Upload 13 files

.gitignore

@@ -0,0 +1,164 @@
+# Byte-compiled / optimized / DLL files
+__pycache__/
+*.py[cod]
+*$py.class
+
+# C extensions
+*.so
+
+# Distribution / packaging
+.Python
+build/
+develop-eggs/
+dist/
+downloads/
+eggs/
+.eggs/
+lib/
+lib64/
+parts/
+sdist/
+var/
+wheels/
+share/python-wheels/
+*.egg-info/
+.installed.cfg
+*.egg
+MANIFEST
+
+# PyInstaller
+#  Usually these files are written by a python script from a template
+#  before PyInstaller builds the exe, so as to inject date/other infos into it.
+*.manifest
+*.spec
+
+# Installer logs
+pip-log.txt
+pip-delete-this-directory.txt
+
+# Unit test / coverage reports
+htmlcov/
+.tox/
+.nox/
+.coverage
+.coverage.*
+.cache
+nosetests.xml
+coverage.xml
+*.cover
+*.py,cover
+.hypothesis/
+.pytest_cache/
+cover/
+
+# Translations
+*.mo
+*.pot
+
+# Django stuff:
+*.log
+local_settings.py
+db.sqlite3
+db.sqlite3-journal
+
+# Flask stuff:
+instance/
+.webassets-cache
+
+# Scrapy stuff:
+.scrapy
+
+# Sphinx documentation
+docs/_build/
+
+# PyBuilder
+.pybuilder/
+target/
+
+# Jupyter Notebook
+.ipynb_checkpoints
+
+# IPython
+profile_default/
+ipython_config.py
+
+# pyenv
+#   For a library or package, you might want to ignore these files since the code is
+#   intended to run in multiple environments; otherwise, check them in:
+# .python-version
+
+# pipenv
+#   According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control.
+#   However, in case of collaboration, if having platform-specific dependencies or dependencies
+#   having no cross-platform support, pipenv may install dependencies that don't work, or not
+#   install all needed dependencies.
+#Pipfile.lock
+
+# poetry
+#   Similar to Pipfile.lock, it is generally recommended to include poetry.lock in version control.
+#   This is especially recommended for binary packages to ensure reproducibility, and is more
+#   commonly ignored for libraries.
+#   https://python-poetry.org/docs/basic-usage/#commit-your-poetrylock-file-to-version-control
+#poetry.lock
+
+# pdm
+#   Similar to Pipfile.lock, it is generally recommended to include pdm.lock in version control.
+#pdm.lock
+#   pdm stores project-wide configurations in .pdm.toml, but it is recommended to not include it
+#   in version control.
+#   https://pdm.fming.dev/#use-with-ide
+.pdm.toml
+
+# PEP 582; used by e.g. github.com/David-OConnor/pyflow and github.com/pdm-project/pdm
+__pypackages__/
+
+# Celery stuff
+celerybeat-schedule
+celerybeat.pid
+
+# SageMath parsed files
+*.sage.py
+
+# Environments
+.env
+.venv
+env/
+venv/
+ENV/
+env.bak/
+venv.bak/
+
+# Spyder project settings
+.spyderproject
+.spyproject
+
+# Rope project settings
+.ropeproject
+
+# mkdocs documentation
+/site
+
+# mypy
+.mypy_cache/
+.dmypy.json
+dmypy.json
+
+# Pyre type checker
+.pyre/
+
+# pytype static type analyzer
+.pytype/
+
+# Cython debug symbols
+cython_debug/
+
+# PyCharm
+#  JetBrains specific template is maintained in a separate JetBrains.gitignore that can
+#  be found at https://github.com/github/gitignore/blob/main/Global/JetBrains.gitignore
+#  and can be added to the global gitignore or merged into this file.  For a more nuclear
+#  option (not recommended) you can uncomment the following to ignore the entire idea folder.
+.idea/
+
+dev.ipynb
+dev/
+wandb/

LICENSE

@@ -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

@@ -1,13 +1 @@
----
-title: Fer Demo 1
-emoji: 📉
-colorFrom: pink
-colorTo: pink
-sdk: gradio
-sdk_version: 4.26.0
-app_file: app.py
-pinned: false
-license: gpl-3.0
----
-
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+# face-expression-recognizer

app.py

@@ -0,0 +1,34 @@
+import gradio as gr
+from fer import FaceExpressionRecognizer
+
+fer = FaceExpressionRecognizer()
+
+webcam_interface = gr.Interface(
+    fer.handle_frame,
+    inputs=gr.Image(type='pil', sources=['webcam'], streaming=True, label='Input webcam'),
+    outputs=gr.Image(label='Output video'),
+    live=True,
+    title='Webcam mode',
+    description='Created by Czarna Magia AI Student Club',
+    theme=gr.themes.Soft(),
+)
+
+img_interface = gr.Interface(
+    fer.handle_frame,
+    inputs=gr.Image(type='pil', sources=['webcam', 'upload'], label='Input image'),
+    outputs=gr.Image(label='Output image'),
+    title='Image upload mode',
+    description='Created by Czarna Magia AI Student Club',
+    theme=gr.themes.Soft(),
+)
+
+app = gr.TabbedInterface(
+    interface_list=[webcam_interface, img_interface],
+    tab_names=['Webcam', 'Image upload'],
+    title='Face Expression Recognizer',
+    theme=gr.themes.Soft(),
+)
+
+
+if __name__ == '__main__':
+    app.launch()

data_filtering.py

@@ -0,0 +1,332 @@
+import os
+from pathlib import Path
+import cv2
+from PIL import Image
+import numpy as np
+import matplotlib.pyplot as plt
+from sklearn.decomposition import PCA
+import imagehash
+from typing import Callable
+from datetime import datetime as dt
+from abc import ABC, abstractmethod
+
+_DATASET_AVG_MEAN = 129.38489987766278
+_DATASET_AVG_STD = 54.084109207654805
+
+
+def save_to_file(location: str = './extracted_paths.txt') -> Callable:
+    def outer_wrapper(fn: Callable) -> Callable:
+        def inner_wrapper(*args, **kwargs):
+            paths: list[str] = fn(*args, **kwargs)
+            if kwargs.get('to_file'):
+                with open(location, 'a') as file:
+                    file.write('\nFiles to remove [TIMESTAMP {}]:\n'.format(dt.now().strftime('%Y%m%d%H%M%S')))
+                    for p in paths:
+                        file.write(f'{p}\n')
+            return paths
+        return inner_wrapper
+    return outer_wrapper
+
+
+def visualize(show_limit: int = -1) -> Callable:
+    def outer_wrapper(fn: Callable) -> Callable:
+        def inner_wrapper(*args, **kwargs):
+            paths: list[str] = fn(*args, **kwargs)
+            if kwargs.get('visualize_'):
+                if show_limit != -1:
+                    paths = paths[:show_limit]
+
+                num_cols = 8
+                num_rows = len(paths) // num_cols + 1
+
+                fig = plt.figure(figsize=(8, 8))
+                for i, path in enumerate(paths, start=1):
+                    plt.subplot(num_rows, num_cols, i)
+                    plt.imshow(Image.open(path), cmap='gray')
+                    plt.title(f'{Path(path).parent.name}', fontsize=7)
+                    plt.axis('off')
+                fig.tight_layout()
+                plt.tight_layout()
+                fig.subplots_adjust(hspace=0.6, top=0.97)
+                plt.show()
+            return paths
+        return inner_wrapper
+    return outer_wrapper
+
+
+class DataFilter(ABC):
+    def __init__(self):
+        self.paths = []
+
+    @abstractmethod
+    def extract(self, data_dir: str | Path, visualize_: bool, to_file: bool) -> list[str]:
+        pass
+
+    @abstractmethod
+    def clear(self) -> None:
+        pass
+
+    @abstractmethod
+    def filter(self) -> bool:
+        pass
+
+    @staticmethod
+    def _load_data(dir_: str) -> tuple[list[np.ndarray], list[str], list[str]]:
+        images = []
+        class_names = []
+        paths = []
+
+        for path in Path(dir_).glob('**/*.jpg'):
+            label = path.parent.name
+            image = cv2.imread(str(path), cv2.IMREAD_GRAYSCALE)
+            if image is not None and label is not None:
+                images.append(np.array(image))
+                class_names.append(label)
+                paths.append(str(path))
+
+        return images, class_names, paths
+
+
+class DataFilterCompose(DataFilter):
+    def __init__(self, components: list[DataFilter]):
+        super().__init__()
+        self.components = components
+
+    @staticmethod
+    def build(components: list[DataFilter]) -> DataFilter:
+        return DataFilterCompose(components)
+
+    def extract(self, data_dir: str | Path, visualize_: bool, to_file: bool) -> list[str]:
+        extracted_paths = []
+        for component in self.components:
+            cur_extracted_paths = component.extract(data_dir,
+                                                    visualize_=visualize_,
+                                                    to_file=to_file)
+            extracted_paths += cur_extracted_paths
+        self.paths += extracted_paths
+        return extracted_paths
+
+    def clear(self) -> None:
+        for component in self.components:
+            component.clear()
+
+    def filter(self):
+        for component in self.components:
+            component.filter()
+
+    def add_component(self, component: DataFilter, position: int) -> None:
+        self.components.insert(position, component)
+
+    def rm_component(self, position: int) -> None:
+        self.components.pop(position)
+
+
+class StatsDataFilter(DataFilter):
+    _OPTIM_MEAN_THRESH = 107
+    _OPTIM_STD_THRESH = 51
+
+    def __init__(self, data_avg_mean: float = None, data_avg_std: float = None, console_output: bool = False):
+        super().__init__()
+        self.data_avg_mean = data_avg_mean
+        self.data_avg_std = data_avg_std
+        self.console_output = console_output
+
+    @visualize()
+    @save_to_file()
+    def extract(self, data_dir: str | Path, visualize_: bool, to_file: bool) -> list[str]:
+        if self.data_avg_mean is None or self.data_avg_std is None:
+            stats = self._compute_dataset_stats(data_dir)
+            self.data_avg_mean = stats['avg_mean']
+            self.data_avg_std = stats['avg_std']
+
+        extracted_paths = self._extract_outliers_by_stats(
+            data_dir,
+            self.data_avg_mean,
+            self.data_avg_std,
+            StatsDataFilter._OPTIM_MEAN_THRESH,
+            StatsDataFilter._OPTIM_STD_THRESH,
+            self.console_output)
+
+        self.paths += extracted_paths
+        return extracted_paths
+
+    def clear(self) -> None:
+        self.paths.clear()
+        if self.console_output:
+            print(f'[{self.__class__.__name__}]: Paths memory cleared.')
+
+    def filter(self) -> bool:
+        has_error = False
+        for path in self.paths:
+            if not Path(path).exists():
+                has_error = True
+                continue
+            os.remove(path)
+            if self.console_output:
+                print(f'[{self.__class__.__name__}]: Removed {path}')
+        return has_error
+
+    @classmethod
+    def _extract_outliers_by_stats(cls,
+                                   data_root: str | Path,
+                                   dataset_avg_mean: float,
+                                   dataset_avg_std: float,
+                                   mean_thresh: float,
+                                   std_thresh: float,
+                                   console_output: bool = False) -> list[str]:
+        outlier_paths = []
+        count = 0
+        _, _, paths = StatsDataFilter._load_data(data_root)
+        total_len = len(paths)
+        for path in iter(paths):
+            img = cv2.imread(path, cv2.IMREAD_GRAYSCALE)
+            if abs(dataset_avg_mean - np.mean(img)) > mean_thresh or abs(
+                    dataset_avg_std - np.std(img)) > std_thresh:
+                outlier_paths.append(path)
+            if console_output:
+                count += 1
+                print(f'[{cls.__name__}]: Computed {count}/{total_len} images ({count / total_len * 100:.2f}%)')
+        return outlier_paths
+
+    @staticmethod
+    def _compute_dataset_stats(data_dir: str) -> dict[str, float]:
+        img_paths = list(Path(data_dir).glob('**/*.jpg'))
+        num_images = len(img_paths)
+        mean_sum = 0
+        std_sum = 0
+
+        for img_path in img_paths:
+            img = cv2.imread(str(img_path), cv2.IMREAD_GRAYSCALE)
+            img_mean = np.mean(img)
+            img_std = np.std(img)
+            mean_sum += img_mean
+            std_sum += img_std
+
+        avg_mean = mean_sum / num_images
+        avg_std = std_sum / num_images
+        stats_dict = {
+            'avg_mean': avg_mean,
+            'avg_std': avg_std,
+        }
+        return stats_dict
+
+
+class PcaDataFilter(DataFilter):
+    _OPTIM_NUM_COMPONENTS = 4
+    _OPTIM_ERROR_THRESH = 87
+
+    def __init__(self, console_output: bool = False):
+        super().__init__()
+        self.console_output = console_output
+
+    @visualize()
+    @save_to_file()
+    def extract(self, data_dir: str | Path, visualize_: bool, to_file: bool) -> list[str]:
+        extracted_paths = self._extract_outliers_with_pca(data_dir)
+        self.paths += extracted_paths
+        return extracted_paths
+
+    def clear(self) -> None:
+        self.paths.clear()
+        if self.console_output:
+            print(f'[{self.__class__.__name__}]: Paths memory cleared.')
+
+    def filter(self) -> bool:
+        has_error = False
+        for path in self.paths:
+            if not Path(path).exists():
+                has_error = True
+                continue
+            os.remove(path)
+            if self.console_output:
+                print(f'[{self.__class__.__name__}]: Removed {path}')
+        return has_error
+
+    @staticmethod
+    def _extract_outliers_with_pca(dir_: str | Path) -> list[str]:
+        x, _, img_paths = PcaDataFilter._load_data(dir_)
+        x = np.array(x)
+        num_samples, height, width = x.shape
+        X_flattened = x.reshape(num_samples, height * width)
+
+        outlier_indices = PcaDataFilter._detect_outliers_with_pca(X_flattened,
+                                                                  PcaDataFilter._OPTIM_NUM_COMPONENTS,
+                                                                  PcaDataFilter._OPTIM_ERROR_THRESH)
+        img_paths_to_remove = [img_paths[i] for i in outlier_indices.tolist()]
+        return img_paths_to_remove
+
+    @staticmethod
+    def _detect_outliers_with_pca(orig_data: np.ndarray,
+                                  num_components: int,
+                                  error_thresh: float) -> np.ndarray:
+        pca = PCA(n_components=num_components)
+        X_reduced = pca.fit_transform(orig_data)
+
+        X_reconstructed = pca.inverse_transform(X_reduced)
+        reconstruction_errors = np.sqrt(np.mean((orig_data - X_reconstructed) ** 2, axis=1))
+
+        outlier_indices = np.where(reconstruction_errors > error_thresh)[0]
+        return outlier_indices
+
+
+class DHashDuplicateFilter(DataFilter):
+    def __init__(self, hash_size: int = 8, console_output: bool = False):
+        super().__init__()
+        self.hash_size = hash_size
+        self.console_output = console_output
+
+    @visualize(60)
+    @save_to_file()
+    def extract(self, data_dir: str | Path, visualize_: bool, to_file: bool) -> list[str]:
+        _, _, paths = self._load_data(data_dir)
+        hashes = set()
+        duplicates = []
+
+        for path in paths:
+            hash_ = imagehash.dhash(Image.open(path), self.hash_size)
+            if hash_ in hashes:
+                duplicates.append(path)
+                if self.console_output:
+                    print(f'[{self.__class__.__name__}]: Duplicate found at {path}')
+            else:
+                hashes.add(hash_)
+
+        self.paths += duplicates
+        return duplicates
+
+    def clear(self) -> None:
+        self.paths.clear()
+        if self.console_output:
+            print(f'[{self.__class__.__name__}]: Paths memory cleared.')
+
+    def filter(self) -> bool:
+        has_error = False
+        for path in self.paths:
+            if not Path(path).exists():
+                has_error = True
+                continue
+            os.remove(path)
+            if self.console_output:
+                print(f'[{self.__class__.__name__}]: Removed {path}')
+        return has_error
+
+
+if __name__ == '__main__':
+    dataset_dir = Path('./dataset')
+
+    stats_filter = StatsDataFilter(_DATASET_AVG_MEAN, _DATASET_AVG_STD, True)
+    pca_filter = PcaDataFilter(console_output=True)
+    duplicate_filter = DHashDuplicateFilter(console_output=True)
+
+    compose = DataFilterCompose.build([
+        stats_filter,
+        pca_filter,
+        duplicate_filter
+    ])
+
+    # You may set the value of visualize_ or to_file parameters to True
+    # to plot extracted images or save paths to a file.
+    stats_filter.extract(dataset_dir, visualize_=False, to_file=False)
+
+    # WARNING: uncommenting the line below will irreversibly remove dataset files
+    # compose.filter()

face_detector.py

@@ -0,0 +1,18 @@
+from facenet_pytorch import MTCNN
+import torch
+from PIL import Image
+import numpy as np
+
+
+class FaceDetector:
+    def __init__(self):
+        device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
+        self.model = MTCNN(keep_all=True, post_process=False, device=device)
+
+    def detect_bboxes(self, img: Image.Image) -> np.ndarray:
+        boxes, probs = self.model.detect(img)
+        return boxes
+    
+    def extract_faces(self, img: Image.Image, bboxes) -> torch.Tensor:
+        face_images_tensors = self.model.extract(img, bboxes, save_path='dev/extracted/extracted.jpg')
+        return face_images_tensors

fer.py

@@ -0,0 +1,59 @@
+from face_detector import FaceDetector
+from model_small import ResNet18
+import numpy as np
+import torch
+from torch import nn
+from PIL import Image
+from util import draw_bboxes, draw_label_on_bbox
+import torchvision.transforms as T
+
+
+class FaceExpressionRecognizer:
+
+    _DATASET_MEAN = 0.5077385902404785
+    _DATASET_STD = 0.255077600479126
+
+    def __init__(self):
+        self.face_detector = FaceDetector()
+        self.fer_classifier = _make_fer_classifier()
+        self.post_process = T.Compose([
+            T.Resize((48, 48)),
+            T.Grayscale(),
+            T.ConvertImageDtype(torch.float32),
+            T.Normalize(FaceExpressionRecognizer._DATASET_MEAN, FaceExpressionRecognizer._DATASET_STD)
+        ])
+        self.idx_to_label = {
+            0: 'angry',
+            1: 'disgust',
+            2: 'fear',
+            3: 'happy',
+            4: 'neutral',
+            5: 'sad',
+            6: 'surprise',
+        }
+
+    def handle_frame(self, image: Image.Image) -> Image.Image:
+        bboxes = self.face_detector.detect_bboxes(image)
+        if bboxes is None:
+            return image
+
+        extracted_faces = self.face_detector.extract_faces(image, bboxes)
+        extracted_faces = self.post_process(extracted_faces)
+        preds = self.fer_classifier(extracted_faces).argmax(dim=1)
+        print(f'Preds: {preds}')
+        preds = preds.tolist()
+
+        img_w_boxes = draw_bboxes(image.copy(), bboxes, (255, 0, 0))
+        image_w_boxes_arr = np.array(img_w_boxes)
+        for bbox, pred in zip(bboxes, preds):
+            image_w_boxes_arr = draw_label_on_bbox(image_w_boxes_arr, bbox, self.idx_to_label[pred])
+        return Image.fromarray(image_w_boxes_arr)
+
+
+def _make_fer_classifier() -> nn.Module:
+    model = ResNet18(1, 7)
+    # fer_fc = nn.Linear(256, 7)
+    # model = nn.Sequential(*list(model.children())[:-1])
+    # model = nn.Sequential(*model, fer_fc)
+    model.load_state_dict(torch.load('./saved_models/weighted_sampler200_fer_model.pth', map_location=torch.device('cpu')))
+    return model

model.py

@@ -0,0 +1,102 @@
+import torch
+from torch import nn
+
+
+class ResNet18(nn.Module):
+    def __init__(self, in_channels: int, num_classes: int):
+        super().__init__()
+        self.initial_conv = nn.Conv2d(
+            in_channels=in_channels,
+            out_channels=64,
+            kernel_size=7,
+            stride=2,
+            padding=3,
+            bias=False,
+        )
+        self.bn = nn.BatchNorm2d(64)
+        self.relu = nn.ReLU(inplace=True)
+        self.max_pool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
+
+        self.layer1 = nn.Sequential(BasicBlock(64, 64), BasicBlock(64, 64))
+        self.layer2 = nn.Sequential(
+            BasicBlock(64, 128, stride=2, downsample=self._downsample(64, 128)),
+            BasicBlock(128, 128),
+        )
+        self.layer3 = nn.Sequential(
+            BasicBlock(128, 256, stride=2, downsample=self._downsample(128, 256)),
+            BasicBlock(256, 256),
+        )
+        self.layer4 = nn.Sequential(
+            BasicBlock(256, 512, stride=2, downsample=self._downsample(256, 512)),
+            BasicBlock(512, 512),
+        )
+
+        self.avg_pool = nn.AdaptiveAvgPool2d((1, 1))
+        self.drop = nn.Dropout(0.15)
+        self.flatten = nn.Flatten(1)
+        self.fc = nn.Linear(512, num_classes)
+
+    @staticmethod
+    def _downsample(in_channels: int, out_channels: int) -> nn.Sequential:
+        return nn.Sequential(
+            nn.Conv2d(in_channels, out_channels, kernel_size=1, stride=2, bias=False),
+            nn.BatchNorm2d(out_channels),
+        )
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        x = self.initial_conv(x)
+        x = self.bn(x)
+        x = self.relu(x)
+        x = self.max_pool(x)
+
+        x = self.layer1(x)
+        x = self.layer2(x)
+        x = self.layer3(x)
+        x = self.layer4(x)
+
+        x = self.avg_pool(x)
+        x = self.drop(x)  # because linear layers tends to overfit more
+        x = self.flatten(x)
+        x = self.fc(x)
+        return x
+
+
+class BasicBlock(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        stride: int = 1,
+        downsample: nn.Module = None,
+    ):
+        super().__init__()
+        self.downsample = downsample
+        self.conv1 = nn.Conv2d(
+            in_channels,
+            out_channels,
+            kernel_size=3,
+            stride=stride,
+            padding=1,
+            bias=False,
+        )
+        self.bn1 = nn.BatchNorm2d(out_channels)
+        self.relu = nn.ReLU(inplace=True)
+        self.conv2 = nn.Conv2d(
+            out_channels, out_channels, kernel_size=3, stride=1, padding=1, bias=False
+        )
+        self.bn2 = nn.BatchNorm2d(out_channels)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        identity = x
+
+        output = self.conv1(x)
+        output = self.bn1(output)
+        output = self.relu(output)
+        output = self.conv2(output)
+        output = self.bn2(output)
+
+        if self.downsample is not None:
+            identity = self.downsample(x)
+        output += identity
+        output = self.relu(output)
+        return output

model_small.py

@@ -0,0 +1,88 @@
+import torch
+from torch import nn
+
+
+class ResNet18(nn.Module):
+    def __init__(self, in_channels: int, num_classes: int):
+        super().__init__()
+        self.initial_conv = nn.Conv2d(in_channels=in_channels,
+                                      out_channels=32,
+                                      kernel_size=5,
+                                      stride=1,
+                                      padding=2,
+                                      bias=False)
+        self.bn = nn.BatchNorm2d(32)
+        self.relu = nn.ReLU(inplace=True)
+        self.max_pool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
+
+        self.layer1 = nn.Sequential(
+            BasicBlock(32, 32),
+            BasicBlock(32, 32)
+        )
+        self.layer2 = nn.Sequential(
+            BasicBlock(32, 64, stride=2, downsample=self._downsample(32, 64)),
+            BasicBlock(64, 64)
+        )
+        self.layer3 = nn.Sequential(
+            BasicBlock(64, 128, stride=2, downsample=self._downsample(64, 128)),
+            BasicBlock(128, 128)
+        )
+        self.layer4 = nn.Sequential(
+            BasicBlock(128, 256, stride=2, downsample=self._downsample(128, 256)),
+            BasicBlock(256, 256)
+        )
+
+        self.avg_pool = nn.AdaptiveAvgPool2d((1, 1))
+        self.drop = nn.Dropout(0.15)
+        self.flatten = nn.Flatten(1)
+        self.fc = nn.Linear(256, num_classes)
+
+    @staticmethod
+    def _downsample(in_channels: int, out_channels: int) -> nn.Sequential:
+        return nn.Sequential(
+            nn.Conv2d(in_channels, out_channels, kernel_size=1, stride=2, bias=False),
+            nn.BatchNorm2d(out_channels)
+        )
+
+    def forward(self, x):
+        x = self.initial_conv(x)
+        x = self.bn(x)
+        x = self.relu(x)
+        x = self.max_pool(x)
+
+        x = self.layer1(x)
+        x = self.layer2(x)
+        x = self.layer3(x)
+        x = self.layer4(x)
+
+        x = self.avg_pool(x)
+        x = self.drop(x)
+        x = self.flatten(x)
+        x = self.fc(x)
+        return x
+
+
+class BasicBlock(nn.Module):
+    def __init__(self, in_channels: int, out_channels: int, stride: int = 1, downsample: nn.Module = None):
+        super().__init__()
+        self.downsample = downsample
+        self.conv1 = nn.Conv2d(in_channels, out_channels, kernel_size=3, stride=stride, padding=1, bias=False)
+        self.bn1 = nn.BatchNorm2d(out_channels)
+        self.relu = nn.ReLU(inplace=True)
+        self.conv2 = nn.Conv2d(out_channels, out_channels, kernel_size=3, stride=1, padding=1, bias=False)
+        self.bn2 = nn.BatchNorm2d(out_channels)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        identity = x
+
+        output = self.conv1(x)
+        output = self.bn1(output)
+        output = self.relu(output)
+        output = self.conv2(output)
+        output = self.bn2(output)
+
+        if self.downsample is not None:
+            identity = self.downsample(x)
+        output += identity
+        output = self.relu(output)
+        return output

preprocessing.py

@@ -0,0 +1,69 @@
+from types import SimpleNamespace
+
+import torch
+from torch.utils.data import DataLoader
+from torchvision.datasets import ImageFolder
+from torchvision.transforms import v2
+
+DATASET_MEAN = 0.5077385902404785
+DATASET_STD = 0.255077600479126
+
+
+class PreprocessedImageFolder(ImageFolder):
+    def __init__(self, root, transform=None):
+        super().__init__(root, transform=transform)
+        self.preprocess = v2.Compose(
+            [
+                v2.Grayscale(),
+                v2.PILToTensor(),
+                v2.ToDtype(torch.float32, scale=True),
+                v2.Normalize(mean=(DATASET_MEAN,), std=(DATASET_STD,)),
+            ]
+        )
+
+        processed_samples = []
+        for path, target in self.samples:
+            sample = self.loader(path)
+            processed_sample = self.preprocess(sample)
+            processed_samples.append((processed_sample, target))
+
+        self.samples = processed_samples
+
+    def __getitem__(self, index):
+        sample, target = self.samples[index]
+        if self.transform is not None:
+            sample = self.transform(sample)
+        return sample, target
+
+
+augmentations = v2.Compose(
+    [
+        v2.ColorJitter(brightness=0.2, contrast=0.2, saturation=0.2),
+        v2.RandomHorizontalFlip(),
+        v2.RandomResizedCrop(size=48, scale=(0.9, 1.1), antialias=True),
+        v2.RandomAffine(degrees=(-10, 10), translate=(0.1, 0.1), scale=(0.9, 1.1)),
+        v2.GaussianBlur(kernel_size=(3, 3), sigma=(0.1, 1.0)),
+    ]
+)
+
+
+def make_dls(train_ds, valid_ds, batch_size=64, num_workers=2):
+    train_dl = DataLoader(
+        train_ds,
+        batch_size=batch_size,
+        shuffle=True,
+        num_workers=num_workers,
+        pin_memory=True,
+        persistent_workers=True,
+    )
+    valid_dl = DataLoader(
+        valid_ds,
+        batch_size=batch_size,
+        shuffle=True,
+        num_workers=num_workers,
+        pin_memory=True,
+        persistent_workers=True,
+    )
+
+    dls = SimpleNamespace(**{"train": train_dl, "valid": valid_dl})
+    return dls

train.py

@@ -0,0 +1,50 @@
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+from model import ResNet18
+from preprocessing import PreprocessedImageFolder, augmentations, make_dls
+from trainer import (
+    LRFinderCB,
+    ActivationStatsCB,
+    AugmentCB,
+    DeviceCB,
+    MultiClassAccuracyCB,
+    ProgressCB,
+    Trainer,
+    WandBCB,
+)
+
+device = "cuda" if torch.cuda.is_available() else "cpu"
+
+train_ds = PreprocessedImageFolder("./dataset/train", None)
+valid_ds = PreprocessedImageFolder("./dataset/test", None)
+dls = make_dls(train_ds, valid_ds, batch_size=32, num_workers=2)
+
+model = ResNet18(in_channels=1, num_classes=len(train_ds.classes))
+
+
+# lr_find = LRFinderCB(min_lr=1e-4, max_lr=0.1, max_mult=3)
+# act_stats = ActivationStatsCB(mod_filter=lambda x: isinstance(x, nn.Conv2d) or isinstance(x, nn.Linear), with_wandb=True) # for debugging purposes
+progress = ProgressCB(in_notebook=False)
+wandb_cb = WandBCB(proj_name="test", model_path="./model.pth")
+augment = AugmentCB(device=device, transform=augmentations)
+acc_cb = MultiClassAccuracyCB(with_wandb=True)
+
+trainer = Trainer(
+    model,
+    dls,
+    F.cross_entropy,
+    torch.optim.SGD,
+    lr=1e-4,
+    cbs=[DeviceCB(device), augment, progress, wandb_cb, acc_cb],
+)  # act_stats, lr_find
+trainer.fit(5, True, True)
+
+# TODO: saving plots to wandb
+progress.plot_losses(save=True)
+# act_stats.plot_stats(save=True)
+# act_stats.color_dim(save=True)
+# act_stats.dead_chart(save=True)
+
+# torch.save(trainer.model.state_dict(), "./model.pth") # done by WandBCB

trainer.py

@@ -0,0 +1,520 @@
+# code inspired by Fastai "Practical Deep Learning Part 2" Learner
+import math
+import os
+from functools import partial
+from operator import attrgetter
+
+import matplotlib.pyplot as plt
+import numpy as np
+import torch
+import torch.nn.functional as F
+import wandb
+
+
+class CancelFitException(Exception):
+    pass
+
+
+class CancelBatchException(Exception):
+    pass
+
+
+class CancelEpochException(Exception):
+    pass
+
+
+class Callback:
+    order = 0
+
+
+class with_cbs:
+    """Decorator that wraps function and calls certain callbacks before/after that function."""
+
+    def __init__(self, nm):
+        self.nm = nm
+
+    def __call__(self, f):
+        def _f(o, *args, **kwargs):
+            try:
+                o.callback(f"before_{self.nm}")
+                f(o, *args, **kwargs)
+                o.callback(f"after_{self.nm}")
+            except globals()[f"Cancel{self.nm.title()}Exception"]:
+                pass
+            finally:
+                o.callback(f"cleanup_{self.nm}")
+
+        return _f
+
+
+def run_cbs(cbs, method_nm, trainer=None):
+    for cb in sorted(cbs, key=attrgetter("order")):  # sort callbacks by 'order'
+        method = getattr(
+            cb, method_nm, None
+        )  # get method from callback e.g. `before_batch`
+        if method is not None:
+            method(trainer)  # if callback has such method then call it
+
+
+class Trainer:
+    """Trainer with callbacks"""
+
+    def __init__(
+        self,
+        model,
+        dls=(0,),
+        loss_func=F.mse_loss,
+        opt_func=torch.optim.SGD,
+        lr=0.1,
+        cbs=[],
+        n_inp=1,
+    ):
+        self.model = model
+        self.dls = dls
+        self.loss_func = loss_func
+        self.opt_func = opt_func
+        self.lr = lr
+        self.cbs = cbs
+        self.n_inp = n_inp
+
+    @with_cbs("batch")
+    def _one_batch(self):
+        self.predict()
+        self.callback("after_predict")
+        self.get_loss()
+        self.callback("after_loss")
+        if self.training:
+            self.backward()
+            self.callback("after_backward")
+            self.step()
+            self.callback("after_step")
+            self.zero_grad()
+
+    @with_cbs("epoch")
+    def _one_epoch(self):
+        for self.iter, self.batch in enumerate(self.dl):
+            self._one_batch()
+
+    def one_epoch(self, training):
+        self.model.train(training)
+        self.dl = self.dls.train if training else self.dls.valid
+        self._one_epoch()
+
+    @with_cbs("fit")
+    def _fit(self, train, valid):
+        for epoch in range(self.n_epochs):
+            if train:
+                self.one_epoch(True)
+            if valid:
+                torch.no_grad()(self.one_epoch)(False)
+
+    def fit(self, n_epochs=1, train=True, valid=True, cbs=None, lr=None):
+        self.n_epochs = n_epochs
+        if lr is not None:
+            self.lr = lr
+        self.opt = self.opt_func(self.model.parameters(), self.lr)
+        self._fit(train, valid)
+
+    def callback(self, method_nm):
+        run_cbs(self.cbs, method_nm, self)
+
+    def predict(self, x=None):
+        if x is not None:
+            return self.model(x)
+        self.preds = self.model(*self.batch[: self.n_inp])
+
+    def get_loss(self):
+        self.loss = self.loss_func(self.preds, *self.batch[self.n_inp :])
+
+    def backward(self):
+        self.loss.backward()
+
+    def step(self):
+        self.opt.step()
+
+    def zero_grad(self):
+        self.opt.zero_grad()
+
+    @property
+    def training(self):
+        return self.model.training
+
+
+class ProgressCB(Callback):
+    """Adds progress bar to Trainer and plotting loss curves after training."""
+
+    def __init__(self, in_notebook=False):
+        super().__init__()
+        self.train_loss = []
+        self.valid_loss = []
+        self.in_notebook = in_notebook
+
+    def before_fit(self, trainer):
+        if self.in_notebook:
+            from tqdm.notebook import tqdm
+        else:
+            from tqdm import tqdm
+        self.pbar = tqdm(total=trainer.n_epochs)
+
+    def after_epoch(self, trainer):
+        if trainer.training:
+            self.pbar.update(1)
+
+    def after_loss(self, trainer):
+        if trainer.training:
+            self.train_loss.append(trainer.loss.item())
+            tmp_train_loss = (
+                np.mean(self.train_loss[-10:]) if len(self.train_loss) > 10 else 0
+            )
+            tmp_valid_loss = (
+                np.mean(self.valid_loss[-len(trainer.dls.valid) :])
+                if len(self.valid_loss) > 0
+                else 0
+            )
+            self.pbar.set_description(
+                f"train loss: {tmp_train_loss:.3f} | valid loss: {tmp_valid_loss:.3f}"
+            )
+        else:
+            self.valid_loss.append(trainer.loss.item())
+
+    def after_fit(self, trainer):
+        self.pbar.close()
+
+    def plot_losses(self, save=True):
+        fig, ax = plt.subplots(1, 2, figsize=(12, 4))
+        ax[0].plot(self.train_loss)
+        ax[0].set_title("train loss")
+        ax[1].plot(self.valid_loss)
+        ax[1].set_title("valid loss")
+        if save:
+            if not os.path.exists("./plots"):
+                os.makedirs("./plots")
+            plt.savefig("./plots/losses.png")
+        else:
+            plt.show()
+
+
+class DeviceCB(Callback):
+    """Moves model and batches to device"""
+
+    def __init__(self, device="cpu"):
+        self.device = device
+
+    def before_fit(self, trainer):
+        if hasattr(trainer.model, "to"):
+            trainer.model.to(self.device)
+
+    def before_batch(self, trainer):
+        trainer.batch = tuple(t.to(self.device) for t in trainer.batch)
+
+
+class Hook:
+    """Registers PyTorch forward hook with provided function"""
+
+    def __init__(self, name, mod, f):
+        self.hook = mod.register_forward_hook(partial(f, self, name))
+
+    def remove(self):
+        self.hook.remove()
+
+    def __del__(self):
+        self.remove()
+
+
+class Hooks(list):
+    """List of hooks"""
+
+    def __init__(self, mods, f):
+        super().__init__([Hook(n, m, f) for n, m in mods])
+
+    def __enter__(self, *args):
+        return self
+
+    def __exit__(self, *args):
+        self.remove()
+
+    def __del__(self):
+        self.remove()
+
+    def __delitem__(self, i):
+        self[i].remove()
+        super().__delitem__(i)
+
+    def remove(self):
+        for h in self:
+            h.remove()
+
+
+class HooksCB(Callback):
+    """Appends hooks with some `hookfunc` to selected layers filtered by `mod_filter`."""
+
+    def __init__(self, hookfunc, mod_filter=lambda x: True):
+        super().__init__()
+        self.hookfunc = hookfunc
+        self.mod_filter = mod_filter
+
+    def before_fit(self, trainer):
+        mods = [
+            (name, mod)
+            for name, mod in trainer.model.named_modules()
+            if self.mod_filter(mod)
+        ]
+        self.hooks = Hooks(mods, partial(self._hookfunc, trainer.training))
+
+    def _hookfunc(self, training, *args, **kwargs):
+        if training:
+            self.hookfunc(*args, **kwargs)
+
+    def after_fit(self, trainer):
+        self.hooks.remove()
+
+    def __iter__(self):
+        return iter(self.hooks)
+
+    def __len__(self):
+        return len(self.hooks)
+
+
+def append_stats(with_wandb, hook, name, mod, inp, outp):
+    if not hasattr(hook, "stats"):
+        hook.stats = {"mean": [], "std": [], "abs": []}
+    acts = outp.detach().cpu()
+    hook.stats["mean"].append(acts.mean().item())
+    hook.stats["std"].append(acts.std().item())
+    hook.stats["abs"].append(acts.abs().histc(40, 0, 10).tolist())
+    if with_wandb:
+        wandb.log(
+            {
+                f"{name}/mean": acts.mean().item(),
+                f"{name}/std": acts.std().item(),
+                f"{name}/abs": wandb.Histogram(acts.abs().histc(40, 0, 10).tolist()),
+            },
+            commit=False,
+        )
+
+
+def get_grid(n, figsize):
+    return plt.subplots(round(n / 2), 2, figsize=figsize)
+
+
+class WandBCB(Callback):
+    """Inits and logs to W&B. Every `wandb.log()` outside this callback should have property `commit=False` because this callback gathers all logs in given batch."""
+
+    order = math.inf  # make sure that this callback will be called last
+
+    def __init__(
+        self, proj_name, model_path, run_name=None, notes=None, **additional_config
+    ):
+        self.proj_name = proj_name
+        self.run_name = run_name
+        self.model_path = model_path
+        self.notes = notes
+        self.additional_config = additional_config
+
+    def before_fit(self, trainer):
+        info = dict(
+            project=self.proj_name,
+            config={"lr": trainer.lr, "n_epochs": trainer.n_epochs},
+        )
+        if self.run_name is not None:
+            info["name"] = self.run_name
+        if self.notes is not None:
+            info["notes"] = self.notes
+        if self.additional_config is not None:
+            info["config"] = {**info["config"], **self.additional_config}
+
+        wandb.init(**info)
+        wandb.watch(trainer.model, log="all")
+
+    def after_loss(self, trainer):
+        if trainer.training:
+            wandb.log({"loss/train": trainer.loss.item()}, commit=False)
+        else:
+            wandb.log({"loss/valid": trainer.loss.item()}, commit=False)
+
+    def after_batch(self, trainer):
+        wandb.log({}, commit=True)
+
+    def after_fit(self, trainer):
+        torch.save(trainer.model.state_dict(), self.model_path)
+        wandb.save(self.model_path)
+        wandb.finish()
+
+
+class ActivationStatsCB(HooksCB):
+    """Stores activation statistics of selected modules. Recommended only for debugging or visualizations, not for actual training because it significantly slows down training."""
+
+    def __init__(self, mod_filter=lambda x: x, with_wandb=False):
+        super().__init__(partial(append_stats, with_wandb), mod_filter)
+
+    def plot_stats(self, save=True):  # plot output means & std devs of each module
+        fig, axes = get_grid(2, figsize=(20, 10))
+        for h in self.hooks:
+            for i, name in enumerate(["mean", "std dev"]):
+                axes[i].plot(h.stats[i])
+                axes[i].set_title(name)
+        plt.legend(range(len(self.hooks)))
+        if save:
+            if not os.path.exists("./plots"):
+                os.makedirs("./plots")
+            plt.savefig("./plots/mean_std_stats.png")
+        else:
+            plt.show()
+
+    # plot "color dim" that shows abs values of outputs through training time (should be normally distributed - uniform gradient)
+    def color_dim(self, save=True):
+        fig, axes = get_grid(len(self.hooks), figsize=(20, 10))
+        for ax, h in zip(axes.flatten(), self.hooks):
+            ax.set_ylim(0, 40)
+            ax.imshow(self.get_hist(h), aspect="auto")
+        if save:
+            if not os.path.exists("./plots"):
+                os.makedirs("./plots")
+            plt.savefig("./plots/color_dim.png")
+        else:
+            plt.show()
+
+    # plot % of dead neurons
+    def dead_chart(self, save=True):
+        fig, axes = get_grid(len(self.hooks), figsize=(20, 10))
+        for ax, h in zip(axes.flatten(), self.hooks):
+            ax.plot(self.get_min(h))
+            ax.set_ylim(0, 1)
+        if save:
+            if not os.path.exists("./plots"):
+                os.makedirs("./plots")
+            plt.savefig("./plots/dead_neurons_perc.png")
+        else:
+            plt.show()
+
+    # ratio of dead neurons (activations near 0)
+    def get_min(self, h):
+        h1 = torch.stack(h.stats[2]).t().float()
+        return h1[0] / h1.sum(0)
+
+    def get_hist(self, h):
+        return torch.stack(h.stats[2]).t().float().log1p()
+
+
+class LRFinderCB(Callback):
+    """Suggests an approx. good LR for a model. Usually you should choose value where loss is still decreasing (steepest slope), not the lowest value."""
+
+    def __init__(self, min_lr=1e-6, max_lr=1, max_mult=3, num_iter=100):
+        self.min_lr = min_lr
+        self.max_lr = max_lr
+        self.max_mult = max_mult
+        self.num_iter = num_iter
+        self.lr_factor = (max_lr / min_lr) ** (1 / num_iter)
+
+    def before_fit(self, trainer):
+        self.lrs, self.losses = [], []
+        self.min = math.inf
+        self.i = 0
+        trainer.opt.param_groups[0]["lr"] = self.min_lr
+
+    def before_batch(self, trainer):
+        trainer.opt.param_groups[0]["lr"] *= self.lr_factor
+
+    def after_batch(self, trainer):
+        if not trainer.training:
+            raise CancelEpochException()
+        self.lrs.append(trainer.opt.param_groups[0]["lr"])
+        loss = trainer.loss.to("cpu").item()
+        self.losses.append(loss)
+        if loss < self.min:
+            self.min = loss
+        self.i += 1
+        if (
+            math.isnan(loss)
+            or (loss > self.min * self.max_mult)
+            or (self.i > self.num_iter)
+        ):
+            raise CancelFitException()
+
+    def plot_lrs(self, log=True, window=None):
+        plt.plot(self.lrs, self.losses)  # original loss curve
+        plt.title("LR finder")
+        if log:
+            plt.xscale("log")
+
+        if window is None:
+            window = self.num_iter // 4
+
+        smoothed_losses = np.convolve(
+            self.losses, np.ones(window) / window, mode="valid"
+        )
+        gradients = np.gradient(smoothed_losses)
+        min_gradient_idx = np.argmin(gradients)
+        self.best_lr = self.lrs[min_gradient_idx + window // 2]
+
+        plt.plot(
+            self.best_lr, smoothed_losses[min_gradient_idx + window // 2], "ro"
+        )  # recomended LR value point
+        plt.text(
+            self.best_lr,
+            smoothed_losses[min_gradient_idx + window // 2],
+            f"LR: {self.best_lr:.1e}",
+            fontsize=12,
+            ha="center",
+            va="bottom",
+            bbox=dict(facecolor="white"),
+        )
+
+        plt.plot(
+            self.lrs[window // 2 : -window // 2 + 1], smoothed_losses, alpha=0.5
+        )  # smoothed loss curve
+
+
+class AugmentCB(Callback):
+    """Computes augmentation transformations on device (e.g. GPU) for faster training."""
+
+    def __init__(self, device="cpu", transform=None):
+        super().__init__()
+        self.device = device
+        self.transform = transform
+
+    def before_batch(self, trainer):
+        trainer.batch = tuple(
+            [
+                *[self.transform(t) for t in trainer.batch[: trainer.n_inp]],
+                *trainer.batch[trainer.n_inp :],
+            ]
+        )
+
+
+class MultiClassAccuracyCB(Callback):
+    def __init__(self, with_wandb=False):
+        self.all_acc = {"train": [], "valid": []}
+        self.with_wandb = with_wandb
+
+    def before_epoch(self, trainer):
+        self.acc = []
+
+    def after_predict(self, trainer):
+        self.acc = []
+        with torch.inference_mode():
+            self.acc.append(
+                (
+                    F.softmax(trainer.preds, dim=1).argmax(1)
+                    == trainer.batch[trainer.n_inp :][0]
+                ).float()
+            )
+
+    def after_epoch(self, trainer):
+        final_acc = torch.hstack(self.acc).mean().item()
+        if trainer.training:
+            if self.with_wandb:
+                wandb.log({"accuracy/train": final_acc}, commit=False)
+            self.all_acc["train"].append(final_acc)
+        else:
+            if self.with_wandb:
+                wandb.log({"accuracy/valid": final_acc}, commit=False)
+            self.all_acc["valid"].append(final_acc)
+        self.acc = []
+
+    def plot_acc(self):
+        fig, axes = get_grid(2, (20, 10))
+        axes[0].plot(self.all_acc["train"])
+        axes[0].set_title("train acc")
+        axes[1].plot(self.all_acc["valid"])
+        axes[1].set_title("valid acc")

util.py

@@ -0,0 +1,24 @@
+import cv2
+import numpy as np
+from PIL import Image, ImageDraw
+
+
+def draw_bboxes(img: Image.Image, boxes, color: tuple[int, int, int]) -> Image.Image:
+    img_draw = ImageDraw.Draw(img)
+    for box in boxes:
+        img_draw.rectangle(box.tolist(), outline=color, width=2)
+    return img
+
+
+def draw_label_on_bbox(image: np.ndarray, bbox: list[float], text: str) -> np.ndarray:
+    image = cv2.cvtColor(image, cv2.COLOR_RGB2BGR)
+    left_pos = bbox[0]
+    bottom_pos = bbox[1] - 5
+    bottom_left_position = (int(left_pos), int(bottom_pos))
+    font = cv2.FONT_HERSHEY_SIMPLEX
+    font_scale = .9
+    color = (255, 0, 0)
+    thickness = 2
+
+    annotated_image = cv2.putText(image, text, bottom_left_position, font, font_scale, color, thickness)
+    return cv2.cvtColor(annotated_image, cv2.COLOR_BGR2RGB)