first commit

.github/workflows/python-app.yml

@@ -0,0 +1,46 @@
+# This workflow will install Python dependencies, run tests and lint with a single version of Python
+# For more information see: https://docs.github.com/en/actions/automating-builds-and-tests/building-and-testing-python
+
+name: Python application
+
+on:
+  push:
+    branches: [ "main" ]
+  pull_request:
+    branches: [ "main" ]
+
+permissions:
+  contents: read
+
+jobs:
+  build:
+
+    runs-on: ubuntu-latest
+
+    steps:
+    - uses: actions/checkout@v4
+    - name: Set up Python 3.10
+      uses: actions/setup-python@v3
+      with:
+        python-version: "3.10"
+    - name: Cache pip
+      uses: actions/cache@v3
+      with:
+        path: ~/.cache/pip
+        key: ${{ runner.os }}-pip-${{ hashFiles('**/requirements.txt') }}
+        restore-keys: |
+          ${{ runner.os }}-pip-
+    - name: Install dependencies
+      run: |
+        python -m pip install --upgrade pip
+        pip install flake8 pytest
+        if [ -f requirements.txt ]; then pip install -r requirements.txt; fi
+    - name: Lint with flake8
+      run: |
+        # stop the build if there are Python syntax errors or undefined names
+        flake8 . --count --select=E9,F63,F7,F82 --show-source --statistics
+        # exit-zero treats all errors as warnings. The GitHub editor is 127 chars wide
+        flake8 . --count --exit-zero --max-complexity=10 --max-line-length=127 --statistics
+    - name: Test with pytest
+      run: |
+        PYTHONPATH=. pytest tests

.gitignore

@@ -0,0 +1 @@
+__pycache__

LICENSE

@@ -0,0 +1,661 @@
+                    GNU AFFERO GENERAL PUBLIC LICENSE
+                       Version 3, 19 November 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 Affero General Public License is a free, copyleft license for
+software and other kinds of works, specifically designed to ensure
+cooperation with the community in the case of network server software.
+
+  The licenses for most software and other practical works are designed
+to take away your freedom to share and change the works.  By contrast,
+our General Public Licenses are 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.
+
+  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.
+
+  Developers that use our General Public Licenses protect your rights
+with two steps: (1) assert copyright on the software, and (2) offer
+you this License which gives you legal permission to copy, distribute
+and/or modify the software.
+
+  A secondary benefit of defending all users' freedom is that
+improvements made in alternate versions of the program, if they
+receive widespread use, become available for other developers to
+incorporate.  Many developers of free software are heartened and
+encouraged by the resulting cooperation.  However, in the case of
+software used on network servers, this result may fail to come about.
+The GNU General Public License permits making a modified version and
+letting the public access it on a server without ever releasing its
+source code to the public.
+
+  The GNU Affero General Public License is designed specifically to
+ensure that, in such cases, the modified source code becomes available
+to the community.  It requires the operator of a network server to
+provide the source code of the modified version running there to the
+users of that server.  Therefore, public use of a modified version, on
+a publicly accessible server, gives the public access to the source
+code of the modified version.
+
+  An older license, called the Affero General Public License and
+published by Affero, was designed to accomplish similar goals.  This is
+a different license, not a version of the Affero GPL, but Affero has
+released a new version of the Affero GPL which permits relicensing under
+this license.
+
+  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 Affero 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. Remote Network Interaction; Use with the GNU General Public License.
+
+  Notwithstanding any other provision of this License, if you modify the
+Program, your modified version must prominently offer all users
+interacting with it remotely through a computer network (if your version
+supports such interaction) an opportunity to receive the Corresponding
+Source of your version by providing access to the Corresponding Source
+from a network server at no charge, through some standard or customary
+means of facilitating copying of software.  This Corresponding Source
+shall include the Corresponding Source for any work covered by version 3
+of the GNU General Public License that is incorporated pursuant to the
+following paragraph.
+
+  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 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 work with which it is combined will remain governed by version
+3 of the GNU General Public License.
+
+  14. Revised Versions of this License.
+
+  The Free Software Foundation may publish revised and/or new versions of
+the GNU Affero 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 Affero 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 Affero 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 Affero 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 Affero 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 Affero General Public License for more details.
+
+    You should have received a copy of the GNU Affero 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 your software can interact with users remotely through a computer
+network, you should also make sure that it provides a way for users to
+get its source.  For example, if your program is a web application, its
+interface could display a "Source" link that leads users to an archive
+of the code.  There are many ways you could offer source, and different
+solutions will be better for different programs; see section 13 for the
+specific requirements.
+
+  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 AGPL, see
+<https://www.gnu.org/licenses/>.

Readme.md

@@ -0,0 +1,77 @@
+## metrics to track
+- loss per epoch per model boost layer
+- number of error per epoch model boost layer
+- number of resolved puzzles per epochs
+- threshold per epochs per model layer
+- number of filled digits per model boost layer per epoch for both pis ans abs
+
+## TODO
+- jupyter notebook to python file
+- threshold compute on test set (with adding a gap) each epoch. and training threshold initialised with test thresholds that evolve each error during training.
+
+## Possible way
+- it might be smart to store the intermitent states as boost layereds "buffers". at the end the first X go to the model layer 0 let write it as puseudo code
+
+### Method threshold
+```
+global init
+th ->  -10
+training step
+  init 
+    pass
+  training loop
+    keep th behind the error limit
+validation step
+  init
+    compute_th =-10
+  validation loop
+    keep compute_th behind error limit + marge
+    but use th
+  end
+    th= compute_th
+```
+
+### Method training
+```
+Xs -> the x initial batch vector 
+Y -> the y batch vector
+Xs' = M0(Xs)
+then we filter Xs'=Y -> resolved sudokus
+Xs'==Xs -> we add the rows to X1 buffer
+and the remaning Xs' is added to X0 buffer.
+```
+
+then we look at each buffers X0 to Xn and we process each of them that are => batch size.
+
+When every buffer are smaller than batch size the process is finished.
+
+object 
+```
+Buffers
+get_batch(limit_batch_size=True) -> idx, Xb # Xb could be none. (Xb should be a shuffled sample of the batch)
+add_batch(Xp, idx)
+```
+
+
+### Loss optimisation
+Both 0 and 1 target are different in the way we should gradient descend them.
+y==0 point is something easy: it should be as low as possible I thing we can use the usual log loss function on it.
+y==1 is different: there is different case possible:
+ - the point could be "unpredictable" in that case the gradient descend should be tuned to low, we expect the predictive function to have a low score.
+ - the point could be well predicted in that case we hope the value is prety hight and we would like to the the gradient descend more heavely.
+This could be applied by using a sigmoid centered on the threshold
+
+
+### Paper writing
+
+Les niveaux supérieurs font appel à divers types de chaînes :
+
+11.6 Dynamic + Dynamic Forcing Chains (145-192 nodes) Cell Forcing Chains
+11.7 Dynamic + Dynamic Forcing Chains (193-288 nodes) Double Forcing Chains
+Ces Dynamic Forcing Chains sont une forme d’essais et erreurs.
+
+### Trial and error solving technique
+We applied trial and error solving technique to reach 100% accuracy over sudoku. The resoning is simple we find the best digit/position to test and produce 2 children grid one with the number the other without. the we process each grid until one of them break sudoku's rules .
+
+The V1 of this algorithm should only stopped at 1 trail and error test (no binary tree search) it should be simpler and feasible and if not : we will se an improve and try the next step.
+

app.py

@@ -0,0 +1,126 @@
+import streamlit as st
+from sudoku.train import SudokuTrialErrorLightning
+from sudoku.helper import display_as_dataframe, get_grid_number_soluce
+import numpy as np
+import re
+import torch
+
+model = SudokuTrialErrorLightning.load_from_checkpoint(checkpoint_path="model_9_m3_5epoch.ckpt")
+
+empty_grid = '''
+0 0 0 | 0 0 0 | 0 0 0
+0 0 0 | 0 0 0 | 0 0 0
+0 0 0 | 0 0 0 | 0 0 0
+---------------------
+0 0 0 | 0 0 0 | 0 0 0
+0 0 0 | 0 0 0 | 0 0 0
+0 0 0 | 0 0 0 | 0 0 0
+---------------------
+0 0 0 | 0 0 0 | 0 0 0
+0 0 0 | 0 0 0 | 0 0 0
+0 0 0 | 0 0 0 | 0 0 0
+'''
+grid_evil_0 = '''
+    [0,0,0, 1,6,0, 2,0,0],
+    [0,3,0, 0,0,5, 0,0,4],
+    [0,0,7, 0,0,0, 0,9,0],
+
+    [0,0,1, 0,0,0, 0,8,2],
+    [3,0,0, 7,0,4, 0,0,5],
+    [4,2,0, 0,0,0, 6,0,0],
+
+    [0,8,0, 0,0,0, 5,0,0],
+    [7,0,0, 8,0,0, 0,2,0],
+    [0,0,6, 0,5,1, 0,0,0],
+'''
+grid_evil_1 = '''
+    [0,0,0, 2,0,0, 0,0,0],
+    [0,0,9, 7,0,0, 3,8,0],
+    [3,6,0, 0,0,0, 0,7,0],
+
+    [1,0,6, 8,0,0, 0,0,0],
+    [0,0,0, 3,5,9, 0,0,0],
+    [0,0,0, 0,0,2, 8,0,5],
+
+    [0,4,0, 0,0,0, 0,9,1],
+    [0,2,5, 0,0,8, 7,0,0],
+    [0,0,0, 0,0,6, 0,0,0],
+]
+'''
+grid_evil_2 = '''
+    [0,2,0, 1,0,0, 0,7,0],
+    [0,0,0, 3,0,2, 0,0,0],
+    [0,0,1, 0,8,0, 0,4,0],
+
+    [9,0,0, 0,1,0, 0,0,7],
+    [0,0,8, 0,6,0, 0,5,0],
+    [0,0,0, 0,0,0, 0,0,0],
+
+    [0,8,0, 0,3,6, 0,0,9],
+    [0,0,5, 0,7,0, 0,0,6],
+    [0,0,2, 0,0,0, 0,0,3],
+]
+'''
+grids = {
+    'empty': empty_grid,
+    'evil_1': grid_evil_0,
+    'evil_2': grid_evil_1, 
+    'evil_3': grid_evil_2
+}
+
+def str_to_row_col_grid(su_str):
+    su_str = re.sub(r'[^\d.]', '', su_str).replace('.','0')
+
+    assert len(su_str)==81
+    return [[int(su_str[j*9+i]) for i in range(9)] for j in range(9)]
+
+def build_tensor_out_of_grid(grid):
+    np_grid = np.asarray(grid) 
+    big_grid = np.zeros((1,2,9,9,9), dtype=np.float32)
+    for i in range(9):
+        big_grid[0,1,:,:,i] = (np_grid==i+1)
+        big_grid[0,0,:,:,i] = ((np_grid!=i+1) & (np_grid!=0))
+    return torch.from_numpy(big_grid).view(1,2,729)
+
+def build_grid_and_display(str_grid):
+    test_x = build_tensor_out_of_grid(str_grid)
+    return test_x, display_as_dataframe(test_x)
+st.markdown('# Deep learning sudoku Solver')
+st.markdown('### Author: Sébastien Guissart')
+
+option = st.selectbox(
+    "select template grids",
+    grids,
+)
+x = st.text_area('grid', value=grids[option], height=300)
+
+tensor_grid, df_styler = build_grid_and_display(str_to_row_col_grid(x))
+# st.table(df_styler)
+# st.dataframe(df_styler)
+html = df_styler.to_html(escape=False, index=False)
+# # st.text(html)
+st.html(html)
+st.markdown('''
+            ## Checking the number of solutions 
+            (with backtracking algorithm)
+            must be equal 1
+            ''')
+n_sol = get_grid_number_soluce(str_to_row_col_grid(x))
+
+st.markdown(f"number of solution: {'>2' if n_sol==2 else n_sol}")
+if n_sol==1:
+    new_X = model.predict(tensor_grid)
+    X_sum= new_X.sum()
+    st.html(display_as_dataframe(new_X).to_html(escape=False, index=False))
+    i=1
+    while new_X.sum()<729:
+        i+=1
+        st.markdown(f'iteration {i}')
+        new_X = model.predict(new_X)
+        st.html(display_as_dataframe(new_X).to_html(escape=False, index=False))
+        new_X_sum = new_X.sum()
+        assert new_X_sum> X_sum
+        X_sum = new_X_sum
+    st.markdown('## Grid solved!')
+
+    

justfile

@@ -0,0 +1,10 @@
+init:
+    rm -rf venv
+    python3 -m venv venv
+    venv/bin/pip install -r requirements.txt
+
+test:
+    PYTHONPATH=. venv/bin/pytest tests
+
+tensorboard:
+    venv/bin/tensorboard --logdir lightning_logs

requirements.txt

@@ -0,0 +1,10 @@
+pytorch_lightning
+lightning_lite
+tensorboard
+pandas
+seaborn
+torch==1.12.1
+jupyter
+scikit-learn
+pytest
+black

sudoku/buffer.py

@@ -0,0 +1,93 @@
+import random
+import torch
+
+# TODO intensively test it!
+# we got an error pop index out of range during 
+
+
+class BufferArray:
+    def __init__(self, array_size, batch_size):
+        self.array_size = array_size
+        self.batch_size = batch_size
+        self.batch_n_vectors = None
+
+    def init_buffers(self):
+        self.buffers = [
+            [[] for _ in range(self.batch_n_vectors)] for _ in range(self.array_size)
+        ]
+
+    def get_batch(self):
+        "return the idx of the first buffer reaching batch_size and a batch"
+        assert hasattr(self, "buffers")
+        for idx, buffer in enumerate(self.buffers):
+            if len(buffer[0]) >= self.batch_size:
+                vectors = [[] for _ in range(self.batch_n_vectors)]
+                for _ in range(self.batch_size):
+                    pop_idx = random.randrange(len(buffer[0]))
+                    for v, b in zip(vectors, buffer):
+                        v.append(b.pop(pop_idx))
+                return idx, tuple([torch.stack(v, dim=0) for v in vectors])
+        return 0, None
+
+    def append(self, idx, batch: tuple):
+        "append batch to the desired buffer. we suppose batch as a tuple such as (x,y)"
+
+        if idx >= self.array_size:
+            return
+        if self.batch_n_vectors is None:
+            self.batch_n_vectors = len(batch)
+            self.init_buffers()
+        else:
+            assert len(batch) == self.batch_n_vectors
+        for i, element_vectors in enumerate(batch):
+            self.buffers[idx][i] = self.buffers[idx][i] + [
+                vector for vector in element_vectors
+            ]
+
+
+#     def append(self, X, Y) -> None:
+#         """Add experience to the buffer.
+
+#         Args:
+#             experience: tuple (state, action, reward, done, new_state)
+#         """
+
+#         X[Y == 0] = 0
+#         mask = ~(X == Y).view(-1, 2 * 729).all(dim=1)
+
+#         for x, y in zip(X[mask], Y[mask]):
+#             self.buffer.append((x, y))
+
+class Buffer:
+    def __init__(self, batch_size):
+        self.batch_size = batch_size
+        self.batch_n_vectors = None
+
+    def init_buffer(self):
+        self.buffer = [[] for _ in range(self.batch_n_vectors)]
+
+    def get_batch(self):
+        "return the idx of the first buffer reaching batch_size and a batch"
+        if not hasattr(self, "buffer"):
+            return None
+        if len(self.buffer[0]) >= self.batch_size:
+            vectors = [[] for _ in range(self.batch_n_vectors)]
+            for _ in range(self.batch_size):
+                pop_idx = random.randrange(len(self.buffer[0]))
+                for v, b in zip(vectors, self.buffer):
+                    v.append(b.pop(pop_idx))
+            return tuple([torch.stack(v, dim=0) for v in vectors])
+        return None
+
+    def append(self, batch: tuple):
+        "append batch to the desired buffer. we suppose batch as a tuple such as (x,y)"
+
+        if self.batch_n_vectors is None:
+            self.batch_n_vectors = len(batch)
+            self.init_buffer()
+        else:
+            assert len(batch) == self.batch_n_vectors
+        for i, element_vectors in enumerate(batch):
+            self.buffer[i] = self.buffer[i] + [
+                vector for vector in element_vectors
+            ]

sudoku/helper.py

@@ -0,0 +1,160 @@
+import numpy as np
+import pandas as pd
+import torch.nn.functional as F
+
+
+x_grid = np.array([x for x in range(1, 10)] * 9 * 9).reshape(9, 9, 3, 3)
+df_grid = pd.DataFrame(x_grid.swapaxes(1, 2).reshape(27, 27))
+
+cell = {  # for row hover use <tr> instead of <td>
+    "selector": "td",
+    "props": [("color", "black"), ("text-align", "center")],
+}
+line_height_1 = {"selector": "tr", "props": [("line-height", "2em")]}
+line_height_2 = {"selector": "td", "props": "line-height: inherit; padding: 0;"}
+border_global = {"selector": "", "props": [("border", "2px solid")]}
+border_left_property_1 = [{"selector": "td", "props": "border-left: 1px solid black"}]
+border_left_property_2 = [{"selector": "td", "props": "border-left: 2px solid black"}]
+
+border_1 = {
+    3: border_left_property_1,
+    6: border_left_property_1,
+    9: border_left_property_2,
+    12: border_left_property_1,
+    15: border_left_property_1,
+    18: border_left_property_2,
+    21: border_left_property_1,
+    24: border_left_property_1,
+}
+border_top_property_1 = [{"selector": "td", "props": "border-top: 1px solid black"}]
+border_top_property_2 = [{"selector": "td", "props": "border-top: 2px solid black"}]
+border_2 = {
+    3: border_top_property_1,
+    6: border_top_property_1,
+    9: border_top_property_2,
+    12: border_top_property_1,
+    15: border_top_property_1,
+    18: border_top_property_2,
+    21: border_top_property_1,
+    24: border_top_property_1,
+}
+
+
+def display_as_dataframe(x_input, output=None, display="neg"):
+    x_mask_pos = (
+        x_input[0, 1, :].numpy().reshape(9, 9, 3, 3).swapaxes(1, 2).reshape(27, 27)
+    )
+    x_mask_is_not_pos = x_input[0, 1, :].numpy().reshape(9, 9, 9).max(axis=2) == 0
+    x_mask_neg_pos = x_input[0, 0, :].numpy().reshape(9, 9, 9).copy()
+    x_mask_neg_pos[x_mask_is_not_pos] = np.array([0 for _ in range(9)])
+    x_mask_full_neg = (
+        x_input[0, 0, :].numpy().reshape(9, 9, 3, 3).swapaxes(1, 2).reshape(27, 27)
+    )
+    x_mask_neg_pos = x_mask_neg_pos.reshape(9, 9, 3, 3).swapaxes(1, 2).reshape(27, 27)
+
+    cell_color = pd.DataFrame("black", index=df_grid.index, columns=df_grid.columns)
+    cell_color[x_mask_pos == 1] = "blue"
+    cell_color[x_mask_full_neg == 1] = "red"
+    cell_color[x_mask_neg_pos == 1] = "white"
+
+    styler = (
+        df_grid.style.hide(axis=1)
+        .hide(axis=0)
+        .set_properties(subset=[x for x in range(27)], **{"width": "2em"})
+        .set_table_styles(
+            [  # create internal CSS classes
+                {"selector": ".red", "props": "color: red; font-weight: bold"},
+                {"selector": ".blue", "props": "color: blue; font-weight: bold"},
+                {"selector": ".white", "props": "color: white; font-weight: bold"},
+            ],
+            overwrite=False,
+        )
+        .set_td_classes(cell_color)
+        .set_table_styles(
+            [cell, border_global, line_height_1, line_height_2], overwrite=False
+        )
+        .set_table_styles(border_1, overwrite=False)
+        .set_table_styles(border_2, overwrite=False, axis=1)
+    )
+    if (output is not None) and (display == "neg"):
+        x_output = (
+            output[0, 0, :].numpy().reshape(9, 9, 3, 3).swapaxes(1, 2).reshape(27, 27)
+        )
+        return styler.background_gradient(
+            axis=None, vmin=-3, vmax=3, cmap="bwr", gmap=x_output
+        )
+    if (output is not None) and (display == "pos"):
+        x_output = (
+            output[0, 1, :].numpy().reshape(9, 9, 3, 3).swapaxes(1, 2).reshape(27, 27)
+        )
+        return styler.background_gradient(
+            axis=None, vmin=-3, vmax=3, cmap="bwr", gmap=x_output
+        )
+
+    return styler
+
+
+def compute_loss(x, y, output, new_x):
+    mask_0_error = (new_x == 1) & (y == 0)
+    mask_error = mask_0_error.view(-1, 2 * 729).any(dim=1)
+    mask_no_improve = new_x.sum(dim=(1, 2)) <= x.sum(dim=(1, 2))
+    mask_no_improve[mask_error] = False
+    mask_1_no_improve = y == 1
+    mask_1_no_improve[~mask_no_improve] = False
+    loss = F.binary_cross_entropy_with_logits(output, y, reduce=False)
+    loss_error = loss[mask_0_error].mean()
+    loss_no_improve = loss[mask_1_no_improve].mean()
+    # loss_error = F.binary_cross_entropy_with_logits(output[mask_0_error], y[mask_0_error])
+    # loss_no_improve = F.binary_cross_entropy_with_logits(output[mask_1_no_improve], y[mask_1_no_improve])
+
+    return loss_error, loss_no_improve, mask_error.sum(), mask_no_improve.sum()
+
+
+# returns 0, 1 or more than 1 depending on whether 0, 1 or more than 1 solutions are found
+def solve(i, j, cells, count):  # initially called with count = 0
+    if i == 9:
+        i = 0
+        j += 1
+        if j == 9:
+            return 1 + count
+    if cells[i][j] != 0:  # skip filled cells
+        return solve(i + 1, j, cells, count)
+    
+    for val in range(1, 10):
+        if count < 2 and legal(i, j, val, cells):
+            cells[i][j] = val
+            count = solve(i + 1, j, cells, count)
+    
+    cells[i][j] = 0  # reset on backtrack
+    return count
+
+
+def legal(row, col, num, grid):
+  
+    # Check if we find the same num
+    # in the similar row , we
+    # return false
+    for x in range(9):
+        if grid[row][x] == num:
+            return False
+
+    # Check if we find the same num in
+    # the similar column , we
+    # return false
+    for x in range(9):
+        if grid[x][col] == num:
+            return False
+
+    # Check if we find the same num in
+    # the particular 3*3 matrix,
+    # we return false
+    startRow = row - row % 3
+    startCol = col - col % 3
+    for i in range(3):
+        for j in range(3):
+            if grid[i + startRow][j + startCol] == num:
+                return False
+    return True
+
+def get_grid_number_soluce(grid):
+    return solve(0,0,grid,0)

sudoku/loader.py

@@ -0,0 +1,190 @@
+import numpy as np
+import torch
+from torch.utils.data import DataLoader
+from sklearn.model_selection import train_test_split
+from torch.utils.data.dataset import IterableDataset
+from collections import deque
+
+from numpy.random import default_rng
+
+DATA = np.load(
+    # "/home/seb/Perso/git/sudoku/sudoku_clean/data/sudoku_reshaped_million.npz"
+    "sudoku_reshaped_3_million.npz"
+)
+
+rng = np.random.default_rng()
+
+
+def get_datasets(
+    add_proba_fill=False, train_size=1280 // 2, test_size=1280 // 2, max_holes=None
+):
+    quizzes = DATA["quizzes"][: train_size + test_size]
+    solutions = DATA["solutions"][: train_size + test_size]
+    X = quizzes
+    if max_holes:
+        while True:
+            x_holes = X[:, 1].sum(-1) == 0
+            x_nb_holes = x_holes.sum((1, 2))
+            mask_x_max_holes = x_nb_holes > max_holes
+            if not any(mask_x_max_holes):
+                break
+            for idx_x in np.nonzero(mask_x_max_holes)[0]:
+                sub_x_holes = x_holes[idx_x]
+                idx_fill = rng.choice(np.transpose(np.nonzero(sub_x_holes)))
+                X[idx_x, :, idx_fill[0], idx_fill[1], :] = solutions[
+                    idx_x, :, idx_fill[0], idx_fill[1], :
+                ]
+    X = X.reshape(X.shape[0], 2, 9 * 9 * 9)
+    solutions = solutions.reshape(solutions.shape[0], 2, 9 * 9 * 9)
+
+    X_train, X_test, solutions_train, solutions_test = train_test_split(
+        X, solutions, test_size=test_size, random_state=42
+    )
+    if add_proba_fill:
+        X_train_bis = X_train.copy()
+        mask = solutions_train == 1
+        X_train_bis[mask] = np.random.randint(0, 2, size=mask.sum())
+        X_train = np.concatenate([X_train, X_train_bis])
+        solutions_train = np.concatenate([solutions_train, solutions_train])
+
+    train = torch.utils.data.TensorDataset(
+        torch.Tensor(X_train), torch.Tensor(solutions_train)
+    )
+    test = torch.utils.data.TensorDataset(
+        torch.Tensor(X_test), torch.Tensor(solutions_test)
+    )
+    return train, test
+
+
+train_dataset, test_dataset = get_datasets()
+
+
+def data_loader(batch_size=32, add_proba_fill=False):
+    train, test = get_datasets(add_proba_fill=add_proba_fill)
+
+    train_loader = torch.utils.data.DataLoader(train, batch_size=batch_size)
+
+    test_loader = torch.utils.data.DataLoader(test, batch_size=batch_size)
+
+    return train_loader, test_loader
+
+
+class DataIterBuffer(IterableDataset):
+    def __init__(self, raw_dataset=[], buffer_optim=50, prop_new=0.1, seed=1):
+        self.raw_dataset = raw_dataset
+        # self.raw_dataset = iter(raw_dataset)
+        self.buffer = deque()
+        self.buffer_optim = buffer_optim
+        self.prop_new = prop_new
+        self.rng = default_rng(seed=seed)
+        self.idx_dataset = 0
+
+    def __iter__(self):
+        # while True:
+        #     if (np.random.random() < self.prop_new) and (
+        #         len(self.buffer) <= self.buffer_optim
+        #     ):
+        #         try:
+        #             yield next(self.raw_dataset)
+        #         except StopIteration:
+        #             if len(self.buffer) != 0:
+        #                 yield self.buffer.popleft()
+        #             else:
+        #                 break
+        #     else:
+        #         if len(self.buffer) != 0:
+        #             yield self.buffer.popleft()
+        #         else:
+        #             try:
+        #                 yield next(self.raw_dataset)
+        #             except StopIteration:
+        #                 break
+        while True:
+            if (np.random.random() < self.prop_new) and (
+                len(self.buffer) <= self.buffer_optim
+            ):
+                if self.idx_dataset >= len(self.raw_dataset):
+                    if len(self.buffer) != 0:
+                        yield self.buffer.popleft()
+                    else:
+                        break
+                else:
+                    yield self.raw_dataset[self.idx_dataset]
+                    self.idx_dataset += 1
+            else:
+                if len(self.buffer) != 0:
+                    yield self.buffer.popleft()
+                else:
+                    if self.idx_dataset >= len(self.raw_dataset):
+                        break
+                    else:
+                        yield self.raw_dataset[self.idx_dataset]
+                        self.idx_dataset += 1
+
+    def append(self, X, Y) -> None:
+        """Add experience to the buffer.
+
+        Args:
+            experience: tuple (state, action, reward, done, new_state)
+        """
+
+        X[Y == 0] = 0
+        mask = ~(X == Y).view(-1, 2 * 729).all(dim=1)
+
+        for x, y in zip(X[mask], Y[mask]):
+            self.buffer.append((x, y))
+
+    def __len__(self):
+        return len(self.buffer) + len(self.raw_dataset)
+
+
+# class DataIterDeepBuffer(IterableDataset):
+#     def __init__(self, raw_dataset=[], buffer_target_size=32, prop_new=0.1, seed=1, prof=6):
+#         self.raw_dataset = iter(raw_dataset)
+#         # self.buffer = deque()
+#         self.buffer_target_size = buffer_target_size
+#         self.prop_new = prop_new
+#         self.rng = default_rng(seed=seed)
+#         self.prof=prof
+#         self.buffers=[deque() for _ in range(prof)]
+
+#     def __iter__(self):
+#         while True:
+#             buffer_sizes = np.array([len(buffer) for buffer in self.buffers])
+#             if any(buffer_sizes>=self.buffer_target_size):
+#                 #
+
+#             if (np.random.random() < self.prop_new) and (
+#                 len(self.buffer) <= self.buffer_optim
+#             ):
+#                 try:
+#                     yield next(self.raw_dataset)
+#                 except StopIteration:
+#                     if len(self.buffer) != 0:
+#                         yield self.buffer.popleft()
+#                     else:
+#                         break
+#             else:
+#                 if len(self.buffer) != 0:
+#                     yield self.buffer.popleft()
+#                 else:
+#                     try:
+#                         yield next(self.raw_dataset)
+#                     except StopIteration:
+#                         break
+
+#     def append(self, X, Y) -> None:
+#         """Add experience to the buffer.
+
+#         Args:
+#             experience: tuple (state, action, reward, done, new_state)
+#         """
+
+#         X[Y == 0] = 0
+#         mask = ~(X == Y).view(-1, 2 * 729).all(dim=1)
+
+#         for x, y in zip(X[mask], Y[mask]):
+#             self.buffer.append((x, y))
+
+#     def __len__(self):
+#         return len(self.buffer) + len(self.raw_dataset)

sudoku/loader_buffer.py

@@ -0,0 +1,28 @@
+
+
+from torch.utils.data.dataset import IterableDataset
+from torch.utils.data import  DataLoader
+import numpy as np
+
+from sudoku.loader import DataIterBuffer, train_dataset, test_dataset, data_loader, get_datasets
+
+
+
+class CustomDataLoader(DataLoader):
+    def __init__(self, data_iters, batch_size):
+        self.data_iters=data_iters
+        self.batch_size=batch_size
+        self.data_loaders = [iter(DataLoader(data_iter, batch_size=batch_size)) for data_iter in data_iters]
+    def __iter__(self):
+        while True:
+            buffer_sizes = np.array([len(buffer) for buffer in self.data_iters])
+            if any(buffer_sizes>=self.batch_size):
+                idx_yield = len(buffer_sizes)-1-np.argmax(buffer_sizes[::-1]>=self.batch_size)
+                
+            else :
+                idx_yield = np.argmax(buffer_sizes)
+            yield [idx_yield]+ next(self.data_loaders[idx_yield])
+    
+    @property
+    def num_workers(self):
+        return 0

sudoku/models.py

@@ -0,0 +1,124 @@
+from torch import nn
+import torch
+from sudoku.symetries import mat_sym
+
+
+class SudokuNet(nn.Module):
+    def __init__(self, n_output=2, coef_hidden=4):
+        super(SudokuNet, self).__init__()
+        self.hidden_neural_number = mat_sym.shape[1]
+
+        self.conv111 = nn.Conv1d(
+            self.hidden_neural_number * 2,
+            self.hidden_neural_number * 2 * coef_hidden,
+            1,
+            groups=self.hidden_neural_number * 2,
+        )
+        self.conv111_last = nn.Conv1d(
+            self.hidden_neural_number * 2 * coef_hidden, n_output, 1
+        )
+
+        sym_tensor = torch.from_numpy(mat_sym).type(torch.FloatTensor)
+        self.sym_tensor = nn.Parameter(sym_tensor, requires_grad=False)
+
+    def forward(self, x):
+        x = torch.tensordot(x, self.sym_tensor, dims=([2], [2]))
+        x = x.view(-1, 2, 9 * 9 * 9, self.hidden_neural_number)
+        x = x.permute(0, 1, 3, 2)
+        x = x.contiguous().view(-1, self.hidden_neural_number * 2, 9 * 9 * 9)
+        x = torch.relu(self.conv111(x))
+        x = self.conv111_last(x)
+        return x
+
+
+class SymPreprocess(nn.Module):
+    def __init__(self):
+        super().__init__()
+        self.hidden_neural_number = mat_sym.shape[1]
+        sym_tensor = torch.from_numpy(mat_sym).type(torch.FloatTensor)
+        self.sym_tensor = nn.Parameter(sym_tensor, requires_grad=False)
+
+    def forward(self, x):
+        n_channel = x.shape[1]
+        x = torch.tensordot(x, self.sym_tensor, dims=([2], [2]))
+        x = x.view(-1, n_channel, 9 * 9 * 9, self.hidden_neural_number)
+        x = x.permute(0, 1, 3, 2)
+        x = x.contiguous().view(-1, self.hidden_neural_number * n_channel, 9 * 9 * 9)
+        return x
+
+
+class SmallNet(nn.Module):
+    def __init__(self, n_output=2, coef_hidden=4, n_input_channel=2):
+        super(SmallNet, self).__init__()
+        self.hidden_neural_number = mat_sym.shape[1]
+
+        self.conv111 = nn.Conv1d(
+            self.hidden_neural_number * n_input_channel,
+            self.hidden_neural_number * n_input_channel * coef_hidden,
+            1,
+            groups=self.hidden_neural_number * n_input_channel,
+        )
+        self.conv111_last = nn.Conv1d(
+            self.hidden_neural_number * n_input_channel * coef_hidden, n_output, 1
+        )
+
+    def forward(self, x):
+        x = torch.relu(self.conv111(x))
+        x = self.conv111_last(x)
+        return x
+
+
+class SmallNetBis(nn.Module):
+    def __init__(self, n_output=2, coef_hidden=2, n_input_channel=2):
+        super(SmallNetBis, self).__init__()
+        self.hidden_neural_number = mat_sym.shape[1] * 2
+
+        self.conv111 = nn.Conv1d(
+            self.hidden_neural_number * n_input_channel,
+            self.hidden_neural_number * n_input_channel * coef_hidden,
+            1,
+            groups=self.hidden_neural_number * n_input_channel,
+        )
+        self.conv111_last = nn.Conv1d(
+            self.hidden_neural_number * n_input_channel * coef_hidden, n_output, 1
+        )
+
+    def forward(self, x):
+        x = torch.cat([x, 1 - x], dim=1)
+        x = torch.relu(self.conv111(x))
+        x = self.conv111_last(x)
+        return x
+
+
+class SplittedSmallNet(nn.Module):
+    def __init__(self, coef_hidden=4, n_input_channel=2):
+        super().__init__()
+        self.hidden_neural_number = mat_sym.shape[1]
+
+        self.conv111_0 = nn.Conv1d(
+            self.hidden_neural_number * n_input_channel,
+            self.hidden_neural_number * n_input_channel * coef_hidden,
+            1,
+            groups=self.hidden_neural_number * n_input_channel,
+        )
+        self.conv111_1 = nn.Conv1d(
+            self.hidden_neural_number * n_input_channel,
+            self.hidden_neural_number * n_input_channel * coef_hidden,
+            1,
+            groups=self.hidden_neural_number * n_input_channel,
+        )
+
+        self.conv111_last_0 = nn.Conv1d(
+            self.hidden_neural_number * n_input_channel * coef_hidden, 1, 1
+        )
+        self.conv111_last_1 = nn.Conv1d(
+            self.hidden_neural_number * n_input_channel * coef_hidden, 1, 1
+        )
+
+    def forward(self, x):
+        x_0 = torch.relu(self.conv111_0(x))
+        x_0 = self.conv111_last_0(x_0)
+
+        x_1 = torch.relu(self.conv111_1(x))
+        x_1 = self.conv111_last_1(x_1)
+        return torch.cat([x_0, x_1], dim=1)

sudoku/symetries.py

@@ -0,0 +1,171 @@
+import numpy as np
+
+
+def generate_symetrie_reduction():
+    # chiffre, ligne, colonne, chiffre ligne colonne
+    a1_array = np.zeros((9, 9, 9, 9, 9, 9), dtype=float)
+    a2_array = np.zeros((9, 9, 9, 9, 9, 9), dtype=float)
+    a3_array = np.zeros((9, 9, 9, 9, 9, 9), dtype=float)
+    a4_array = np.zeros((9, 9, 9, 9, 9, 9), dtype=float)
+    a5_array = np.zeros((9, 9, 9, 9, 9, 9), dtype=float)
+    a6_array = np.zeros((9, 9, 9, 9, 9, 9), dtype=float)
+    a7_array = np.zeros((9, 9, 9, 9, 9, 9), dtype=float)
+    a8_array = np.zeros((9, 9, 9, 9, 9, 9), dtype=float)
+
+    b1_array = np.zeros((9, 9, 9, 9, 9, 9), dtype=float)
+    b2_array = np.zeros((9, 9, 9, 9, 9, 9), dtype=float)
+    b3_array = np.zeros((9, 9, 9, 9, 9, 9), dtype=float)
+    b4_array = np.zeros((9, 9, 9, 9, 9, 9), dtype=float)
+    b5_array = np.zeros((9, 9, 9, 9, 9, 9), dtype=float)
+    b6_array = np.zeros((9, 9, 9, 9, 9, 9), dtype=float)
+    b7_array = np.zeros((9, 9, 9, 9, 9, 9), dtype=float)
+    b8_array = np.zeros((9, 9, 9, 9, 9, 9), dtype=float)
+
+    for i in range(9):
+        for j in range(9):
+            for n in range(9):
+                a1_array[i, j, n, i, j, n] = 1.0
+
+                a2_array[autre_chiffre(i), j, n, i, j, n] = 1.0 / 8.0
+
+                a3_array[i, autre_chiffre(j), n, i, j, n] = 1.0 / 8.0
+
+                a4_array[np.ix_([i], meme_bloc(j), [n], [i], [j], [n])] = 1.0 / 8.0
+                a4_array[np.ix_(meme_bloc(i), [j], [n], [i], [j], [n])] = 1.0 / 8.0
+                a4_array[np.ix_(meme_bloc(i), meme_bloc(j), [n], [i], [j], [n])] = (
+                    1.0 / 8.0
+                )
+
+                a5_array[np.ix_(meme_bloc(i), autre_bloc_b(j), [n], [i], [j], [n])] = (
+                    1.0 / 6.0
+                )
+
+                a6_array[np.ix_(meme_bloc(i), autre_bloc_h(j), [n], [i], [j], [n])] = (
+                    1.0 / 6.0
+                )
+
+                a7_array[np.ix_(autre_bloc_b(i), meme_bloc(j), [n], [i], [j], [n])] = (
+                    1.0 / 6.0
+                )
+
+                a8_array[np.ix_(autre_bloc_h(i), meme_bloc(j), [n], [i], [j], [n])] = (
+                    1.0 / 6.0
+                )
+
+                b1_array[i, j, n, i, j, autre_chiffre(n)] = 1.0 / 8.0
+
+                b2_array[
+                    np.ix_(autre_chiffre(i), [j], autre_chiffre(n), [i], [j], [n])
+                ] = (1.0 / 64.0)
+
+                b3_array[
+                    np.ix_([i], autre_chiffre(j), autre_chiffre(n), [i], [j], [n])
+                ] = (1.0 / 64.0)
+
+                b4_array[np.ix_([i], meme_bloc(j), autre_chiffre(n), [i], [j], [n])] = (
+                    1.0 / 64.0
+                )
+                b4_array[np.ix_(meme_bloc(i), [j], autre_chiffre(n), [i], [j], [n])] = (
+                    1.0 / 64.0
+                )
+                b4_array[
+                    np.ix_(meme_bloc(i), meme_bloc(j), autre_chiffre(n), [i], [j], [n])
+                ] = (1.0 / 64.0)
+
+                b5_array[
+                    np.ix_(
+                        meme_bloc(i), autre_bloc_b(j), autre_chiffre(n), [i], [j], [n]
+                    )
+                ] = (1.0 / 48.0)
+
+                b6_array[
+                    np.ix_(
+                        meme_bloc(i), autre_bloc_h(j), autre_chiffre(n), [i], [j], [n]
+                    )
+                ] = (1.0 / 48.0)
+
+                b7_array[
+                    np.ix_(
+                        autre_bloc_b(i), meme_bloc(j), autre_chiffre(n), [i], [j], [n]
+                    )
+                ] = (1.0 / 48.0)
+
+                b8_array[
+                    np.ix_(
+                        autre_bloc_h(i), meme_bloc(j), autre_chiffre(n), [i], [j], [n]
+                    )
+                ] = (1.0 / 48.0)
+
+    list_arr = [
+        a1_array,
+        a2_array,
+        a3_array,
+        a4_array,
+        a5_array,
+        a6_array,
+        a7_array,
+        a8_array,
+        b1_array,
+        b2_array,
+        b3_array,
+        b4_array,
+        b5_array,
+        b6_array,
+        b7_array,
+        b8_array,
+    ]
+    full_sym = np.zeros((9, 9, 9, 9, 9, 9, len(list_arr)))
+    for i, arr in enumerate(list_arr):
+        full_sym[:, :, :, :, :, :, i] = arr
+
+    full_sym = full_sym.reshape(9 * 9 * 9, 9 * 9 * 9, len(list_arr))
+    return full_sym
+
+
+def meme_bloc(i):
+    bloc_start = 3 * (i // 3)
+    bloc_end = bloc_start + 2
+    if i == bloc_start:
+        return [bloc_start + 1, bloc_end]
+    elif i == bloc_end:
+        return [bloc_start, bloc_start + 1]
+    else:
+        return [bloc_start, bloc_end]
+
+
+def autre_bloc(i):
+    bloc_start = 3 * (i // 3)
+    all_in_bloc = [bloc_start, bloc_start + 1, bloc_start + 2]
+    list_to_return = []
+    for n in range(9):
+        if not (n in all_in_bloc):
+            list_to_return.append(n)
+    return list_to_return
+
+
+def autre_bloc_b(i):
+    bloc_start = 3 * (i // 3)
+    all_in_bloc = [bloc_start, bloc_start + 1, bloc_start + 2]
+    list_to_return = []
+    for n in range(9):
+        if not (n in all_in_bloc):
+            list_to_return.append(n)
+    return list_to_return[:3]
+
+
+def autre_bloc_h(i):
+    bloc_start = 3 * (i // 3)
+    all_in_bloc = [bloc_start, bloc_start + 1, bloc_start + 2]
+    list_to_return = []
+    for n in range(9):
+        if not (n in all_in_bloc):
+            list_to_return.append(n)
+    return list_to_return[3:]
+
+
+def autre_chiffre(i):
+    return [n for n in range(9) if n != i]
+
+
+mat_sym = generate_symetrie_reduction()
+mat_sym = np.moveaxis(mat_sym, (0, 1, 2), (2, 0, 1))

sudoku/train.py

@@ -0,0 +1,885 @@
+from math import copysign
+import torch
+from torch import nn
+from torch.optim.lr_scheduler import ReduceLROnPlateau
+import pytorch_lightning as pl
+from sudoku.models import SmallNetBis, SymPreprocess
+import torch.nn.functional as F
+
+from sudoku.buffer import BufferArray, Buffer
+from sudoku.trial_grid import TrialGrid
+
+from copy import deepcopy
+
+
+class SudokuLightning(pl.LightningModule):
+    def __init__(
+        self,
+        lr=0.1,
+        margin=0.1,  # th marge?
+        coef_0=10,
+        nets_number=6,
+        nets_training_number=1,
+        batch_size=32,
+    ):
+        super().__init__()
+        self.nets_number = nets_number
+        self.batch_size = batch_size
+        self.nets_training_number = nets_training_number
+        # self.nets=[SmallNetBis() for _ in range(self.nets_number)]
+        self.nets = nn.ModuleList([SmallNetBis() for _ in range(self.nets_number)])
+        self.buffer = BufferArray(self.nets_number, self.batch_size)
+        self.sym_preprocess = SymPreprocess()
+        pos_weight = torch.ones((2, 9 * 9 * 9))
+        pos_weight[0, :] = 1.0 / 8.0
+        pos_weight[1, :] = 1.0
+        pos_weight /= coef_0
+        weight = torch.ones((2, 9 * 9 * 9))
+        weight[0, :] = 8.0
+        weight[1, :] = 1.0
+        weight *= coef_0
+
+        self.bcewll = nn.BCEWithLogitsLoss(
+            pos_weight=pos_weight, weight=weight, reduce=False
+        )
+        self.lr = lr
+        # self.auroc = AUROC(task='binary')
+
+        self.margin = margin
+        self.th_epsilon = margin * 0.01
+        self.threshold_pres = torch.tensor([-10.0 for _ in range(nets_number)])
+        self.threshold_abs = torch.tensor([-10.0 for _ in range(nets_number)])
+
+        self.automatic_optimization = False
+        self.reset_threshold_on_validation = True
+
+    def configure_optimizers(self):
+        # no need config scheduler -> manual optimisation
+        optimizers = []
+        for net in self.nets:
+            opti = torch.optim.Adam(net.parameters(), lr=self.lr)
+            optimizers.append(
+                {
+                    "optimizer": opti,
+                    "lr_scheduler": ReduceLROnPlateau(opti, "min"),
+                }
+            )
+        return optimizers
+
+    # def configure_optimizers(self):
+    #     optimizer1 = Adam(...)
+    #     optimizer2 = SGD(...)
+    #     scheduler1 = ReduceLROnPlateau(optimizer1, ...)
+    #     scheduler2 = LambdaLR(optimizer2, ...)
+    #     return (
+    #         {
+    #             "optimizer": optimizer1,
+    #             "lr_scheduler": {
+    #                 "scheduler": scheduler1,
+    #                 "monitor": "metric_to_track",
+    #             },
+    #         },
+    #         {"optimizer": optimizer2, "lr_scheduler": scheduler2},
+    #     )
+
+    #     lr_scheduler_config = {
+    #     # REQUIRED: The scheduler instance
+    #     "scheduler": lr_scheduler,
+    #     # The unit of the scheduler's step size, could also be 'step'.
+    #     # 'epoch' updates the scheduler on epoch end whereas 'step'
+    #     # updates it after a optimizer update.
+    #     "interval": "epoch",
+    #     # How many epochs/steps should pass between calls to
+    #     # `scheduler.step()`. 1 corresponds to updating the learning
+    #     # rate after every epoch/step.
+    #     "frequency": 1,
+    #     # Metric to to monitor for schedulers like `ReduceLROnPlateau`
+    #     "monitor": "val_loss",
+    #     # If set to `True`, will enforce that the value specified 'monitor'
+    #     # is available when the scheduler is updated, thus stopping
+    #     # training if not found. If set to `False`, it will only produce a warning
+    #     "strict": True,
+    #     # If using the `LearningRateMonitor` callback to monitor the
+    #     # learning rate progress, this keyword can be used to specify
+    #     # a custom logged name
+    #     "name": None,
+    # }
+    # lr_scheduler_config = {'scheduler: lr_sch, interval: epoch, frequency: 1, monitor: 'val_loss'}
+
+    def forward_layer(self, x, idx=0):
+        x = self.sym_preprocess.forward(x)
+        return self.nets[idx](x)
+
+    def forward(self, x):
+        for idx in range(self.nets_number):
+            output = self.forward_layer(x, idx)
+            new_X = self.compute_new_X(output, x, idx, None, train=False)
+            improved_mask = ((new_X == 1) & (x == 0)).any(dim=1).any(dim=1)
+            if improved_mask.sum() > 0:
+                return idx, new_X
+        return idx, new_X
+
+    def predict_from_net(self, x, net, th_abs, th_pres):
+        x = self.sym_preprocess.forward(x)
+        x = net(x)
+        new_x = torch.empty(x.shape, device=x.device)
+        new_x[:, 0] = (x[:, 0] > th_abs).float()
+        new_x[:, 1] = (x[:, 1] > th_pres).float()
+        return new_x
+
+    @staticmethod
+    def mask_uncomplete(x, y):
+        mask_uncomplete = x.reshape(-1, 2, 9, 9, 9).sum(-1) < torch.tensor((8, 1)).to(
+            x
+        ).reshape(1, 2, 1, 1)
+        mask_uncomplete = mask_uncomplete.reshape(-1, 2, 9, 9, 1)
+        mask = ((x == 0).reshape(-1, 2, 9, 9, 9) * mask_uncomplete).reshape(
+            -1, 2, 9**3
+        )
+        mask = mask.float()
+        return mask
+
+    def computing_loss(self, x, y, output):
+        loss = self.bcewll(output, y)
+        mask = self.mask_uncomplete(x, y)
+        loss = (loss * mask).sum()
+
+        return loss
+
+    def training_step(self, batch, batch_idx):
+        self.log(
+            "train_grid_count",
+            batch[0].shape[0],
+            reduce_fx=torch.sum,
+            on_epoch=True,
+            on_step=False,
+        )
+
+        self.layer_training_step(0, batch)
+        while True:
+            idx, batch = self.buffer.get_batch()
+            if batch is None:
+                break
+            # check if the train should be done by comparing lr from sch = self.lr_schedulers()
+            # if self.lr != sch[idx].get_last_lr():
+            self.layer_training_step(idx, batch)
+
+    def validation_step(self, batch, batch_idx):
+        self.layer_training_step(0, batch, train=False)
+        while True:
+            idx, batch = self.buffer.get_batch()
+            if batch is None:
+                break
+            # check if the train should be done by comparing lr from sch = self.lr_schedulers()
+            # if self.lr != sch[idx].get_last_lr():
+            self.layer_training_step(idx, batch, train=False)
+
+    def layer_training_step(
+        self, idx, batch, train=True
+    ):  # to rename to layer_training_step
+        x, y = batch
+
+        prefix = "train" if train else "val"
+        self.log(
+            f"{prefix}_grid_count_{idx}",
+            batch[0].shape[0],
+            reduce_fx=torch.sum,
+            on_epoch=True,
+            on_step=False,
+        )
+
+        output = self.forward_layer(x, idx)
+        loss = self.computing_loss(x, y, output)
+        if train:
+
+            opt = self.optimizers()#[idx]
+            if isinstance(opt, list):
+                opt=opt[idx]
+            opt.zero_grad()
+            self.manual_backward(loss)
+            opt.step()
+
+        loss_0 = F.binary_cross_entropy_with_logits(output[:, [0], :], y[:, [0], :])
+        loss_1 = F.binary_cross_entropy_with_logits(output[:, [1], :], y[:, [1], :])
+        self.log_dict(
+            {f"{prefix}_loss_pos": loss_1, f"{prefix}_loss_neg": loss_0}, on_epoch=True
+        )
+
+        # accuracy_1 = torch.mean(torch.eq(transform_to_number_1(output), transform_to_number_1(x)).type(torch.float))
+        # accuracy_0 = torch.mean(torch.eq(transform_to_number_0(output), transform_to_number_0(x)).type(torch.float))
+        # self.log_dict({'accuracy_1': accuracy_1, 'accuracy_0': accuracy_0}, on_epoch=True)
+        self.log(f"{prefix}_loss_{idx}", loss)
+        # add a count log on (X and x == y)
+
+        new_X = self.compute_new_X(output, x, idx, y, train=train)
+        solved_mask = (new_X == y).all(dim=1).all(dim=1)
+        new_X = new_X[~solved_mask]
+        x = x[~solved_mask]
+        y = y[~solved_mask]
+        self.log(
+            f"{prefix}_resolved_grid_count",
+            solved_mask.sum(),
+            on_epoch=True,
+            on_step=False,
+            reduce_fx=torch.sum,
+        )
+        mask_no_improve = new_X.sum(dim=(1, 2)) <= x.sum(dim=(1, 2))
+        self.log(
+            f"{prefix}_improved_grid_count_{idx}",
+            (~mask_no_improve).sum(),
+            on_epoch=True,
+            on_step=False,
+            reduce_fx=torch.sum,
+        )
+        # store_new_x
+        # TODO keep the log in this method
+        # loss per epoch per model boost layer
+        # number of error per epoch model boost layer
+        # number of resolved puzzles per epochs
+        # threshold per epochs per model layer
+        # number of sudoku grid
+
+        # number of filled digits per model boost layer per epoch for both pis ans abs
+        # add parameter reduce_fx=torch.sum() to numbers
+        # th -> on_epoch=False
+
+        self.store_new_x(idx, new_X, x, y)
+
+    def store_new_x(self, idx, new_X, x, y):
+        mask_improve = new_X.sum(dim=(1, 2)) > x.sum(dim=(1, 2))
+        self.buffer.append(
+            idx + 1, (new_X[~mask_improve].clone(), y[~mask_improve].clone())
+        )
+        self.buffer.append(0, (new_X[mask_improve].clone(), y[mask_improve].clone()))
+        # TODO if improve on no improvments -> add one digit from y to new_X and ad it to idx=0
+
+    def compute_new_X(self, output, x, idx, y=None, train=True, mask_adapt_th=None):
+        # y could be None
+        prefix = "train" if train else "val"
+        new_X = torch.empty(output.shape, device=output.device)
+        # we could try to make evolv threshold here
+        if y is not None:
+            # max_th_abs = (
+            #     output[:, 0][(x[:, 0] == 0) & (y[:, 0] == 0)].max().item()
+            #     + self.th_epsilon
+            # )
+            max_th_abs = output[:, 0][(y[:, 0] == 0)].max().item() + self.th_epsilon
+            max_th_pres = (
+                output[:, 1][(x[:, 1] == 0) & (y[:, 1] == 0)].max().item()
+                + self.th_epsilon
+            )
+            if mask_adapt_th is None or (mask_adapt_th.sum()>0):
+                if mask_adapt_th is not None and (mask_adapt_th.sum()>0):
+                    max_th_abs = output[mask_adapt_th, 0][(y[mask_adapt_th, 0] == 0)].max().item() + self.th_epsilon
+                    max_th_pres = (
+                        output[mask_adapt_th, 1][(x[mask_adapt_th, 1] == 0) & (y[mask_adapt_th, 1] == 0)].max().item()
+                        + self.th_epsilon
+                    )
+                self.threshold_abs[idx] = max(max_th_abs, self.threshold_abs[idx])
+                self.threshold_pres[idx] = max(max_th_pres, self.threshold_pres[idx])
+                self.log_dict(
+                    {
+                        f"{prefix}_th_abs_{idx}": self.threshold_abs[idx],
+                        f"{prefix}_th_pres_{idx}": self.threshold_pres[idx],
+                    },
+                    on_step=True,
+                )
+                if not train:
+                    self.threshold_abs_compute[idx] = max(
+                        max_th_abs + self.margin, self.threshold_abs_compute[idx]
+                    )
+                    self.threshold_pres_compute[idx] = max(
+                        max_th_pres + self.margin, self.threshold_pres_compute[idx]
+                    )
+
+        if self.training:
+            new_X[:, 0] = (output[:, 0].detach() > self.threshold_abs[idx]).float()
+            new_X[:, 1] = (output[:, 1].detach() > self.threshold_pres[idx]).float()
+        else:
+            new_X[:, 0] = (output[:, 0].detach() > self.threshold_abs[idx]).float()
+            new_X[:, 1] = (output[:, 1].detach() > self.threshold_pres[idx]).float()
+        new_X[x.detach() == 1] = 1
+        if y is not None:
+            self.log(
+                f"{prefix}_count_error_grid_{idx}",
+                ((new_X == 1) & (y == 0)).any(dim=1).any(dim=1).sum(),
+                on_epoch=True,
+                on_step=False,
+                reduce_fx=torch.sum,
+            )
+            if mask_adapt_th is None:
+                new_X[y.detach() == 0] = 0 # do not remove the error!!!!!!
+            else:
+                y_bis = y.detach().clone()
+                y_bis[~mask_adapt_th]=1
+                new_X[y_bis==0] = 0
+        return new_X
+
+    # TODO add idx stuff (one lr scheduler per net)
+    # def on_train_epoch_end(self):
+    #     sch = self.lr_schedulers()
+
+    #     # If the selected scheduler is a ReduceLROnPlateau scheduler.
+    #     if isinstance(sch, torch.optim.lr_scheduler.ReduceLROnPlateau):
+    #         sch.step(self.trainer.callback_metrics["loss"])
+
+    def on_validation_epoch_start(self) -> None:
+        if self.reset_threshold_on_validation:
+            self.threshold_abs_compute = torch.tensor(
+                [-10.0 for _ in range(self.nets_number)]
+            )
+            self.threshold_pres_compute = torch.tensor(
+                [-10.0 for _ in range(self.nets_number)]
+            )
+        else:
+            self.threshold_abs_compute = self.threshold_abs
+            self.threshold_pres_compute = self.threshold_pres
+            
+        self.buffer = BufferArray(self.nets_number, self.batch_size)
+
+    def on_train_epoch_start(self) -> None:
+        self.buffer = BufferArray(self.nets_number, self.batch_size)
+        return super().on_train_epoch_start()
+
+    def on_validation_epoch_end(self):
+        # tensorboard = self.logger.experiment
+        self.threshold_abs = self.threshold_abs_compute
+        self.threshold_pres = self.threshold_pres_compute
+
+        schs = self.lr_schedulers()
+        if not isinstance(schs, list):
+            schs=[schs]
+        for idx, sch in enumerate(schs):
+            # sch.step(self.validation.callback_metrics["val_loss_{idx}"])
+            try:
+                sch.step(self.trainer.callback_metrics[f"val_loss_{idx}"])
+            except:
+                # print(f"val_loss_{idx} not found")
+                pass
+            # sch.step(self.trainer.callback_metrics["val_loss_"])
+
+    def on_save_checkpoint(self, checkpoint) -> None:
+        "Objects to include in checkpoint file"
+        checkpoint["ths_abs"] = self.threshold_abs
+        checkpoint["ths_pres"] = self.threshold_pres
+
+    def on_load_checkpoint(self, checkpoint) -> None:
+        "Objects to retrieve from checkpoint file"
+        self.threshold_abs = checkpoint["ths_abs"]
+        self.threshold_pres = checkpoint["ths_pres"]
+        self.nets = nn.ModuleList([SmallNetBis() for _ in self.threshold_abs])
+
+    def validate_grids(self, x) -> "torch.tensor":
+        return ~(
+            (self.sym_preprocess(x)[:, 17].max(dim=1).values > (1 / 8))
+            | (self.sym_preprocess(x)[:, 18].max(dim=1).values > (1 / 8))
+            | (self.sym_preprocess(x)[:, 19].max(dim=1).values > (1 / 8))
+        )
+
+    # steps to trial error
+    #  - get stops
+    #    - choose a number -> store it
+    #  - process to get either a new stop either a a validation grid fail
+    # if validation grid fail back propagate
+    #  else choose a number
+
+    # add counter to each grid,
+    # add id to each grid id=batch_id + position
+    # add validation
+
+    # if non improvment stop ->
+    #   - check if id already exist, if true add non improve counter
+    #     if non improve counter = 2 -> add grid to trial_error_model_buffer with 1000 step target.
+    #   - store the grid to trial_error_model deep search dict
+    #   - create two grids with counter to 0, same id
+    #     add them in the buffer
+
+    # - when validation fail ->
+    #    - check if id already exist
+    #       if true: add grid to trial_error_model with the counter
+    #       if false: raise error
+
+# TODO adapt training to something softer
+# 
+class SudokuTrialErrorLightning(SudokuLightning):
+    def __init__(self, **kwargs):
+        super().__init__(**kwargs)
+        self.deep_backtrack_regressor = SmallNetBis(n_output=1)
+        self.trial_error_buffer = Buffer(self.batch_size)
+        self.trial_grids = [None]
+        # schema:
+        # [
+        #   idx:
+        #    "tried_pos": [
+        #   ]
+        #   "pos": pos
+        #   "no_improve_counter": 0
+        # ]
+        #
+        # self.tracking_grid = []
+
+    def copy_from_model(self, model):
+        self.nets = model.nets
+        self.threshold_pres = model.threshold_pres
+        self.threshold_abs = model.threshold_abs
+
+    def reg(self, x):
+        x_reg = self.sym_preprocess.forward(x)
+        x_reg = self.deep_backtrack_regressor(x_reg)
+        return torch.softmax(x_reg, dim=1)
+
+    def configure_optimizers(self):
+        # no need config scheduler -> manual optimisation
+        # optimizers = [torch.optim.Adam(net.parameters(), lr=self.lr) for net in self.nets]
+        optimizers = []
+        for net in self.nets:
+            opti = torch.optim.Adam(net.parameters(), lr=self.lr)
+            optimizers.append(
+                {
+                    "optimizer": opti,
+                    "lr_scheduler": ReduceLROnPlateau(opti, "min"),
+                }
+            )
+        optimizers.append(
+            {
+                'optimizer': torch.optim.Adam(self.deep_backtrack_regressor.parameters(), lr=self.lr), 
+                "lr_scheduler": ReduceLROnPlateau(opti, "min"),
+             }
+        )
+        return optimizers
+
+    def training_step(self, batch, batch_idx):
+        self.log(
+            "train_grid_count",
+            batch[0].shape[0],
+            reduce_fx=torch.sum,
+            on_epoch=True,
+            on_step=False,
+        )
+        x, y = batch
+        x_idx = torch.zeros(self.batch_size)  # if we are not on trial error x_idx=0
+        counters = torch.zeros(self.batch_size)
+
+        self.layer_training_step(0, (x, y, x_idx, counters))
+        idx_while=0
+        while True:
+            idx_while+=1
+            if idx_while ==10000:
+                print('a while')
+            idx, batch = self.buffer.get_batch()
+            if batch is None:
+                break
+            # check if the train should be done by comparing lr from sch = self.lr_schedulers()
+            # if self.lr != sch[idx].get_last_lr():
+            self.layer_training_step(idx, batch)
+
+        while True:
+            trial_error_batch = self.trial_error_buffer.get_batch()
+            if trial_error_batch is None:
+                break
+            self.trial_error_training_step(trial_error_batch)
+            
+    def validation_step(self, batch, batch_idx):
+        x, y = batch
+        x_idx = torch.zeros(x.shape[0], dtype=torch.long)  # if we are not on trial error x_idx=0
+        counters = torch.zeros(x.shape[0])
+
+        self.layer_training_step(0, (x, y, x_idx, counters), train=False)
+        while True:
+            idx, batch = self.buffer.get_batch()
+            if batch is None:
+                break
+            # check if the train should be done by comparing lr from sch = self.lr_schedulers()
+            # if self.lr != sch[idx].get_last_lr():
+            self.layer_training_step(idx, batch, train=False)
+
+        while True:
+            trial_error_batch = self.trial_error_buffer.get_batch()
+            if trial_error_batch is None:
+                break
+            self.trial_error_training_step(trial_error_batch, train=False)
+
+
+    def layer_training_step(
+        self, idx, batch, train=True
+    ):  # to rename to layer_training_step
+        x, y, x_idx, counters = batch
+
+        prefix = "train" if train else "val"
+        self.log(
+            f"{prefix}_grid_count_{idx}",
+            batch[0].shape[0],
+            reduce_fx=torch.sum,
+            on_epoch=True,
+            on_step=False,
+        )
+        output = self.forward_layer(x, idx)
+        loss = self.computing_loss(x[x_idx==0], y[x_idx==0], output[x_idx==0])
+        if train:
+            pass
+            opt = self.optimizers()[idx]
+            opt.zero_grad()
+            self.manual_backward(loss)
+            opt.step()
+
+        loss_0 = F.binary_cross_entropy_with_logits(output[:, [0], :], y[:, [0], :])
+        loss_1 = F.binary_cross_entropy_with_logits(output[:, [1], :], y[:, [1], :])
+        self.log_dict(
+            {f"{prefix}_loss_pos": loss_1, f"{prefix}_loss_neg": loss_0}, on_epoch=True
+        )
+        self.log(f"{prefix}_loss_{idx}", loss)
+
+        mask_bad_x = ((x==1)&(y==0)).any(dim=1).any(dim=1)
+        new_X = self.compute_new_X(output, x, idx, y, train=train, mask_adapt_th=(~mask_bad_x))
+        solved_mask = (new_X == y).all(dim=1).all(dim=1)
+
+        new_X = new_X[~solved_mask]
+        x = x[~solved_mask]
+        y = y[~solved_mask]
+        x_idx = x_idx[~solved_mask]
+        counters = counters[~solved_mask]
+        
+        self.log(
+            f"{prefix}_resolved_grid_count",
+            solved_mask.sum(),
+            on_epoch=True,
+            on_step=False,
+            reduce_fx=torch.sum,
+        )
+        mask_no_improve = new_X.sum(dim=(1, 2)) <= x.sum(dim=(1, 2))
+        self.log(
+            f"{prefix}_improved_grid_count_{idx}",
+            (~mask_no_improve).sum(),
+            on_epoch=True,
+            on_step=False,
+            reduce_fx=torch.sum,
+        )
+        # self.store_new_x(idx, new_X, x, y) # TODO create another function (need to increment counter and validate)
+        self.process_validation(idx, new_X, x, y, x_idx, counters)
+
+    def process_validation(self, idx, new_X, x, y, x_idx, counters):
+        new_X = self.redresse_new_X(new_X,y,x)
+        mask_validated = self.validate_grids(new_X)
+        # mask_not_validated = (~self.validate_grids(new_X)) & ((x==0)&(y==1)).any(dim=(1,2))
+        mask_improve = (new_X.sum(dim=(1, 2)) > x.sum(dim=(1, 2))) & mask_validated
+        mask_not_improved = (new_X.sum(dim=(1, 2)) == x.sum(dim=(1, 2))) & mask_validated
+
+
+        for i, (failed_idx, failed_counter, s_new_X, s_y) in enumerate(zip(
+            x_idx[~mask_validated],
+            counters[~mask_validated],
+            new_X[~mask_validated],
+            y[~mask_validated],
+        )):
+            # when we find failed:
+            #  - we store good grid to continue the process # /!\ it is not necessary, the second half will continue to process.
+            #  - we store the initial grid with the score (to traine the regressor)
+            if failed_idx == 0:
+                self.failed_batch = (x[~mask_validated][i], s_y)
+                raise ValueError("validation error on no trial-error grid")
+            if not ((x[~mask_validated][i]==0)&(s_y==1)).any():
+                raise ValueError()
+            is_pos = copysign(1, failed_idx)==1
+            trial_grid: TrialGrid = self.trial_grids[int(abs(failed_idx))]
+            if is_pos:
+                trial_grid.pos_result = 'fail'
+            else:
+                trial_grid.neg_result = 'fail'
+            
+            self.process_search_store_grid(int(abs(failed_idx)), trial_grid, s_y)
+
+        if idx == self.nets_number - 1:
+            for no_improved_idx, s_new_X, s_y in zip(
+                x_idx[mask_not_improved], new_X[mask_not_improved], y[mask_not_improved]
+            ):
+                if no_improved_idx == 0:
+                    self.search_trial_buffer_trials(s_new_X, s_y)
+                    continue
+
+                is_pos = copysign(1, no_improved_idx)==1
+                trial_grid: TrialGrid = self.trial_grids[int(abs(no_improved_idx.item()))]
+                if is_pos:
+                    trial_grid.pos_result = 'no_improved'
+                else:
+                    trial_grid.neg_result = 'no_improved'
+                assert s_new_X.sum()> trial_grid.initial_grid.sum()
+                
+                self.process_search_store_grid(int(abs(no_improved_idx)), trial_grid, s_y)
+
+        self.buffer.append(
+            idx + 1,
+            (new_X[mask_not_improved].clone(), y[mask_not_improved].clone(), x_idx[mask_not_improved].clone(), counters[mask_not_improved].clone()),
+        )
+        # assert mask_improve.sum()>0
+        if ((new_X[mask_improve & (x_idx.to(self.device)==0)]==1) & (y[mask_improve & (x_idx.to(self.device)==0)]==0)).any():
+            self.failed_batch=(x[mask_improve & (x_idx.to(self.device)==0)],y[mask_improve & (x_idx.to(self.device)==0)] )
+            raise ValueError()
+        self.buffer.append(
+            0,
+            (new_X[mask_improve].clone(), y[mask_improve].clone(), x_idx[mask_improve].clone(), counters[mask_improve].clone() + 1),
+        )
+
+    def process_search_store_grid(self, idx, trial_grid: TrialGrid, s_y):
+        """_summary_
+        if score is 1:
+            great 
+                if fail -> the second one should continue (it has his id if it stopped)
+                    so do nothing
+                if no_improved ->
+                    trial_error and reset trial_error_grid
+        if score is not None -> store the new grid in the trial_error_buffer
+        
+        if score is -1 => also search_trial and store.
+        if store is 1 =>
+            if both result are here:
+                get the no_improved -> search_trial and store on a new grid
+                if one complete grid -> set grid place to None
+            else: wait
+        
+        
+        if score is None:
+          if grid_idx==-1 or oposite_grid failed:
+              we create a new_idx, and store stuff.
+          else:
+              we store the grid (in case the second grid fail)
+          we increment the non_improvement counter
+
+          if non_improvement counter = 2:
+              we add the initial grid to the search training buffer
+              we process the search training engine to find another grid postion
+        else:
+          we add the initial grid to the search training buffer
+          if a non improved grid is store we create a new_idx and store stuf.
+
+        Args:
+            grid_idx (_type_): _description_
+            score (_type_): _description_
+            s_new_X (_type_): _description_
+            s_y (_type_): _description_
+        """
+        score = trial_grid.score()
+        if score is None:
+            self.trial_grids[idx]=trial_grid
+            return
+        # add grid to buffer (initial_grid, score)
+        self.trial_error_buffer.append((
+            trial_grid.initial_grid.view(-1,2,729),
+            torch.tensor([score,],dtype=torch.float).to(self.device),
+            torch.tensor([trial_grid.row_col_digit_position,], dtype=torch.long).to(self.device),
+        ))
+        
+        # find the no_improve_grid ~and search_trial~ and add it to buffer
+        if trial_grid.neg_result == 'no_improved':
+            if trial_grid.pos_result == 'no_improved':
+                trial_grid.tried_grid.append(trial_grid.row_col_digit_position)
+                trial_grid.neg_result= None
+                trial_grid.pos_result= None
+                self.trial_grids[idx] = trial_grid
+                self.search_trial_buffer_trials(None, s_y, idx)
+                # new trial with same idx
+                return
+            # add to buffer neg grid     
+            # we get back the initial grid
+            # set the correct row col digit
+            # add it the buffer
+            # set trial_grids to None
+            grid_neg = deepcopy(trial_grid.initial_grid)
+            grid_neg[0,trial_grid.row_col_digit_position] = 1
+            if ((grid_neg==1) & (s_y==0)).any():
+                raise ValueError()
+            self.buffer.append(
+                0,
+                (
+                    grid_neg.view(-1,2,729),
+                    s_y.clone().view(-1,2,729),
+                    torch.tensor([0]),
+                    torch.tensor([0]),
+                )
+            )
+            self.trial_grids[idx] = None
+            return
+        if trial_grid.pos_result == 'no_improved':
+            grid_pos = deepcopy(trial_grid.initial_grid)
+            grid_pos[1,trial_grid.row_col_digit_position] = 1
+            if ((grid_pos==1) & (s_y==0)).any():
+                raise ValueError()
+            self.buffer.append(
+                0,
+                (
+                    grid_pos.view(-1,2,729),
+                    s_y.clone().view(-1,2,729),
+                    torch.tensor([0]),
+                    torch.tensor([0]),
+                )
+            )
+            self.trial_grids[idx] = None
+            # add to buffer pos grid
+            return 
+
+        
+        # if complete: replace grid by none.
+        if "complete" in [trial_grid.neg_result, trial_grid.pos_result]:
+            self.trial_grids[idx]=None
+
+    # def store_new_trial_error_grid(self, new_X, y):
+    #     """build a new idx add the grid in the tracking stuff
+    #         and add grid in the buffer
+
+    #     Args:
+    #         new_X (_type_): _description_
+    #         y (_type_): _description_
+    #     """
+    #     ...
+
+    # def store_training_trail_search_batch(self, grid, score):
+    #     """store grid to train trial_search nn model
+
+    #     Args:
+    #         grid (_type_): _description_
+    #         score (_type_): _description_
+    #     """
+    #     ...
+
+    def search_trial(self, s_new_X, tried_pos):
+        """use the trail_search nn model to probe a new
+
+        Args:
+            s_new_X (_type_): _description_
+            tried_pos (_type_): _description_
+        """
+        mask_possibility = s_new_X.sum(dim=0)==0
+        for pos in tried_pos:
+            mask_possibility[pos]=False
+        if  mask_possibility.sum()==0:
+            print('mask_possible=0')
+            raise ValueError()
+            
+        with torch.no_grad():
+            x_reg = self.sym_preprocess.forward(s_new_X.view(1,2,-1))
+            output = self.deep_backtrack_regressor(x_reg)
+        # shape (1, 729)
+        # can be regression -> i want the smallest
+        # can be logistic regression -> i want the smallest
+        # if i do softmax -> i can add 1 to each tried pos
+        output = torch.softmax(output[0][0],dim=0)
+        # for pos in tried_pos:
+        #     output[pos]=1
+        output[~mask_possibility]+=1
+        return torch.argmin(output, dim=0).item()
+
+    def search_trial_buffer_trials(self, s_new_X, s_y, idx_trial_grids=None):
+        
+        if idx_trial_grids is None:
+            row_col_digit_trial = self.search_trial(s_new_X, [])
+            trial_grid = TrialGrid(s_new_X, row_col_digit_trial)
+            self.trial_grids.append(TrialGrid(s_new_X, row_col_digit_trial))
+            idx_trial_grids = len(self.trial_grids)-1
+        else:
+            trial_grid = self.trial_grids[idx_trial_grids]
+            s_new_X = trial_grid.initial_grid
+            row_col_digit_trial = self.search_trial(s_new_X, trial_grid.tried_grid)
+            trial_grid.row_col_digit_position = row_col_digit_trial
+            self.trial_grids[idx_trial_grids] = trial_grid
+            
+        # and we add both into buffer.
+        grid_pos = deepcopy(s_new_X)
+        grid_neg = deepcopy(s_new_X)
+        grid_pos[1,row_col_digit_trial] = 1
+        grid_neg[0,row_col_digit_trial] = 1
+        self.buffer.append(
+            0,
+            (
+                torch.stack([grid_pos,grid_neg], dim=0),
+                torch.stack([s_y.clone(),s_y.clone()], dim=0),
+                torch.tensor([idx_trial_grids, -idx_trial_grids]),
+                torch.tensor([0, 0]),
+            )
+        )
+
+    def trial_error_training_step(self, batch, train=True):
+        x, y, row_col_digit = batch
+        prefix = "train" if train else "val"
+        self.log(
+            f"{prefix}_grid_count_trial_error_training",
+            batch[0].shape[0],
+            reduce_fx=torch.sum,
+            on_epoch=True,
+            on_step=False,
+        )
+
+        
+        x_reg = self.sym_preprocess.forward(x)
+        output = self.deep_backtrack_regressor(x_reg)
+        loss = nn.functional.binary_cross_entropy_with_logits(output[[i for i in range(self.batch_size)], 0, row_col_digit], y, weight=None, size_average=None, reduce=None, reduction='mean', pos_weight=None) 
+        # # depending the distribution of the target, the weight could be different
+        # loss = binary (output[:,0,row_col_digit], y)
+        # loss = self.computing_loss(x, y, output)
+        if train:
+            opt = self.optimizers()[-1]
+            opt.zero_grad()
+            self.manual_backward(loss)
+            opt.step()
+            
+        self.log(f"{prefix}_loss_trial_error", loss)
+        self.log(f"{prefix}_loss_{self.nets_number}", loss)
+        self.log(f"{prefix}_y_pos_trial_error", y.sum())
+        self.log(f"{prefix}_y_neg_trial_eror", y.shape[0]-y.sum())
+
+    def predict(self, x):
+        """ return an improvement of x
+        
+        """
+        
+        idx, new_X = self.forward(x.view(-1,2,729))
+        if (new_X.sum()>x.sum()) or (new_X.sum()==729):
+            return new_X
+        else:
+            # call trial error until we find a solution
+            tried_position = []
+            while True:
+                pos = self.search_trial(x.view(2,729), tried_position)
+                tried_position.append(pos)
+                # creat pos neg tensor
+                grid_pos = deepcopy(x.view(2,729))
+                grid_neg = deepcopy(x.view(2,729))
+                grid_pos[1,pos] = 1
+                grid_neg[0,pos] = 1
+                X_tried = torch.stack([grid_neg, grid_pos], dim=0)
+                # process it 
+                while True:
+                    idx, new_X = self.forward(X_tried)
+                    mask_validated = self.validate_grids(new_X)
+                    if mask_validated.sum()<2:
+                        x[0, mask_validated, pos] = 1 # TODO check if it work
+                        return x
+                    if X_tried.sum()==new_X.sum():
+                        # if both stop to improve -> break it will tried an new pos
+                        break
+                    mask_complete = (X_tried.sum(dim=1)==729)# check if it works
+                    if mask_complete.sum()>0:
+                        x[0, mask_complete, pos] = 1
+                        return x
+                    X_tried = new_X
+                    # if one of X_tried is complete (weird but possible) -> return x with tried_position mask_complet set to 1 (cause we still want a step by step resolution)
+            
+    def on_validation_epoch_start(self) -> None:
+        # self.buffer = BufferArray(self.nets_number, self.batch_size)
+        self.trial_error_buffer = Buffer(self.batch_size)
+        self.trial_grids = [None]
+        return super().on_validation_epoch_start()
+        
+    def on_train_epoch_start(self) -> None:
+        self.trial_error_buffer = Buffer(self.batch_size)
+        self.trial_grids = [None]
+        return super().on_train_epoch_start()
+
+    def redresse_new_X(self, new_X,y,x):
+        mask_bad_x = ((x==1)&(y==0)).any(dim=1).any(dim=1)
+        y_bis = y.clone()
+        y_bis[mask_bad_x]=1
+        new_X[y_bis==0]=0        
+        return new_X
+
+
+# ADD threshold adjustment during prediction
+# or maybe validate? on it? bah oui!

sudoku/trial_grid.py

@@ -0,0 +1,19 @@
+
+
+class TrialGrid():
+  def __init__(self, initial_grid, row_col_digit_position):
+    self.initial_grid=initial_grid
+    self.row_col_digit_position = row_col_digit_position
+    self.neg_result=None
+    self.pos_result=None
+    self.tried_grid = []
+
+  def score(self):
+    # currently return 0 if find a good path else 1 (I'm too lazy)
+    if "fail" in [self.neg_result, self.pos_result]:
+      return 0
+    if "complete" in [self.neg_result, self.pos_result]:
+      return 0
+    if None in [self.neg_result, self.pos_result]:
+      return None
+    return 1

tensorboard.sh

@@ -0,0 +1 @@
+~/git/prose-dl/venv/bin/tensorboard --logdir lightning_logs/

tests/test_buffer.py

@@ -0,0 +1,73 @@
+import torch
+import pytest
+import numpy as np
+
+from sudoku.buffer import BufferArray
+
+
+# test fait to get batch
+def test_get_batch_on_empty():
+    ba = BufferArray(array_size=3, batch_size=2)
+    with pytest.raises(AssertionError):
+        ba.get_batch()
+
+
+def test_append_get_batch():
+    x = torch.tensor(
+        [
+            [[1, 2, 3], [4, 5, 6]],
+            [[11, 12, 13], [14, 15, 16]],
+        ]
+    )
+    y = torch.tensor(
+        [
+            1,
+            2,
+        ]
+    )
+    ba = BufferArray(array_size=3, batch_size=2)
+    ba.append(0, (x, y))
+    idx, batch = ba.get_batch()
+    assert batch[0].shape == x.shape
+    assert batch[1].shape == y.shape
+    assert idx == 0
+    assert len(ba.buffers[0][0]) == 0
+
+
+@pytest.mark.parametrize("idx", [0, 1])
+def test_append_get_batch_2(idx):
+    x = torch.tensor(
+        [
+            [[1, 2, 3], [4, 5, 6]],
+            [[11, 12, 13], [14, 15, 16]],
+        ]
+    )
+    y = torch.tensor(
+        [
+            1,
+            2,
+        ]
+    )
+    ba = BufferArray(array_size=3, batch_size=2)
+    ba.append(idx, (x, y))
+    ba.append(idx, (x, y))
+    batch_idx, batch = ba.get_batch()
+    assert batch[0].shape == x.shape
+    assert batch[1].shape == y.shape
+    assert idx == batch_idx
+    assert len(ba.buffers[idx][0]) == 2
+
+
+# test buffer array ->
+
+# test -> faire des scenarios
+# append
+# append
+# append
+# batch batch
+
+# verifier que chaque batch est bien séparer
+# que la limite de batch est bien respecté
+# que le sampling se fait.
+
+# que les assert pops

tests/test_helper.py

@@ -0,0 +1,16 @@
+from sudoku.helper import compute_loss
+import torch
+
+
+def test_compute_loss():
+    x = torch.zeros((3, 2, 729))
+    y = torch.zeros((3, 2, 729))
+    output = torch.zeros((3, 2, 729))
+    y[:, 0, 0] = 1
+    output[0, 0, 0] = 0.1
+    output[1, 0, 0] = 0.1
+    output[2, 0, 1] = 0.1
+    new_x = (output > 0).type("torch.FloatTensor")
+    loss_error, loss_no_improve, n_error, n_no_improve = compute_loss(
+        x, y, output, new_x
+    )

tests/test_loader.py

@@ -0,0 +1,32 @@
+from sudoku.loader import data_loader, DataIterBuffer, get_datasets, train_dataset
+from torch.utils.data import DataLoader
+
+
+def test_data_loader():
+    train_loader, test_loader = data_loader()
+
+    X, Y = next(iter(train_loader))
+    assert X.shape == Y.shape == (32, 2, 9 * 9 * 9)
+    assert (Y.sum(-1)[:, 0] == (9 * 9 * 8)).all()
+    assert (Y.sum(-1)[:, 1] == (9 * 9)).all()
+    assert (X.sum(-1)[:, 0] < (9 * 9 * 8)).all()
+    assert (X.sum(-1)[:, 1] < (9 * 9)).all()
+
+
+def test_data_iter_buffer():
+    i = 0
+    data_iter = DataIterBuffer(train_dataset)
+    data_loader_buffer = DataLoader(data_iter, 32)
+    for X, Y in data_loader_buffer:
+        i += 1
+        if i % 3 == 0:
+            data_iter.append(X, Y)
+    assert i == 29
+
+
+def test_max_holes():
+    train_dataset, test_dataset = get_datasets(train_size=2, test_size=2, max_holes=2)
+    X, Y = next(iter(train_dataset))
+    x_holes = X[1].reshape(9, 9, 9).sum(-1) == 0
+    print(X.shape, x_holes.shape)
+    assert x_holes.sum() == 2

tests/test_models.py

@@ -0,0 +1,49 @@
+import torch
+
+from sudoku.models import SudokuNet
+
+
+def test_same_output_under_rotation():
+    model = SudokuNet()
+    arr1 = torch.zeros((1, 2, 9, 9, 9))
+    arr1[0, 0, 1, 2, 3] = 1
+
+    output_1 = model.forward(arr1.view(1, 2, 9 * 9 * 9))
+    assert output_1.shape == (1, 2, 9 * 9 * 9), output_1
+    arr2 = torch.zeros((1, 2, 9, 9, 9))
+    arr2[0, 0, 2, 3, 4] = 1
+
+    output_2 = model.forward(arr2.view(1, 2, 9 * 9 * 9))
+    assert (
+        output_1.view(1, 2, 9, 9, 9)[0, 0, 1, 2, 3]
+        == output_2.view(1, 2, 9, 9, 9)[0, 0, 2, 3, 4]
+    )
+    assert (
+        output_1.view(1, 2, 9, 9, 9)[0, 0, 1, 2, 4]
+        == output_2.view(1, 2, 9, 9, 9)[0, 0, 2, 3, 6]
+    )
+    assert (
+        output_1.view(1, 2, 9, 9, 9)[0, 1, 1, 2, 4]
+        == output_2.view(1, 2, 9, 9, 9)[0, 1, 2, 3, 6]
+    )
+    assert (
+        output_1.view(1, 2, 9, 9, 9)[0, 1, 2, 2, 4]
+        == output_2.view(1, 2, 9, 9, 9)[0, 1, 1, 3, 6]
+    )
+    assert (
+        output_1.view(1, 2, 9, 9, 9)[0, 1, 2, 3, 4]
+        == output_2.view(1, 2, 9, 9, 9)[0, 1, 1, 2, 6]
+    )
+
+
+#  0, 1, 2 | 3, 4, 5 | 6, 7, 8
+#  0, 1, a | 3, 4, 5 | 6, 7, 8
+#  0, 1, 2 | b, 4, 5 | 6, 7, 8
+# ----------------------------
+#  0, 1, 2 | 3, 4, 5 | 6, 7, 8
+#  0, 1, 2 | 3, 4, 5 | 6, 7, 8
+#  0, 1, 2 | 3, 4, 5 | 6, 7, 8
+# ----------------------------
+#  0, 1, 2 | 3, 4, 5 | 6, 7, 8
+#  0, 1, 2 | 3, 4, 5 | 6, 7, 8
+#  0, 1, 2 | 3, 4, 5 | 6, 7, 8

tests/test_symetries.py

@@ -0,0 +1,21 @@
+import pytest
+import numpy as np
+
+from sudoku.symetries import mat_sym
+
+
+def test_same_output_under_rotation():
+    assert mat_sym.shape == (9 * 9 * 9, 16, 9 * 9 * 9)
+
+    arr1 = np.zeros((9, 9, 9))
+    arr1[0, 1, 2] = 1
+    arr1_sym = np.dot(mat_sym, arr1.reshape(9 * 9 * 9, -1)).reshape(9, 9, 9, -1)
+    arr2 = np.zeros((9, 9, 9))
+    arr2[1, 2, 3] = 1
+    arr2_sym = np.dot(mat_sym, arr2.reshape(9 * 9 * 9, -1)).reshape(9, 9, 9, -1)
+    assert arr1_sym.shape == (9, 9, 9, 16)
+    assert (arr1_sym[0, 1, 2] == arr2_sym[1, 2, 3]).all()
+    assert (arr1_sym[5, 1, 2] == arr2_sym[5, 2, 3]).all()
+    assert (arr1_sym[0, 1, 1] == arr2_sym[1, 2, 4]).all()
+    assert (arr1_sym[0, 5, 1] == arr2_sym[1, 8, 4]).all()
+    assert (arr1_sym[6, 5, 1] == arr2_sym[6, 8, 4]).all()