modified readme

.gitignore

@@ -0,0 +1,4 @@
+data
+results
+!results/17_06_2025_12_19_40
+.ipynb_checkpoints

LICENSE.txt

@@ -0,0 +1,674 @@
+                    GNU GENERAL PUBLIC LICENSE
+                       Version 3, 29 June 2007
+
+ Copyright (C) 2007 Free Software Foundation, Inc. <https://fsf.org/>
+ Everyone is permitted to copy and distribute verbatim copies
+ of this license document, but changing it is not allowed.
+
+                            Preamble
+
+  The GNU General Public License is a free, copyleft license for
+software and other kinds of works.
+
+  The licenses for most software and other practical works are designed
+to take away your freedom to share and change the works.  By contrast,
+the GNU General Public License is intended to guarantee your freedom to
+share and change all versions of a program--to make sure it remains free
+software for all its users.  We, the Free Software Foundation, use the
+GNU General Public License for most of our software; it applies also to
+any other work released this way by its authors.  You can apply it to
+your programs, too.
+
+  When we speak of free software, we are referring to freedom, not
+price.  Our General Public Licenses are designed to make sure that you
+have the freedom to distribute copies of free software (and charge for
+them if you wish), that you receive source code or can get it if you
+want it, that you can change the software or use pieces of it in new
+free programs, and that you know you can do these things.
+
+  To protect your rights, we need to prevent others from denying you
+these rights or asking you to surrender the rights.  Therefore, you have
+certain responsibilities if you distribute copies of the software, or if
+you modify it: responsibilities to respect the freedom of others.
+
+  For example, if you distribute copies of such a program, whether
+gratis or for a fee, you must pass on to the recipients the same
+freedoms that you received.  You must make sure that they, too, receive
+or can get the source code.  And you must show them these terms so they
+know their rights.
+
+  Developers that use the GNU GPL protect your rights with two steps:
+(1) assert copyright on the software, and (2) offer you this License
+giving you legal permission to copy, distribute and/or modify it.
+
+  For the developers' and authors' protection, the GPL clearly explains
+that there is no warranty for this free software.  For both users' and
+authors' sake, the GPL requires that modified versions be marked as
+changed, so that their problems will not be attributed erroneously to
+authors of previous versions.
+
+  Some devices are designed to deny users access to install or run
+modified versions of the software inside them, although the manufacturer
+can do so.  This is fundamentally incompatible with the aim of
+protecting users' freedom to change the software.  The systematic
+pattern of such abuse occurs in the area of products for individuals to
+use, which is precisely where it is most unacceptable.  Therefore, we
+have designed this version of the GPL to prohibit the practice for those
+products.  If such problems arise substantially in other domains, we
+stand ready to extend this provision to those domains in future versions
+of the GPL, as needed to protect the freedom of users.
+
+  Finally, every program is threatened constantly by software patents.
+States should not allow patents to restrict development and use of
+software on general-purpose computers, but in those that do, we wish to
+avoid the special danger that patents applied to a free program could
+make it effectively proprietary.  To prevent this, the GPL assures that
+patents cannot be used to render the program non-free.
+
+  The precise terms and conditions for copying, distribution and
+modification follow.
+
+                       TERMS AND CONDITIONS
+
+  0. Definitions.
+
+  "This License" refers to version 3 of the GNU General Public License.
+
+  "Copyright" also means copyright-like laws that apply to other kinds of
+works, such as semiconductor masks.
+
+  "The Program" refers to any copyrightable work licensed under this
+License.  Each licensee is addressed as "you".  "Licensees" and
+"recipients" may be individuals or organizations.
+
+  To "modify" a work means to copy from or adapt all or part of the work
+in a fashion requiring copyright permission, other than the making of an
+exact copy.  The resulting work is called a "modified version" of the
+earlier work or a work "based on" the earlier work.
+
+  A "covered work" means either the unmodified Program or a work based
+on the Program.
+
+  To "propagate" a work means to do anything with it that, without
+permission, would make you directly or secondarily liable for
+infringement under applicable copyright law, except executing it on a
+computer or modifying a private copy.  Propagation includes copying,
+distribution (with or without modification), making available to the
+public, and in some countries other activities as well.
+
+  To "convey" a work means any kind of propagation that enables other
+parties to make or receive copies.  Mere interaction with a user through
+a computer network, with no transfer of a copy, is not conveying.
+
+  An interactive user interface displays "Appropriate Legal Notices"
+to the extent that it includes a convenient and prominently visible
+feature that (1) displays an appropriate copyright notice, and (2)
+tells the user that there is no warranty for the work (except to the
+extent that warranties are provided), that licensees may convey the
+work under this License, and how to view a copy of this License.  If
+the interface presents a list of user commands or options, such as a
+menu, a prominent item in the list meets this criterion.
+
+  1. Source Code.
+
+  The "source code" for a work means the preferred form of the work
+for making modifications to it.  "Object code" means any non-source
+form of a work.
+
+  A "Standard Interface" means an interface that either is an official
+standard defined by a recognized standards body, or, in the case of
+interfaces specified for a particular programming language, one that
+is widely used among developers working in that language.
+
+  The "System Libraries" of an executable work include anything, other
+than the work as a whole, that (a) is included in the normal form of
+packaging a Major Component, but which is not part of that Major
+Component, and (b) serves only to enable use of the work with that
+Major Component, or to implement a Standard Interface for which an
+implementation is available to the public in source code form.  A
+"Major Component", in this context, means a major essential component
+(kernel, window system, and so on) of the specific operating system
+(if any) on which the executable work runs, or a compiler used to
+produce the work, or an object code interpreter used to run it.
+
+  The "Corresponding Source" for a work in object code form means all
+the source code needed to generate, install, and (for an executable
+work) run the object code and to modify the work, including scripts to
+control those activities.  However, it does not include the work's
+System Libraries, or general-purpose tools or generally available free
+programs which are used unmodified in performing those activities but
+which are not part of the work.  For example, Corresponding Source
+includes interface definition files associated with source files for
+the work, and the source code for shared libraries and dynamically
+linked subprograms that the work is specifically designed to require,
+such as by intimate data communication or control flow between those
+subprograms and other parts of the work.
+
+  The Corresponding Source need not include anything that users
+can regenerate automatically from other parts of the Corresponding
+Source.
+
+  The Corresponding Source for a work in source code form is that
+same work.
+
+  2. Basic Permissions.
+
+  All rights granted under this License are granted for the term of
+copyright on the Program, and are irrevocable provided the stated
+conditions are met.  This License explicitly affirms your unlimited
+permission to run the unmodified Program.  The output from running a
+covered work is covered by this License only if the output, given its
+content, constitutes a covered work.  This License acknowledges your
+rights of fair use or other equivalent, as provided by copyright law.
+
+  You may make, run and propagate covered works that you do not
+convey, without conditions so long as your license otherwise remains
+in force.  You may convey covered works to others for the sole purpose
+of having them make modifications exclusively for you, or provide you
+with facilities for running those works, provided that you comply with
+the terms of this License in conveying all material for which you do
+not control copyright.  Those thus making or running the covered works
+for you must do so exclusively on your behalf, under your direction
+and control, on terms that prohibit them from making any copies of
+your copyrighted material outside their relationship with you.
+
+  Conveying under any other circumstances is permitted solely under
+the conditions stated below.  Sublicensing is not allowed; section 10
+makes it unnecessary.
+
+  3. Protecting Users' Legal Rights From Anti-Circumvention Law.
+
+  No covered work shall be deemed part of an effective technological
+measure under any applicable law fulfilling obligations under article
+11 of the WIPO copyright treaty adopted on 20 December 1996, or
+similar laws prohibiting or restricting circumvention of such
+measures.
+
+  When you convey a covered work, you waive any legal power to forbid
+circumvention of technological measures to the extent such circumvention
+is effected by exercising rights under this License with respect to
+the covered work, and you disclaim any intention to limit operation or
+modification of the work as a means of enforcing, against the work's
+users, your or third parties' legal rights to forbid circumvention of
+technological measures.
+
+  4. Conveying Verbatim Copies.
+
+  You may convey verbatim copies of the Program's source code as you
+receive it, in any medium, provided that you conspicuously and
+appropriately publish on each copy an appropriate copyright notice;
+keep intact all notices stating that this License and any
+non-permissive terms added in accord with section 7 apply to the code;
+keep intact all notices of the absence of any warranty; and give all
+recipients a copy of this License along with the Program.
+
+  You may charge any price or no price for each copy that you convey,
+and you may offer support or warranty protection for a fee.
+
+  5. Conveying Modified Source Versions.
+
+  You may convey a work based on the Program, or the modifications to
+produce it from the Program, in the form of source code under the
+terms of section 4, provided that you also meet all of these conditions:
+
+    a) The work must carry prominent notices stating that you modified
+    it, and giving a relevant date.
+
+    b) The work must carry prominent notices stating that it is
+    released under this License and any conditions added under section
+    7.  This requirement modifies the requirement in section 4 to
+    "keep intact all notices".
+
+    c) You must license the entire work, as a whole, under this
+    License to anyone who comes into possession of a copy.  This
+    License will therefore apply, along with any applicable section 7
+    additional terms, to the whole of the work, and all its parts,
+    regardless of how they are packaged.  This License gives no
+    permission to license the work in any other way, but it does not
+    invalidate such permission if you have separately received it.
+
+    d) If the work has interactive user interfaces, each must display
+    Appropriate Legal Notices; however, if the Program has interactive
+    interfaces that do not display Appropriate Legal Notices, your
+    work need not make them do so.
+
+  A compilation of a covered work with other separate and independent
+works, which are not by their nature extensions of the covered work,
+and which are not combined with it such as to form a larger program,
+in or on a volume of a storage or distribution medium, is called an
+"aggregate" if the compilation and its resulting copyright are not
+used to limit the access or legal rights of the compilation's users
+beyond what the individual works permit.  Inclusion of a covered work
+in an aggregate does not cause this License to apply to the other
+parts of the aggregate.
+
+  6. Conveying Non-Source Forms.
+
+  You may convey a covered work in object code form under the terms
+of sections 4 and 5, provided that you also convey the
+machine-readable Corresponding Source under the terms of this License,
+in one of these ways:
+
+    a) Convey the object code in, or embodied in, a physical product
+    (including a physical distribution medium), accompanied by the
+    Corresponding Source fixed on a durable physical medium
+    customarily used for software interchange.
+
+    b) Convey the object code in, or embodied in, a physical product
+    (including a physical distribution medium), accompanied by a
+    written offer, valid for at least three years and valid for as
+    long as you offer spare parts or customer support for that product
+    model, to give anyone who possesses the object code either (1) a
+    copy of the Corresponding Source for all the software in the
+    product that is covered by this License, on a durable physical
+    medium customarily used for software interchange, for a price no
+    more than your reasonable cost of physically performing this
+    conveying of source, or (2) access to copy the
+    Corresponding Source from a network server at no charge.
+
+    c) Convey individual copies of the object code with a copy of the
+    written offer to provide the Corresponding Source.  This
+    alternative is allowed only occasionally and noncommercially, and
+    only if you received the object code with such an offer, in accord
+    with subsection 6b.
+
+    d) Convey the object code by offering access from a designated
+    place (gratis or for a charge), and offer equivalent access to the
+    Corresponding Source in the same way through the same place at no
+    further charge.  You need not require recipients to copy the
+    Corresponding Source along with the object code.  If the place to
+    copy the object code is a network server, the Corresponding Source
+    may be on a different server (operated by you or a third party)
+    that supports equivalent copying facilities, provided you maintain
+    clear directions next to the object code saying where to find the
+    Corresponding Source.  Regardless of what server hosts the
+    Corresponding Source, you remain obligated to ensure that it is
+    available for as long as needed to satisfy these requirements.
+
+    e) Convey the object code using peer-to-peer transmission, provided
+    you inform other peers where the object code and Corresponding
+    Source of the work are being offered to the general public at no
+    charge under subsection 6d.
+
+  A separable portion of the object code, whose source code is excluded
+from the Corresponding Source as a System Library, need not be
+included in conveying the object code work.
+
+  A "User Product" is either (1) a "consumer product", which means any
+tangible personal property which is normally used for personal, family,
+or household purposes, or (2) anything designed or sold for incorporation
+into a dwelling.  In determining whether a product is a consumer product,
+doubtful cases shall be resolved in favor of coverage.  For a particular
+product received by a particular user, "normally used" refers to a
+typical or common use of that class of product, regardless of the status
+of the particular user or of the way in which the particular user
+actually uses, or expects or is expected to use, the product.  A product
+is a consumer product regardless of whether the product has substantial
+commercial, industrial or non-consumer uses, unless such uses represent
+the only significant mode of use of the product.
+
+  "Installation Information" for a User Product means any methods,
+procedures, authorization keys, or other information required to install
+and execute modified versions of a covered work in that User Product from
+a modified version of its Corresponding Source.  The information must
+suffice to ensure that the continued functioning of the modified object
+code is in no case prevented or interfered with solely because
+modification has been made.
+
+  If you convey an object code work under this section in, or with, or
+specifically for use in, a User Product, and the conveying occurs as
+part of a transaction in which the right of possession and use of the
+User Product is transferred to the recipient in perpetuity or for a
+fixed term (regardless of how the transaction is characterized), the
+Corresponding Source conveyed under this section must be accompanied
+by the Installation Information.  But this requirement does not apply
+if neither you nor any third party retains the ability to install
+modified object code on the User Product (for example, the work has
+been installed in ROM).
+
+  The requirement to provide Installation Information does not include a
+requirement to continue to provide support service, warranty, or updates
+for a work that has been modified or installed by the recipient, or for
+the User Product in which it has been modified or installed.  Access to a
+network may be denied when the modification itself materially and
+adversely affects the operation of the network or violates the rules and
+protocols for communication across the network.
+
+  Corresponding Source conveyed, and Installation Information provided,
+in accord with this section must be in a format that is publicly
+documented (and with an implementation available to the public in
+source code form), and must require no special password or key for
+unpacking, reading or copying.
+
+  7. Additional Terms.
+
+  "Additional permissions" are terms that supplement the terms of this
+License by making exceptions from one or more of its conditions.
+Additional permissions that are applicable to the entire Program shall
+be treated as though they were included in this License, to the extent
+that they are valid under applicable law.  If additional permissions
+apply only to part of the Program, that part may be used separately
+under those permissions, but the entire Program remains governed by
+this License without regard to the additional permissions.
+
+  When you convey a copy of a covered work, you may at your option
+remove any additional permissions from that copy, or from any part of
+it.  (Additional permissions may be written to require their own
+removal in certain cases when you modify the work.)  You may place
+additional permissions on material, added by you to a covered work,
+for which you have or can give appropriate copyright permission.
+
+  Notwithstanding any other provision of this License, for material you
+add to a covered work, you may (if authorized by the copyright holders of
+that material) supplement the terms of this License with terms:
+
+    a) Disclaiming warranty or limiting liability differently from the
+    terms of sections 15 and 16 of this License; or
+
+    b) Requiring preservation of specified reasonable legal notices or
+    author attributions in that material or in the Appropriate Legal
+    Notices displayed by works containing it; or
+
+    c) Prohibiting misrepresentation of the origin of that material, or
+    requiring that modified versions of such material be marked in
+    reasonable ways as different from the original version; or
+
+    d) Limiting the use for publicity purposes of names of licensors or
+    authors of the material; or
+
+    e) Declining to grant rights under trademark law for use of some
+    trade names, trademarks, or service marks; or
+
+    f) Requiring indemnification of licensors and authors of that
+    material by anyone who conveys the material (or modified versions of
+    it) with contractual assumptions of liability to the recipient, for
+    any liability that these contractual assumptions directly impose on
+    those licensors and authors.
+
+  All other non-permissive additional terms are considered "further
+restrictions" within the meaning of section 10.  If the Program as you
+received it, or any part of it, contains a notice stating that it is
+governed by this License along with a term that is a further
+restriction, you may remove that term.  If a license document contains
+a further restriction but permits relicensing or conveying under this
+License, you may add to a covered work material governed by the terms
+of that license document, provided that the further restriction does
+not survive such relicensing or conveying.
+
+  If you add terms to a covered work in accord with this section, you
+must place, in the relevant source files, a statement of the
+additional terms that apply to those files, or a notice indicating
+where to find the applicable terms.
+
+  Additional terms, permissive or non-permissive, may be stated in the
+form of a separately written license, or stated as exceptions;
+the above requirements apply either way.
+
+  8. Termination.
+
+  You may not propagate or modify a covered work except as expressly
+provided under this License.  Any attempt otherwise to propagate or
+modify it is void, and will automatically terminate your rights under
+this License (including any patent licenses granted under the third
+paragraph of section 11).
+
+  However, if you cease all violation of this License, then your
+license from a particular copyright holder is reinstated (a)
+provisionally, unless and until the copyright holder explicitly and
+finally terminates your license, and (b) permanently, if the copyright
+holder fails to notify you of the violation by some reasonable means
+prior to 60 days after the cessation.
+
+  Moreover, your license from a particular copyright holder is
+reinstated permanently if the copyright holder notifies you of the
+violation by some reasonable means, this is the first time you have
+received notice of violation of this License (for any work) from that
+copyright holder, and you cure the violation prior to 30 days after
+your receipt of the notice.
+
+  Termination of your rights under this section does not terminate the
+licenses of parties who have received copies or rights from you under
+this License.  If your rights have been terminated and not permanently
+reinstated, you do not qualify to receive new licenses for the same
+material under section 10.
+
+  9. Acceptance Not Required for Having Copies.
+
+  You are not required to accept this License in order to receive or
+run a copy of the Program.  Ancillary propagation of a covered work
+occurring solely as a consequence of using peer-to-peer transmission
+to receive a copy likewise does not require acceptance.  However,
+nothing other than this License grants you permission to propagate or
+modify any covered work.  These actions infringe copyright if you do
+not accept this License.  Therefore, by modifying or propagating a
+covered work, you indicate your acceptance of this License to do so.
+
+  10. Automatic Licensing of Downstream Recipients.
+
+  Each time you convey a covered work, the recipient automatically
+receives a license from the original licensors, to run, modify and
+propagate that work, subject to this License.  You are not responsible
+for enforcing compliance by third parties with this License.
+
+  An "entity transaction" is a transaction transferring control of an
+organization, or substantially all assets of one, or subdividing an
+organization, or merging organizations.  If propagation of a covered
+work results from an entity transaction, each party to that
+transaction who receives a copy of the work also receives whatever
+licenses to the work the party's predecessor in interest had or could
+give under the previous paragraph, plus a right to possession of the
+Corresponding Source of the work from the predecessor in interest, if
+the predecessor has it or can get it with reasonable efforts.
+
+  You may not impose any further restrictions on the exercise of the
+rights granted or affirmed under this License.  For example, you may
+not impose a license fee, royalty, or other charge for exercise of
+rights granted under this License, and you may not initiate litigation
+(including a cross-claim or counterclaim in a lawsuit) alleging that
+any patent claim is infringed by making, using, selling, offering for
+sale, or importing the Program or any portion of it.
+
+  11. Patents.
+
+  A "contributor" is a copyright holder who authorizes use under this
+License of the Program or a work on which the Program is based.  The
+work thus licensed is called the contributor's "contributor version".
+
+  A contributor's "essential patent claims" are all patent claims
+owned or controlled by the contributor, whether already acquired or
+hereafter acquired, that would be infringed by some manner, permitted
+by this License, of making, using, or selling its contributor version,
+but do not include claims that would be infringed only as a
+consequence of further modification of the contributor version.  For
+purposes of this definition, "control" includes the right to grant
+patent sublicenses in a manner consistent with the requirements of
+this License.
+
+  Each contributor grants you a non-exclusive, worldwide, royalty-free
+patent license under the contributor's essential patent claims, to
+make, use, sell, offer for sale, import and otherwise run, modify and
+propagate the contents of its contributor version.
+
+  In the following three paragraphs, a "patent license" is any express
+agreement or commitment, however denominated, not to enforce a patent
+(such as an express permission to practice a patent or covenant not to
+sue for patent infringement).  To "grant" such a patent license to a
+party means to make such an agreement or commitment not to enforce a
+patent against the party.
+
+  If you convey a covered work, knowingly relying on a patent license,
+and the Corresponding Source of the work is not available for anyone
+to copy, free of charge and under the terms of this License, through a
+publicly available network server or other readily accessible means,
+then you must either (1) cause the Corresponding Source to be so
+available, or (2) arrange to deprive yourself of the benefit of the
+patent license for this particular work, or (3) arrange, in a manner
+consistent with the requirements of this License, to extend the patent
+license to downstream recipients.  "Knowingly relying" means you have
+actual knowledge that, but for the patent license, your conveying the
+covered work in a country, or your recipient's use of the covered work
+in a country, would infringe one or more identifiable patents in that
+country that you have reason to believe are valid.
+
+  If, pursuant to or in connection with a single transaction or
+arrangement, you convey, or propagate by procuring conveyance of, a
+covered work, and grant a patent license to some of the parties
+receiving the covered work authorizing them to use, propagate, modify
+or convey a specific copy of the covered work, then the patent license
+you grant is automatically extended to all recipients of the covered
+work and works based on it.
+
+  A patent license is "discriminatory" if it does not include within
+the scope of its coverage, prohibits the exercise of, or is
+conditioned on the non-exercise of one or more of the rights that are
+specifically granted under this License.  You may not convey a covered
+work if you are a party to an arrangement with a third party that is
+in the business of distributing software, under which you make payment
+to the third party based on the extent of your activity of conveying
+the work, and under which the third party grants, to any of the
+parties who would receive the covered work from you, a discriminatory
+patent license (a) in connection with copies of the covered work
+conveyed by you (or copies made from those copies), or (b) primarily
+for and in connection with specific products or compilations that
+contain the covered work, unless you entered into that arrangement,
+or that patent license was granted, prior to 28 March 2007.
+
+  Nothing in this License shall be construed as excluding or limiting
+any implied license or other defenses to infringement that may
+otherwise be available to you under applicable patent law.
+
+  12. No Surrender of Others' Freedom.
+
+  If conditions are imposed on you (whether by court order, agreement or
+otherwise) that contradict the conditions of this License, they do not
+excuse you from the conditions of this License.  If you cannot convey a
+covered work so as to satisfy simultaneously your obligations under this
+License and any other pertinent obligations, then as a consequence you may
+not convey it at all.  For example, if you agree to terms that obligate you
+to collect a royalty for further conveying from those to whom you convey
+the Program, the only way you could satisfy both those terms and this
+License would be to refrain entirely from conveying the Program.
+
+  13. Use with the GNU Affero General Public License.
+
+  Notwithstanding any other provision of this License, you have
+permission to link or combine any covered work with a work licensed
+under version 3 of the GNU Affero General Public License into a single
+combined work, and to convey the resulting work.  The terms of this
+License will continue to apply to the part which is the covered work,
+but the special requirements of the GNU Affero General Public License,
+section 13, concerning interaction through a network will apply to the
+combination as such.
+
+  14. Revised Versions of this License.
+
+  The Free Software Foundation may publish revised and/or new versions of
+the GNU General Public License from time to time.  Such new versions will
+be similar in spirit to the present version, but may differ in detail to
+address new problems or concerns.
+
+  Each version is given a distinguishing version number.  If the
+Program specifies that a certain numbered version of the GNU General
+Public License "or any later version" applies to it, you have the
+option of following the terms and conditions either of that numbered
+version or of any later version published by the Free Software
+Foundation.  If the Program does not specify a version number of the
+GNU General Public License, you may choose any version ever published
+by the Free Software Foundation.
+
+  If the Program specifies that a proxy can decide which future
+versions of the GNU General Public License can be used, that proxy's
+public statement of acceptance of a version permanently authorizes you
+to choose that version for the Program.
+
+  Later license versions may give you additional or different
+permissions.  However, no additional obligations are imposed on any
+author or copyright holder as a result of your choosing to follow a
+later version.
+
+  15. Disclaimer of Warranty.
+
+  THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY
+APPLICABLE LAW.  EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT
+HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY
+OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO,
+THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
+PURPOSE.  THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM
+IS WITH YOU.  SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF
+ALL NECESSARY SERVICING, REPAIR OR CORRECTION.
+
+  16. Limitation of Liability.
+
+  IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING
+WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS
+THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY
+GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE
+USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF
+DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD
+PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS),
+EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF
+SUCH DAMAGES.
+
+  17. Interpretation of Sections 15 and 16.
+
+  If the disclaimer of warranty and limitation of liability provided
+above cannot be given local legal effect according to their terms,
+reviewing courts shall apply local law that most closely approximates
+an absolute waiver of all civil liability in connection with the
+Program, unless a warranty or assumption of liability accompanies a
+copy of the Program in return for a fee.
+
+                     END OF TERMS AND CONDITIONS
+
+            How to Apply These Terms to Your New Programs
+
+  If you develop a new program, and you want it to be of the greatest
+possible use to the public, the best way to achieve this is to make it
+free software which everyone can redistribute and change under these terms.
+
+  To do so, attach the following notices to the program.  It is safest
+to attach them to the start of each source file to most effectively
+state the exclusion of warranty; and each file should have at least
+the "copyright" line and a pointer to where the full notice is found.
+
+    <one line to give the program's name and a brief idea of what it does.>
+    Copyright (C) <year>  <name of author>
+
+    This program is free software: you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation, either version 3 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License
+    along with this program.  If not, see <https://www.gnu.org/licenses/>.
+
+Also add information on how to contact you by electronic and paper mail.
+
+  If the program does terminal interaction, make it output a short
+notice like this when it starts in an interactive mode:
+
+    <program>  Copyright (C) <year>  <name of author>
+    This program comes with ABSOLUTELY NO WARRANTY; for details type `show w'.
+    This is free software, and you are welcome to redistribute it
+    under certain conditions; type `show c' for details.
+
+The hypothetical commands `show w' and `show c' should show the appropriate
+parts of the General Public License.  Of course, your program's commands
+might be different; for a GUI interface, you would use an "about box".
+
+  You should also get your employer (if you work as a programmer) or school,
+if any, to sign a "copyright disclaimer" for the program, if necessary.
+For more information on this, and how to apply and follow the GNU GPL, see
+<https://www.gnu.org/licenses/>.
+
+  The GNU General Public License does not permit incorporating your program
+into proprietary programs.  If your program is a subroutine library, you
+may consider it more useful to permit linking proprietary applications with
+the library.  If this is what you want to do, use the GNU Lesser General
+Public License instead of this License.  But first, please read
+<https://www.gnu.org/licenses/why-not-lgpl.html>.

PCAM-pipeline.py

@@ -0,0 +1,973 @@
+#!/usr/bin/env python
+# coding: utf-8
+
+# # 🧬 PCam Dataset: Tumor Detection via Binary Image Classification
+# 
+# For full dataset details, visit the official repository:  
+# 🔗 [github.com/basveeling/pcam](https://github.com/basveeling/pcam)
+# 
+# 
+# ## 📊 Dataset Overview
+# 
+# The **PatchCamelyon (PCam)** benchmark is a challenging image classification dataset designed for breast cancer detection tasks.
+# 
+# - 📦 **Total images**: 327,680 color patches  
+# - 🖼️ **Image size**: 96 × 96 pixels
+# - 🧪 **Source**: Histopathologic scans of lymph node sections  
+# - 🏷️ **Labels**: Binary — A positive (1) label indicates that the center 32x32px region of a patch contains at least one pixel of tumor tissue. Tumor tissue in the outer region of the patch does not influence the label.
+# 
+# 
+# ## 🧠 Solution to Implement
+# 
+# In this notebook, we implement a solution inspired by the following research paper:
+# 
+# > 📄 [**Cancer Image Classification Based on DenseNet Model**](https://arxiv.org/abs/2011.11186)  
+# > _by Zhong, Ziliang; Zheng, Muhang; Mai, Huafeng; Zhao, Jianan; Liu, Xinyi_
+# 
+# This study explores the application of DenseNet architectures to the PCam dataset for accurate cancer classification.
+# 
+# ---
+# 
+# ## Results
+# 
+# The submission on kaggle with the best model trained on this notebook is 
+# 
+# ```Score: 0.9648```
+# ```Private score: 0.9702```
+# 
+
+# # 1. Load the dataset
+# Load the training, test and validation datasets from PCAM.
+# 
+# We are going to use the kaggle version that is a cleaned version of the official PCAM dataset. 
+# 
+# In the kaggle version duplicates ar removed and there is no leakage between training and test datasets.
+
+# In[1]:
+
+
+import typing as tp
+import numpy as np
+import torch
+import torchvision
+from torch import nn
+from torch.utils.data import Dataset, DataLoader, ConcatDataset
+from torchvision.transforms import ToTensor 
+from torchvision import datasets
+from torch.utils.tensorboard import SummaryWriter
+
+
+# We need to use GPU if available
+
+# In[2]:
+
+
+from torch.optim import Optimizer, lr_scheduler
+from torch.optim.lr_scheduler import LRScheduler
+
+if torch.cuda.is_available():
+    device = torch.device("cuda")
+else:
+    device = torch.device("cpu")
+print("Using device", device)
+
+
+# Let's download the kaggle dataset.
+# For this you need your credentials.
+# If you did not set already your ```~/.kaggle/kaggle.json``` key:
+#  - Go to your kaggle account setting and create a new API token if needed.
+#  - Then feel in this part with your information ```creds = '{"username":"xxxxx","key":"xxxxx"}'```
+
+# In[3]:
+
+
+get_ipython().system('pip install kaggle')
+creds = '{"username":"xxxxx","key":"xxxxx"}'
+from pathlib import Path
+
+cred_path = Path('~/.kaggle/kaggle.json').expanduser()
+if not cred_path.exists():
+    cred_path.parent.mkdir(exist_ok=True)
+    cred_path.write_text(creds)
+    cred_path.chmod(0o600)
+
+
+# In[4]:
+
+
+import os
+import zipfile
+
+root = "data/"
+dataset_dir = "data/histopathologic-cancer-detection"
+zip_file = "histopathologic-cancer-detection.zip"
+train_path = os.path.join(dataset_dir, "train")
+
+if not os.path.exists(root):
+  os.mkdir(root)
+
+if not os.path.exists('results'):
+  os.mkdir('results')
+
+if not os.path.exists(train_path):
+    print("Downloading Histopathologic Cancer Detection dataset...")
+    get_ipython().system('kaggle competitions download -c histopathologic-cancer-detection -p {root} --force')
+else:
+    print("Dataset zip already downloaded.")
+
+if not os.path.exists(train_path):
+    print("Unzipping dataset...")
+    with zipfile.ZipFile(os.path.join(root, zip_file), 'r') as zip_ref:
+        zip_ref.extractall(dataset_dir)
+else:
+    print("Dataset already unzipped.")
+
+
+# Know Let's create our pytorch dataset class.
+# I have used train_test_split from sklearn to have a stratified dataset (The kaggle PCAM dataset is unbalanced)
+
+# In[5]:
+
+
+from sklearn.model_selection import train_test_split
+from PIL import Image
+import pandas as pd
+
+class PcamDatasetKaggle(torchvision.datasets.VisionDataset):
+    def __init__(self, root, split, transform, target_transform = None):
+         super().__init__(root, transform=transform, target_transform=target_transform)
+         self.root = root
+         self.split = split
+         self.transform = transform
+         self.img_path = os.path.join(self.root, "train")
+
+         self.full_labels = pd.read_csv(self.root+'/train_labels.csv')
+         X_train, X_test, y_train, y_test = train_test_split(self.full_labels['id'],
+                                                             self.full_labels['label'],
+                                                             test_size = 0.2, 
+                                                             train_size = 0.8,
+                                                             random_state=30,
+                                                             shuffle=True,
+                                                             stratify=self.full_labels['label'])
+
+         if (split == "train"):
+             self.imgs = X_train + ".tif"
+             self.labels = y_train
+         elif (split == "val"):
+             self.imgs = X_test + ".tif"
+             self.labels = y_test
+         else:
+             self.img_path = os.path.join(self.root, self.split)
+             self.imgs = pd.Series(list(sorted(os.listdir(self.img_path))))
+             self.labels = pd.Series(torch.full((len(self.imgs),), -10))      
+         assert len(self.labels) == len(self.imgs)
+         print("Split", split, "Negative/Positive samples % " , 100.0*(self.labels.value_counts() / self.labels.shape[0]))
+
+    def __getitem__(self, idx):
+        assert idx < len(self.imgs)
+        img_pil = Image.open(os.path.join(self.img_path, self.imgs.iloc[idx]))
+        img = self.transform(img_pil)
+        label = self.labels.iloc[idx]
+        return img, label
+    def __len__(self) :
+        return len(self.imgs)
+
+def check_dataset_leakage(dataset1, dataset2):
+    duplicates = set(dataset1.imgs) & set(dataset2.imgs)
+    assert len(duplicates) == 0
+
+def check_same_imgs(dataset1, dataset2):
+    duplicates = set(dataset1.imgs) & set(dataset2.imgs)
+    assert len(duplicates) == len(dataset1.imgs)
+    assert len(duplicates) == len(dataset2.imgs)
+
+
+# Let's define some transforms for dataloading and data augmentation
+# 
+# An improvment could be to use [albumentation](https://albumentations.ai/) to define a more refined ```transform_data_augment```
+
+# In[6]:
+
+
+import torchvision.transforms as transforms
+
+torch.manual_seed(30)
+torch.cuda.manual_seed_all(30)
+
+# Preprocess images with transforms
+transform = transforms.Compose([
+    transforms.Resize((224, 224)), #Match resnet original input size            
+    transforms.ToTensor()
+])
+
+# For augmenting data
+transform_data_augment = transforms.Compose([ 
+    transforms.Resize((300, 300)),
+    transforms.RandomHorizontalFlip(),
+    transforms.RandomVerticalFlip(),
+    transforms.GaussianBlur(kernel_size = (5,5),sigma=(0.1, 0.5)),
+    transforms.RandomRotation(degrees=25),
+    transforms.ColorJitter(
+        brightness=0.1, 
+        contrast=0.1, 
+        saturation=0.01, 
+        hue=0.005
+    ),
+    transforms.CenterCrop((224, 224)),
+    transforms.RandomResizedCrop(size = (224, 224), scale = (0.9, 1.0)),
+    transforms.ToTensor()
+])
+
+
+
+# In[7]:
+
+
+from copy import deepcopy
+
+""" PCAM pytorch version but the dataset is not clean 
+training_set_original = datasets.PCAM(root="data", split="train",download = True, transform = transform) 
+training_set_augment = datasets.PCAM(root="data", split="train",download = True, transform = transform_data_augment)
+val_set = datasets.PCAM(root="data", split="val", download=True, transform = transform)
+test_set = datasets.PCAM(root="data", split="test", download=True, transform = transform)
+"""
+
+training_set_original = PcamDatasetKaggle(root=dataset_dir, split="train", transform = deepcopy(transform)) 
+training_set_augment = PcamDatasetKaggle(root=dataset_dir, split="train", transform = deepcopy(transform_data_augment)) 
+
+val_set = PcamDatasetKaggle(root=dataset_dir, split="val", transform = deepcopy(transform)) 
+val_set_augment = PcamDatasetKaggle(root=dataset_dir, split="val", transform = deepcopy(transform_data_augment)) 
+
+test_set = PcamDatasetKaggle(root=dataset_dir, split="test", transform = deepcopy(transform))
+test_set_augment = PcamDatasetKaggle(root=dataset_dir, split="test", transform = deepcopy(transform_data_augment)) #For TTA
+
+check_dataset_leakage(training_set_original, val_set)
+check_dataset_leakage(training_set_original, test_set)
+check_dataset_leakage(val_set, test_set)
+check_same_imgs(training_set_original, training_set_augment)
+check_same_imgs(val_set, val_set_augment)
+check_same_imgs(test_set, test_set_augment)
+
+
+# # 2. Plot and visualize original and augmented the data
+# Each (3,96,96) image is associated with a binary label indicates the presence of a tumor.
+# 
+# Let's define a function to plot some images with their label.
+# 
+# Let's save the plots in an experiment directory for logging purposes
+# 
+
+# In[8]:
+
+
+import matplotlib.pyplot as plt
+
+def plot_training_set_sample(training_set, 
+                             file_name = "results/pcam/data.png", 
+                             rows = 5, 
+                             cols = 5, 
+                             mean_stdev = torch.Tensor([[0.0, 0.0, 0.0], [1.0, 1.0, 1.0]])):
+    mean = mean_stdev[0].numpy()
+    std  = mean_stdev[1].numpy()
+    fig = plt.figure(figsize=(2*cols, 2*rows))
+    for i in range(1, rows*cols + 1):
+        random_idx = torch.randint(len(training_set), (1,)).item()
+        fig.add_subplot(rows, cols, i)
+        img = training_set[random_idx][0].permute(1,2,0).numpy() 
+        img_unnormalized = img*std + mean
+        img_unnormalized = np.clip(img_unnormalized, 0, 1)
+        plt.imshow(img_unnormalized)
+        plt.axis("off")
+        plt.title(training_set[random_idx][1])
+    plt.savefig(file_name)
+    plt.show()
+
+
+
+# In[9]:
+
+
+import os
+from datetime import datetime
+exp_dir = "results/pcam/"+datetime.now().strftime("%d_%m_%Y_%H_%M_%S")
+os.mkdir(exp_dir)
+
+
+# In[10]:
+
+
+print("Original Training Set")
+plot_training_set_sample(training_set_original, exp_dir + "/training_set_original.png",rows=2, cols=5)
+
+
+# In[11]:
+
+
+print("Augmented Training Set")
+plot_training_set_sample(training_set_augment, exp_dir + "/training_set_augment.png",rows=2, cols=5)
+
+
+# # 3.Normalize and create augmented dataset
+
+# Let's create a function that computes mean, standard deviation and class balance for a pytorch DataLoader.
+# 
+# Normalize the datasets accordingly
+
+# In[12]:
+
+
+def compute_dataset_mean_stdev_class_balance(dataloader: DataLoader, device: torch.cuda.device) -> tp.List[float]:
+    mean = 0.0
+    stdev = 0.0
+    y_full = torch.Tensor([]).to(device)
+    for batch, (X,y) in enumerate(dataloader):
+        X = X.to(device)
+        y = y.to(device)
+        batch_samples = X.size(0)
+        mean += torch.mean(X, dim = (0,2,3)) * batch_samples
+        stdev += torch.std(X, dim = (0,2,3)) * batch_samples
+        y_full = torch.cat([y_full, y])
+    positive_labels = (y_full == torch.Tensor([1]).to(device)).sum()
+    negative_labels = (y_full == torch.Tensor([0]).to(device)).sum()
+    return [mean.detach().cpu() / len(dataloader.dataset), stdev.detach().cpu() / len(dataloader.dataset)], positive_labels.detach().cpu(), negative_labels.detach().cpu()
+
+
+
+# In[13]:
+
+
+# Create DataLoader
+batch_size = 128
+training_set_original_dataloader = DataLoader(training_set_original, batch_size=batch_size, shuffle=True, pin_memory=True, num_workers=6, persistent_workers = True)
+training_set_augment_dataloader = DataLoader(training_set_augment, batch_size=batch_size, shuffle=True, pin_memory=True, num_workers=6, persistent_workers = True)
+
+# Compute Mean and Std to normalize images if not already done
+COMPUTE_NORMALIZATION_AGAIN = False
+
+mean_stdev_original =[torch.Tensor([0.7022, 0.5459, 0.6962]), torch.Tensor([0.2218, 0.2668, 0.1982])]
+mean_stdev_augment = [torch.Tensor([0.6939, 0.5397, 0.6904]), torch.Tensor([0.2225, 0.2661, 0.1988])]
+
+pos = 71294
+neg = 104726
+apos = 71294
+aneg = 104726
+
+if (COMPUTE_NORMALIZATION_AGAIN):
+    mean_stdev_original, pos, neg = compute_dataset_mean_stdev_class_balance(training_set_original_dataloader, device)
+    mean_stdev_augment, apos, aneg = compute_dataset_mean_stdev_class_balance(training_set_augment_dataloader, device)
+
+
+def combine_std(mean1_stdev1: torch.torch.Tensor, mean2_stdev2: torch.Tensor):
+    mean1, stdev1 = mean1_stdev1[0], mean1_stdev1[1]
+    mean2, stdev2 = mean2_stdev2[0], mean2_stdev2[1]
+
+    mean3 = (mean1 + mean2) * 0.5
+
+    var1 = stdev1 ** 2
+    var2 = stdev2 ** 2
+    var3 = 0.5 * (var1 + (mean1 - mean3) ** 2 + var2 + (mean2 - mean3) ** 2)
+
+    stdev3 = torch.sqrt(var3)
+    return [mean3, stdev3]
+
+new_mean_stdev = combine_std(mean_stdev_original, mean_stdev_augment)
+new_mean_stdev = torch.stack(new_mean_stdev).cpu().detach()
+
+print("Normalization done with")
+print("training_set [mean, stdev]: ", new_mean_stdev)
+
+training_set_original_transform = transforms.Compose([*training_set_original.transforms.transform.transforms,
+                                              transforms.Normalize(new_mean_stdev[0], new_mean_stdev[1])])
+
+training_set_augment_transform = transforms.Compose([*training_set_augment.transforms.transform.transforms,
+                                                      transforms.Normalize(new_mean_stdev[0], new_mean_stdev[1])])
+
+
+training_set_original = PcamDatasetKaggle(root=dataset_dir, split="train", transform = deepcopy(training_set_original_transform)) 
+training_set_augment = PcamDatasetKaggle(root=dataset_dir, split="train", transform = deepcopy(training_set_augment_transform)) 
+val_set = PcamDatasetKaggle(root=dataset_dir, split="val", transform = deepcopy(training_set_original_transform)) 
+val_set_augment = PcamDatasetKaggle(root=dataset_dir, split="val", transform = deepcopy(training_set_augment_transform)) 
+test_set = PcamDatasetKaggle(root=dataset_dir, split="test", transform = deepcopy(training_set_original_transform))
+
+
+# Create Augmented Training Dataset
+training_set = ConcatDataset([training_set_original, training_set_augment])
+
+# Create Final DataLoaders
+training_dataloader = DataLoader(training_set, batch_size=batch_size, shuffle=True, pin_memory=True, num_workers=6, persistent_workers = True)
+val_dataloader = DataLoader(val_set, batch_size=batch_size, shuffle=False, pin_memory=True, num_workers=6, persistent_workers = True)
+val_dataloader_augment = DataLoader(val_set_augment, batch_size=batch_size, shuffle=False, pin_memory=True, num_workers=6, persistent_workers = True)
+test_dataloader = DataLoader(test_set, batch_size=batch_size, shuffle=False, pin_memory=True, num_workers=6, persistent_workers = True)
+
+
+# In[14]:
+
+
+print("Full Training Set Normalized")
+plot_training_set_sample(training_set, exp_dir + "/training_set_final.png", rows = 2, cols = 5, mean_stdev=new_mean_stdev)
+
+
+# # 3. Defining a training loop over one epoch and a metric
+# The dataset is not balance thus it is better to use roc_auc_score than accuracy
+
+# In[15]:
+
+
+def compute_metrics(full_y: torch.Tensor, 
+                    full_logits: torch.Tensor,  
+                    full_pred: torch.Tensor,  
+                    sk_learn_metrics_logits: tp.List[tp.Callable],
+                    sk_learn_metrics_pred: tp.List[tp.Callable]) -> tp.Dict:
+    full_y = full_y.detach().cpu().numpy()
+    full_logits = torch.sigmoid(full_logits).detach().cpu().numpy()
+    full_pred = full_pred.detach().cpu().numpy()
+
+    results = {}
+    for metric in sk_learn_metrics_logits:
+        results[metric.__name__] = metric(full_y, full_logits)
+    for metric in sk_learn_metrics_pred:
+        results[metric.__name__] = metric(full_y, full_pred)
+    return results
+
+
+# In[16]:
+
+
+def run_one_epoch(model : nn.Module, 
+                   training_dataloader: DataLoader,
+                   optimizer: Optimizer,
+                   loss_function: nn.Module,
+                   scheduler : LRScheduler,
+                   device: torch.cuda.device,
+                   writer: SummaryWriter,
+                   epoch: int,
+                   sk_learn_metrics_logits: tp.List[tp.Callable],
+                   sk_learn_metrics_pred: tp.List[tp.Callable],
+                   threshold: float = 0.5):
+    running_loss = 0.0
+    num_batch = len(training_dataloader)
+    full_y = torch.Tensor([]).to(device)
+    full_logits = torch.Tensor([]).to(device)
+    full_pred = torch.Tensor([]).to(device)
+
+    model.train()
+    scaler = torch.amp.GradScaler("cuda")
+    for batch, (X, y) in enumerate(training_dataloader):
+        optimizer.zero_grad()
+        X = X.to(device, non_blocking=True)
+        y = y.to(device, non_blocking=True)
+        with torch.amp.autocast("cuda"):
+            logits = model(X).squeeze()
+            loss = loss_function(logits, y.float())
+        scaler.scale(loss).backward()
+        scaler.step(optimizer)
+        scaler.update()
+
+        with torch.no_grad():
+            preds = (torch.sigmoid(logits) > threshold).float()
+            full_y = torch.cat([full_y, y])
+            full_logits = torch.cat([full_logits, logits])
+            full_pred = torch.cat([full_pred, preds])
+
+        running_loss += loss.item()
+        avg_loss = running_loss / (batch + 1.)
+        if batch % 250 == 0:
+            writer.add_scalar('Training Loss(avg)', avg_loss, batch + epoch*num_batch)
+            writer.add_scalar('Training Loss (raw)', loss.item(), batch + epoch*num_batch)
+    scheduler.step()
+    writer.flush()
+    return compute_metrics(full_y, full_logits, full_pred, sk_learn_metrics_logits, sk_learn_metrics_pred)
+
+
+# In[17]:
+
+
+def eval_model(model: nn.Module,
+               dataloader: DataLoader, 
+               sk_learn_metrics_logits: tp.List[tp.Callable],
+               sk_learn_metrics_pred: tp.List[tp.Callable],
+               device: torch.cuda.device,
+               threshold: float = 0.5) -> tp.Dict:
+
+    model.eval()
+    full_y = torch.Tensor([]).to(device)
+    full_logits = torch.Tensor([]).to(device)
+    full_pred = torch.Tensor([]).to(device)
+
+    with torch.no_grad():
+        for X, y in dataloader:
+            X = X.to(device)
+            y = y.to(device)
+            logits = model(X).squeeze()
+            preds = (torch.sigmoid(logits) > threshold).float()
+
+            full_y = torch.cat([full_y, y])
+            full_logits = torch.cat([full_logits, logits])
+            full_pred = torch.cat([full_pred, preds])
+    return compute_metrics(full_y, full_logits, full_pred, sk_learn_metrics_logits, sk_learn_metrics_pred)
+
+
+# # 4. Setup tensorboard for monitoring
+
+# In[18]:
+
+
+import threading 
+import tensorboard
+from tensorboard import program
+
+def start_tensorboard(logdir):
+    tb = program.TensorBoard()
+    tb.configure(argv=[None, '--logdir', logdir])
+    url = tb.launch()
+    print(f"TensorBoard is running at {url}")
+
+# Replace 'logs' with your actual log directory
+logdir = exp_dir
+tb_thread = threading.Thread(target=start_tensorboard, args=(logdir,), daemon=True)
+tb_thread.start()
+
+
+# In[19]:
+
+
+from PIL import Image
+
+def load_image(path):
+    img = Image.open(path)
+    # Convert to numpy array and add batch dimension (C, H, W)
+    img_array = np.array(img)
+    if len(img_array.shape) == 2:  # Grayscale image
+        img_array = np.expand_dims(img_array, axis=0)  # (1, H, W)
+    else:  # Color image
+        img_array = img_array.transpose(2, 0, 1)  # (C, H, W)
+    return img_array
+
+writer = SummaryWriter(exp_dir + '/tensorboard')
+writer.add_image('training_set_original', load_image(exp_dir + "/training_set_original.png"), 0)
+writer.flush()
+writer.add_image('training_set_augment',  load_image(exp_dir + "/training_set_augment.png"), 0)
+writer.flush()
+writer.add_image('training_set_final',  load_image(exp_dir + "/training_set_final.png"), 0)
+writer.flush()
+
+
+# # 5. Find best learning rate
+# 
+# > 📄 [**Cancer Image Classification Based on DenseNet Model**](https://arxiv.org/abs/2011.11186)  
+# > _by Zhong, Ziliang; Zheng, Muhang; Mai, Huafeng; Zhao, Jianan; Liu, Xinyi_
+# 
+# Suggest to use a learning rate lr = 1e-4 for densenet201. 
+# 
+# You can also plot the loss with respect to the lr evaluated on a few batches.
+# 
+# It gives insight on which lr to take: between 1e-4 and 1e-3
+
+# In[20]:
+
+
+from torchvision.models import densenet201, DenseNet201_Weights
+model = densenet201(weights=DenseNet201_Weights.DEFAULT)
+
+for params in model.parameters():
+    params.requires_grad = False
+
+model.classifier = nn.Sequential(nn.Linear(1920, 1, bias= True))
+
+for param in model.classifier.parameters():
+    param.requires_grad = True
+
+model = model.to(device)
+
+def custom_lr_find(model : nn.Module, 
+                   dataloader: DataLoader,
+                   loss_function: nn.Module,
+                   device: str,
+                   start_lr = 1e-7,
+                   end_lr = 1.0,
+                   num_iteration = 200):
+    rates = []
+    lossses = []
+    model = model.to(device)
+    optimizer = torch.optim.Adam(model.parameters(),lr=start_lr)
+
+
+    def lr_lambda(iteration):
+        return (end_lr / start_lr) ** (iteration / num_iteration)
+
+    scheduler = lr_scheduler.LambdaLR(optimizer, lr_lambda)
+    initial_weights = model.state_dict()
+    model.train()
+
+    X_full = torch.Tensor([]).to(device)
+    y_full = torch.Tensor([]).to(device)
+
+    for h in range (0, 5):
+        X, y = next(iter(dataloader))
+        X = X.to(device)
+        y = y.to(device)
+        X_full = torch.cat([X_full, X])
+        y_full = torch.cat([y_full, y])
+
+    for i in range(0, num_iteration):
+        optimizer.zero_grad()
+
+        pred = model(X_full).squeeze()
+        loss = loss_function(pred, y_full.float())
+        lossses.append(loss.item())
+        rates.append(scheduler.get_last_lr()[0])
+        loss.backward()
+        optimizer.step()
+        scheduler.step()
+        model.load_state_dict(initial_weights)
+        if(scheduler.get_last_lr()[0] > end_lr):
+            break
+    return rates, lossses
+
+def plot_lr_find(rates, losses, file_name):
+    fig = plt.Figure()
+    plt.plot(rates, losses)
+    plt.xscale('log')
+    plt.xlabel('learning_rate')
+    plt.ylabel('loss')
+    plt.ylim(0.0, 1.0)
+    plt.title('lr_find_results')
+    plt.legend()
+    plt.savefig(file_name)
+    plt.figure()
+
+pos_weight = torch.Tensor([float(neg) / float(pos)]).to(device)# models class imbalance.
+#rates, losses = custom_lr_find(model, training_dataloader, torch.nn.BCEWithLogitsLoss(pos_weight=pos_weight), device)
+rates, losses = custom_lr_find(model, training_dataloader, torch.nn.BCEWithLogitsLoss(), device)
+plot_lr_find(rates, losses, exp_dir + '/lr_find.jpg')
+writer.add_image('lr_find', load_image(exp_dir + "/lr_find.jpg"), 0)
+writer.flush()
+
+
+# # 6. Using already trained networks: Train the head only
+# 
+# First train the head and freeze all other layers
+
+# In[21]:
+
+
+from torchvision.models import densenet201, DenseNet201_Weights, densenet121, DenseNet121_Weights
+model = densenet201(weights=DenseNet201_Weights.DEFAULT)
+
+for params in model.parameters():
+    params.requires_grad = False
+
+#Replace the last layer (to output a 1d prediction)
+model.classifier = nn.Sequential(nn.Linear(model.classifier.in_features, 1, bias= True))
+
+for param in model.classifier.parameters():
+    param.requires_grad = True
+
+model = model.to(device)
+
+
+# In[22]:
+
+
+#optionnaly load from checkpoint
+"""
+model = torch.load('results/pcam/14_06_2025_10_25_48/model_'+str(19)+'.pt', weights_only = False)
+for params in model.parameters():
+    params.requires_grad = False
+for param in model.classifier.parameters():
+    param.requires_grad = True
+model = model.to(device)
+"""
+
+
+# In[23]:
+
+
+lr = 1e-4
+
+optimizer = torch.optim.Adam(model.parameters(), lr=lr, weight_decay=1e-4)
+#loss_func = torch.nn.BCEWithLogitsLoss(pos_weight=pos_weight)
+loss_func = torch.nn.BCEWithLogitsLoss()
+scheduler = lr_scheduler.StepLR(optimizer, step_size=1000, gamma=0.01)
+
+
+# In[24]:
+
+
+from sklearn.metrics import classification_report, roc_auc_score, f1_score, precision_score, recall_score, accuracy_score, classification_report
+import time
+epoch_num = 2
+sk_learn_metrics_logits = [roc_auc_score]
+sk_learn_metrics_pred = [f1_score, accuracy_score]
+for i in range(0, epoch_num):
+    start_time = time.time()
+    train_res = run_one_epoch(model,
+                  training_dataloader,
+                  optimizer,
+                  loss_func,
+                  scheduler,
+                  device,
+                  writer,
+                  i,
+                  sk_learn_metrics_logits,
+                  sk_learn_metrics_pred)
+    end_time = time.time()
+    print("epoch n°: ", i, " training time : ", end_time-start_time, " sec")
+    start_time = time.time()
+    val_res = eval_model(model, val_dataloader, sk_learn_metrics_logits, sk_learn_metrics_pred, device)
+    for key in train_res.keys():
+        writer.add_scalars(key, {"Train " + key: train_res[key], "Val "+ key : val_res[key]}, i*len(training_dataloader))
+    end_time = time.time()
+    print("epoch n°: ", i, " evaluation time : ", end_time-start_time, " sec")
+    torch.save(model, exp_dir+"/model_" + str(i) + ".pt")
+
+
+# # 7. Using already trained networks: Fine Tune a few layers
+# I did not use it in the end, this is optional
+
+# In[25]:
+
+
+'''
+for name, param in model.features.denseblock4.denselayer32.conv1.named_parameters():
+    param.requires_grad = True
+
+for name, param in model.features.denseblock4.denselayer32.conv2.named_parameters():
+    param.requires_grad = True
+'''
+
+
+# In[26]:
+
+
+# Unfreeze last two blocks (features.6 and features.7)
+'''
+lr = 1e-4
+#optimizer = torch.optim.Adam(model.parameters(), lr=lr, weight_decay=1e-4)
+#loss_func = torch.nn.BCEWithLogitsLoss(pos_weight=pos_weight)
+loss_func = torch.nn.BCEWithLogitsLoss()
+# Use lower LR for fine-tuning
+optimizer = torch.optim.Adam([
+    {"params": model.classifier.parameters(), "lr": 1e-4},
+     {"params": model.features.denseblock4.denselayer32.conv1.parameters(), "lr": 1e-5},
+     {"params": model.features.denseblock4.denselayer32.conv2.parameters(), "lr": 1e-5},
+ ])
+'''
+
+
+# In[27]:
+
+
+'''
+from sklearn.metrics import classification_report, roc_auc_score, f1_score, precision_score, recall_score, accuracy_score, classification_report
+import time
+sk_learn_metrics_logits = [roc_auc_score]
+sk_learn_metrics_pred = [f1_score, accuracy_score]
+epoch_num = 2
+finetune_epoch_num = 6
+for i in range(epoch_num, epoch_num + finetune_epoch_num):
+    start_time = time.time()
+    train_res = run_one_epoch(model,
+                  training_dataloader,
+                  optimizer,
+                  loss_func,
+                  scheduler,
+                  device,
+                  writer,
+                  i,
+                  sk_learn_metrics_logits,
+                  sk_learn_metrics_pred)
+    end_time = time.time()
+    print("epoch n°: ", i, " training time : ", end_time-start_time, " sec")
+    start_time = time.time()
+    val_res = eval_model(model, val_dataloader, sk_learn_metrics_logits, sk_learn_metrics_pred, device)
+    for key in train_res.keys():
+        writer.add_scalars(key, {"Train " + key: train_res[key], "Val "+ key : val_res[key]}, i*len(training_dataloader))
+    end_time = time.time()
+    print("epoch n°: ", i, " evaluation time : ", end_time-start_time, " sec")
+    torch.save(model, exp_dir+"/model_" + str(i) + ".pt")
+
+'''
+
+
+# # 8. Fine tune the entire model
+
+# In[28]:
+
+
+for params in model.parameters():
+    params.requires_grad = True
+
+lr = 1e-5
+optimizer = torch.optim.Adam(model.parameters(), lr=lr, weight_decay=1e-4)
+#loss_func = torch.nn.BCEWithLogitsLoss(pos_weight=pos_weight)
+loss_func = torch.nn.BCEWithLogitsLoss()
+scheduler = lr_scheduler.StepLR(optimizer, step_size=1000, gamma=0.01)
+
+
+# In[ ]:
+
+
+from sklearn.metrics import classification_report, roc_auc_score, f1_score, precision_score, recall_score, accuracy_score, classification_report
+import time
+sk_learn_metrics_logits = [roc_auc_score]
+sk_learn_metrics_pred = [f1_score, accuracy_score]
+epoch_num = 2
+finetune_epoch_num = 4
+for i in range(epoch_num, epoch_num + finetune_epoch_num):
+    start_time = time.time()
+    train_res = run_one_epoch(model,
+                  training_dataloader,
+                  optimizer,
+                  loss_func,
+                  scheduler,
+                  device,
+                  writer,
+                  i,
+                  sk_learn_metrics_logits,
+                  sk_learn_metrics_pred)
+    end_time = time.time()
+    print("epoch n°: ", i, " training time : ", end_time-start_time, " sec")
+    start_time = time.time()
+    val_res = eval_model(model, val_dataloader, sk_learn_metrics_logits, sk_learn_metrics_pred, device)
+    for key in train_res.keys():
+        writer.add_scalars(key, {"Train " + key: train_res[key], "Val "+ key : val_res[key]}, i*len(training_dataloader))
+    end_time = time.time()
+    print("epoch n°: ", i, " evaluation time : ", end_time-start_time, " sec")
+    torch.save(model, exp_dir+"/model_" + str(i) + ".pt")
+
+
+# # 9. Compute test set prediction and submit to kaggle
+# 
+# We will use TTA (Test Time with Augmentation).
+# We can also optionally use several models to make a prediction and average the results
+
+# In[30]:
+
+
+def run_inference(model: nn.Module,
+                     dataloader: DataLoader, 
+                     device: torch.cuda.device):
+
+    model.eval()
+    full_y = torch.Tensor([]).to(device)
+    full_logits = torch.Tensor([]).to(device)
+
+    with torch.no_grad():
+        for X, y in dataloader:
+            X = X.to(device)
+            y = y.to(device)
+            logits = model(X).squeeze()
+
+            full_y = torch.cat([full_y, y])
+            full_logits = torch.cat([full_logits, logits])
+
+    return full_y, full_logits
+
+
+# In[54]:
+
+
+models_paths = ['results/pcam/17_06_2025_12_19_40/model_5.pt']
+
+# First create tta_num augmented dataloaders
+tta_num = 5
+logits = []
+for i in range(0, tta_num):
+    test_set_augment = PcamDatasetKaggle(root=dataset_dir, split="test", transform = deepcopy(transform_data_augment)) #For TTA
+    test_dataloader_augment = DataLoader(test_set_augment, batch_size=batch_size, shuffle=False, pin_memory=True, num_workers=6, persistent_workers = True)
+    for model in models_paths:
+        pcam_model = torch.load(models_paths[0], weights_only = False)
+        pcam_model = pcam_model.to(device)
+        test_y, test_logits = run_inference(pcam_model, test_dataloader, device)
+        logits.append(test_logits)
+        test_y_augm, test_logits_aum = run_inference(pcam_model, test_dataloader_augment, device)
+        logits.append(test_logits_aum)
+
+
+# In[55]:
+
+
+# Average logits
+logits_stacked = torch.stack(logits)
+mean_logits = torch.mean(logits_stacked, dim = 0, keepdims=True)
+
+
+# In[56]:
+
+
+#Create submission file with final predictions
+image_ids = [img.replace('.tif', '') for img in test_set.imgs.tolist()]
+test_preds = torch.sigmoid(mean_logits)
+
+submission_df = pd.DataFrame({
+    'id': image_ids,
+    'label': test_preds.squeeze().detach().cpu().numpy()
+})
+
+submission_df.to_csv(exp_dir+'/submission.csv', index=False)
+
+
+# In[57]:
+
+
+sub_path = exp_dir + '/submission.csv'
+get_ipython().system('kaggle competitions submit -c histopathologic-cancer-detection -f {sub_path} -m "DenseNet201 + correct normalization + no ensemble, no 42*42 crop pytorch "')
+
+
+# # 11. Find best threshold for prediction on validation set
+
+# In[40]:
+
+
+models_paths = ['results/pcam/17_06_2025_12_19_40/model_4.pt']
+pcam_model = torch.load(models_paths[0], weights_only = False)
+pcam_model = pcam_model.to(device)
+test_y, test_logits = run_inference(pcam_model, val_dataloader, device)
+test_y_augment, test_logits_augment = run_inference(pcam_model, val_dataloader_augment, device)
+full_y = torch.cat([test_y, test_y_augment])
+full_logits = torch.cat([test_logits, test_logits_augment])
+
+
+# In[41]:
+
+
+from sklearn.metrics import roc_curve, auc
+fpr, tpr, thresholds = roc_curve(full_y.detach().cpu().numpy(), torch.sigmoid(full_logits).detach().cpu().numpy())
+roc_auc = auc(fpr, tpr)
+
+
+# In[42]:
+
+
+plt.figure(figsize=(8,6))
+plt.plot(fpr, tpr, color='orange', lw=2, label=f'ROC curve (AUC = {roc_auc})')
+plt.xlim([0.0, 1.0])
+plt.ylim([0.0, 1.0])
+plt.xlabel('False Positive Rate')
+plt.ylabel('True Positive Rate')
+plt.title('Receiver Operating Characteristic')
+plt.grid(alpha=0.3)
+plt.show()
+
+
+# In[43]:
+
+
+# Find best threshold index (maximize TPR-FPR).
+j_scores = tpr - fpr
+best_idx = np.argmax(j_scores)
+best_threshold = thresholds[best_idx]
+
+
+# In[44]:
+
+
+best_threshold
+
+
+# In[ ]:
+
+
+
+

app.py

@@ -0,0 +1,98 @@
+import torch
+import torchvision
+from torch import nn
+from torchvision import transforms
+from torchvision.datasets import PCAM
+import gradio as gr
+from PIL import Image
+
+# ---------------------------------
+# 1. Load model
+# ---------------------------------
+torch.manual_seed(42)
+torch.cuda.manual_seed_all(42)
+model = torch.load("results/pcam/17_06_2025_12_19_40/model_5.pt", map_location="cpu", weights_only=False)
+model.eval()
+
+# ---------------------------------
+# 2. Define transform and dataset
+# ---------------------------------
+mean_stdev = [torch.Tensor([0.6981, 0.5428, 0.6933]), torch.Tensor([0.2222, 0.2665, 0.1985])]
+
+models_paths = ['results/pcam/16_06_2025_21_34_05/model_5.pt']
+transform = transforms.Compose([
+    transforms.Resize((224, 224)),
+    transforms.ToTensor(),
+    transforms.Normalize(mean_stdev[0], mean_stdev[1])
+])
+
+test_dataset = PCAM(root="data/", split="val", download=True, transform=transform)
+original_dataset = PCAM(root="data/", split="val", download=True)
+
+# ---------------------------------
+# 3. Prepare choices for dropdown
+# ---------------------------------
+MAX_SAMPLES = 100  # Change to show more
+sample_choices = [f"Sample {i}" for i in range(min(len(test_dataset), MAX_SAMPLES))]
+
+# ---------------------------------
+# 4. Prediction function
+# ---------------------------------
+def get_sample(index: int):
+    index = max(0, min(index, MAX_SAMPLES - 1))  # clamp index
+    image_tensor, ground_truth = test_dataset[index]
+    image_pil, _ = original_dataset[index]  # Untransformed image for display
+
+    with torch.no_grad():
+        output = model(image_tensor.unsqueeze(0)).squeeze()
+        probability = torch.sigmoid(output)
+        predicted_label = "Tumor" if probability >= 0.45 else "No Tumor"
+        true_label = "Tumor" if ground_truth == 1 else "No Tumor"
+        error_label = ""
+        if predicted_label != true_label:
+            error_label = "Error !"
+
+    return image_pil, predicted_label, probability.numpy(), true_label, index, error_label, index
+
+# ---------------------------------
+# 4. Navigation functions
+# ---------------------------------
+def next_sample(index):
+    return get_sample(index + 1)
+
+def prev_sample(index):
+    return get_sample(index - 1)
+
+# ---------------------------------
+# 5. UI elements
+# ---------------------------------
+with gr.Blocks() as demo:
+    gr.Markdown("## 🧬 PCAM Tumor Classifier")
+    gr.Markdown("Use **Next** or **Previous** to browse samples and see model predictions vs ground truth.")
+    gr.Markdown("This is done on the validation set.")
+
+    state = gr.State(0)  # holds current index
+
+    with gr.Row():
+        prev_btn = gr.Button("⬅️ Prev")
+        next_btn = gr.Button("Next ➡️")
+
+    image_output = gr.Image(label="Image")
+    pred_label = gr.Text(label="Predicted")
+    confidence = gr.Text(label="Confidence")
+    true_label = gr.Text(label="Ground Truth")
+    error_label = gr.Text(label="Prediction error")
+    index = gr.Text(label="Image Number")
+
+    # Connect navigation
+    prev_btn.click(fn=prev_sample, inputs=state, outputs=[image_output, pred_label, confidence, true_label, state, error_label, index])
+    next_btn.click(fn=next_sample, inputs=state, outputs=[image_output, pred_label, confidence, true_label, state, error_label, index])
+
+    # Load initial image
+    demo.load(fn=get_sample, inputs=state, outputs=[image_output, pred_label, confidence, true_label, state])
+
+# ---------------------------------
+# 6. Run
+# ---------------------------------
+if __name__ == "__main__":
+    demo.launch()

requirements.txt

@@ -0,0 +1,8 @@
+gradio==5.34.1
+matplotlib==3.10.3
+numpy==2.3.0
+pandas==2.3.0
+Pillow==11.2.1
+scikit_learn==1.7.0
+torch==2.7.0
+torchvision==0.22.0