty, or other charge for exercise of
rights granted under this License, and you may not initiate litigation
(including a cross-claim or counterclaim in a lawsuit) alleging that
any patent claim is infringed by making, using, selling, offering for
sale, or importing the Program or any portion of it.
11. Patents.
A "contributor" is a copyright holder who authorizes use under this
License of the Program or a work on which the Program is based. The
work thus licensed is called the contributor's "contributor version".
A contributor's "essential patent claims" are all patent claims
owned or controlled by the contributor, whether already acquired or
hereafter acquired, that would be infringed by some manner, permitted
by this License, of making, using, or selling its contributor version,
but do not include claims that would be infringed only as a
consequence of further modification of the contributor version. For
purposes of this definition, "control" includes the right to grant
patent sublicenses in a manner consistent with the requirements of
this License.
Each contributor grants you a non-exclusive, worldwide, royalty-free
patent license under the contributor's essential patent claims, to
make, use, sell, offer for sale, import and otherwise run, modify and
propagate the contents of its contributor version.
In the following three paragraphs, a "patent license" is any express
agreement or commitment, however denominated, not to enforce a patent
(such as an express permission to practice a patent or covenant not to
sue for patent infringement). To "grant" such a patent license to a
party means to make such an agreement or commitment not to enforce a
patent against the party.
If you convey a covered work, knowingly relying on a patent license,
and the Corresponding Source of the work is not available for anyone
to copy, free of charge and under the terms of this License, through a
publicly available network server or other readily accessible means,
then you must either (1) cause the Corresponding Source to be so
available, or (2) arrange to deprive yourself of the benefit of the
patent license for this particular work, or (3) arrange, in a manner
consistent with the requirements of this License, to extend the patent
license to downstream recipients. "Knowingly relying" means you have
actual knowledge that, but for the patent license, your conveying the
covered work in a country, or your recipient's use of the covered work
in a country, would infringe one or more identifiable patents in that
country that you have reason to believe are valid.
If, pursuant to or in connection with a single transaction or
arrangement, you convey, or propagate by procuring conveyance of, a
covered work, and grant a patent license to some of the parties
receiving the covered work authorizing them to use, propagate, modify
or convey a specific copy of the covered work, then the patent license
you grant is automatically extended to all recipients of the covered
work and works based on it.
A patent license is "discriminatory" if it does not include within
the scope of its coverage, prohibits the exercise of, or is
conditioned on the non-exercise of one or more of the rights that are
specifically granted under this License. You may not convey a covered
work if you are a party to an arrangement with a third party that is
in the business of distributing software, under which you make payment
to the third party based on the extent of your activity of conveying
the work, and under which the third party grants, to any of the
parties who would receive the covered work from you, a discriminatory
patent license (a) in connection with copies of the covered work
conveyed by you (or copies made from those copies), or (b) primarily
for and in connection with specific products or compilations that
contain the covered work, unless you entered into that arrangement,
or that patent license was granted, prior to 28 March 2007.
Nothing in this License shall be construed as excluding or limiting
any implied license or other defenses to infringement that may
otherwise be available to you under applicable patent law.
12. No Surrender of Others' Freedom.
If conditions are imposed on you (whether by court order, agreement or
otherwise) that contradict the conditions of this License, they do not
excuse you from the conditions of this License. If you cannot convey a
covered work so as to satisfy simultaneously your obligations under this
License and any other pertinent obligations, then as a consequence you may
not convey it at all. For example, if you agree to terms that obligate you
to collect a royalty for further conveying from those to whom you convey
the Program, the only way you could satisfy both those terms and this
License would be to refrain entirely from conveying the Program.
13. Use with the GNU Affero General Public License.
Notwithstanding any other provision of this License, you have
permission to link or combine any covered work with a work licensed
under version 3 of the GNU Affero General Public License into a single
combined work, and to convey the resulting work. The terms of this
License will continue to apply to the part which is the covered work,
but the special requirements of the GNU Affero General Public License,
section 13, concerning interaction through a network will apply to the
combination as such.
14. Revised Versions of this License.
The Free Software Foundation may publish revised and/or new versions of
the GNU General Public License from time to time. Such new versions will
be similar in spirit to the present version, but may differ in detail to
address new problems or concerns.
Each version is given a distinguishing version number. If the
Program specifies that a certain numbered version of the GNU General
Public License "or any later version" applies to it, you have the
option of following the terms and conditions either of that numbered
version or of any later version published by the Free Software
Foundation. If the Program does not specify a version number of the
GNU General Public License, you may choose any version ever published
by the Free Software Foundation.
If the Program specifies that a proxy can decide which future
versions of the GNU General Public License can be used, that proxy's
public statement of acceptance of a version permanently authorizes you
to choose that version for the Program.
Later license versions may give you additional or different
permissions. However, no additional obligations are imposed on any
author or copyright holder as a result of your choosing to follow a
later version.
15. Disclaimer of Warranty.
THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY
APPLICABLE LAW. EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT
HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY
OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO,
THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
PURPOSE. THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM
IS WITH YOU. SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF
ALL NECESSARY SERVICING, REPAIR OR CORRECTION.
16. Limitation of Liability.
IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING
WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS
THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY
GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE
USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF
DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD
PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS),
EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF
SUCH DAMAGES.
17. Interpretation of Sections 15 and 16.
If the disclaimer of warranty and limitation of liability provided
above cannot be given local legal effect according to their terms,
reviewing courts shall apply local law that most closely approximates
an absolute waiver of all civil liability in connection with the
Program, unless a warranty or assumption of liability accompanies a
copy of the Program in return for a fee.
END OF TERMS AND CONDITIONS
How to Apply These Terms to Your New Programs
If you develop a new program, and you want it to be of the greatest
possible use to the public, the best way to achieve this is to make it
free software which everyone can redistribute and change under these terms.
To do so, attach the following notices to the program. It is safest
to attach them to the start of each source file to most effectively
state the exclusion of warranty; and each file should have at least
the "copyright" line and a pointer to where the full notice is found.
<one line to give the program's name and a brief idea of what it does.>
Copyright (C) <year> <name of author>
This program is free software: you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation, either version 3 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License
along with this program. If not, see <https://www.gnu.org/licenses/>.
Also add information on how to contact you by electronic and paper mail.
If the program does terminal interaction, make it output a short
notice like this when it starts in an interactive mode:
<program> Copyright (C) <year> <name of author>
This program comes with ABSOLUTELY NO WARRANTY; for details type `show w'.
This is free software, and you are welcome to redistribute it
under certain conditions; type `show c' for details.
The hypothetical commands `show w' and `show c' should show the appropriate
parts of the General Public License. Of course, your program's commands
might be different; for a GUI interface, you would use an "about box".
You should also get your employer (if you work as a programmer) or school,
if any, to sign a "copyright disclaimer" for the program, if necessary.
For more information on this, and how to apply and follow the GNU GPL, see
<https://www.gnu.org/licenses/>.
The GNU General Public License does not permit incorporating your program
into proprietary programs. If your program is a subroutine library, you
may consider it more useful to permit linking proprietary applications with
the library. If this is what you want to do, use the GNU Lesser General
Public License instead of this License. But first, please read
<https://www.gnu.org/licenses/why-not-lgpl.html>.
README.md
+199
View File
@@ -0,0 +1,199 @@
# Node to use APISR upscale models in ComfyUI
# Original repository:
https://github.com/Kiteretsu77/APISR
<p align="center">
<img src="__assets__/logo.png" height="100">
</p>
## APISR: Anime Production Inspired Real-World Anime Super-Resolution (CVPR 2024)
APISR aims at restoring and enhancing low-quality low-resolution anime images and video sources with various degradations from real-world scenarios.
[![Arxiv](https://img.shields.io/badge/Arxiv-<COLOR>.svg)](https://arxiv.org/abs/2403.01598) &ensp; [![HF Demo](https://img.shields.io/static/v1?label=Demo&message=HuggingFace&color=orange)](https://huggingface.co/spaces/HikariDawn/APISR)
👀 [**Visualization**](#Visualization) **|** 🔥 [Update](#Update) **|** 🔧 [Installation](#installation) **|** 🏰 [**Model Zoo**](docs/model_zoo.md) **|** ⚡ [Inference](#inference) **|** 🧩 [Dataset Curation](#dataset_curation) **|** 💻 [Train](#train)
<p align="center">
<img src="__assets__/workflow.png" style="border-radius: 15px">
</p>
:star: If you like APISR, please help star this repo. Thanks! :hugs:
<!---------------------------------------- Visualization ---------------------------------------->
## <a name="Visualization"></a> Visualization (Click them for the best view!) 👀
<!-- Kiteret: https://imgsli.com/MjQ1NzE0 -->
<!-- EVA: https://imgsli.com/MjQ1NzIx -->
<!-- Pokemon: https://imgsli.com/MjQ1NzIy -->
<!-- Pokemon2: https://imgsli.com/MjQ1NzM5 -->
<!-- Gundam0079: https://imgsli.com/MjQ1NzIz -->
<!-- Gundam0079 #2: https://imgsli.com/MjQ1NzMw -->
<!-- f91: https://imgsli.com/MjQ1NzMx -->
<!-- wataru: https://imgsli.com/MjQ1NzMy -->
[<img src="__assets__/visual_results/0079_visual.png" height="223px"/>](https://imgsli.com/MjQ1NzIz) [<img src="__assets__/visual_results/0079_2_visual.png" height="223px"/>](https://imgsli.com/MjQ1NzMw)
[<img src="__assets__/visual_results/pokemon_visual.png" height="223px"/>](https://imgsli.com/MjQ1NzIy) [<img src="__assets__/visual_results/pokemon2_visual.png" height="223px"/>](https://imgsli.com/MjQ1NzM5)
[<img src="__assets__/visual_results/eva_visual.png" height="223px"/>](https://imgsli.com/MjQ1NzIx) [<img src="__assets__/visual_results/kiteret_visual.png" height="223px"/>](https://imgsli.com/MjQ1NzE0)
[<img src="__assets__/visual_results/f91_visual.png" height="223px"/>](https://imgsli.com/MjQ1NzMx) [<img src="__assets__/visual_results/wataru_visual.png" height="223px"/>](https://imgsli.com/MjQ1NzMy)
<p align="center">
<img src="__assets__/AVC_RealLQ_comparison.png">
</p>
<!-------------------------------------------- --------------------------------------------------->
## <a name="Update"></a>Update 🔥🔥🔥
- [x] Release Paper version implementation of APISR
- [x] Release different upscaler factor weight (for 2x, 4x and more)
- [x] Gradio demo (maybe online)
## <a name="installation"></a> Installation 🔧
```shell
git clone git@github.com:Kiteretsu77/APISR.git
cd APISR
# Create conda env
conda create -n APISR python=3.10
conda activate APISR
# Install Pytorch and other packages needed
pip install torch==2.1.1 torchvision==0.16.1 torchaudio==2.1.1 --index-url https://download.pytorch.org/whl/cu118
pip install -r requirements.txt
# To be absolutely sure that the tensorboard can execute. I recommend the following CMD from "https://github.com/pytorch/pytorch/issues/22676#issuecomment-534882021"
pip uninstall tb-nightly tensorboard tensorflow-estimator tensorflow-gpu tf-estimator-nightly
pip install tensorflow
# Install FFMPEG [Only needed for training and dataset curation stage; inference only does not need ffmpeg] (the following is for the linux system, Windows users can download ffmpeg from https://ffmpeg.org/download.html)
sudo apt install ffmpeg
```
## <a name="inference"></a> Gradio Fast Inference ⚡⚡⚡
Gradio option doesn't need to prepare the weight from the user side but they can only process one image each time.
An online demo can be found at https://huggingface.co/spaces/HikariDawn/APISR.
```shell
python gradio_apisr.py
```
## <a name="regular_inference"></a> Regular Inference ⚡⚡
1. Download the model weight from [**model zoo**](docs/model_zoo.md) and **put the weight to "pretrained" folder**.
2. Then, Execute
```shell
python test_code/inference.py --input_dir XXX --weight_path XXX --store_dir XXX
```
If the weight you download is paper weight, the default argument of test_code/inference.py is capable of executing sample images from "__assets__" folder
## <a name="dataset_curation"></a> Dataset Curation 🧩
Our dataset curation pipeline is under **dataset_curation_pipeline** folder.
You can collect your own dataset by sending videos into the pipeline and get the least compressed and the most informative images from the video sources.
1. Download [IC9600](https://github.com/tinglyfeng/IC9600?tab=readme-ov-file) weight (ck.pth) from https://drive.google.com/drive/folders/1N3FSS91e7FkJWUKqT96y_zcsG9CRuIJw and place it at "pretrained/" folder (else, you can define a different **--IC9600_pretrained_weight_path** in the following collect.py execution)
2. With a folder with video sources, you can execute the following to get a basic dataset (with **ffmpeg** installed):
```shell
python dataset_curation_pipeline/collect.py --video_folder_dir XXX --save_dir XXX
```
3. Once you get an image dataset with various aspect ratios and resolutions, you can run the following scripts
Be careful to check **full_patch_source** && **degrade_hr_dataset_path** && **train_hr_dataset_path** (we will use these path in **opt.py** setting during training stage)
In order to decrease memory utilization and increase training efficiency, we pre-process all time-consuming pseudo-GT (**train_hr_dataset_path**) at the dataset preparation stage.
But in order to create a natural input for prediction-oriented compression, in every epoch, the degradation started from the uncropped GT (**full_patch_source**), and LR synthetic images are concurrently stored. The cropped HR GT dataset (**degrade_hr_dataset_path**) and cropped pseudo-GT (**train_hr_dataset_path**) are fixed in the dataset preparation stage and won't be modified during training.
```shell
bash scripts/prepare_datasets.sh
```
## <a name="train"></a> Train 💻
**The whole training process can be done in one RTX3090/4090!**
1. Prepare a dataset (AVC/API) which follows step 2 & 3 in [**Dataset Curation**](#dataset_curation)
In total, you will have 3 folders prepared before executing the following commands:
--> **full_patch_source**: uncropped GT
--> **degrade_hr_dataset_path**: cropped GT
--> **train_hr_dataset_path**: cropped Pseudo-GT
2. Train: Please check **opt.py** carefully to setup parameters you want (modifying **Frequently Changed Setting** is usually enough)
**Step1** (Net **L1** loss training): Run
```shell
python train_code/train.py
```
The trained model weights will be inside the folder 'saved_models' (same to checkpoints)
**Step2** (GAN **Adversarial** Training):
1. Change opt['architecture'] in **opt.py** to "GRLGAN" and change **batch size** if you need. BTW, I don't think that, for personal training, it is needed to train 300K iter for GAN. I did that in order to follow the same setting as in AnimeSR and VQDSR, but **100K ~ 130K** should have a decent visual result.
2. Following previous works, GAN should start from L1 loss pre-trained network, so please carry a **pretrained_path** (the default path below should be fine)
```shell
python train_code/train.py --pretrained_path saved_models/grl_best_generator.pth
```
## Related Projects
1. Fast Anime SR acceleration: https://github.com/Kiteretsu77/FAST_Anime_VSR
2. My previous paper (VCISR - WACV2024) as the baseline method: https://github.com/Kiteretsu77/VCISR-official
## Citation
Please cite us if our work is useful for your research.
```
@article{wang2024apisr,
title={APISR: Anime Production Inspired Real-World Anime Super-Resolution},
author={Wang, Boyang and Yang, Fengyu and Yu, Xihang and Zhang, Chao and Zhao, Hanbin},
journal={arXiv preprint arXiv:2403.01598},
year={2024}
}
```
## Disclaimer
This project is released for academic use only. We disclaim responsibility for the distribution of the dataset. Users are solely liable for their actions.
The project contributors are not legally affiliated with, nor accountable for, users' behaviors.
## License
This project is released under the [GPL 3.0 license](LICENSE).
## Contact
If you have any questions, please feel free to contact me at hikaridawn412316@gmail.com or boyangwa@umich.edu.
__init__.py
+3
View File
@@ -0,0 +1,3 @@
from .nodes import NODE_CLASS_MAPPINGS, NODE_DISPLAY_NAME_MAPPINGS
__all__ = ["NODE_CLASS_MAPPINGS", "NODE_DISPLAY_NAME_MAPPINGS"]
architecture/cunet.py
+189
View File
@@ -0,0 +1,189 @@
# Github Repository: https://github.com/bilibili/ailab/blob/main/Real-CUGAN/README_EN.md
# Code snippet (with certain modificaiton) from: https://github.com/bilibili/ailab/blob/main/Real-CUGAN/VapourSynth/upcunet_v3_vs.py
import torch
from torch import nn as nn
from torch.nn import functional as F
import os, sys
import numpy as np
from time import time as ttime, sleep
class UNet_Full(nn.Module):
def __init__(self):
super(UNet_Full, self).__init__()
self.unet1 = UNet1(3, 3, deconv=True)
self.unet2 = UNet2(3, 3, deconv=False)
def forward(self, x):
n, c, h0, w0 = x.shape
ph = ((h0 - 1) // 2 + 1) * 2
pw = ((w0 - 1) // 2 + 1) * 2
x = F.pad(x, (18, 18 + pw - w0, 18, 18 + ph - h0), 'reflect') # In order to ensure that it can be divided by 2
x1 = self.unet1(x)
x2 = self.unet2(x1)
x1 = F.pad(x1, (-20, -20, -20, -20))
output = torch.add(x2, x1)
if (w0 != pw or h0 != ph):
output = output[:, :, :h0 * 2, :w0 * 2]
return output
class SEBlock(nn.Module):
def __init__(self, in_channels, reduction=8, bias=False):
super(SEBlock, self).__init__()
self.conv1 = nn.Conv2d(in_channels, in_channels // reduction, 1, 1, 0, bias=bias)
self.conv2 = nn.Conv2d(in_channels // reduction, in_channels, 1, 1, 0, bias=bias)
def forward(self, x):
if ("Half" in x.type()): # torch.HalfTensor/torch.cuda.HalfTensor
x0 = torch.mean(x.float(), dim=(2, 3), keepdim=True).half()
else:
x0 = torch.mean(x, dim=(2, 3), keepdim=True)
x0 = self.conv1(x0)
x0 = F.relu(x0, inplace=True)
x0 = self.conv2(x0)
x0 = torch.sigmoid(x0)
x = torch.mul(x, x0)
return x
class UNetConv(nn.Module):
def __init__(self, in_channels, mid_channels, out_channels, se):
super(UNetConv, self).__init__()
self.conv = nn.Sequential(
nn.Conv2d(in_channels, mid_channels, 3, 1, 0),
nn.LeakyReLU(0.1, inplace=True),
nn.Conv2d(mid_channels, out_channels, 3, 1, 0),
nn.LeakyReLU(0.1, inplace=True),
)
if se:
self.seblock = SEBlock(out_channels, reduction=8, bias=True)
else:
self.seblock = None
def forward(self, x):
z = self.conv(x)
if self.seblock is not None:
z = self.seblock(z)
return z
class UNet1(nn.Module):
def __init__(self, in_channels, out_channels, deconv):
super(UNet1, self).__init__()
self.conv1 = UNetConv(in_channels, 32, 64, se=False)
self.conv1_down = nn.Conv2d(64, 64, 2, 2, 0)
self.conv2 = UNetConv(64, 128, 64, se=True)
self.conv2_up = nn.ConvTranspose2d(64, 64, 2, 2, 0)
self.conv3 = nn.Conv2d(64, 64, 3, 1, 0)
if deconv:
self.conv_bottom = nn.ConvTranspose2d(64, out_channels, 4, 2, 3)
else:
self.conv_bottom = nn.Conv2d(64, out_channels, 3, 1, 0)
for m in self.modules():
if isinstance(m, (nn.Conv2d, nn.ConvTranspose2d)):
nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu')
elif isinstance(m, nn.Linear):
nn.init.normal_(m.weight, 0, 0.01)
if m.bias is not None:
nn.init.constant_(m.bias, 0)
def forward(self, x):
x1 = self.conv1(x)
x2 = self.conv1_down(x1)
x2 = F.leaky_relu(x2, 0.1, inplace=True)
x2 = self.conv2(x2)
x2 = self.conv2_up(x2)
x2 = F.leaky_relu(x2, 0.1, inplace=True)
x1 = F.pad(x1, (-4, -4, -4, -4))
x3 = self.conv3(x1 + x2)
x3 = F.leaky_relu(x3, 0.1, inplace=True)
z = self.conv_bottom(x3)
return z
class UNet2(nn.Module):
def __init__(self, in_channels, out_channels, deconv):
super(UNet2, self).__init__()
self.conv1 = UNetConv(in_channels, 32, 64, se=False)
self.conv1_down = nn.Conv2d(64, 64, 2, 2, 0)
self.conv2 = UNetConv(64, 64, 128, se=True)
self.conv2_down = nn.Conv2d(128, 128, 2, 2, 0)
self.conv3 = UNetConv(128, 256, 128, se=True)
self.conv3_up = nn.ConvTranspose2d(128, 128, 2, 2, 0)
self.conv4 = UNetConv(128, 64, 64, se=True)
self.conv4_up = nn.ConvTranspose2d(64, 64, 2, 2, 0)
self.conv5 = nn.Conv2d(64, 64, 3, 1, 0)
if deconv:
self.conv_bottom = nn.ConvTranspose2d(64, out_channels, 4, 2, 3)
else:
self.conv_bottom = nn.Conv2d(64, out_channels, 3, 1, 0)
for m in self.modules():
if isinstance(m, (nn.Conv2d, nn.ConvTranspose2d)):
nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu')
elif isinstance(m, nn.Linear):
nn.init.normal_(m.weight, 0, 0.01)
if m.bias is not None:
nn.init.constant_(m.bias, 0)
def forward(self, x):
x1 = self.conv1(x)
x2 = self.conv1_down(x1)
x2 = F.leaky_relu(x2, 0.1, inplace=True)
x2 = self.conv2(x2)
x3 = self.conv2_down(x2)
x3 = F.leaky_relu(x3, 0.1, inplace=True)
x3 = self.conv3(x3)
x3 = self.conv3_up(x3)
x3 = F.leaky_relu(x3, 0.1, inplace=True)
x2 = F.pad(x2, (-4, -4, -4, -4))
x4 = self.conv4(x2 + x3)
x4 = self.conv4_up(x4)
x4 = F.leaky_relu(x4, 0.1, inplace=True)
x1 = F.pad(x1, (-16, -16, -16, -16))
x5 = self.conv5(x1 + x4)
x5 = F.leaky_relu(x5, 0.1, inplace=True)
z = self.conv_bottom(x5)
return z
def main():
root_path = os.path.abspath('.')
sys.path.append(root_path)
from opt import opt # Manage GPU to choose
import time
model = UNet_Full().cuda()
pytorch_total_params = sum(p.numel() for p in model.parameters())
print(f"CuNet has param {pytorch_total_params//1000} K params")
# Count the number of FLOPs to double check
x = torch.randn((1, 3, 180, 180)).cuda()
start = time.time()
x = model(x)
print("output size is ", x.shape)
total = time.time() - start
print(total)
if __name__ == "__main__":
main()
architecture/discriminator.py
+241
View File
@@ -0,0 +1,241 @@
# -*- coding: utf-8 -*-
from torch import nn as nn
from torch.nn import functional as F
from torch.nn.utils import spectral_norm
import torch
import functools
class UNetDiscriminatorSN(nn.Module):
"""Defines a U-Net discriminator with spectral normalization (SN)
It is used in Real-ESRGAN: Training Real-World Blind Super-Resolution with Pure Synthetic Data.
Arg:
num_in_ch (int): Channel number of inputs. Default: 3.
num_feat (int): Channel number of base intermediate features. Default: 64.
skip_connection (bool): Whether to use skip connections between U-Net. Default: True.
"""
def __init__(self, num_in_ch, num_feat=64, skip_connection=True):
super(UNetDiscriminatorSN, self).__init__()
self.skip_connection = skip_connection
norm = spectral_norm
# the first convolution
self.conv0 = nn.Conv2d(num_in_ch, num_feat, kernel_size=3, stride=1, padding=1)
# downsample
self.conv1 = norm(nn.Conv2d(num_feat, num_feat * 2, 4, 2, 1, bias=False))
self.conv2 = norm(nn.Conv2d(num_feat * 2, num_feat * 4, 4, 2, 1, bias=False))
self.conv3 = norm(nn.Conv2d(num_feat * 4, num_feat * 8, 4, 2, 1, bias=False))
# upsample
self.conv4 = norm(nn.Conv2d(num_feat * 8, num_feat * 4, 3, 1, 1, bias=False))
self.conv5 = norm(nn.Conv2d(num_feat * 4, num_feat * 2, 3, 1, 1, bias=False))
self.conv6 = norm(nn.Conv2d(num_feat * 2, num_feat, 3, 1, 1, bias=False))
# extra convolutions
self.conv7 = norm(nn.Conv2d(num_feat, num_feat, 3, 1, 1, bias=False))
self.conv8 = norm(nn.Conv2d(num_feat, num_feat, 3, 1, 1, bias=False))
self.conv9 = nn.Conv2d(num_feat, 1, 3, 1, 1)
def forward(self, x):
# downsample
x0 = F.leaky_relu(self.conv0(x), negative_slope=0.2, inplace=True)
x1 = F.leaky_relu(self.conv1(x0), negative_slope=0.2, inplace=True)
x2 = F.leaky_relu(self.conv2(x1), negative_slope=0.2, inplace=True)
x3 = F.leaky_relu(self.conv3(x2), negative_slope=0.2, inplace=True)
# upsample
x3 = F.interpolate(x3, scale_factor=2, mode='bilinear', align_corners=False)
x4 = F.leaky_relu(self.conv4(x3), negative_slope=0.2, inplace=True)
if self.skip_connection:
x4 = x4 + x2
x4 = F.interpolate(x4, scale_factor=2, mode='bilinear', align_corners=False)
x5 = F.leaky_relu(self.conv5(x4), negative_slope=0.2, inplace=True)
if self.skip_connection:
x5 = x5 + x1
x5 = F.interpolate(x5, scale_factor=2, mode='bilinear', align_corners=False)
x6 = F.leaky_relu(self.conv6(x5), negative_slope=0.2, inplace=True)
if self.skip_connection:
x6 = x6 + x0
# extra convolutions
out = F.leaky_relu(self.conv7(x6), negative_slope=0.2, inplace=True)
out = F.leaky_relu(self.conv8(out), negative_slope=0.2, inplace=True)
out = self.conv9(out)
return out
def get_conv_layer(input_nc, ndf, kernel_size, stride, padding, bias=True, use_sn=False):
if not use_sn:
return nn.Conv2d(input_nc, ndf, kernel_size=kernel_size, stride=stride, padding=padding, bias=bias)
return spectral_norm(nn.Conv2d(input_nc, ndf, kernel_size=kernel_size, stride=stride, padding=padding, bias=bias))
class PatchDiscriminator(nn.Module):
"""Defines a PatchGAN discriminator, the receptive field of default config is 70x70.
Args:
use_sn (bool): Use spectra_norm or not, if use_sn is True, then norm_type should be none.
"""
def __init__(self,
num_in_ch,
num_feat=64,
num_layers=3,
max_nf_mult=8,
norm_type='batch',
use_sigmoid=False,
use_sn=False):
super(PatchDiscriminator, self).__init__()
norm_layer = self._get_norm_layer(norm_type)
if type(norm_layer) == functools.partial: # no need to use bias as BatchNorm2d has affine parameters
use_bias = norm_layer.func != nn.BatchNorm2d
else:
use_bias = norm_layer != nn.BatchNorm2d
kw = 4
padw = 1
sequence = [
get_conv_layer(num_in_ch, num_feat, kernel_size=kw, stride=2, padding=padw, use_sn=use_sn),
nn.LeakyReLU(0.2, True)
]
nf_mult = 1
nf_mult_prev = 1
for n in range(1, num_layers): # gradually increase the number of filters
nf_mult_prev = nf_mult
nf_mult = min(2**n, max_nf_mult)
sequence += [
get_conv_layer(
num_feat * nf_mult_prev,
num_feat * nf_mult,
kernel_size=kw,
stride=2,
padding=padw,
bias=use_bias,
use_sn=use_sn),
norm_layer(num_feat * nf_mult),
nn.LeakyReLU(0.2, True)
]
nf_mult_prev = nf_mult
nf_mult = min(2**num_layers, max_nf_mult)
sequence += [
get_conv_layer(
num_feat * nf_mult_prev,
num_feat * nf_mult,
kernel_size=kw,
stride=1,
padding=padw,
bias=use_bias,
use_sn=use_sn),
norm_layer(num_feat * nf_mult),
nn.LeakyReLU(0.2, True)
]
# output 1 channel prediction map 我觉得这个应该就是pixel by pixel的feedback反馈
sequence += [get_conv_layer(num_feat * nf_mult, 1, kernel_size=kw, stride=1, padding=padw, use_sn=use_sn)]
if use_sigmoid:
sequence += [nn.Sigmoid()]
self.model = nn.Sequential(*sequence)
def _get_norm_layer(self, norm_type='batch'):
if norm_type == 'batch':
norm_layer = functools.partial(nn.BatchNorm2d, affine=True)
elif norm_type == 'instance':
norm_layer = functools.partial(nn.InstanceNorm2d, affine=False)
elif norm_type == 'batchnorm2d':
norm_layer = nn.BatchNorm2d
elif norm_type == 'none':
norm_layer = nn.Identity
else:
raise NotImplementedError(f'normalization layer [{norm_type}] is not found')
return norm_layer
def forward(self, x):
return self.model(x)
class MultiScaleDiscriminator(nn.Module):
"""Define a multi-scale discriminator, each discriminator is a instance of PatchDiscriminator.
Args:
num_layers (int or list): If the type of this variable is int, then degrade to PatchDiscriminator.
If the type of this variable is list, then the length of the list is
the number of discriminators.
use_downscale (bool): Progressive downscale the input to feed into different discriminators.
If set to True, then the discriminators are usually the same.
"""
def __init__(self,
num_in_ch,
num_feat=64,
num_layers=[3, 3, 3],
max_nf_mult=8,
norm_type='none',
use_sigmoid=False,
use_sn=True,
use_downscale=True):
super(MultiScaleDiscriminator, self).__init__()
if isinstance(num_layers, int):
num_layers = [num_layers]
# check whether the discriminators are the same
if use_downscale:
assert len(set(num_layers)) == 1
self.use_downscale = use_downscale
self.num_dis = len(num_layers)
self.dis_list = nn.ModuleList()
for nl in num_layers:
self.dis_list.append(
PatchDiscriminator(
num_in_ch,
num_feat=num_feat,
num_layers=nl,
max_nf_mult=max_nf_mult,
norm_type=norm_type,
use_sigmoid=use_sigmoid,
use_sn=use_sn,
))
def forward(self, x):
outs = []
h, w = x.size()[2:]
y = x
for i in range(self.num_dis):
if i != 0 and self.use_downscale:
y = F.interpolate(y, size=(h // 2, w // 2), mode='bilinear', align_corners=True)
h, w = y.size()[2:]
outs.append(self.dis_list[i](y))
return outs
#def main():
#from pthflops import count_ops
#from torchsummary import summary
#model = UNetDiscriminatorSN(3)
#pytorch_total_params = sum(p.numel() for p in model.parameters())
# Create a network and a corresponding input
#device = 'cuda'
#inp = torch.rand(1, 3, 400, 400)
# Count the number of FLOPs
#count_ops(model, inp)
#summary(model.cuda(), (3, 400, 400), batch_size=1)
# print(f"pathGAN has param {pytorch_total_params//1000} K params")
#if __name__ == "__main__":
# main()
architecture/grl.py
+616
View File
@@ -0,0 +1,616 @@
"""
Efficient and Explicit Modelling of Image Hierarchies for Image Restoration
Image restoration transformers with global, regional, and local modelling
A clean version of the.
Shared buffers are used for relative_coords_table, relative_position_index, and attn_mask.
"""
#import cv2
import torch
import torch.nn as nn
import torch.nn.functional as F
from torchvision.transforms import ToTensor
from torchvision.utils import save_image
#from fairscale.nn import checkpoint_wrapper
from omegaconf import OmegaConf
from timm.models.layers import to_2tuple, trunc_normal_
# Import files from local folder
import os, sys
root_path = os.path.abspath('.')
sys.path.append(root_path)
from .grl_common import Upsample, UpsampleOneStep
from .grl_common.mixed_attn_block_efficient import (
_get_stripe_info,
EfficientMixAttnTransformerBlock,
)
from .grl_common.ops import (
bchw_to_blc,
blc_to_bchw,
calculate_mask,
calculate_mask_all,
get_relative_coords_table_all,
get_relative_position_index_simple,
)
from .grl_common.swin_v1_block import (
build_last_conv,
)
class TransformerStage(nn.Module):
"""Transformer stage.
Args:
dim (int): Number of input channels.
input_resolution (tuple[int]): Input resolution.
depth (int): Number of blocks.
num_heads_window (list[int]): Number of window attention heads in different layers.
num_heads_stripe (list[int]): Number of stripe attention heads in different layers.
stripe_size (list[int]): Stripe size. Default: [8, 8]
stripe_groups (list[int]): Number of stripe groups. Default: [None, None].
stripe_shift (bool): whether to shift the stripes. This is used as an ablation study.
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
qkv_proj_type (str): QKV projection type. Default: linear. Choices: linear, separable_conv.
anchor_proj_type (str): Anchor projection type. Default: avgpool. Choices: avgpool, maxpool, conv2d, separable_conv, patchmerging.
anchor_one_stage (bool): Whether to use one operator or multiple progressive operators to reduce feature map resolution. Default: True.
anchor_window_down_factor (int): The downscale factor used to get the anchors.
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
pretrained_window_size (list[int]): pretrained window size. This is actually not used. Default: [0, 0].
pretrained_stripe_size (list[int]): pretrained stripe size. This is actually not used. Default: [0, 0].
conv_type: The convolutional block before residual connection.
init_method: initialization method of the weight parameters used to train large scale models.
Choices: n, normal -- Swin V1 init method.
l, layernorm -- Swin V2 init method. Zero the weight and bias in the post layer normalization layer.
r, res_rescale -- EDSR rescale method. Rescale the residual blocks with a scaling factor 0.1
w, weight_rescale -- MSRResNet rescale method. Rescale the weight parameter in residual blocks with a scaling factor 0.1
t, trunc_normal_ -- nn.Linear, trunc_normal; nn.Conv2d, weight_rescale
fairscale_checkpoint (bool): Whether to use fairscale checkpoint.
offload_to_cpu (bool): used by fairscale_checkpoint
args:
out_proj_type (str): Type of the output projection in the self-attention modules. Default: linear. Choices: linear, conv2d.
local_connection (bool): Whether to enable the local modelling module (two convs followed by Channel attention). For GRL base model, this is used. "local_connection": local_connection,
euclidean_dist (bool): use Euclidean distance or inner product as the similarity metric. An ablation study.
"""
def __init__(
self,
dim,
input_resolution,
depth,
num_heads_window,
num_heads_stripe,
window_size,
stripe_size,
stripe_groups,
stripe_shift,
mlp_ratio=4.0,
qkv_bias=True,
qkv_proj_type="linear",
anchor_proj_type="avgpool",
anchor_one_stage=True,
anchor_window_down_factor=1,
drop=0.0,
attn_drop=0.0,
drop_path=0.0,
norm_layer=nn.LayerNorm,
pretrained_window_size=[0, 0],
pretrained_stripe_size=[0, 0],
conv_type="1conv",
init_method="",
fairscale_checkpoint=False,
offload_to_cpu=False,
args=None,
):
super().__init__()
self.dim = dim
self.input_resolution = input_resolution
self.init_method = init_method
self.blocks = nn.ModuleList()
for i in range(depth):
block = EfficientMixAttnTransformerBlock(
dim=dim,
input_resolution=input_resolution,
num_heads_w=num_heads_window,
num_heads_s=num_heads_stripe,
window_size=window_size,
window_shift=i % 2 == 0,
stripe_size=stripe_size,
stripe_groups=stripe_groups,
stripe_type="H" if i % 2 == 0 else "W",
stripe_shift=i % 4 in [2, 3] if stripe_shift else False,
mlp_ratio=mlp_ratio,
qkv_bias=qkv_bias,
qkv_proj_type=qkv_proj_type,
anchor_proj_type=anchor_proj_type,
anchor_one_stage=anchor_one_stage,
anchor_window_down_factor=anchor_window_down_factor,
drop=drop,
attn_drop=attn_drop,
drop_path=drop_path[i] if isinstance(drop_path, list) else drop_path,
norm_layer=norm_layer,
pretrained_window_size=pretrained_window_size,
pretrained_stripe_size=pretrained_stripe_size,
res_scale=0.1 if init_method == "r" else 1.0,
args=args,
)
# print(fairscale_checkpoint, offload_to_cpu)
if fairscale_checkpoint:
block = checkpoint_wrapper(block, offload_to_cpu=offload_to_cpu)
self.blocks.append(block)
self.conv = build_last_conv(conv_type, dim)
def _init_weights(self):
for n, m in self.named_modules():
if self.init_method == "w":
if isinstance(m, (nn.Linear, nn.Conv2d)) and n.find("cpb_mlp") < 0:
print("nn.Linear and nn.Conv2d weight initilization")
m.weight.data *= 0.1
elif self.init_method == "l":
if isinstance(m, nn.LayerNorm):
print("nn.LayerNorm initialization")
nn.init.constant_(m.bias, 0)
nn.init.constant_(m.weight, 0)
elif self.init_method.find("t") >= 0:
scale = 0.1 ** (len(self.init_method) - 1) * int(self.init_method[-1])
if isinstance(m, nn.Linear) and n.find("cpb_mlp") < 0:
trunc_normal_(m.weight, std=scale)
elif isinstance(m, nn.Conv2d):
m.weight.data *= 0.1
print(
"Initialization nn.Linear - trunc_normal; nn.Conv2d - weight rescale."
)
else:
raise NotImplementedError(
f"Parameter initialization method {self.init_method} not implemented in TransformerStage."
)
def forward(self, x, x_size, table_index_mask):
res = x
for blk in self.blocks:
res = blk(res, x_size, table_index_mask)
res = bchw_to_blc(self.conv(blc_to_bchw(res, x_size)))
return res + x
def flops(self):
pass
class GRL(nn.Module):
r"""Image restoration transformer with global, non-local, and local connections
Args:
img_size (int | list[int]): Input image size. Default 64
in_channels (int): Number of input image channels. Default: 3
out_channels (int): Number of output image channels. Default: None
embed_dim (int): Patch embedding dimension. Default: 96
upscale (int): Upscale factor. 2/3/4/8 for image SR, 1 for denoising and compress artifact reduction
img_range (float): Image range. 1. or 255.
upsampler (str): The reconstruction reconstruction module. 'pixelshuffle'/'pixelshuffledirect'/'nearest+conv'/None
depths (list[int]): Depth of each Swin Transformer layer.
num_heads_window (list[int]): Number of window attention heads in different layers.
num_heads_stripe (list[int]): Number of stripe attention heads in different layers.
window_size (int): Window size. Default: 8.
stripe_size (list[int]): Stripe size. Default: [8, 8]
stripe_groups (list[int]): Number of stripe groups. Default: [None, None].
stripe_shift (bool): whether to shift the stripes. This is used as an ablation study.
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
qkv_proj_type (str): QKV projection type. Default: linear. Choices: linear, separable_conv.
anchor_proj_type (str): Anchor projection type. Default: avgpool. Choices: avgpool, maxpool, conv2d, separable_conv, patchmerging.
anchor_one_stage (bool): Whether to use one operator or multiple progressive operators to reduce feature map resolution. Default: True.
anchor_window_down_factor (int): The downscale factor used to get the anchors.
out_proj_type (str): Type of the output projection in the self-attention modules. Default: linear. Choices: linear, conv2d.
local_connection (bool): Whether to enable the local modelling module (two convs followed by Channel attention). For GRL base model, this is used.
drop_rate (float): Dropout rate. Default: 0
attn_drop_rate (float): Attention dropout rate. Default: 0
drop_path_rate (float): Stochastic depth rate. Default: 0.1
pretrained_window_size (list[int]): pretrained window size. This is actually not used. Default: [0, 0].
pretrained_stripe_size (list[int]): pretrained stripe size. This is actually not used. Default: [0, 0].
norm_layer (nn.Module): Normalization layer. Default: nn.LayerNorm.
conv_type (str): The convolutional block before residual connection. Default: 1conv. Choices: 1conv, 3conv, 1conv1x1, linear
init_method: initialization method of the weight parameters used to train large scale models.
Choices: n, normal -- Swin V1 init method.
l, layernorm -- Swin V2 init method. Zero the weight and bias in the post layer normalization layer.
r, res_rescale -- EDSR rescale method. Rescale the residual blocks with a scaling factor 0.1
w, weight_rescale -- MSRResNet rescale method. Rescale the weight parameter in residual blocks with a scaling factor 0.1
t, trunc_normal_ -- nn.Linear, trunc_normal; nn.Conv2d, weight_rescale
fairscale_checkpoint (bool): Whether to use fairscale checkpoint.
offload_to_cpu (bool): used by fairscale_checkpoint
euclidean_dist (bool): use Euclidean distance or inner product as the similarity metric. An ablation study.
"""
def __init__(
self,
img_size=64,
in_channels=3,
out_channels=None,
embed_dim=96,
upscale=2,
img_range=1.0,
upsampler="",
depths=[6, 6, 6, 6, 6, 6],
num_heads_window=[3, 3, 3, 3, 3, 3],
num_heads_stripe=[3, 3, 3, 3, 3, 3],
window_size=8,
stripe_size=[8, 8], # used for stripe window attention
stripe_groups=[None, None],
stripe_shift=False,
mlp_ratio=4.0,
qkv_bias=True,
qkv_proj_type="linear",
anchor_proj_type="avgpool",
anchor_one_stage=True,
anchor_window_down_factor=1,
out_proj_type="linear",
local_connection=False,
drop_rate=0.0,
attn_drop_rate=0.0,
drop_path_rate=0.1,
norm_layer=nn.LayerNorm,
pretrained_window_size=[0, 0],
pretrained_stripe_size=[0, 0],
conv_type="1conv",
init_method="n", # initialization method of the weight parameters used to train large scale models.
fairscale_checkpoint=False, # fairscale activation checkpointing
offload_to_cpu=False,
euclidean_dist=False,
**kwargs,
):
super(GRL, self).__init__()
# Process the input arguments
out_channels = out_channels or in_channels
self.in_channels = in_channels
self.out_channels = out_channels
num_out_feats = 64
self.embed_dim = embed_dim
self.upscale = upscale
self.upsampler = upsampler
self.img_range = img_range
if in_channels == 3:
rgb_mean = (0.4488, 0.4371, 0.4040)
self.mean = torch.Tensor(rgb_mean).view(1, 3, 1, 1)
else:
self.mean = torch.zeros(1, 1, 1, 1)
max_stripe_size = max([0 if s is None else s for s in stripe_size])
max_stripe_groups = max([0 if s is None else s for s in stripe_groups])
max_stripe_groups *= anchor_window_down_factor
self.pad_size = max(window_size, max_stripe_size, max_stripe_groups)
# if max_stripe_size >= window_size:
# self.pad_size *= anchor_window_down_factor
# if stripe_groups[0] is None and stripe_groups[1] is None:
# self.pad_size = max(stripe_size)
# else:
# self.pad_size = window_size
self.input_resolution = to_2tuple(img_size)
self.window_size = to_2tuple(window_size)
self.shift_size = [w // 2 for w in self.window_size]
self.stripe_size = stripe_size
self.stripe_groups = stripe_groups
self.pretrained_window_size = pretrained_window_size
self.pretrained_stripe_size = pretrained_stripe_size
self.anchor_window_down_factor = anchor_window_down_factor
# Head of the network. First convolution.
self.conv_first = nn.Conv2d(in_channels, embed_dim, 3, 1, 1)
# Body of the network
self.norm_start = norm_layer(embed_dim)
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
args = OmegaConf.create(
{
"out_proj_type": out_proj_type,
"local_connection": local_connection,
"euclidean_dist": euclidean_dist,
}
)
for k, v in self.set_table_index_mask(self.input_resolution).items():
self.register_buffer(k, v)
self.layers = nn.ModuleList()
for i in range(len(depths)):
layer = TransformerStage(
dim=embed_dim,
input_resolution=self.input_resolution,
depth=depths[i],
num_heads_window=num_heads_window[i],
num_heads_stripe=num_heads_stripe[i],
window_size=self.window_size,
stripe_size=stripe_size,
stripe_groups=stripe_groups,
stripe_shift=stripe_shift,
mlp_ratio=mlp_ratio,
qkv_bias=qkv_bias,
qkv_proj_type=qkv_proj_type,
anchor_proj_type=anchor_proj_type,
anchor_one_stage=anchor_one_stage,
anchor_window_down_factor=anchor_window_down_factor,
drop=drop_rate,
attn_drop=attn_drop_rate,
drop_path=dpr[
sum(depths[:i]) : sum(depths[: i + 1])
], # no impact on SR results
norm_layer=norm_layer,
pretrained_window_size=pretrained_window_size,
pretrained_stripe_size=pretrained_stripe_size,
conv_type=conv_type,
init_method=init_method,
fairscale_checkpoint=fairscale_checkpoint,
offload_to_cpu=offload_to_cpu,
args=args,
)
self.layers.append(layer)
self.norm_end = norm_layer(embed_dim)
# Tail of the network
self.conv_after_body = build_last_conv(conv_type, embed_dim)
#####################################################################################################
################################ 3, high quality image reconstruction ################################
if self.upsampler == "pixelshuffle":
# for classical SR
self.conv_before_upsample = nn.Sequential(
nn.Conv2d(embed_dim, num_out_feats, 3, 1, 1), nn.LeakyReLU(inplace=True)
)
self.upsample = Upsample(upscale, num_out_feats)
self.conv_last = nn.Conv2d(num_out_feats, out_channels, 3, 1, 1)
elif self.upsampler == "pixelshuffledirect":
# for lightweight SR (to save parameters)
self.upsample = UpsampleOneStep(
upscale,
embed_dim,
out_channels,
)
elif self.upsampler == "nearest+conv":
# for real-world SR (less artifacts)
assert self.upscale == 4, "only support x4 now."
self.conv_before_upsample = nn.Sequential(
nn.Conv2d(embed_dim, num_out_feats, 3, 1, 1), nn.LeakyReLU(inplace=True)
)
self.conv_up1 = nn.Conv2d(num_out_feats, num_out_feats, 3, 1, 1)
self.conv_up2 = nn.Conv2d(num_out_feats, num_out_feats, 3, 1, 1)
self.conv_hr = nn.Conv2d(num_out_feats, num_out_feats, 3, 1, 1)
self.conv_last = nn.Conv2d(num_out_feats, out_channels, 3, 1, 1)
self.lrelu = nn.LeakyReLU(negative_slope=0.2, inplace=True)
else:
# for image denoising and JPEG compression artifact reduction
self.conv_last = nn.Conv2d(embed_dim, out_channels, 3, 1, 1)
self.apply(self._init_weights)
if init_method in ["l", "w"] or init_method.find("t") >= 0:
for layer in self.layers:
layer._init_weights()
def set_table_index_mask(self, x_size):
"""
Two used cases:
1) At initialization: set the shared buffers.
2) During forward pass: get the new buffers if the resolution of the input changes
"""
# ss - stripe_size, sss - stripe_shift_size
ss, sss = _get_stripe_info(self.stripe_size, self.stripe_groups, True, x_size)
df = self.anchor_window_down_factor
table_w = get_relative_coords_table_all(
self.window_size, self.pretrained_window_size
)
table_sh = get_relative_coords_table_all(ss, self.pretrained_stripe_size, df)
table_sv = get_relative_coords_table_all(
ss[::-1], self.pretrained_stripe_size, df
)
index_w = get_relative_position_index_simple(self.window_size)
index_sh_a2w = get_relative_position_index_simple(ss, df, False)
index_sh_w2a = get_relative_position_index_simple(ss, df, True)
index_sv_a2w = get_relative_position_index_simple(ss[::-1], df, False)
index_sv_w2a = get_relative_position_index_simple(ss[::-1], df, True)
mask_w = calculate_mask(x_size, self.window_size, self.shift_size)
mask_sh_a2w = calculate_mask_all(x_size, ss, sss, df, False)
mask_sh_w2a = calculate_mask_all(x_size, ss, sss, df, True)
mask_sv_a2w = calculate_mask_all(x_size, ss[::-1], sss[::-1], df, False)
mask_sv_w2a = calculate_mask_all(x_size, ss[::-1], sss[::-1], df, True)
return {
"table_w": table_w,
"table_sh": table_sh,
"table_sv": table_sv,
"index_w": index_w,
"index_sh_a2w": index_sh_a2w,
"index_sh_w2a": index_sh_w2a,
"index_sv_a2w": index_sv_a2w,
"index_sv_w2a": index_sv_w2a,
"mask_w": mask_w,
"mask_sh_a2w": mask_sh_a2w,
"mask_sh_w2a": mask_sh_w2a,
"mask_sv_a2w": mask_sv_a2w,
"mask_sv_w2a": mask_sv_w2a,
}
def get_table_index_mask(self, device=None, input_resolution=None):
# Used during forward pass
if input_resolution == self.input_resolution:
return {
"table_w": self.table_w,
"table_sh": self.table_sh,
"table_sv": self.table_sv,
"index_w": self.index_w,
"index_sh_a2w": self.index_sh_a2w,
"index_sh_w2a": self.index_sh_w2a,
"index_sv_a2w": self.index_sv_a2w,
"index_sv_w2a": self.index_sv_w2a,
"mask_w": self.mask_w,
"mask_sh_a2w": self.mask_sh_a2w,
"mask_sh_w2a": self.mask_sh_w2a,
"mask_sv_a2w": self.mask_sv_a2w,
"mask_sv_w2a": self.mask_sv_w2a,
}
else:
table_index_mask = self.set_table_index_mask(input_resolution)
for k, v in table_index_mask.items():
table_index_mask[k] = v.to(device)
return table_index_mask
def _init_weights(self, m):
if isinstance(m, nn.Linear):
# Only used to initialize linear layers
# weight_shape = m.weight.shape
# if weight_shape[0] > 256 and weight_shape[1] > 256:
# std = 0.004
# else:
# std = 0.02
# print(f"Standard deviation during initialization {std}.")
trunc_normal_(m.weight, std=0.02)
if isinstance(m, nn.Linear) and m.bias is not None:
nn.init.constant_(m.bias, 0)
elif isinstance(m, nn.LayerNorm):
nn.init.constant_(m.bias, 0)
nn.init.constant_(m.weight, 1.0)
@torch.jit.ignore
def no_weight_decay(self):
return {"absolute_pos_embed"}
@torch.jit.ignore
def no_weight_decay_keywords(self):
return {"relative_position_bias_table"}
def check_image_size(self, x):
_, _, h, w = x.size()
mod_pad_h = (self.pad_size - h % self.pad_size) % self.pad_size
mod_pad_w = (self.pad_size - w % self.pad_size) % self.pad_size
# print("padding size", h, w, self.pad_size, mod_pad_h, mod_pad_w)
try:
x = F.pad(x, (0, mod_pad_w, 0, mod_pad_h), "reflect")
except BaseException:
x = F.pad(x, (0, mod_pad_w, 0, mod_pad_h), "constant")
return x
def forward_features(self, x):
x_size = (x.shape[2], x.shape[3])
x = bchw_to_blc(x)
x = self.norm_start(x)
x = self.pos_drop(x)
table_index_mask = self.get_table_index_mask(x.device, x_size)
for layer in self.layers:
x = layer(x, x_size, table_index_mask)
x = self.norm_end(x) # B L C
x = blc_to_bchw(x, x_size)
return x
def forward(self, x):
H, W = x.shape[2:]
x = self.check_image_size(x)
self.mean = self.mean.type_as(x)
x = (x - self.mean) * self.img_range
if self.upsampler == "pixelshuffle":
# for classical SR
x = self.conv_first(x)
x = self.conv_after_body(self.forward_features(x)) + x
x = self.conv_before_upsample(x)
x = self.conv_last(self.upsample(x))
elif self.upsampler == "pixelshuffledirect":
# for lightweight SR
x = self.conv_first(x)
x = self.conv_after_body(self.forward_features(x)) + x
x = self.upsample(x)
elif self.upsampler == "nearest+conv":
# for real-world SR (claimed to have less artifacts)
x = self.conv_first(x)
x = self.conv_after_body(self.forward_features(x)) + x
x = self.conv_before_upsample(x)
x = self.lrelu(
self.conv_up1(
torch.nn.functional.interpolate(x, scale_factor=2, mode="nearest")
)
)
x = self.lrelu(
self.conv_up2(
torch.nn.functional.interpolate(x, scale_factor=2, mode="nearest")
)
)
x = self.conv_last(self.lrelu(self.conv_hr(x)))
else:
# for image denoising and JPEG compression artifact reduction
x_first = self.conv_first(x)
res = self.conv_after_body(self.forward_features(x_first)) + x_first
if self.in_channels == self.out_channels:
x = x + self.conv_last(res)
else:
x = self.conv_last(res)
x = x / self.img_range + self.mean
return x[:, :, : H * self.upscale, : W * self.upscale]
def flops(self):
pass
def convert_checkpoint(self, state_dict):
for k in list(state_dict.keys()):
if (
k.find("relative_coords_table") >= 0
or k.find("relative_position_index") >= 0
or k.find("attn_mask") >= 0
or k.find("model.table_") >= 0
or k.find("model.index_") >= 0
or k.find("model.mask_") >= 0
# or k.find(".upsample.") >= 0
):
state_dict.pop(k)
print(k)
return state_dict
if __name__ == "__main__":
# The version of GRL we use
model = GRL(
upscale = 4,
img_size = 64,
window_size = 8,
depths = [4, 4, 4, 4],
embed_dim = 64,
num_heads_window = [2, 2, 2, 2],
num_heads_stripe = [2, 2, 2, 2],
mlp_ratio = 2,
qkv_proj_type = "linear",
anchor_proj_type = "avgpool",
anchor_window_down_factor = 2,
out_proj_type = "linear",
conv_type = "1conv",
upsampler = "nearest+conv", # Change
).cuda()
# Parameter analysis
num_params = 0
for p in model.parameters():
if p.requires_grad:
num_params += p.numel()
print(f"Number of parameters {num_params / 10 ** 6: 0.2f}")
# Print param
for name, param in model.named_parameters():
print(name, param.dtype)
# Count the number of FLOPs to double check
x = torch.randn((1, 3, 180, 180)).cuda() # Don't use input size that is too big (we don't have @torch.no_grad here)
x = model(x)
print("output size is ", x.shape)
architecture/grl_common/__init__.py
+8
View File
@@ -0,0 +1,8 @@
from .resblock import ResBlock
from .upsample import (
Upsample,
UpsampleOneStep,
)
__all__ = ["Upsample", "UpsampleOneStep", "ResBlock"]
architecture/grl_common/common_edsr.py
+227
View File
@@ -0,0 +1,227 @@
"""
EDSR common.py
Since a lot of models are developed on top of EDSR, here we include some common functions from EDSR.
In this repository, the common functions is used by edsr_esa.py and ipt.py
"""
import math
import torch
import torch.nn as nn
import torch.nn.functional as F
def default_conv(in_channels, out_channels, kernel_size, bias=True):
return nn.Conv2d(
in_channels, out_channels, kernel_size, padding=(kernel_size // 2), bias=bias
)
class MeanShift(nn.Conv2d):
def __init__(
self,
rgb_range,
rgb_mean=(0.4488, 0.4371, 0.4040),
rgb_std=(1.0, 1.0, 1.0),
sign=-1,
):
super(MeanShift, self).__init__(3, 3, kernel_size=1)
std = torch.Tensor(rgb_std)
self.weight.data = torch.eye(3).view(3, 3, 1, 1) / std.view(3, 1, 1, 1)
self.bias.data = sign * rgb_range * torch.Tensor(rgb_mean) / std
for p in self.parameters():
p.requires_grad = False
class BasicBlock(nn.Sequential):
def __init__(
self,
conv,
in_channels,
out_channels,
kernel_size,
stride=1,
bias=False,
bn=True,
act=nn.ReLU(True),
):
m = [conv(in_channels, out_channels, kernel_size, bias=bias)]
if bn:
m.append(nn.BatchNorm2d(out_channels))
if act is not None:
m.append(act)
super(BasicBlock, self).__init__(*m)
class ESA(nn.Module):
def __init__(self, esa_channels, n_feats):
super(ESA, self).__init__()
f = esa_channels
self.conv1 = nn.Conv2d(n_feats, f, kernel_size=1)
self.conv_f = nn.Conv2d(f, f, kernel_size=1)
# self.conv_max = conv(f, f, kernel_size=3, padding=1)
self.conv2 = nn.Conv2d(f, f, kernel_size=3, stride=2, padding=0)
self.conv3 = nn.Conv2d(f, f, kernel_size=3, padding=1)
# self.conv3_ = conv(f, f, kernel_size=3, padding=1)
self.conv4 = nn.Conv2d(f, n_feats, kernel_size=1)
self.sigmoid = nn.Sigmoid()
# self.relu = nn.ReLU(inplace=True)
def forward(self, x):
c1_ = self.conv1(x)
c1 = self.conv2(c1_)
v_max = F.max_pool2d(c1, kernel_size=7, stride=3)
c3 = self.conv3(v_max)
# v_range = self.relu(self.conv_max(v_max))
# c3 = self.relu(self.conv3(v_range))
# c3 = self.conv3_(c3)
c3 = F.interpolate(
c3, (x.size(2), x.size(3)), mode="bilinear", align_corners=False
)
cf = self.conv_f(c1_)
c4 = self.conv4(c3 + cf)
m = self.sigmoid(c4)
return x * m
# class ESA(nn.Module):
# def __init__(self, esa_channels, n_feats, conv=nn.Conv2d):
# super(ESA, self).__init__()
# f = n_feats // 4
# self.conv1 = conv(n_feats, f, kernel_size=1)
# self.conv_f = conv(f, f, kernel_size=1)
# self.conv_max = conv(f, f, kernel_size=3, padding=1)
# self.conv2 = conv(f, f, kernel_size=3, stride=2, padding=0)
# self.conv3 = conv(f, f, kernel_size=3, padding=1)
# self.conv3_ = conv(f, f, kernel_size=3, padding=1)
# self.conv4 = conv(f, n_feats, kernel_size=1)
# self.sigmoid = nn.Sigmoid()
# self.relu = nn.ReLU(inplace=True)
#
# def forward(self, x):
# c1_ = (self.conv1(x))
# c1 = self.conv2(c1_)
# v_max = F.max_pool2d(c1, kernel_size=7, stride=3)
# v_range = self.relu(self.conv_max(v_max))
# c3 = self.relu(self.conv3(v_range))
# c3 = self.conv3_(c3)
# c3 = F.interpolate(c3, (x.size(2), x.size(3)), mode='bilinear', align_corners=False)
# cf = self.conv_f(c1_)
# c4 = self.conv4(c3 + cf)
# m = self.sigmoid(c4)
#
# return x * m
class ResBlock(nn.Module):
def __init__(
self,
conv,
n_feats,
kernel_size,
bias=True,
bn=False,
act=nn.ReLU(True),
res_scale=1,
esa_block=True,
depth_wise_kernel=7,
):
super(ResBlock, self).__init__()
m = []
for i in range(2):
m.append(conv(n_feats, n_feats, kernel_size, bias=bias))
if bn:
m.append(nn.BatchNorm2d(n_feats))
if i == 0:
m.append(act)
self.body = nn.Sequential(*m)
self.esa_block = esa_block
if self.esa_block:
esa_channels = 16
self.c5 = nn.Conv2d(
n_feats,
n_feats,
depth_wise_kernel,
padding=depth_wise_kernel // 2,
groups=n_feats,
bias=True,
)
self.esa = ESA(esa_channels, n_feats)
self.res_scale = res_scale
def forward(self, x):
res = self.body(x).mul(self.res_scale)
res += x
if self.esa_block:
res = self.esa(self.c5(res))
return res
class Upsampler(nn.Sequential):
def __init__(self, conv, scale, n_feats, bn=False, act=False, bias=True):
m = []
if (scale & (scale - 1)) == 0: # Is scale = 2^n?
for _ in range(int(math.log(scale, 2))):
m.append(conv(n_feats, 4 * n_feats, 3, bias))
m.append(nn.PixelShuffle(2))
if bn:
m.append(nn.BatchNorm2d(n_feats))
if act == "relu":
m.append(nn.ReLU(True))
elif act == "prelu":
m.append(nn.PReLU(n_feats))
elif scale == 3:
m.append(conv(n_feats, 9 * n_feats, 3, bias))
m.append(nn.PixelShuffle(3))
if bn:
m.append(nn.BatchNorm2d(n_feats))
if act == "relu":
m.append(nn.ReLU(True))
elif act == "prelu":
m.append(nn.PReLU(n_feats))
else:
raise NotImplementedError
super(Upsampler, self).__init__(*m)
class LiteUpsampler(nn.Sequential):
def __init__(self, conv, scale, n_feats, n_out=3, bn=False, act=False, bias=True):
m = []
m.append(conv(n_feats, n_out * (scale**2), 3, bias))
m.append(nn.PixelShuffle(scale))
# if (scale & (scale - 1)) == 0: # Is scale = 2^n?
# for _ in range(int(math.log(scale, 2))):
# m.append(conv(n_feats, 4 * n_out, 3, bias))
# m.append(nn.PixelShuffle(2))
# if bn:
# m.append(nn.BatchNorm2d(n_out))
# if act == 'relu':
# m.append(nn.ReLU(True))
# elif act == 'prelu':
# m.append(nn.PReLU(n_out))
# elif scale == 3:
# m.append(conv(n_feats, 9 * n_out, 3, bias))
# m.append(nn.PixelShuffle(3))
# if bn:
# m.append(nn.BatchNorm2d(n_out))
# if act == 'relu':
# m.append(nn.ReLU(True))
# elif act == 'prelu':
# m.append(nn.PReLU(n_out))
# else:
# raise NotImplementedError
super(LiteUpsampler, self).__init__(*m)
architecture/grl_common/mixed_attn_block.py
+1126
View File
File diff suppressed because it is too large Load Diff
architecture/grl_common/mixed_attn_block_efficient.py
+568
View File
@@ -0,0 +1,568 @@
import math
from abc import ABC
from math import prod
import torch
import torch.nn as nn
import torch.nn.functional as F
from timm.models.layers import DropPath
from .mixed_attn_block import (
AnchorProjection,
CAB,
CPB_MLP,
QKVProjection,
)
from .ops import (
window_partition,
window_reverse,
)
from .swin_v1_block import Mlp
class AffineTransform(nn.Module):
r"""Affine transformation of the attention map.
The window could be a square window or a stripe window. Supports attention between different window sizes
"""
def __init__(self, num_heads):
super(AffineTransform, self).__init__()
logit_scale = torch.log(10 * torch.ones((num_heads, 1, 1)))
self.logit_scale = nn.Parameter(logit_scale, requires_grad=True)
# mlp to generate continuous relative position bias
self.cpb_mlp = CPB_MLP(2, num_heads)
def forward(self, attn, relative_coords_table, relative_position_index, mask):
B_, H, N1, N2 = attn.shape
# logit scale
attn = attn * torch.clamp(self.logit_scale, max=math.log(1.0 / 0.01)).exp()
bias_table = self.cpb_mlp(relative_coords_table) # 2*Wh-1, 2*Ww-1, num_heads
bias_table = bias_table.view(-1, H)
bias = bias_table[relative_position_index.view(-1)]
bias = bias.view(N1, N2, -1).permute(2, 0, 1).contiguous()
# nH, Wh*Ww, Wh*Ww
bias = 16 * torch.sigmoid(bias)
attn = attn + bias.unsqueeze(0)
# W-MSA/SW-MSA
# shift attention mask
if mask is not None:
nW = mask.shape[0]
mask = mask.unsqueeze(1).unsqueeze(0)
attn = attn.view(B_ // nW, nW, H, N1, N2) + mask
attn = attn.view(-1, H, N1, N2)
return attn
def _get_stripe_info(stripe_size_in, stripe_groups_in, stripe_shift, input_resolution):
stripe_size, shift_size = [], []
for s, g, d in zip(stripe_size_in, stripe_groups_in, input_resolution):
if g is None:
stripe_size.append(s)
shift_size.append(s // 2 if stripe_shift else 0)
else:
stripe_size.append(d // g)
shift_size.append(0 if g == 1 else d // (g * 2))
return stripe_size, shift_size
class Attention(ABC, nn.Module):
def __init__(self):
super(Attention, self).__init__()
def attn(self, q, k, v, attn_transform, table, index, mask, reshape=True):
# q, k, v: # nW*B, H, wh*ww, dim
# cosine attention map
B_, _, H, head_dim = q.shape
if self.euclidean_dist:
# print("use euclidean distance")
attn = torch.norm(q.unsqueeze(-2) - k.unsqueeze(-3), dim=-1)
else:
attn = F.normalize(q, dim=-1) @ F.normalize(k, dim=-1).transpose(-2, -1)
attn = attn_transform(attn, table, index, mask)
# attention
attn = self.softmax(attn)
attn = self.attn_drop(attn)
x = attn @ v # B_, H, N1, head_dim
if reshape:
x = x.transpose(1, 2).reshape(B_, -1, H * head_dim)
# B_, N, C
return x
class WindowAttention(Attention):
r"""Window attention. QKV is the input to the forward method.
Args:
num_heads (int): Number of attention heads.
attn_drop (float, optional): Dropout ratio of attention weight. Default: 0.0
pretrained_window_size (tuple[int]): The height and width of the window in pre-training.
"""
def __init__(
self,
input_resolution,
window_size,
num_heads,
window_shift=False,
attn_drop=0.0,
pretrained_window_size=[0, 0],
args=None,
):
super(WindowAttention, self).__init__()
self.input_resolution = input_resolution
self.window_size = window_size
self.pretrained_window_size = pretrained_window_size
self.num_heads = num_heads
self.shift_size = window_size[0] // 2 if window_shift else 0
self.euclidean_dist = args.euclidean_dist
self.attn_transform = AffineTransform(num_heads)
self.attn_drop = nn.Dropout(attn_drop)
self.softmax = nn.Softmax(dim=-1)
def forward(self, qkv, x_size, table, index, mask):
"""
Args:
qkv: input QKV features with shape of (B, L, 3C)
x_size: use x_size to determine whether the relative positional bias table and index
need to be regenerated.
"""
H, W = x_size
B, L, C = qkv.shape
qkv = qkv.view(B, H, W, C)
# cyclic shift
if self.shift_size > 0:
qkv = torch.roll(
qkv, shifts=(-self.shift_size, -self.shift_size), dims=(1, 2)
)
# partition windows
qkv = window_partition(qkv, self.window_size) # nW*B, wh, ww, C
qkv = qkv.view(-1, prod(self.window_size), C) # nW*B, wh*ww, C
B_, N, _ = qkv.shape
qkv = qkv.reshape(B_, N, 3, self.num_heads, -1).permute(2, 0, 3, 1, 4)
q, k, v = qkv[0], qkv[1], qkv[2] # nW*B, H, wh*ww, dim
# attention
x = self.attn(q, k, v, self.attn_transform, table, index, mask)
# merge windows
x = x.view(-1, *self.window_size, C // 3)
x = window_reverse(x, self.window_size, x_size) # B, H, W, C/3
# reverse cyclic shift
if self.shift_size > 0:
x = torch.roll(x, shifts=(self.shift_size, self.shift_size), dims=(1, 2))
x = x.view(B, L, C // 3)
return x
def extra_repr(self) -> str:
return (
f"window_size={self.window_size}, shift_size={self.shift_size}, "
f"pretrained_window_size={self.pretrained_window_size}, num_heads={self.num_heads}"
)
def flops(self, N):
pass
class AnchorStripeAttention(Attention):
r"""Stripe attention
Args:
stripe_size (tuple[int]): The height and width of the stripe.
num_heads (int): Number of attention heads.
attn_drop (float, optional): Dropout ratio of attention weight. Default: 0.0
pretrained_stripe_size (tuple[int]): The height and width of the stripe in pre-training.
"""
def __init__(
self,
input_resolution,
stripe_size,
stripe_groups,
stripe_shift,
num_heads,
attn_drop=0.0,
pretrained_stripe_size=[0, 0],
anchor_window_down_factor=1,
args=None,
):
super(AnchorStripeAttention, self).__init__()
self.input_resolution = input_resolution
self.stripe_size = stripe_size # Wh, Ww
self.stripe_groups = stripe_groups
self.stripe_shift = stripe_shift
self.num_heads = num_heads
self.pretrained_stripe_size = pretrained_stripe_size
self.anchor_window_down_factor = anchor_window_down_factor
self.euclidean_dist = args.euclidean_dist
self.attn_transform1 = AffineTransform(num_heads)
self.attn_transform2 = AffineTransform(num_heads)
self.attn_drop = nn.Dropout(attn_drop)
self.softmax = nn.Softmax(dim=-1)
def forward(
self, qkv, anchor, x_size, table, index_a2w, index_w2a, mask_a2w, mask_w2a
):
"""
Args:
qkv: input features with shape of (B, L, C)
anchor:
x_size: use stripe_size to determine whether the relative positional bias table and index
need to be regenerated.
"""
H, W = x_size
B, L, C = qkv.shape
qkv = qkv.view(B, H, W, C)
stripe_size, shift_size = _get_stripe_info(
self.stripe_size, self.stripe_groups, self.stripe_shift, x_size
)
anchor_stripe_size = [s // self.anchor_window_down_factor for s in stripe_size]
anchor_shift_size = [s // self.anchor_window_down_factor for s in shift_size]
# cyclic shift
if self.stripe_shift:
qkv = torch.roll(qkv, shifts=(-shift_size[0], -shift_size[1]), dims=(1, 2))
anchor = torch.roll(
anchor,
shifts=(-anchor_shift_size[0], -anchor_shift_size[1]),
dims=(1, 2),
)
# partition windows
qkv = window_partition(qkv, stripe_size) # nW*B, wh, ww, C
qkv = qkv.view(-1, prod(stripe_size), C) # nW*B, wh*ww, C
anchor = window_partition(anchor, anchor_stripe_size)
anchor = anchor.view(-1, prod(anchor_stripe_size), C // 3)
B_, N1, _ = qkv.shape
N2 = anchor.shape[1]
qkv = qkv.reshape(B_, N1, 3, self.num_heads, -1).permute(2, 0, 3, 1, 4)
q, k, v = qkv[0], qkv[1], qkv[2]
anchor = anchor.reshape(B_, N2, self.num_heads, -1).permute(0, 2, 1, 3)
# attention
x = self.attn(
anchor, k, v, self.attn_transform1, table, index_a2w, mask_a2w, False
)
x = self.attn(q, anchor, x, self.attn_transform2, table, index_w2a, mask_w2a)
# merge windows
x = x.view(B_, *stripe_size, C // 3)
x = window_reverse(x, stripe_size, x_size) # B H' W' C
# reverse the shift
if self.stripe_shift:
x = torch.roll(x, shifts=shift_size, dims=(1, 2))
x = x.view(B, H * W, C // 3)
return x
def extra_repr(self) -> str:
return (
f"stripe_size={self.stripe_size}, stripe_groups={self.stripe_groups}, stripe_shift={self.stripe_shift}, "
f"pretrained_stripe_size={self.pretrained_stripe_size}, num_heads={self.num_heads}, anchor_window_down_factor={self.anchor_window_down_factor}"
)
def flops(self, N):
pass
class MixedAttention(nn.Module):
r"""Mixed window attention and stripe attention
Args:
dim (int): Number of input channels.
stripe_size (tuple[int]): The height and width of the stripe.
num_heads (int): Number of attention heads.
qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True
attn_drop (float, optional): Dropout ratio of attention weight. Default: 0.0
proj_drop (float, optional): Dropout ratio of output. Default: 0.0
pretrained_stripe_size (tuple[int]): The height and width of the stripe in pre-training.
"""
def __init__(
self,
dim,
input_resolution,
num_heads_w,
num_heads_s,
window_size,
window_shift,
stripe_size,
stripe_groups,
stripe_shift,
qkv_bias=True,
qkv_proj_type="linear",
anchor_proj_type="separable_conv",
anchor_one_stage=True,
anchor_window_down_factor=1,
attn_drop=0.0,
proj_drop=0.0,
pretrained_window_size=[0, 0],
pretrained_stripe_size=[0, 0],
args=None,
):
super(MixedAttention, self).__init__()
self.dim = dim
self.input_resolution = input_resolution
self.args = args
# print(args)
self.qkv = QKVProjection(dim, qkv_bias, qkv_proj_type, args)
# anchor is only used for stripe attention
self.anchor = AnchorProjection(
dim, anchor_proj_type, anchor_one_stage, anchor_window_down_factor, args
)
self.window_attn = WindowAttention(
input_resolution,
window_size,
num_heads_w,
window_shift,
attn_drop,
pretrained_window_size,
args,
)
self.stripe_attn = AnchorStripeAttention(
input_resolution,
stripe_size,
stripe_groups,
stripe_shift,
num_heads_s,
attn_drop,
pretrained_stripe_size,
anchor_window_down_factor,
args,
)
self.proj = nn.Linear(dim, dim)
self.proj_drop = nn.Dropout(proj_drop)
def forward(self, x, x_size, table_index_mask):
"""
Args:
x: input features with shape of (B, L, C)
stripe_size: use stripe_size to determine whether the relative positional bias table and index
need to be regenerated.
"""
B, L, C = x.shape
# qkv projection
qkv = self.qkv(x, x_size)
qkv_window, qkv_stripe = torch.split(qkv, C * 3 // 2, dim=-1)
# anchor projection
anchor = self.anchor(x, x_size)
# attention
x_window = self.window_attn(
qkv_window, x_size, *self._get_table_index_mask(table_index_mask, True)
)
x_stripe = self.stripe_attn(
qkv_stripe,
anchor,
x_size,
*self._get_table_index_mask(table_index_mask, False),
)
x = torch.cat([x_window, x_stripe], dim=-1)
# output projection
x = self.proj(x)
x = self.proj_drop(x)
return x
def _get_table_index_mask(self, table_index_mask, window_attn=True):
if window_attn:
return (
table_index_mask["table_w"],
table_index_mask["index_w"],
table_index_mask["mask_w"],
)
else:
return (
table_index_mask["table_s"],
table_index_mask["index_a2w"],
table_index_mask["index_w2a"],
table_index_mask["mask_a2w"],
table_index_mask["mask_w2a"],
)
def extra_repr(self) -> str:
return f"dim={self.dim}, input_resolution={self.input_resolution}"
def flops(self, N):
pass
class EfficientMixAttnTransformerBlock(nn.Module):
r"""Mix attention transformer block with shared QKV projection and output projection for mixed attention modules.
Args:
dim (int): Number of input channels.
input_resolution (tuple[int]): Input resulotion.
num_heads (int): Number of attention heads.
window_size (int): Window size.
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
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
pretrained_stripe_size (int): Window size in pre-training.
attn_type (str, optional): Attention type. Default: cwhv.
c: residual blocks
w: window attention
h: horizontal stripe attention
v: vertical stripe attention
"""
def __init__(
self,
dim,
input_resolution,
num_heads_w,
num_heads_s,
window_size=7,
window_shift=False,
stripe_size=[8, 8],
stripe_groups=[None, None],
stripe_shift=False,
stripe_type="H",
mlp_ratio=4.0,
qkv_bias=True,
qkv_proj_type="linear",
anchor_proj_type="separable_conv",
anchor_one_stage=True,
anchor_window_down_factor=1,
drop=0.0,
attn_drop=0.0,
drop_path=0.0,
act_layer=nn.GELU,
norm_layer=nn.LayerNorm,
pretrained_window_size=[0, 0],
pretrained_stripe_size=[0, 0],
res_scale=1.0,
args=None,
):
super().__init__()
self.dim = dim
self.input_resolution = input_resolution
self.num_heads_w = num_heads_w
self.num_heads_s = num_heads_s
self.window_size = window_size
self.window_shift = window_shift
self.stripe_shift = stripe_shift
self.stripe_type = stripe_type
self.args = args
if self.stripe_type == "W":
self.stripe_size = stripe_size[::-1]
self.stripe_groups = stripe_groups[::-1]
else:
self.stripe_size = stripe_size
self.stripe_groups = stripe_groups
self.mlp_ratio = mlp_ratio
self.res_scale = res_scale
self.attn = MixedAttention(
dim,
input_resolution,
num_heads_w,
num_heads_s,
window_size,
window_shift,
self.stripe_size,
self.stripe_groups,
stripe_shift,
qkv_bias,
qkv_proj_type,
anchor_proj_type,
anchor_one_stage,
anchor_window_down_factor,
attn_drop,
drop,
pretrained_window_size,
pretrained_stripe_size,
args,
)
self.norm1 = norm_layer(dim)
if self.args.local_connection:
self.conv = CAB(dim)
self.drop_path = DropPath(drop_path) if drop_path > 0.0 else nn.Identity()
self.mlp = Mlp(
in_features=dim,
hidden_features=int(dim * mlp_ratio),
act_layer=act_layer,
drop=drop,
)
self.norm2 = norm_layer(dim)
def _get_table_index_mask(self, all_table_index_mask):
table_index_mask = {
"table_w": all_table_index_mask["table_w"],
"index_w": all_table_index_mask["index_w"],
}
if self.stripe_type == "W":
table_index_mask["table_s"] = all_table_index_mask["table_sv"]
table_index_mask["index_a2w"] = all_table_index_mask["index_sv_a2w"]
table_index_mask["index_w2a"] = all_table_index_mask["index_sv_w2a"]
else:
table_index_mask["table_s"] = all_table_index_mask["table_sh"]
table_index_mask["index_a2w"] = all_table_index_mask["index_sh_a2w"]
table_index_mask["index_w2a"] = all_table_index_mask["index_sh_w2a"]
if self.window_shift:
table_index_mask["mask_w"] = all_table_index_mask["mask_w"]
else:
table_index_mask["mask_w"] = None
if self.stripe_shift:
if self.stripe_type == "W":
table_index_mask["mask_a2w"] = all_table_index_mask["mask_sv_a2w"]
table_index_mask["mask_w2a"] = all_table_index_mask["mask_sv_w2a"]
else:
table_index_mask["mask_a2w"] = all_table_index_mask["mask_sh_a2w"]
table_index_mask["mask_w2a"] = all_table_index_mask["mask_sh_w2a"]
else:
table_index_mask["mask_a2w"] = None
table_index_mask["mask_w2a"] = None
return table_index_mask
def forward(self, x, x_size, all_table_index_mask):
# Mixed attention
table_index_mask = self._get_table_index_mask(all_table_index_mask)
if self.args.local_connection:
x = (
x
+ self.res_scale
* self.drop_path(self.norm1(self.attn(x, x_size, table_index_mask)))
+ self.conv(x, x_size)
)
else:
x = x + self.res_scale * self.drop_path(
self.norm1(self.attn(x, x_size, table_index_mask))
)
# FFN
x = x + self.res_scale * self.drop_path(self.norm2(self.mlp(x)))
return x
def extra_repr(self) -> str:
return (
f"dim={self.dim}, input_resolution={self.input_resolution}, num_heads=({self.num_heads_w}, {self.num_heads_s}), "
f"window_size={self.window_size}, window_shift={self.window_shift}, "
f"stripe_size={self.stripe_size}, stripe_groups={self.stripe_groups}, stripe_shift={self.stripe_shift}, self.stripe_type={self.stripe_type}, "
f"mlp_ratio={self.mlp_ratio}, res_scale={self.res_scale}"
)
def flops(self):
pass
architecture/grl_common/ops.py
+551
View File
@@ -0,0 +1,551 @@
from math import prod
from typing import Tuple
import numpy as np
import torch
from timm.models.layers import to_2tuple
def bchw_to_bhwc(x: torch.Tensor) -> torch.Tensor:
"""Permutes a tensor from the shape (B, C, H, W) to (B, H, W, C)."""
return x.permute(0, 2, 3, 1)
def bhwc_to_bchw(x: torch.Tensor) -> torch.Tensor:
"""Permutes a tensor from the shape (B, H, W, C) to (B, C, H, W)."""
return x.permute(0, 3, 1, 2)
def bchw_to_blc(x: torch.Tensor) -> torch.Tensor:
"""Rearrange a tensor from the shape (B, C, H, W) to (B, L, C)."""
return x.flatten(2).transpose(1, 2)
def blc_to_bchw(x: torch.Tensor, x_size: Tuple) -> torch.Tensor:
"""Rearrange a tensor from the shape (B, L, C) to (B, C, H, W)."""
B, L, C = x.shape
return x.transpose(1, 2).view(B, C, *x_size)
def blc_to_bhwc(x: torch.Tensor, x_size: Tuple) -> torch.Tensor:
"""Rearrange a tensor from the shape (B, L, C) to (B, H, W, C)."""
B, L, C = x.shape
return x.view(B, *x_size, C)
def window_partition(x, window_size: Tuple[int, int]):
"""
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[0], window_size[0], W // window_size[1], window_size[1], C
)
windows = (
x.permute(0, 1, 3, 2, 4, 5)
.contiguous()
.view(-1, window_size[0], window_size[1], C)
)
return windows
def window_reverse(windows, window_size: Tuple[int, int], img_size: Tuple[int, int]):
"""
Args:
windows: (num_windows * B, window_size[0], window_size[1], C)
window_size (Tuple[int, int]): Window size
img_size (Tuple[int, int]): Image size
Returns:
x: (B, H, W, C)
"""
H, W = img_size
B = int(windows.shape[0] / (H * W / window_size[0] / window_size[1]))
x = windows.view(
B, H // window_size[0], W // window_size[1], window_size[0], window_size[1], -1
)
x = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(B, H, W, -1)
return x
def _fill_window(input_resolution, window_size, shift_size=None):
if shift_size is None:
shift_size = [s // 2 for s in window_size]
img_mask = torch.zeros((1, *input_resolution, 1)) # 1 H W 1
h_slices = (
slice(0, -window_size[0]),
slice(-window_size[0], -shift_size[0]),
slice(-shift_size[0], None),
)
w_slices = (
slice(0, -window_size[1]),
slice(-window_size[1], -shift_size[1]),
slice(-shift_size[1], 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, window_size)
# nW, window_size, window_size, 1
mask_windows = mask_windows.view(-1, prod(window_size))
return mask_windows
#####################################
# Different versions of the functions
# 1) Swin Transformer, SwinIR, Square window attention in GRL;
# 2) Early development of the decomposition-based efficient attention mechanism (efficient_win_attn.py);
# 3) GRL. Window-anchor attention mechanism.
# 1) & 3) are still useful
#####################################
def calculate_mask(input_resolution, window_size, shift_size):
"""
Use case: 1)
"""
# calculate attention mask for SW-MSA
if isinstance(shift_size, int):
shift_size = to_2tuple(shift_size)
mask_windows = _fill_window(input_resolution, window_size, shift_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)
) # nW, window_size**2, window_size**2
return attn_mask
def calculate_mask_all(
input_resolution,
window_size,
shift_size,
anchor_window_down_factor=1,
window_to_anchor=True,
):
"""
Use case: 3)
"""
# calculate attention mask for SW-MSA
anchor_resolution = [s // anchor_window_down_factor for s in input_resolution]
aws = [s // anchor_window_down_factor for s in window_size]
anchor_shift = [s // anchor_window_down_factor for s in shift_size]
# mask of window1: nW, Wh**Ww
mask_windows = _fill_window(input_resolution, window_size, shift_size)
# mask of window2: nW, AWh*AWw
mask_anchor = _fill_window(anchor_resolution, aws, anchor_shift)
if window_to_anchor:
attn_mask = mask_windows.unsqueeze(2) - mask_anchor.unsqueeze(1)
else:
attn_mask = mask_anchor.unsqueeze(2) - mask_windows.unsqueeze(1)
attn_mask = attn_mask.masked_fill(attn_mask != 0, float(-100.0)).masked_fill(
attn_mask == 0, float(0.0)
) # nW, Wh**Ww, AWh*AWw
return attn_mask
def calculate_win_mask(
input_resolution1, input_resolution2, window_size1, window_size2
):
"""
Use case: 2)
"""
# calculate attention mask for SW-MSA
# mask of window1: nW, Wh**Ww
mask_windows1 = _fill_window(input_resolution1, window_size1)
# mask of window2: nW, AWh*AWw
mask_windows2 = _fill_window(input_resolution2, window_size2)
attn_mask = mask_windows1.unsqueeze(2) - mask_windows2.unsqueeze(1)
attn_mask = attn_mask.masked_fill(attn_mask != 0, float(-100.0)).masked_fill(
attn_mask == 0, float(0.0)
) # nW, Wh**Ww, AWh*AWw
return attn_mask
def _get_meshgrid_coords(start_coords, end_coords):
coord_h = torch.arange(start_coords[0], end_coords[0])
coord_w = torch.arange(start_coords[1], end_coords[1])
coords = torch.stack(torch.meshgrid([coord_h, coord_w], indexing="ij")) # 2, Wh, Ww
coords = torch.flatten(coords, 1) # 2, Wh*Ww
return coords
def get_relative_coords_table(
window_size, pretrained_window_size=[0, 0], anchor_window_down_factor=1
):
"""
Use case: 1)
"""
# get relative_coords_table
ws = window_size
aws = [w // anchor_window_down_factor for w in window_size]
pws = pretrained_window_size
paws = [w // anchor_window_down_factor for w in pretrained_window_size]
ts = [(w1 + w2) // 2 for w1, w2 in zip(ws, aws)]
pts = [(w1 + w2) // 2 for w1, w2 in zip(pws, paws)]
# TODO: pretrained window size and pretrained anchor window size is only used here.
# TODO: Investigate whether it is really important to use this setting when finetuning large window size
# TODO: based on pretrained weights with small window size.
coord_h = torch.arange(-(ts[0] - 1), ts[0], dtype=torch.float32)
coord_w = torch.arange(-(ts[1] - 1), ts[1], dtype=torch.float32)
table = torch.stack(torch.meshgrid([coord_h, coord_w], indexing="ij")).permute(
1, 2, 0
)
table = table.contiguous().unsqueeze(0) # 1, Wh+AWh-1, Ww+AWw-1, 2
if pts[0] > 0:
table[:, :, :, 0] /= pts[0] - 1
table[:, :, :, 1] /= pts[1] - 1
else:
table[:, :, :, 0] /= ts[0] - 1
table[:, :, :, 1] /= ts[1] - 1
table *= 8 # normalize to -8, 8
table = torch.sign(table) * torch.log2(torch.abs(table) + 1.0) / np.log2(8)
return table
def get_relative_coords_table_all(
window_size, pretrained_window_size=[0, 0], anchor_window_down_factor=1
):
"""
Use case: 3)
Support all window shapes.
Args:
window_size:
pretrained_window_size:
anchor_window_down_factor:
Returns:
"""
# get relative_coords_table
ws = window_size
aws = [w // anchor_window_down_factor for w in window_size]
pws = pretrained_window_size
paws = [w // anchor_window_down_factor for w in pretrained_window_size]
# positive table size: (Ww - 1) - (Ww - AWw) // 2
ts_p = [w1 - 1 - (w1 - w2) // 2 for w1, w2 in zip(ws, aws)]
# negative table size: -(AWw - 1) - (Ww - AWw) // 2
ts_n = [-(w2 - 1) - (w1 - w2) // 2 for w1, w2 in zip(ws, aws)]
pts = [w1 - 1 - (w1 - w2) // 2 for w1, w2 in zip(pws, paws)]
# TODO: pretrained window size and pretrained anchor window size is only used here.
# TODO: Investigate whether it is really important to use this setting when finetuning large window size
# TODO: based on pretrained weights with small window size.
coord_h = torch.arange(ts_n[0], ts_p[0] + 1, dtype=torch.float32)
coord_w = torch.arange(ts_n[1], ts_p[1] + 1, dtype=torch.float32)
table = torch.stack(torch.meshgrid([coord_h, coord_w], indexing="ij")).permute(
1, 2, 0
)
table = table.contiguous().unsqueeze(0) # 1, Wh+AWh-1, Ww+AWw-1, 2
if pts[0] > 0:
table[:, :, :, 0] /= pts[0]
table[:, :, :, 1] /= pts[1]
else:
table[:, :, :, 0] /= ts_p[0]
table[:, :, :, 1] /= ts_p[1]
table *= 8 # normalize to -8, 8
table = torch.sign(table) * torch.log2(torch.abs(table) + 1.0) / np.log2(8)
# 1, Wh+AWh-1, Ww+AWw-1, 2
return table
def coords_diff(coords1, coords2, max_diff):
# The coordinates starts from (-start_coord[0], -start_coord[1])
coords = coords1[:, :, None] - coords2[:, None, :] # 2, Wh*Ww, AWh*AWw
coords = coords.permute(1, 2, 0).contiguous() # Wh*Ww, AWh*AWw, 2
coords[:, :, 0] += max_diff[0] - 1 # shift to start from 0
coords[:, :, 1] += max_diff[1] - 1
coords[:, :, 0] *= 2 * max_diff[1] - 1
idx = coords.sum(-1) # Wh*Ww, AWh*AWw
return idx
def get_relative_position_index(
window_size, anchor_window_down_factor=1, window_to_anchor=True
):
"""
Use case: 1)
"""
# get pair-wise relative position index for each token inside the window
ws = window_size
aws = [w // anchor_window_down_factor for w in window_size]
coords_anchor_end = [(w1 + w2) // 2 for w1, w2 in zip(ws, aws)]
coords_anchor_start = [(w1 - w2) // 2 for w1, w2 in zip(ws, aws)]
coords = _get_meshgrid_coords((0, 0), window_size) # 2, Wh*Ww
coords_anchor = _get_meshgrid_coords(coords_anchor_start, coords_anchor_end)
# 2, AWh*AWw
if window_to_anchor:
idx = coords_diff(coords, coords_anchor, max_diff=coords_anchor_end)
else:
idx = coords_diff(coords_anchor, coords, max_diff=coords_anchor_end)
return idx # Wh*Ww, AWh*AWw or AWh*AWw, Wh*Ww
def coords_diff_odd(coords1, coords2, start_coord, max_diff):
# The coordinates starts from (-start_coord[0], -start_coord[1])
coords = coords1[:, :, None] - coords2[:, None, :] # 2, Wh*Ww, AWh*AWw
coords = coords.permute(1, 2, 0).contiguous() # Wh*Ww, AWh*AWw, 2
coords[:, :, 0] += start_coord[0] # shift to start from 0
coords[:, :, 1] += start_coord[1]
coords[:, :, 0] *= max_diff
idx = coords.sum(-1) # Wh*Ww, AWh*AWw
return idx
def get_relative_position_index_all(
window_size, anchor_window_down_factor=1, window_to_anchor=True
):
"""
Use case: 3)
Support all window shapes:
square window - square window
rectangular window - rectangular window
window - anchor
anchor - window
[8, 8] - [8, 8]
[4, 86] - [2, 43]
"""
# get pair-wise relative position index for each token inside the window
ws = window_size
aws = [w // anchor_window_down_factor for w in window_size]
coords_anchor_start = [(w1 - w2) // 2 for w1, w2 in zip(ws, aws)]
coords_anchor_end = [s + w2 for s, w2 in zip(coords_anchor_start, aws)]
coords = _get_meshgrid_coords((0, 0), window_size) # 2, Wh*Ww
coords_anchor = _get_meshgrid_coords(coords_anchor_start, coords_anchor_end)
# 2, AWh*AWw
max_horizontal_diff = aws[1] + ws[1] - 1
if window_to_anchor:
offset = [w2 + s - 1 for s, w2 in zip(coords_anchor_start, aws)]
idx = coords_diff_odd(coords, coords_anchor, offset, max_horizontal_diff)
else:
offset = [w1 - s - 1 for s, w1 in zip(coords_anchor_start, ws)]
idx = coords_diff_odd(coords_anchor, coords, offset, max_horizontal_diff)
return idx # Wh*Ww, AWh*AWw or AWh*AWw, Wh*Ww
def get_relative_position_index_simple(
window_size, anchor_window_down_factor=1, window_to_anchor=True
):
"""
Use case: 3)
This is a simplified version of get_relative_position_index_all
The start coordinate of anchor window is also (0, 0)
get pair-wise relative position index for each token inside the window
"""
ws = window_size
aws = [w // anchor_window_down_factor for w in window_size]
coords = _get_meshgrid_coords((0, 0), window_size) # 2, Wh*Ww
coords_anchor = _get_meshgrid_coords((0, 0), aws)
# 2, AWh*AWw
max_horizontal_diff = aws[1] + ws[1] - 1
if window_to_anchor:
offset = [w2 - 1 for w2 in aws]
idx = coords_diff_odd(coords, coords_anchor, offset, max_horizontal_diff)
else:
offset = [w1 - 1 for w1 in ws]
idx = coords_diff_odd(coords_anchor, coords, offset, max_horizontal_diff)
return idx # Wh*Ww, AWh*AWw or AWh*AWw, Wh*Ww
# def get_relative_position_index(window_size):
# # This is a very early version
# # get pair-wise relative position index for each token inside the window
# coords = _get_meshgrid_coords(start_coords=(0, 0), end_coords=window_size)
# coords = coords[:, :, None] - coords[:, None, :] # 2, Wh*Ww, Wh*Ww
# coords = coords.permute(1, 2, 0).contiguous() # Wh*Ww, Wh*Ww, 2
# coords[:, :, 0] += window_size[0] - 1 # shift to start from 0
# coords[:, :, 1] += window_size[1] - 1
# coords[:, :, 0] *= 2 * window_size[1] - 1
# idx = coords.sum(-1) # Wh*Ww, Wh*Ww
# return idx
def get_relative_win_position_index(window_size, anchor_window_size):
"""
Use case: 2)
"""
# get pair-wise relative position index for each token inside the window
ws = window_size
aws = anchor_window_size
coords_anchor_end = [(w1 + w2) // 2 for w1, w2 in zip(ws, aws)]
coords_anchor_start = [(w1 - w2) // 2 for w1, w2 in zip(ws, aws)]
coords = _get_meshgrid_coords((0, 0), window_size) # 2, Wh*Ww
coords_anchor = _get_meshgrid_coords(coords_anchor_start, coords_anchor_end)
# 2, AWh*AWw
coords = coords[:, :, None] - coords_anchor[:, None, :] # 2, Wh*Ww, AWh*AWw
coords = coords.permute(1, 2, 0).contiguous() # Wh*Ww, AWh*AWw, 2
coords[:, :, 0] += coords_anchor_end[0] - 1 # shift to start from 0
coords[:, :, 1] += coords_anchor_end[1] - 1
coords[:, :, 0] *= 2 * coords_anchor_end[1] - 1
idx = coords.sum(-1) # Wh*Ww, AWh*AWw
return idx
# def get_relative_coords_table(window_size, pretrained_window_size):
# # This is a very early version
# # get relative_coords_table
# ws = window_size
# pws = pretrained_window_size
# coord_h = torch.arange(-(ws[0] - 1), ws[0], dtype=torch.float32)
# coord_w = torch.arange(-(ws[1] - 1), ws[1], dtype=torch.float32)
# table = torch.stack(torch.meshgrid([coord_h, coord_w], indexing='ij')).permute(1, 2, 0)
# table = table.contiguous().unsqueeze(0) # 1, 2*Wh-1, 2*Ww-1, 2
# if pws[0] > 0:
# table[:, :, :, 0] /= pws[0] - 1
# table[:, :, :, 1] /= pws[1] - 1
# else:
# table[:, :, :, 0] /= ws[0] - 1
# table[:, :, :, 1] /= ws[1] - 1
# table *= 8 # normalize to -8, 8
# table = torch.sign(table) * torch.log2(torch.abs(table) + 1.0) / np.log2(8)
# return table
def get_relative_win_coords_table(
window_size,
anchor_window_size,
pretrained_window_size=[0, 0],
pretrained_anchor_window_size=[0, 0],
):
"""
Use case: 2)
"""
# get relative_coords_table
ws = window_size
aws = anchor_window_size
pws = pretrained_window_size
paws = pretrained_anchor_window_size
# TODO: pretrained window size and pretrained anchor window size is only used here.
# TODO: Investigate whether it is really important to use this setting when finetuning large window size
# TODO: based on pretrained weights with small window size.
table_size = [(wsi + awsi) // 2 for wsi, awsi in zip(ws, aws)]
table_size_pretrained = [(pwsi + pawsi) // 2 for pwsi, pawsi in zip(pws, paws)]
coord_h = torch.arange(-(table_size[0] - 1), table_size[0], dtype=torch.float32)
coord_w = torch.arange(-(table_size[1] - 1), table_size[1], dtype=torch.float32)
table = torch.stack(torch.meshgrid([coord_h, coord_w], indexing="ij")).permute(
1, 2, 0
)
table = table.contiguous().unsqueeze(0) # 1, Wh+AWh-1, Ww+AWw-1, 2
if table_size_pretrained[0] > 0:
table[:, :, :, 0] /= table_size_pretrained[0] - 1
table[:, :, :, 1] /= table_size_pretrained[1] - 1
else:
table[:, :, :, 0] /= table_size[0] - 1
table[:, :, :, 1] /= table_size[1] - 1
table *= 8 # normalize to -8, 8
table = torch.sign(table) * torch.log2(torch.abs(table) + 1.0) / np.log2(8)
return table
if __name__ == "__main__":
table = get_relative_coords_table_all((4, 86), anchor_window_down_factor=2)
table = table.view(-1, 2)
index1 = get_relative_position_index_all((4, 86), 2, False)
index2 = get_relative_position_index_simple((4, 86), 2, False)
print(index2)
index3 = get_relative_position_index_all((4, 86), 2)
index4 = get_relative_position_index_simple((4, 86), 2)
print(index4)
print(
table.shape,
index2.shape,
index2.max(),
index2.min(),
index4.shape,
index4.max(),
index4.min(),
torch.allclose(index1, index2),
torch.allclose(index3, index4),
)
table = get_relative_coords_table_all((4, 86), anchor_window_down_factor=1)
table = table.view(-1, 2)
index1 = get_relative_position_index_all((4, 86), 1, False)
index2 = get_relative_position_index_simple((4, 86), 1, False)
# print(index1)
index3 = get_relative_position_index_all((4, 86), 1)
index4 = get_relative_position_index_simple((4, 86), 1)
# print(index2)
print(
table.shape,
index2.shape,
index2.max(),
index2.min(),
index4.shape,
index4.max(),
index4.min(),
torch.allclose(index1, index2),
torch.allclose(index3, index4),
)
table = get_relative_coords_table_all((8, 8), anchor_window_down_factor=2)
table = table.view(-1, 2)
index1 = get_relative_position_index_all((8, 8), 2, False)
index2 = get_relative_position_index_simple((8, 8), 2, False)
# print(index1)
index3 = get_relative_position_index_all((8, 8), 2)
index4 = get_relative_position_index_simple((8, 8), 2)
# print(index2)
print(
table.shape,
index2.shape,
index2.max(),
index2.min(),
index4.shape,
index4.max(),
index4.min(),
torch.allclose(index1, index2),
torch.allclose(index3, index4),
)
table = get_relative_coords_table_all((8, 8), anchor_window_down_factor=1)
table = table.view(-1, 2)
index1 = get_relative_position_index_all((8, 8), 1, False)
index2 = get_relative_position_index_simple((8, 8), 1, False)
# print(index1)
index3 = get_relative_position_index_all((8, 8), 1)
index4 = get_relative_position_index_simple((8, 8), 1)
# print(index2)
print(
table.shape,
index2.shape,
index2.max(),
index2.min(),
index4.shape,
index4.max(),
index4.min(),
torch.allclose(index1, index2),
torch.allclose(index3, index4),
)
architecture/grl_common/resblock.py
+61
View File
@@ -0,0 +1,61 @@
import torch.nn as nn
class ResBlock(nn.Module):
"""Residual block without BN.
It has a style of:
::
---Conv-ReLU-Conv-+-
|________________|
Args:
num_feats (int): Channel number of intermediate features.
Default: 64.
res_scale (float): Used to scale the residual before addition.
Default: 1.0.
"""
def __init__(self, num_feats=64, res_scale=1.0, bias=True, shortcut=True):
super().__init__()
self.res_scale = res_scale
self.shortcut = shortcut
self.conv1 = nn.Conv2d(num_feats, num_feats, 3, 1, 1, bias=bias)
self.conv2 = nn.Conv2d(num_feats, num_feats, 3, 1, 1, bias=bias)
self.relu = nn.ReLU(inplace=True)
def forward(self, x):
"""Forward function.
Args:
x (Tensor): Input tensor with shape (n, c, h, w).
Returns:
Tensor: Forward results.
"""
identity = x
out = self.conv2(self.relu(self.conv1(x)))
if self.shortcut:
return identity + out * self.res_scale
else:
return out * self.res_scale
class ResBlockWrapper(ResBlock):
"Used for transformers"
def __init__(self, num_feats, bias=True, shortcut=True):
super(ResBlockWrapper, self).__init__(
num_feats=num_feats, bias=bias, shortcut=shortcut
)
def forward(self, x, x_size):
H, W = x_size
B, L, C = x.shape
x = x.view(B, H, W, C).permute(0, 3, 1, 2)
x = super(ResBlockWrapper, self).forward(x)
x = x.flatten(2).permute(0, 2, 1)
return x
architecture/grl_common/swin_v1_block.py
+602
View File
@@ -0,0 +1,602 @@
from math import prod
import torch
import torch.nn as nn
from .ops import (
bchw_to_blc,
blc_to_bchw,
calculate_mask,
window_partition,
window_reverse,
)
from timm.models.layers import DropPath, to_2tuple, trunc_normal_
class Mlp(nn.Module):
"""MLP as used in Vision Transformer, MLP-Mixer and related networks"""
def __init__(
self,
in_features,
hidden_features=None,
out_features=None,
act_layer=nn.GELU,
drop=0.0,
):
super().__init__()
out_features = out_features or in_features
hidden_features = hidden_features or in_features
drop_probs = to_2tuple(drop)
self.fc1 = nn.Linear(in_features, hidden_features)
self.act = act_layer()
self.drop1 = nn.Dropout(drop_probs[0])
self.fc2 = nn.Linear(hidden_features, out_features)
self.drop2 = nn.Dropout(drop_probs[1])
def forward(self, x):
x = self.fc1(x)
x = self.act(x)
x = self.drop1(x)
x = self.fc2(x)
x = self.drop2(x)
return x
class WindowAttentionV1(nn.Module):
r"""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=0.0,
proj_drop=0.0,
use_pe=True,
):
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
self.use_pe = use_pe
if self.use_pe:
# define a parameter table of relative position bias
ws = self.window_size
table = torch.zeros((2 * ws[0] - 1) * (2 * ws[1] - 1), num_heads)
self.relative_position_bias_table = nn.Parameter(table)
# 2*Wh-1 * 2*Ww-1, nH
trunc_normal_(self.relative_position_bias_table, std=0.02)
self.get_relative_position_index(self.window_size)
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)
self.softmax = nn.Softmax(dim=-1)
def get_relative_position_index(self, window_size):
# get pair-wise relative position index for each token inside the window
coord_h = torch.arange(window_size[0])
coord_w = torch.arange(window_size[1])
coords = torch.stack(torch.meshgrid([coord_h, coord_w])) # 2, Wh, Ww
coords = torch.flatten(coords, 1) # 2, Wh*Ww
coords = coords[:, :, None] - coords[:, None, :] # 2, Wh*Ww, Wh*Ww
coords = coords.permute(1, 2, 0).contiguous() # Wh*Ww, Wh*Ww, 2
coords[:, :, 0] += window_size[0] - 1 # shift to start from 0
coords[:, :, 1] += window_size[1] - 1
coords[:, :, 0] *= 2 * window_size[1] - 1
relative_position_index = coords.sum(-1) # Wh*Ww, Wh*Ww
self.register_buffer("relative_position_index", relative_position_index)
def forward(self, x, mask=None):
"""
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 projection
qkv = self.qkv(x).reshape(B_, N, 3, self.num_heads, -1).permute(2, 0, 3, 1, 4)
q, k, v = qkv[0], qkv[1], qkv[2]
# attention map
q = q * self.scale
attn = q @ k.transpose(-2, -1)
# positional encoding
if self.use_pe:
win_dim = prod(self.window_size)
bias = self.relative_position_bias_table[
self.relative_position_index.view(-1)
]
bias = bias.view(win_dim, win_dim, -1).permute(2, 0, 1).contiguous()
# nH, Wh*Ww, Wh*Ww
attn = attn + bias.unsqueeze(0)
# shift attention mask
if mask is not None:
nW = mask.shape[0]
mask = mask.unsqueeze(1).unsqueeze(0)
attn = attn.view(B_ // nW, nW, self.num_heads, N, N) + mask
attn = attn.view(-1, self.num_heads, N, N)
# attention
attn = self.softmax(attn)
attn = self.attn_drop(attn)
x = (attn @ v).transpose(1, 2).reshape(B_, N, C)
# output projection
x = self.proj(x)
x = self.proj_drop(x)
return x
def extra_repr(self) -> str:
return f"dim={self.dim}, window_size={self.window_size}, num_heads={self.num_heads}"
def flops(self, N):
# calculate flops for 1 window with token length of N
flops = 0
# qkv = self.qkv(x)
flops += N * self.dim * 3 * self.dim
# attn = (q @ k.transpose(-2, -1))
flops += self.num_heads * N * (self.dim // self.num_heads) * N
# x = (attn @ v)
flops += self.num_heads * N * N * (self.dim // self.num_heads)
# x = self.proj(x)
flops += N * self.dim * self.dim
return flops
class WindowAttentionWrapperV1(WindowAttentionV1):
def __init__(self, shift_size, input_resolution, **kwargs):
super(WindowAttentionWrapperV1, self).__init__(**kwargs)
self.shift_size = shift_size
self.input_resolution = input_resolution
if self.shift_size > 0:
attn_mask = calculate_mask(input_resolution, self.window_size, shift_size)
else:
attn_mask = None
self.register_buffer("attn_mask", attn_mask)
def forward(self, x, x_size):
H, W = x_size
B, L, C = x.shape
x = x.view(B, H, W, C)
# cyclic shift
if self.shift_size > 0:
x = torch.roll(x, shifts=(-self.shift_size, -self.shift_size), dims=(1, 2))
# partition windows
x = window_partition(x, self.window_size) # nW*B, wh, ww, C
x = x.view(-1, prod(self.window_size), C) # nW*B, wh*ww, C
# W-MSA/SW-MSA (to be compatible for testing on images whose shapes are the multiple of window size
if self.input_resolution == x_size:
attn_mask = self.attn_mask
else:
attn_mask = calculate_mask(x_size, self.window_size, self.shift_size)
attn_mask = attn_mask.to(x.device)
# attention
x = super(WindowAttentionWrapperV1, self).forward(x, mask=attn_mask)
# nW*B, wh*ww, C
# merge windows
x = x.view(-1, *self.window_size, C)
x = window_reverse(x, self.window_size, x_size) # B, H, W, C
# reverse cyclic shift
if self.shift_size > 0:
x = torch.roll(x, shifts=(self.shift_size, self.shift_size), dims=(1, 2))
x = x.view(B, H * W, C)
return x
class SwinTransformerBlockV1(nn.Module):
r"""Swin Transformer Block.
Args:
dim (int): Number of input channels.
input_resolution (tuple[int]): Input resulotion.
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,
input_resolution,
num_heads,
window_size=7,
shift_size=0,
mlp_ratio=4.0,
qkv_bias=True,
qk_scale=None,
drop=0.0,
attn_drop=0.0,
drop_path=0.0,
act_layer=nn.GELU,
norm_layer=nn.LayerNorm,
use_pe=True,
res_scale=1.0,
):
super().__init__()
self.dim = dim
self.input_resolution = input_resolution
self.num_heads = num_heads
self.window_size = window_size
self.shift_size = shift_size
self.mlp_ratio = mlp_ratio
if min(self.input_resolution) <= self.window_size:
# if window size is larger than input resolution, we don't partition windows
self.shift_size = 0
self.window_size = min(self.input_resolution)
assert (
0 <= self.shift_size < self.window_size
), "shift_size must in 0-window_size"
self.res_scale = res_scale
self.norm1 = norm_layer(dim)
self.attn = WindowAttentionWrapperV1(
shift_size=self.shift_size,
input_resolution=self.input_resolution,
dim=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,
use_pe=use_pe,
)
self.drop_path = DropPath(drop_path) if drop_path > 0.0 else nn.Identity()
self.norm2 = norm_layer(dim)
self.mlp = Mlp(
in_features=dim,
hidden_features=int(dim * mlp_ratio),
act_layer=act_layer,
drop=drop,
)
def forward(self, x, x_size):
# Window attention
x = x + self.res_scale * self.drop_path(self.attn(self.norm1(x), x_size))
# FFN
x = x + self.res_scale * self.drop_path(self.mlp(self.norm2(x)))
return x
def extra_repr(self) -> str:
return (
f"dim={self.dim}, input_resolution={self.input_resolution}, num_heads={self.num_heads}, "
f"window_size={self.window_size}, shift_size={self.shift_size}, mlp_ratio={self.mlp_ratio}, res_scale={self.res_scale}"
)
def flops(self):
flops = 0
H, W = self.input_resolution
# norm1
flops += self.dim * H * W
# W-MSA/SW-MSA
nW = H * W / self.window_size / self.window_size
flops += nW * self.attn.flops(self.window_size * self.window_size)
# mlp
flops += 2 * H * W * self.dim * self.dim * self.mlp_ratio
# norm2
flops += self.dim * H * W
return flops
class PatchMerging(nn.Module):
r"""Patch Merging Layer.
Args:
input_resolution (tuple[int]): Resolution of input feature.
dim (int): Number of input channels.
norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm
"""
def __init__(self, input_resolution, dim, norm_layer=nn.LayerNorm):
super().__init__()
self.input_resolution = input_resolution
self.dim = dim
self.reduction = nn.Linear(4 * dim, 2 * dim, bias=False)
self.norm = norm_layer(4 * dim)
def forward(self, x):
"""
x: B, H*W, C
"""
H, W = self.input_resolution
B, L, C = x.shape
assert L == H * W, "input feature has wrong size"
assert H % 2 == 0 and W % 2 == 0, f"x size ({H}*{W}) are not even."
x = x.view(B, H, W, C)
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
def extra_repr(self) -> str:
return f"input_resolution={self.input_resolution}, dim={self.dim}"
def flops(self):
H, W = self.input_resolution
flops = H * W * self.dim
flops += (H // 2) * (W // 2) * 4 * self.dim * 2 * self.dim
return flops
class PatchEmbed(nn.Module):
r"""Image to Patch Embedding
Args:
img_size (int): Image size. Default: 224.
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, img_size=224, patch_size=4, in_chans=3, embed_dim=96, norm_layer=None
):
super().__init__()
img_size = to_2tuple(img_size)
patch_size = to_2tuple(patch_size)
patches_resolution = [
img_size[0] // patch_size[0],
img_size[1] // patch_size[1],
]
self.img_size = img_size
self.patch_size = patch_size
self.patches_resolution = patches_resolution
self.num_patches = patches_resolution[0] * patches_resolution[1]
self.in_chans = in_chans
self.embed_dim = embed_dim
if norm_layer is not None:
self.norm = norm_layer(embed_dim)
else:
self.norm = None
def forward(self, x):
x = x.flatten(2).transpose(1, 2) # B Ph*Pw C
if self.norm is not None:
x = self.norm(x)
return x
def flops(self):
flops = 0
H, W = self.img_size
if self.norm is not None:
flops += H * W * self.embed_dim
return flops
class PatchUnEmbed(nn.Module):
r"""Image to Patch Unembedding
Args:
img_size (int): Image size. Default: 224.
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, img_size=224, patch_size=4, in_chans=3, embed_dim=96, norm_layer=None
):
super().__init__()
img_size = to_2tuple(img_size)
patch_size = to_2tuple(patch_size)
patches_resolution = [
img_size[0] // patch_size[0],
img_size[1] // patch_size[1],
]
self.img_size = img_size
self.patch_size = patch_size
self.patches_resolution = patches_resolution
self.num_patches = patches_resolution[0] * patches_resolution[1]
self.in_chans = in_chans
self.embed_dim = embed_dim
def forward(self, x, x_size):
B, HW, C = x.shape
x = x.transpose(1, 2).view(B, self.embed_dim, x_size[0], x_size[1]) # B Ph*Pw C
return x
def flops(self):
flops = 0
return flops
class Linear(nn.Linear):
def __init__(self, in_features, out_features, bias=True):
super(Linear, self).__init__(in_features, out_features, bias)
def forward(self, x):
B, C, H, W = x.shape
x = bchw_to_blc(x)
x = super(Linear, self).forward(x)
x = blc_to_bchw(x, (H, W))
return x
def build_last_conv(conv_type, dim):
if conv_type == "1conv":
block = nn.Conv2d(dim, dim, 3, 1, 1)
elif conv_type == "3conv":
# to save parameters and memory
block = nn.Sequential(
nn.Conv2d(dim, dim // 4, 3, 1, 1),
nn.LeakyReLU(negative_slope=0.2, inplace=True),
nn.Conv2d(dim // 4, dim // 4, 1, 1, 0),
nn.LeakyReLU(negative_slope=0.2, inplace=True),
nn.Conv2d(dim // 4, dim, 3, 1, 1),
)
elif conv_type == "1conv1x1":
block = nn.Conv2d(dim, dim, 1, 1, 0)
elif conv_type == "linear":
block = Linear(dim, dim)
return block
# class BasicLayer(nn.Module):
# """A basic Swin Transformer layer for one stage.
# Args:
# dim (int): Number of input channels.
# input_resolution (tuple[int]): Input resolution.
# depth (int): Number of blocks.
# num_heads (int): Number of attention heads.
# window_size (int): Local window size.
# 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 | 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
# args: Additional arguments
# """
# def __init__(
# self,
# dim,
# input_resolution,
# depth,
# num_heads,
# window_size,
# mlp_ratio=4.0,
# qkv_bias=True,
# qk_scale=None,
# drop=0.0,
# attn_drop=0.0,
# drop_path=0.0,
# norm_layer=nn.LayerNorm,
# downsample=None,
# args=None,
# ):
# super().__init__()
# self.dim = dim
# self.input_resolution = input_resolution
# self.depth = depth
# # build blocks
# self.blocks = nn.ModuleList(
# [
# _parse_block(
# dim=dim,
# input_resolution=input_resolution,
# num_heads=num_heads,
# window_size=window_size,
# shift_size=0
# if args.no_shift
# else (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,
# stripe_type="H" if (i % 2 == 0) else "W",
# args=args,
# )
# for i in range(depth)
# ]
# )
# # self.blocks = nn.ModuleList(
# # [
# # STV1Block(
# # dim=dim,
# # input_resolution=input_resolution,
# # 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(
# input_resolution, dim=dim, norm_layer=norm_layer
# )
# else:
# self.downsample = None
# def forward(self, x, x_size):
# for blk in self.blocks:
# x = blk(x, x_size)
# if self.downsample is not None:
# x = self.downsample(x)
# return x
# def extra_repr(self) -> str:
# return f"dim={self.dim}, input_resolution={self.input_resolution}, depth={self.depth}"
# def flops(self):
# flops = 0
# for blk in self.blocks:
# flops += blk.flops()
# if self.downsample is not None:
# flops += self.downsample.flops()
# return flops
architecture/grl_common/swin_v2_block.py
+306
View File
@@ -0,0 +1,306 @@
import math
from math import prod
import torch
import torch.nn as nn
import torch.nn.functional as F
from .ops import (
calculate_mask,
get_relative_coords_table,
get_relative_position_index,
window_partition,
window_reverse,
)
from .swin_v1_block import Mlp
from timm.models.layers import DropPath, to_2tuple
class WindowAttentionV2(nn.Module):
r"""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
attn_drop (float, optional): Dropout ratio of attention weight. Default: 0.0
proj_drop (float, optional): Dropout ratio of output. Default: 0.0
pretrained_window_size (tuple[int]): The height and width of the window in pre-training.
"""
def __init__(
self,
dim,
window_size,
num_heads,
qkv_bias=True,
attn_drop=0.0,
proj_drop=0.0,
pretrained_window_size=[0, 0],
use_pe=True,
):
super().__init__()
self.dim = dim
self.window_size = window_size # Wh, Ww
self.pretrained_window_size = pretrained_window_size
self.num_heads = num_heads
self.use_pe = use_pe
self.logit_scale = nn.Parameter(
torch.log(10 * torch.ones((num_heads, 1, 1))), requires_grad=True
)
if self.use_pe:
# mlp to generate continuous relative position bias
self.cpb_mlp = nn.Sequential(
nn.Linear(2, 512, bias=True),
nn.ReLU(inplace=True),
nn.Linear(512, num_heads, bias=False),
)
table = get_relative_coords_table(window_size, pretrained_window_size)
index = get_relative_position_index(window_size)
self.register_buffer("relative_coords_table", table)
self.register_buffer("relative_position_index", index)
self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
# self.qkv = nn.Linear(dim, dim * 3, bias=False)
# if qkv_bias:
# self.q_bias = nn.Parameter(torch.zeros(dim))
# self.v_bias = nn.Parameter(torch.zeros(dim))
# else:
# self.q_bias = None
# self.v_bias = None
self.attn_drop = nn.Dropout(attn_drop)
self.proj = nn.Linear(dim, dim)
self.proj_drop = nn.Dropout(proj_drop)
self.softmax = nn.Softmax(dim=-1)
def forward(self, x, mask=None):
"""
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 projection
# qkv_bias = None
# if self.q_bias is not None:
# qkv_bias = torch.cat(
# (
# self.q_bias,
# torch.zeros_like(self.v_bias, requires_grad=False),
# self.v_bias,
# )
# )
# qkv = F.linear(input=x, weight=self.qkv.weight, bias=qkv_bias)
qkv = self.qkv(x)
qkv = qkv.reshape(B_, N, 3, self.num_heads, -1).permute(2, 0, 3, 1, 4)
q, k, v = qkv[0], qkv[1], qkv[2]
# cosine attention map
attn = F.normalize(q, dim=-1) @ F.normalize(k, dim=-1).transpose(-2, -1)
logit_scale = torch.clamp(self.logit_scale, max=math.log(1.0 / 0.01)).exp()
attn = attn * logit_scale
# positional encoding
if self.use_pe:
bias_table = self.cpb_mlp(self.relative_coords_table)
bias_table = bias_table.view(-1, self.num_heads)
win_dim = prod(self.window_size)
bias = bias_table[self.relative_position_index.view(-1)]
bias = bias.view(win_dim, win_dim, -1).permute(2, 0, 1).contiguous()
# nH, Wh*Ww, Wh*Ww
bias = 16 * torch.sigmoid(bias)
attn = attn + bias.unsqueeze(0)
# shift attention mask
if mask is not None:
nW = mask.shape[0]
mask = mask.unsqueeze(1).unsqueeze(0)
attn = attn.view(B_ // nW, nW, self.num_heads, N, N) + mask
attn = attn.view(-1, self.num_heads, N, N)
# attention
attn = self.softmax(attn)
attn = self.attn_drop(attn)
x = (attn @ v).transpose(1, 2).reshape(B_, N, C)
# output projection
x = self.proj(x)
x = self.proj_drop(x)
return x
def extra_repr(self) -> str:
return (
f"dim={self.dim}, window_size={self.window_size}, "
f"pretrained_window_size={self.pretrained_window_size}, num_heads={self.num_heads}"
)
def flops(self, N):
# calculate flops for 1 window with token length of N
flops = 0
# qkv = self.qkv(x)
flops += N * self.dim * 3 * self.dim
# attn = (q @ k.transpose(-2, -1))
flops += self.num_heads * N * (self.dim // self.num_heads) * N
# x = (attn @ v)
flops += self.num_heads * N * N * (self.dim // self.num_heads)
# x = self.proj(x)
flops += N * self.dim * self.dim
return flops
class WindowAttentionWrapperV2(WindowAttentionV2):
def __init__(self, shift_size, input_resolution, **kwargs):
super(WindowAttentionWrapperV2, self).__init__(**kwargs)
self.shift_size = shift_size
self.input_resolution = input_resolution
if self.shift_size > 0:
attn_mask = calculate_mask(input_resolution, self.window_size, shift_size)
else:
attn_mask = None
self.register_buffer("attn_mask", attn_mask)
def forward(self, x, x_size):
H, W = x_size
B, L, C = x.shape
x = x.view(B, H, W, C)
# cyclic shift
if self.shift_size > 0:
x = torch.roll(x, shifts=(-self.shift_size, -self.shift_size), dims=(1, 2))
# partition windows
x = window_partition(x, self.window_size) # nW*B, wh, ww, C
x = x.view(-1, prod(self.window_size), C) # nW*B, wh*ww, C
# W-MSA/SW-MSA
if self.input_resolution == x_size:
attn_mask = self.attn_mask
else:
attn_mask = calculate_mask(x_size, self.window_size, self.shift_size)
attn_mask = attn_mask.to(x.device)
# attention
x = super(WindowAttentionWrapperV2, self).forward(x, mask=attn_mask)
# nW*B, wh*ww, C
# merge windows
x = x.view(-1, *self.window_size, C)
x = window_reverse(x, self.window_size, x_size) # B, H, W, C
# reverse cyclic shift
if self.shift_size > 0:
x = torch.roll(x, shifts=(self.shift_size, self.shift_size), dims=(1, 2))
x = x.view(B, H * W, C)
return x
class SwinTransformerBlockV2(nn.Module):
r"""Swin Transformer Block.
Args:
dim (int): Number of input channels.
input_resolution (tuple[int]): Input resulotion.
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
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
pretrained_window_size (int): Window size in pre-training.
"""
def __init__(
self,
dim,
input_resolution,
num_heads,
window_size=7,
shift_size=0,
mlp_ratio=4.0,
qkv_bias=True,
drop=0.0,
attn_drop=0.0,
drop_path=0.0,
act_layer=nn.GELU,
norm_layer=nn.LayerNorm,
pretrained_window_size=0,
use_pe=True,
res_scale=1.0,
):
super().__init__()
self.dim = dim
self.input_resolution = input_resolution
self.num_heads = num_heads
self.window_size = window_size
self.shift_size = shift_size
self.mlp_ratio = mlp_ratio
if min(self.input_resolution) <= self.window_size:
# if window size is larger than input resolution, we don't partition windows
self.shift_size = 0
self.window_size = min(self.input_resolution)
assert (
0 <= self.shift_size < self.window_size
), "shift_size must in 0-window_size"
self.res_scale = res_scale
self.attn = WindowAttentionWrapperV2(
shift_size=self.shift_size,
input_resolution=self.input_resolution,
dim=dim,
window_size=to_2tuple(self.window_size),
num_heads=num_heads,
qkv_bias=qkv_bias,
attn_drop=attn_drop,
proj_drop=drop,
pretrained_window_size=to_2tuple(pretrained_window_size),
use_pe=use_pe,
)
self.norm1 = norm_layer(dim)
self.drop_path = DropPath(drop_path) if drop_path > 0.0 else nn.Identity()
self.mlp = Mlp(
in_features=dim,
hidden_features=int(dim * mlp_ratio),
act_layer=act_layer,
drop=drop,
)
self.norm2 = norm_layer(dim)
def forward(self, x, x_size):
# Window attention
x = x + self.res_scale * self.drop_path(self.norm1(self.attn(x, x_size)))
# FFN
x = x + self.res_scale * self.drop_path(self.norm2(self.mlp(x)))
return x
def extra_repr(self) -> str:
return (
f"dim={self.dim}, input_resolution={self.input_resolution}, num_heads={self.num_heads}, "
f"window_size={self.window_size}, shift_size={self.shift_size}, mlp_ratio={self.mlp_ratio}, res_scale={self.res_scale}"
)
def flops(self):
flops = 0
H, W = self.input_resolution
# norm1
flops += self.dim * H * W
# W-MSA/SW-MSA
nW = H * W / self.window_size / self.window_size
flops += nW * self.attn.flops(self.window_size * self.window_size)
# mlp
flops += 2 * H * W * self.dim * self.dim * self.mlp_ratio
# norm2
flops += self.dim * H * W
return flops
architecture/grl_common/upsample.py
+50
View File
@@ -0,0 +1,50 @@
import math
import torch.nn as nn
class Upsample(nn.Module):
"""Upsample module.
Args:
scale (int): Scale factor. Supported scales: 2^n and 3.
num_feat (int): Channel number of intermediate features.
"""
def __init__(self, scale, num_feat):
super(Upsample, self).__init__()
m = []
if (scale & (scale - 1)) == 0: # scale = 2^n
for _ in range(int(math.log(scale, 2))):
m.append(nn.Conv2d(num_feat, 4 * num_feat, 3, 1, 1))
m.append(nn.PixelShuffle(2))
elif scale == 3:
m.append(nn.Conv2d(num_feat, 9 * num_feat, 3, 1, 1))
m.append(nn.PixelShuffle(3))
else:
raise ValueError(
f"scale {scale} is not supported. " "Supported scales: 2^n and 3."
)
self.up = nn.Sequential(*m)
def forward(self, x):
return self.up(x)
class UpsampleOneStep(nn.Module):
"""UpsampleOneStep module (the difference with Upsample is that it always only has 1conv + 1pixelshuffle)
Used in lightweight SR to save parameters.
Args:
scale (int): Scale factor. Supported scales: 2^n and 3.
num_feat (int): Channel number of intermediate features.
"""
def __init__(self, scale, num_feat, num_out_ch):
super(UpsampleOneStep, self).__init__()
self.num_feat = num_feat
m = []
m.append(nn.Conv2d(num_feat, (scale**2) * num_out_ch, 3, 1, 1))
m.append(nn.PixelShuffle(scale))
self.up = nn.Sequential(*m)
def forward(self, x):
return self.up(x)
architecture/rrdb.py
+218
View File
@@ -0,0 +1,218 @@
# -*- coding: utf-8 -*-
# Paper Github Repository: https://github.com/xinntao/Real-ESRGAN
# Code snippet from: https://github.com/XPixelGroup/BasicSR/blob/master/basicsr/archs/rrdbnet_arch.py
# Paper: https://arxiv.org/pdf/2107.10833.pdf
import os, sys
import torch
from torch import nn as nn
from torch.nn import functional as F
from itertools import repeat
from torch.nn import init as init
from torch.nn.modules.batchnorm import _BatchNorm
def pixel_unshuffle(x, scale):
""" Pixel unshuffle.
Args:
x (Tensor): Input feature with shape (b, c, hh, hw).
scale (int): Downsample ratio.
Returns:
Tensor: the pixel unshuffled feature.
"""
b, c, hh, hw = x.size()
out_channel = c * (scale**2)
assert hh % scale == 0 and hw % scale == 0
h = hh // scale
w = hw // scale
x_view = x.view(b, c, h, scale, w, scale)
return x_view.permute(0, 1, 3, 5, 2, 4).reshape(b, out_channel, h, w)
def make_layer(basic_block, num_basic_block, **kwarg):
"""Make layers by stacking the same blocks.
Args:
basic_block (nn.module): nn.module class for basic block.
num_basic_block (int): number of blocks.
Returns:
nn.Sequential: Stacked blocks in nn.Sequential.
"""
layers = []
for _ in range(num_basic_block):
layers.append(basic_block(**kwarg))
return nn.Sequential(*layers)
def default_init_weights(module_list, scale=1, bias_fill=0, **kwargs):
"""Initialize network weights.
Args:
module_list (list[nn.Module] | nn.Module): Modules to be initialized.
scale (float): Scale initialized weights, especially for residual
blocks. Default: 1.
bias_fill (float): The value to fill bias. Default: 0
kwargs (dict): Other arguments for initialization function.
"""
if not isinstance(module_list, list):
module_list = [module_list]
for module in module_list:
for m in module.modules():
if isinstance(m, nn.Conv2d):
init.kaiming_normal_(m.weight, **kwargs)
m.weight.data *= scale
if m.bias is not None:
m.bias.data.fill_(bias_fill)
elif isinstance(m, nn.Linear):
init.kaiming_normal_(m.weight, **kwargs)
m.weight.data *= scale
if m.bias is not None:
m.bias.data.fill_(bias_fill)
elif isinstance(m, _BatchNorm):
init.constant_(m.weight, 1)
if m.bias is not None:
m.bias.data.fill_(bias_fill)
class ResidualDenseBlock(nn.Module):
"""Residual Dense Block.
Used in RRDB block in ESRGAN.
Args:
num_feat (int): Channel number of intermediate features.
num_grow_ch (int): Channels for each growth.
"""
def __init__(self, num_feat=64, num_grow_ch=32):
super(ResidualDenseBlock, self).__init__()
self.conv1 = nn.Conv2d(num_feat, num_grow_ch, 3, 1, 1)
self.conv2 = nn.Conv2d(num_feat + num_grow_ch, num_grow_ch, 3, 1, 1)
self.conv3 = nn.Conv2d(num_feat + 2 * num_grow_ch, num_grow_ch, 3, 1, 1)
self.conv4 = nn.Conv2d(num_feat + 3 * num_grow_ch, num_grow_ch, 3, 1, 1)
self.conv5 = nn.Conv2d(num_feat + 4 * num_grow_ch, num_feat, 3, 1, 1)
self.lrelu = nn.LeakyReLU(negative_slope=0.2, inplace=True)
# initialization
default_init_weights([self.conv1, self.conv2, self.conv3, self.conv4, self.conv5], 0.1)
def forward(self, x):
x1 = self.lrelu(self.conv1(x))
x2 = self.lrelu(self.conv2(torch.cat((x, x1), 1)))
x3 = self.lrelu(self.conv3(torch.cat((x, x1, x2), 1)))
x4 = self.lrelu(self.conv4(torch.cat((x, x1, x2, x3), 1)))
x5 = self.conv5(torch.cat((x, x1, x2, x3, x4), 1))
# Empirically, we use 0.2 to scale the residual for better performance
return x5 * 0.2 + x
class RRDB(nn.Module):
"""Residual in Residual Dense Block.
Used in RRDB-Net in ESRGAN.
Args:
num_feat (int): Channel number of intermediate features.
num_grow_ch (int): Channels for each growth.
"""
def __init__(self, num_feat, num_grow_ch=32):
super(RRDB, self).__init__()
self.rdb1 = ResidualDenseBlock(num_feat, num_grow_ch)
self.rdb2 = ResidualDenseBlock(num_feat, num_grow_ch)
self.rdb3 = ResidualDenseBlock(num_feat, num_grow_ch)
def forward(self, x):
out = self.rdb1(x)
out = self.rdb2(out)
out = self.rdb3(out)
# Empirically, we use 0.2 to scale the residual for better performance
return out * 0.2 + x
class RRDBNet(nn.Module):
"""Networks consisting of Residual in Residual Dense Block, which is used
in ESRGAN.
ESRGAN: Enhanced Super-Resolution Generative Adversarial Networks.
We extend ESRGAN for scale x2 and scale x1.
Note: This is one option for scale 1, scale 2 in RRDBNet.
We first employ the pixel-unshuffle (an inverse operation of pixelshuffle to reduce the spatial size
and enlarge the channel size before feeding inputs into the main ESRGAN architecture.
Args:
num_in_ch (int): Channel number of inputs.
num_out_ch (int): Channel number of outputs.
num_feat (int): Channel number of intermediate features.
Default: 64
num_block (int): Block number in the trunk network. Defaults: 6 for our Anime training cases
num_grow_ch (int): Channels for each growth. Default: 32.
"""
def __init__(self, num_in_ch, num_out_ch, scale, num_feat=64, num_block=6, num_grow_ch=32):
super(RRDBNet, self).__init__()
self.scale = scale
if scale == 2:
num_in_ch = num_in_ch * 4
elif scale == 1:
num_in_ch = num_in_ch * 16
self.conv_first = nn.Conv2d(num_in_ch, num_feat, 3, 1, 1)
self.body = make_layer(RRDB, num_block, num_feat=num_feat, num_grow_ch=num_grow_ch)
self.conv_body = nn.Conv2d(num_feat, num_feat, 3, 1, 1)
# upsample
self.conv_up1 = nn.Conv2d(num_feat, num_feat, 3, 1, 1)
self.conv_up2 = nn.Conv2d(num_feat, num_feat, 3, 1, 1)
self.conv_hr = nn.Conv2d(num_feat, num_feat, 3, 1, 1)
self.conv_last = nn.Conv2d(num_feat, num_out_ch, 3, 1, 1)
self.lrelu = nn.LeakyReLU(negative_slope=0.2, inplace=True)
def forward(self, x):
if self.scale == 2:
feat = pixel_unshuffle(x, scale=2)
elif self.scale == 1:
feat = pixel_unshuffle(x, scale=4)
else:
feat = x
feat = self.conv_first(feat)
body_feat = self.conv_body(self.body(feat))
feat = feat + body_feat
# upsample
feat = self.lrelu(self.conv_up1(F.interpolate(feat, scale_factor=2, mode='nearest')))
feat = self.lrelu(self.conv_up2(F.interpolate(feat, scale_factor=2, mode='nearest')))
out = self.conv_last(self.lrelu(self.conv_hr(feat)))
return out
# def main():
# root_path = os.path.abspath('.')
# sys.path.append(root_path)
# from opt import opt # Manage GPU to choose
# from pthflops import count_ops
# #from torchsummary import summary
# import time
# # We use RRDB 6Blocks by default.
# model = RRDBNet(3, 3)
# pytorch_total_params = sum(p.numel() for p in model.parameters())
# print(f"RRDB has param {pytorch_total_params//1000} K params")
# # Count the number of FLOPs to double check
# x = torch.randn((1, 3, 180, 180))
# start = time.time()
# x = model(x)
# print("output size is ", x.shape)
# total = time.time() - start
# print(total)
# if __name__ == "__main__":
# main()
architecture/swinir.py
+874
View File
@@ -0,0 +1,874 @@
# -----------------------------------------------------------------------------------
# SwinIR: Image Restoration Using Swin Transformer, https://arxiv.org/abs/2108.10257
# Originally Written by Ze Liu, Modified by Jingyun Liang.
# -----------------------------------------------------------------------------------
import math
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.utils.checkpoint as checkpoint
from timm.models.layers import DropPath, to_2tuple, trunc_normal_
class Mlp(nn.Module):
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):
r""" 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])) # 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):
"""
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
def extra_repr(self) -> str:
return f'dim={self.dim}, window_size={self.window_size}, num_heads={self.num_heads}'
def flops(self, N):
# calculate flops for 1 window with token length of N
flops = 0
# qkv = self.qkv(x)
flops += N * self.dim * 3 * self.dim
# attn = (q @ k.transpose(-2, -1))
flops += self.num_heads * N * (self.dim // self.num_heads) * N
# x = (attn @ v)
flops += self.num_heads * N * N * (self.dim // self.num_heads)
# x = self.proj(x)
flops += N * self.dim * self.dim
return flops
class SwinTransformerBlock(nn.Module):
r""" Swin Transformer Block.
Args:
dim (int): Number of input channels.
input_resolution (tuple[int]): Input resulotion.
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, input_resolution, 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.input_resolution = input_resolution
self.num_heads = num_heads
self.window_size = window_size
self.shift_size = shift_size
self.mlp_ratio = mlp_ratio
if min(self.input_resolution) <= self.window_size:
# if window size is larger than input resolution, we don't partition windows
self.shift_size = 0
self.window_size = min(self.input_resolution)
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)
if self.shift_size > 0:
attn_mask = self.calculate_mask(self.input_resolution)
else:
attn_mask = None
self.register_buffer("attn_mask", attn_mask)
def calculate_mask(self, x_size):
# calculate attention mask for SW-MSA
H, W = x_size
img_mask = torch.zeros((1, H, W, 1)) # 1 H W 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))
return attn_mask
def forward(self, x, x_size):
H, W = x_size
B, L, C = x.shape
# assert L == H * W, "input feature has wrong size"
shortcut = x
x = self.norm1(x)
x = x.view(B, H, W, C)
# cyclic shift
if self.shift_size > 0:
shifted_x = torch.roll(x, shifts=(-self.shift_size, -self.shift_size), dims=(1, 2))
else:
shifted_x = x
# 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 (to be compatible for testing on images whose shapes are the multiple of window size
if self.input_resolution == x_size:
attn_windows = self.attn(x_windows, mask=self.attn_mask) # nW*B, window_size*window_size, C
else:
attn_windows = self.attn(x_windows, mask=self.calculate_mask(x_size).to(x.device))
# merge windows
attn_windows = attn_windows.view(-1, self.window_size, self.window_size, C)
shifted_x = window_reverse(attn_windows, self.window_size, H, W) # 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
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
def extra_repr(self) -> str:
return f"dim={self.dim}, input_resolution={self.input_resolution}, num_heads={self.num_heads}, " \
f"window_size={self.window_size}, shift_size={self.shift_size}, mlp_ratio={self.mlp_ratio}"
def flops(self):
flops = 0
H, W = self.input_resolution
# norm1
flops += self.dim * H * W
# W-MSA/SW-MSA
nW = H * W / self.window_size / self.window_size
flops += nW * self.attn.flops(self.window_size * self.window_size)
# mlp
flops += 2 * H * W * self.dim * self.dim * self.mlp_ratio
# norm2
flops += self.dim * H * W
return flops
class PatchMerging(nn.Module):
r""" Patch Merging Layer.
Args:
input_resolution (tuple[int]): Resolution of input feature.
dim (int): Number of input channels.
norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm
"""
def __init__(self, input_resolution, dim, norm_layer=nn.LayerNorm):
super().__init__()
self.input_resolution = input_resolution
self.dim = dim
self.reduction = nn.Linear(4 * dim, 2 * dim, bias=False)
self.norm = norm_layer(4 * dim)
def forward(self, x):
"""
x: B, H*W, C
"""
H, W = self.input_resolution
B, L, C = x.shape
assert L == H * W, "input feature has wrong size"
assert H % 2 == 0 and W % 2 == 0, f"x size ({H}*{W}) are not even."
x = x.view(B, H, W, C)
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
def extra_repr(self) -> str:
return f"input_resolution={self.input_resolution}, dim={self.dim}"
def flops(self):
H, W = self.input_resolution
flops = H * W * self.dim
flops += (H // 2) * (W // 2) * 4 * self.dim * 2 * self.dim
return flops
class BasicLayer(nn.Module):
""" A basic Swin Transformer layer for one stage.
Args:
dim (int): Number of input channels.
input_resolution (tuple[int]): Input resolution.
depth (int): Number of blocks.
num_heads (int): Number of attention heads.
window_size (int): Local window size.
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 | 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, input_resolution, depth, num_heads, window_size,
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.dim = dim
self.input_resolution = input_resolution
self.depth = depth
self.use_checkpoint = use_checkpoint
# build blocks
self.blocks = nn.ModuleList([
SwinTransformerBlock(dim=dim, input_resolution=input_resolution,
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(input_resolution, dim=dim, norm_layer=norm_layer)
else:
self.downsample = None
def forward(self, x, x_size):
for blk in self.blocks:
if self.use_checkpoint:
x = checkpoint.checkpoint(blk, x, x_size)
else:
x = blk(x, x_size)
if self.downsample is not None:
x = self.downsample(x)
return x
def extra_repr(self) -> str:
return f"dim={self.dim}, input_resolution={self.input_resolution}, depth={self.depth}"
def flops(self):
flops = 0
for blk in self.blocks:
flops += blk.flops()
if self.downsample is not None:
flops += self.downsample.flops()
return flops
class RSTB(nn.Module):
"""Residual Swin Transformer Block (RSTB).
Args:
dim (int): Number of input channels.
input_resolution (tuple[int]): Input resolution.
depth (int): Number of blocks.
num_heads (int): Number of attention heads.
window_size (int): Local window size.
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 | 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.
img_size: Input image size.
patch_size: Patch size.
resi_connection: The convolutional block before residual connection.
"""
def __init__(self, dim, input_resolution, depth, num_heads, window_size,
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,
img_size=224, patch_size=4, resi_connection='1conv'):
super(RSTB, self).__init__()
self.dim = dim
self.input_resolution = input_resolution
self.residual_group = BasicLayer(dim=dim,
input_resolution=input_resolution,
depth=depth,
num_heads=num_heads,
window_size=window_size,
mlp_ratio=mlp_ratio,
qkv_bias=qkv_bias, qk_scale=qk_scale,
drop=drop, attn_drop=attn_drop,
drop_path=drop_path,
norm_layer=norm_layer,
downsample=downsample,
use_checkpoint=use_checkpoint)
if resi_connection == '1conv':
self.conv = nn.Conv2d(dim, dim, 3, 1, 1)
elif resi_connection == '3conv':
# to save parameters and memory
self.conv = nn.Sequential(nn.Conv2d(dim, dim // 4, 3, 1, 1), nn.LeakyReLU(negative_slope=0.2, inplace=True),
nn.Conv2d(dim // 4, dim // 4, 1, 1, 0),
nn.LeakyReLU(negative_slope=0.2, inplace=True),
nn.Conv2d(dim // 4, dim, 3, 1, 1))
self.patch_embed = PatchEmbed(
img_size=img_size, patch_size=patch_size, in_chans=0, embed_dim=dim,
norm_layer=None)
self.patch_unembed = PatchUnEmbed(
img_size=img_size, patch_size=patch_size, in_chans=0, embed_dim=dim,
norm_layer=None)
def forward(self, x, x_size):
return self.patch_embed(self.conv(self.patch_unembed(self.residual_group(x, x_size), x_size))) + x
def flops(self):
flops = 0
flops += self.residual_group.flops()
H, W = self.input_resolution
flops += H * W * self.dim * self.dim * 9
flops += self.patch_embed.flops()
flops += self.patch_unembed.flops()
return flops
class PatchEmbed(nn.Module):
r""" Image to Patch Embedding
Args:
img_size (int): Image size. Default: 224.
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, img_size=224, patch_size=4, in_chans=3, embed_dim=96, norm_layer=None):
super().__init__()
img_size = to_2tuple(img_size)
patch_size = to_2tuple(patch_size)
patches_resolution = [img_size[0] // patch_size[0], img_size[1] // patch_size[1]]
self.img_size = img_size
self.patch_size = patch_size
self.patches_resolution = patches_resolution
self.num_patches = patches_resolution[0] * patches_resolution[1]
self.in_chans = in_chans
self.embed_dim = embed_dim
if norm_layer is not None:
self.norm = norm_layer(embed_dim)
else:
self.norm = None
def forward(self, x):
x = x.flatten(2).transpose(1, 2) # B Ph*Pw C
if self.norm is not None:
x = self.norm(x)
return x
def flops(self):
flops = 0
H, W = self.img_size
if self.norm is not None:
flops += H * W * self.embed_dim
return flops
class PatchUnEmbed(nn.Module):
r""" Image to Patch Unembedding
Args:
img_size (int): Image size. Default: 224.
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, img_size=224, patch_size=4, in_chans=3, embed_dim=96, norm_layer=None):
super().__init__()
img_size = to_2tuple(img_size)
patch_size = to_2tuple(patch_size)
patches_resolution = [img_size[0] // patch_size[0], img_size[1] // patch_size[1]]
self.img_size = img_size
self.patch_size = patch_size
self.patches_resolution = patches_resolution
self.num_patches = patches_resolution[0] * patches_resolution[1]
self.in_chans = in_chans
self.embed_dim = embed_dim
def forward(self, x, x_size):
B, HW, C = x.shape
x = x.transpose(1, 2).view(B, self.embed_dim, x_size[0], x_size[1]) # B Ph*Pw C
return x
def flops(self):
flops = 0
return flops
class Upsample(nn.Sequential):
"""Upsample module.
Args:
scale (int): Scale factor. Supported scales: 2^n and 3.
num_feat (int): Channel number of intermediate features.
"""
def __init__(self, scale, num_feat):
m = []
if (scale & (scale - 1)) == 0: # scale = 2^n
for _ in range(int(math.log(scale, 2))):
m.append(nn.Conv2d(num_feat, 4 * num_feat, 3, 1, 1))
m.append(nn.PixelShuffle(2))
elif scale == 3:
m.append(nn.Conv2d(num_feat, 9 * num_feat, 3, 1, 1))
m.append(nn.PixelShuffle(3))
else:
raise ValueError(f'scale {scale} is not supported. ' 'Supported scales: 2^n and 3.')
super(Upsample, self).__init__(*m)
class UpsampleOneStep(nn.Sequential):
"""UpsampleOneStep module (the difference with Upsample is that it always only has 1conv + 1pixelshuffle)
Used in lightweight SR to save parameters.
Args:
scale (int): Scale factor. Supported scales: 2^n and 3.
num_feat (int): Channel number of intermediate features.
"""
def __init__(self, scale, num_feat, num_out_ch, input_resolution=None):
self.num_feat = num_feat
self.input_resolution = input_resolution
m = []
m.append(nn.Conv2d(num_feat, (scale ** 2) * num_out_ch, 3, 1, 1))
m.append(nn.PixelShuffle(scale))
super(UpsampleOneStep, self).__init__(*m)
def flops(self):
H, W = self.input_resolution
flops = H * W * self.num_feat * 3 * 9
return flops
class SwinIR(nn.Module):
r""" SwinIR
A PyTorch impl of : `SwinIR: Image Restoration Using Swin Transformer`, based on Swin Transformer.
Args:
img_size (int | tuple(int)): Input image size. Default 64
patch_size (int | tuple(int)): Patch size. Default: 1
in_chans (int): Number of input image channels. Default: 3
embed_dim (int): Patch embedding dimension. Default: 96
depths (tuple(int)): Depth of each Swin Transformer layer.
num_heads (tuple(int)): Number of attention heads in different layers.
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. Default: None
drop_rate (float): Dropout rate. Default: 0
attn_drop_rate (float): Attention dropout rate. Default: 0
drop_path_rate (float): Stochastic depth rate. Default: 0.1
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
use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False
upscale: Upscale factor. 2/3/4/8 for image SR, 1 for denoising and compress artifact reduction
img_range: Image range. 1. or 255.
upsampler: The reconstruction reconstruction module. 'pixelshuffle'/'pixelshuffledirect'/'nearest+conv'/None
resi_connection: The convolutional block before residual connection. '1conv'/'3conv'
"""
def __init__(self, img_size=64, patch_size=1, in_chans=3,
embed_dim=96, depths=[6, 6, 6, 6], num_heads=[6, 6, 6, 6],
window_size=7, mlp_ratio=4., qkv_bias=True, qk_scale=None,
drop_rate=0., attn_drop_rate=0., drop_path_rate=0.1,
norm_layer=nn.LayerNorm, ape=False, patch_norm=True,
use_checkpoint=False, upscale=2, img_range=1., upsampler='', resi_connection='1conv',
**kwargs):
super(SwinIR, self).__init__()
num_in_ch = in_chans
num_out_ch = in_chans
num_feat = 64
self.img_range = img_range
if in_chans == 3:
rgb_mean = (0.4488, 0.4371, 0.4040)
self.mean = torch.Tensor(rgb_mean).view(1, 3, 1, 1)
else:
self.mean = torch.zeros(1, 1, 1, 1)
self.upscale = upscale
self.upsampler = upsampler
self.window_size = window_size
#####################################################################################################
################################### 1, shallow feature extraction ###################################
self.conv_first = nn.Conv2d(num_in_ch, embed_dim, 3, 1, 1)
#####################################################################################################
################################### 2, deep feature extraction ######################################
self.num_layers = len(depths)
self.embed_dim = embed_dim
self.ape = ape
self.patch_norm = patch_norm
self.num_features = embed_dim
self.mlp_ratio = mlp_ratio
# split image into non-overlapping patches
self.patch_embed = PatchEmbed(
img_size=img_size, patch_size=patch_size, in_chans=embed_dim, embed_dim=embed_dim,
norm_layer=norm_layer if self.patch_norm else None)
num_patches = self.patch_embed.num_patches
patches_resolution = self.patch_embed.patches_resolution
self.patches_resolution = patches_resolution
# merge non-overlapping patches into image
self.patch_unembed = PatchUnEmbed(
img_size=img_size, patch_size=patch_size, in_chans=embed_dim, embed_dim=embed_dim,
norm_layer=norm_layer if self.patch_norm else None)
# absolute position embedding
if self.ape:
self.absolute_pos_embed = nn.Parameter(torch.zeros(1, num_patches, embed_dim))
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 Residual Swin Transformer blocks (RSTB)
self.layers = nn.ModuleList()
for i_layer in range(self.num_layers):
layer = RSTB(dim=embed_dim,
input_resolution=(patches_resolution[0],
patches_resolution[1]),
depth=depths[i_layer],
num_heads=num_heads[i_layer],
window_size=window_size,
mlp_ratio=self.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])], # no impact on SR results
norm_layer=norm_layer,
downsample=None,
use_checkpoint=use_checkpoint,
img_size=img_size,
patch_size=patch_size,
resi_connection=resi_connection
)
self.layers.append(layer)
self.norm = norm_layer(self.num_features)
# build the last conv layer in deep feature extraction
if resi_connection == '1conv':
self.conv_after_body = nn.Conv2d(embed_dim, embed_dim, 3, 1, 1)
elif resi_connection == '3conv':
# to save parameters and memory
self.conv_after_body = nn.Sequential(nn.Conv2d(embed_dim, embed_dim // 4, 3, 1, 1),
nn.LeakyReLU(negative_slope=0.2, inplace=True),
nn.Conv2d(embed_dim // 4, embed_dim // 4, 1, 1, 0),
nn.LeakyReLU(negative_slope=0.2, inplace=True),
nn.Conv2d(embed_dim // 4, embed_dim, 3, 1, 1))
#####################################################################################################
################################ 3, high quality image reconstruction ################################
if self.upsampler == 'pixelshuffle':
# for classical SR
self.conv_before_upsample = nn.Sequential(nn.Conv2d(embed_dim, num_feat, 3, 1, 1),
nn.LeakyReLU(inplace=True))
self.upsample = Upsample(upscale, num_feat)
self.conv_last = nn.Conv2d(num_feat, num_out_ch, 3, 1, 1)
elif self.upsampler == 'pixelshuffledirect':
# for lightweight SR (to save parameters)
self.upsample = UpsampleOneStep(upscale, embed_dim, num_out_ch,
(patches_resolution[0], patches_resolution[1]))
elif self.upsampler == 'nearest+conv':
# for real-world SR (less artifacts)
self.conv_before_upsample = nn.Sequential(nn.Conv2d(embed_dim, num_feat, 3, 1, 1),
nn.LeakyReLU(inplace=True))
self.conv_up1 = nn.Conv2d(num_feat, num_feat, 3, 1, 1)
if self.upscale == 4:
self.conv_up2 = nn.Conv2d(num_feat, num_feat, 3, 1, 1)
self.conv_hr = nn.Conv2d(num_feat, num_feat, 3, 1, 1)
self.conv_last = nn.Conv2d(num_feat, num_out_ch, 3, 1, 1)
self.lrelu = nn.LeakyReLU(negative_slope=0.2, inplace=True)
else:
# for image denoising and JPEG compression artifact reduction
self.conv_last = nn.Conv2d(embed_dim, num_out_ch, 3, 1, 1)
self.apply(self._init_weights)
def _init_weights(self, m):
if isinstance(m, nn.Linear):
trunc_normal_(m.weight, std=.02)
if isinstance(m, nn.Linear) and m.bias is not None:
nn.init.constant_(m.bias, 0)
elif isinstance(m, nn.LayerNorm):
nn.init.constant_(m.bias, 0)
nn.init.constant_(m.weight, 1.0)
@torch.jit.ignore
def no_weight_decay(self):
return {'absolute_pos_embed'}
@torch.jit.ignore
def no_weight_decay_keywords(self):
return {'relative_position_bias_table'}
def check_image_size(self, x):
_, _, h, w = x.size()
mod_pad_h = (self.window_size - h % self.window_size) % self.window_size
mod_pad_w = (self.window_size - w % self.window_size) % self.window_size
x = F.pad(x, (0, mod_pad_w, 0, mod_pad_h), 'reflect')
return x
def forward_features(self, x):
x_size = (x.shape[2], x.shape[3])
x = self.patch_embed(x)
if self.ape:
x = x + self.absolute_pos_embed
x = self.pos_drop(x)
for layer in self.layers:
x = layer(x, x_size)
x = self.norm(x) # B L C
x = self.patch_unembed(x, x_size)
return x
def forward(self, x):
H, W = x.shape[2:]
x = self.check_image_size(x)
self.mean = self.mean.type_as(x)
x = (x - self.mean) * self.img_range
if self.upsampler == 'pixelshuffle':
# for classical SR
x = self.conv_first(x)
x = self.conv_after_body(self.forward_features(x)) + x
x = self.conv_before_upsample(x)
x = self.conv_last(self.upsample(x))
elif self.upsampler == 'pixelshuffledirect':
# for lightweight SR
x = self.conv_first(x)
x = self.conv_after_body(self.forward_features(x)) + x
x = self.upsample(x)
elif self.upsampler == 'nearest+conv':
# for real-world SR
x = self.conv_first(x)
x = self.conv_after_body(self.forward_features(x)) + x
x = self.conv_before_upsample(x)
x = self.lrelu(self.conv_up1(torch.nn.functional.interpolate(x, scale_factor=2, mode='nearest')))
if self.upscale == 4:
x = self.lrelu(self.conv_up2(torch.nn.functional.interpolate(x, scale_factor=2, mode='nearest')))
x = self.conv_last(self.lrelu(self.conv_hr(x)))
else:
# for image denoising and JPEG compression artifact reduction
x_first = self.conv_first(x)
res = self.conv_after_body(self.forward_features(x_first)) + x_first
x = x + self.conv_last(res)
x = x / self.img_range + self.mean
return x[:, :, :H*self.upscale, :W*self.upscale]
def flops(self):
flops = 0
H, W = self.patches_resolution
flops += H * W * 3 * self.embed_dim * 9
flops += self.patch_embed.flops()
for i, layer in enumerate(self.layers):
flops += layer.flops()
flops += H * W * 3 * self.embed_dim * self.embed_dim
flops += self.upsample.flops()
return flops
if __name__ == '__main__':
upscale = 4
window_size = 8
height = (1024 // upscale // window_size + 1) * window_size
width = (720 // upscale // window_size + 1) * window_size
model = SwinIR(upscale=2, img_size=(height, width),
window_size=window_size, img_range=1., depths=[6, 6, 6, 6],
embed_dim=60, num_heads=[6, 6, 6, 6], mlp_ratio=2, upsampler='pixelshuffledirect').cuda()
print(model)
pytorch_total_params = sum(p.numel() for p in model.parameters())
print(f"pathGAN has param {pytorch_total_params//1000} K params")
# Count the time
import time
x = torch.randn((1, 3, 180, 180)).cuda()
start = time.time()
x = model(x)
total = time.time() - start
print("total time spent is ", total)
docs/model_zoo.md
+24
View File
@@ -0,0 +1,24 @@
# :european_castle: Model Zoo
- [For Paper weight](#for-paper-weight)
- [For Diverse Upscaler](#for-diverse-upscaler)
## For Paper Weight
| Models | Scale | Description |
| ------------------------------------------------------------------------------------------------------------------------------- | :---- | :------------------------------------------- |
| [4x_APISR_GRL_GAN_generator](https://github.com/Kiteretsu77/APISR/releases/download/v0.1.0/4x_APISR_GRL_GAN_generator.pth) | 4X | 4X GRL model used in the paper |
## For Diverse Upscaler
Actually, I am not that much like GRL. Though they can have the smallest param size with higher numerical results, they are not very memory efficient and the processing speed is slow for Transformer model. One more concern come from the TensorRT deployment, where Transformer architecture is hard to be adapted (needless to say for a modified version of Transformer like GRL).
Thus, for other weights, I will not train a GRL network and also real-world SR of GRL only supports 4x.
| Models | Scale | Description |
| ------------------------------------------------------------------------------------------------------------------------------- | :---- | :------------------------------------------- |
| [2x_APISR_RRDB_GAN_generator](https://github.com/Kiteretsu77/APISR/releases/download/v0.1.0/2x_APISR_RRDB_GAN_generator.pth) | 2X | 2X upscaler by RRDB-6blocks |
nodes.py
+196
View File
@@ -0,0 +1,196 @@
import folder_paths
import os
import torch
import torch.nn.functional as F
from .architecture.rrdb import RRDBNet
from .architecture.grl import GRL
import comfy.model_management as mm
import comfy.utils
def convert_dtype(dtype_str):
if dtype_str == 'fp32':
return torch.float32
elif dtype_str == 'fp16':
return torch.float16
elif dtype_str == 'bf16':
return torch.bfloat16
else:
raise NotImplementedError
def load_rrdb(generator_weight_PATH, scale, print_options=False):
''' A simpler API to load RRDB model from Real-ESRGAN
Args:
generator_weight_PATH (str): The path to the weight
scale (int): the scaling factor
print_options (bool): whether to print options to show what kinds of setting is used
Returns:
generator (torch): the generator instance of the model
'''
# Load the checkpoint
checkpoint_g = torch.load(generator_weight_PATH)
# Find the generator weight
if 'params_ema' in checkpoint_g:
# For official ESRNET/ESRGAN weight
weight = checkpoint_g['params_ema']
generator = RRDBNet(3, 3, scale=scale) # Default blocks num is 6
elif 'params' in checkpoint_g:
# For official ESRNET/ESRGAN weight
weight = checkpoint_g['params']
generator = RRDBNet(3, 3, scale=scale)
elif 'model_state_dict' in checkpoint_g:
# For my personal trained weight
weight = checkpoint_g['model_state_dict']
generator = RRDBNet(3, 3, scale=scale)
else:
print("This weight is not supported")
os._exit(0)
# Handle torch.compile weight key rename
old_keys = [key for key in weight]
for old_key in old_keys:
if old_key[:10] == "_orig_mod.":
new_key = old_key[10:]
weight[new_key] = weight[old_key]
del weight[old_key]
generator.load_state_dict(weight)
generator = generator.eval()
# Print options to show what kinds of setting is used
if print_options:
if 'opt' in checkpoint_g:
for key in checkpoint_g['opt']:
value = checkpoint_g['opt'][key]
print(f'{key} : {value}')
return generator
def load_grl(generator_weight_PATH, scale=4):
''' A simpler API to load GRL model
Args:
generator_weight_PATH (str): The path to the weight
scale (int): Scale Factor (Usually Set as 4)
Returns:
generator (torch): the generator instance of the model
'''
# Load the checkpoint
checkpoint_g = torch.load(generator_weight_PATH)
# Find the generator weight
if 'model_state_dict' in checkpoint_g:
weight = checkpoint_g['model_state_dict']
# GRL tiny model (Note: tiny2 version)
generator = GRL(
upscale = scale,
img_size = 64,
window_size = 8,
depths = [4, 4, 4, 4],
embed_dim = 64,
num_heads_window = [2, 2, 2, 2],
num_heads_stripe = [2, 2, 2, 2],
mlp_ratio = 2,
qkv_proj_type = "linear",
anchor_proj_type = "avgpool",
anchor_window_down_factor = 2,
out_proj_type = "linear",
conv_type = "1conv",
upsampler = "nearest+conv", # Change
)
else:
print("This weight is not supported")
os._exit(0)
generator.load_state_dict(weight)
generator = generator.eval()
num_params = 0
for p in generator.parameters():
if p.requires_grad:
num_params += p.numel()
print(f"Number of parameters {num_params / 10 ** 6: 0.2f}")
return generator
class APISR_upscale:
@classmethod
def INPUT_TYPES(s):
return {"required": {
"ckpt_name": (folder_paths.get_filename_list("upscale_models"), ),
"images": ("IMAGE",),
"per_batch": ("INT", {"default": 16, "min": 1, "max": 4096, "step": 1}),
"dtype": (
[
'fp32',
'fp16',
], {
"default": 'fp32'
}),
},
}
RETURN_TYPES = ("IMAGE", )
RETURN_NAMES = ("images", )
FUNCTION = "upscale"
CATEGORY = "ASPIR"
def upscale(self, ckpt_name, dtype, images, per_batch):
device = mm.get_torch_device()
model_path = folder_paths.get_full_path("upscale_models", ckpt_name)
custom_config = {
'dtype': dtype,
'ckpt_name': ckpt_name,
}
dtype = (convert_dtype(dtype))
if not hasattr(self, 'model') or self.model == None or custom_config != self.current_config:
self.model = None
self.current_config = custom_config
if "RRDB" in ckpt_name:
self.model = load_rrdb(model_path, scale=2)
elif "GRL" in ckpt_name:
self.model = load_grl(model_path, scale=4)
self.model = self.model.to(dtype).to(device)
images = images.permute(0, 3, 1, 2)
B, C, H, W = images.shape
H = (H // 8) * 8
W = (W // 8) * 8
if images.shape[2] != H or images.shape[3] != W:
images = F.interpolate(images, size=(H, W), mode="bicubic")
images = images.to(device = device, dtype = dtype)
self.model.to(device)
pbar = comfy.utils.ProgressBar(B)
t = []
for start_idx in range(0, B, per_batch):
sub_images = self.model(images[start_idx:start_idx+per_batch])
t.append(sub_images.cpu())
# Calculate the number of images processed in this batch
batch_count = sub_images.shape[0]
# Update the progress bar by the number of images processed in this batch
pbar.update(batch_count)
self.model.cpu()
t = torch.cat(t, dim=0).permute(0, 2, 3, 1).cpu().to(torch.float32)
return (t,)
NODE_CLASS_MAPPINGS = {
"APISR_upscale": APISR_upscale,
}
NODE_DISPLAY_NAME_MAPPINGS = {
"APISR_upscale": "APISR Upscale",
}
requirements.txt
+4
View File
@@ -0,0 +1,4 @@
numpy>=1.26.0
scipy>=1.11.3
omegaconf>=2.3.0
timm>=0.9.7