diff --git a/.gitignore b/.gitignore new file mode 100644 index 0000000..999e295 --- /dev/null +++ b/.gitignore @@ -0,0 +1,107 @@ +# Byte-compiled / optimized / DLL files +__pycache__/ +*.py[cod] +*$py.class + +# C extensions +*.so + +# Distribution / packaging +.Python +env/ +build/ +develop-eggs/ +dist/ +downloads/ +eggs/ +.eggs/ +lib/ +lib64/ +parts/ +sdist/ +var/ +wheels/ +*.egg-info/ +.installed.cfg +*.egg + +# PyInstaller +# Usually these files are written by a python script from a template +# before PyInstaller builds the exe, so as to inject date/other infos into it. +*.manifest +*.spec + +# Installer logs +pip-log.txt +pip-delete-this-directory.txt + +# Unit test / coverage reports +htmlcov/ +.tox/ +.coverage +.coverage.* +.cache +nosetests.xml +coverage.xml +*.cover +.hypothesis/ + +# Translations +*.mo +*.pot + +# Django stuff: +*.log +local_settings.py + +# Flask stuff: +instance/ +.webassets-cache + +# Scrapy stuff: +.scrapy + +# Sphinx documentation +docs/_build/ + +# PyBuilder +target/ + +# Jupyter Notebook +.ipynb_checkpoints + +# pyenv +.python-version + +# celery beat schedule file +celerybeat-schedule + +# SageMath parsed files +*.sage.py + +# dotenv +.env + +# virtualenv +.venv +venv/ +ENV/ + +# Spyder project settings +.spyderproject +.spyproject + +# Rope project settings +.ropeproject + +# mkdocs documentation +/site + +# mypy +.mypy_cache/ + +# Experiments data +xp/ + +# Saved models +*.pkl \ No newline at end of file diff --git a/.gitmodules b/.gitmodules new file mode 100644 index 0000000..067a76d --- /dev/null +++ b/.gitmodules @@ -0,0 +1,3 @@ +[submodule "src/InferSent"] + path = src/InferSent + url = https://github.com/lberrada/InferSent diff --git a/LICENSE b/LICENSE new file mode 100644 index 0000000..94a9ed0 --- /dev/null +++ b/LICENSE @@ -0,0 +1,674 @@ + GNU GENERAL PUBLIC LICENSE + Version 3, 29 June 2007 + + Copyright (C) 2007 Free Software Foundation, Inc. + Everyone is permitted to copy and distribute verbatim copies + of this license document, but changing it is not allowed. + + Preamble + + The GNU General Public License is a free, copyleft license for +software and other kinds of works. + + The licenses for most software and other practical works are designed +to take away your freedom to share and change the works. By contrast, +the GNU General Public License is intended to guarantee your freedom to +share and change all versions of a program--to make sure it remains free +software for all its users. We, the Free Software Foundation, use the +GNU General Public License for most of our software; it applies also to +any other work released this way by its authors. You can apply it to +your programs, too. + + When we speak of free software, we are referring to freedom, not +price. Our General Public Licenses are designed to make sure that you +have the freedom to distribute copies of free software (and charge for +them if you wish), that you receive source code or can get it if you +want it, that you can change the software or use pieces of it in new +free programs, and that you know you can do these things. + + To protect your rights, we need to prevent others from denying you +these rights or asking you to surrender the rights. Therefore, you have +certain responsibilities if you distribute copies of the software, or if +you modify it: responsibilities to respect the freedom of others. + + For example, if you distribute copies of such a program, whether +gratis or for a fee, you must pass on to the recipients the same +freedoms that you received. You must make sure that they, too, receive +or can get the source code. And you must show them these terms so they +know their rights. + + Developers that use the GNU GPL protect your rights with two steps: +(1) assert copyright on the software, and (2) offer you this License +giving you legal permission to copy, distribute and/or modify it. + + For the developers' and authors' protection, the GPL clearly explains +that there is no warranty for this free software. For both users' and +authors' sake, the GPL requires that modified versions be marked as +changed, so that their problems will not be attributed erroneously to +authors of previous versions. + + Some devices are designed to deny users access to install or run +modified versions of the software inside them, although the manufacturer +can do so. This is fundamentally incompatible with the aim of +protecting users' freedom to change the software. The systematic +pattern of such abuse occurs in the area of products for individuals to +use, which is precisely where it is most unacceptable. Therefore, we +have designed this version of the GPL to prohibit the practice for those +products. If such problems arise substantially in other domains, we +stand ready to extend this provision to those domains in future versions +of the GPL, as needed to protect the freedom of users. + + Finally, every program is threatened constantly by software patents. +States should not allow patents to restrict development and use of +software on general-purpose computers, but in those that do, we wish to +avoid the special danger that patents applied to a free program could +make it effectively proprietary. To prevent this, the GPL assures that +patents cannot be used to render the program non-free. + + The precise terms and conditions for copying, distribution and +modification follow. + + TERMS AND CONDITIONS + + 0. Definitions. + + "This License" refers to version 3 of the GNU General Public License. + + "Copyright" also means copyright-like laws that apply to other kinds of +works, such as semiconductor masks. + + "The Program" refers to any copyrightable work licensed under this +License. Each licensee is addressed as "you". "Licensees" and +"recipients" may be individuals or organizations. + + To "modify" a work means to copy from or adapt all or part of the work +in a fashion requiring copyright permission, other than the making of an +exact copy. The resulting work is called a "modified version" of the +earlier work or a work "based on" the earlier work. + + A "covered work" means either the unmodified Program or a work based +on the Program. + + To "propagate" a work means to do anything with it that, without +permission, would make you directly or secondarily liable for +infringement under applicable copyright law, except executing it on a +computer or modifying a private copy. Propagation includes copying, +distribution (with or without modification), making available to the +public, and in some countries other activities as well. + + To "convey" a work means any kind of propagation that enables other +parties to make or receive copies. Mere interaction with a user through +a computer network, with no transfer of a copy, is not conveying. + + An interactive user interface displays "Appropriate Legal Notices" +to the extent that it includes a convenient and prominently visible +feature that (1) displays an appropriate copyright notice, and (2) +tells the user that there is no warranty for the work (except to the +extent that warranties are provided), that licensees may convey the +work under this License, and how to view a copy of this License. If +the interface presents a list of user commands or options, such as a +menu, a prominent item in the list meets this criterion. + + 1. Source Code. + + The "source code" for a work means the preferred form of the work +for making modifications to it. "Object code" means any non-source +form of a work. + + A "Standard Interface" means an interface that either is an official +standard defined by a recognized standards body, or, in the case of +interfaces specified for a particular programming language, one that +is widely used among developers working in that language. + + The "System Libraries" of an executable work include anything, other +than the work as a whole, that (a) is included in the normal form of +packaging a Major Component, but which is not part of that Major +Component, and (b) serves only to enable use of the work with that +Major Component, or to implement a Standard Interface for which an +implementation is available to the public in source code form. A +"Major Component", in this context, means a major essential component +(kernel, window system, and so on) of the specific operating system +(if any) on which the executable work runs, or a compiler used to +produce the work, or an object code interpreter used to run it. + + The "Corresponding Source" for a work in object code form means all +the source code needed to generate, install, and (for an executable +work) run the object code and to modify the work, including scripts to +control those activities. However, it does not include the work's +System Libraries, or general-purpose tools or generally available free +programs which are used unmodified in performing those activities but +which are not part of the work. For example, Corresponding Source +includes interface definition files associated with source files for +the work, and the source code for shared libraries and dynamically +linked subprograms that the work is specifically designed to require, +such as by intimate data communication or control flow between those +subprograms and other parts of the work. + + The Corresponding Source need not include anything that users +can regenerate automatically from other parts of the Corresponding +Source. + + The Corresponding Source for a work in source code form is that +same work. + + 2. Basic Permissions. + + All rights granted under this License are granted for the term of +copyright on the Program, and are irrevocable provided the stated +conditions are met. This License explicitly affirms your unlimited +permission to run the unmodified Program. The output from running a +covered work is covered by this License only if the output, given its +content, constitutes a covered work. This License acknowledges your +rights of fair use or other equivalent, as provided by copyright law. + + You may make, run and propagate covered works that you do not +convey, without conditions so long as your license otherwise remains +in force. You may convey covered works to others for the sole purpose +of having them make modifications exclusively for you, or provide you +with facilities for running those works, provided that you comply with +the terms of this License in conveying all material for which you do +not control copyright. Those thus making or running the covered works +for you must do so exclusively on your behalf, under your direction +and control, on terms that prohibit them from making any copies of +your copyrighted material outside their relationship with you. + + Conveying under any other circumstances is permitted solely under +the conditions stated below. Sublicensing is not allowed; section 10 +makes it unnecessary. + + 3. Protecting Users' Legal Rights From Anti-Circumvention Law. + + No covered work shall be deemed part of an effective technological +measure under any applicable law fulfilling obligations under article +11 of the WIPO copyright treaty adopted on 20 December 1996, or +similar laws prohibiting or restricting circumvention of such +measures. + + When you convey a covered work, you waive any legal power to forbid +circumvention of technological measures to the extent such circumvention +is effected by exercising rights under this License with respect to +the covered work, and you disclaim any intention to limit operation or +modification of the work as a means of enforcing, against the work's +users, your or third parties' legal rights to forbid circumvention of +technological measures. + + 4. Conveying Verbatim Copies. + + You may convey verbatim copies of the Program's source code as you +receive it, in any medium, provided that you conspicuously and +appropriately publish on each copy an appropriate copyright notice; +keep intact all notices stating that this License and any +non-permissive terms added in accord with section 7 apply to the code; +keep intact all notices of the absence of any warranty; and give all +recipients a copy of this License along with the Program. + + You may charge any price or no price for each copy that you convey, +and you may offer support or warranty protection for a fee. + + 5. Conveying Modified Source Versions. + + You may convey a work based on the Program, or the modifications to +produce it from the Program, in the form of source code under the +terms of section 4, provided that you also meet all of these conditions: + + a) The work must carry prominent notices stating that you modified + it, and giving a relevant date. + + b) The work must carry prominent notices stating that it is + released under this License and any conditions added under section + 7. This requirement modifies the requirement in section 4 to + "keep intact all notices". + + c) You must license the entire work, as a whole, under this + License to anyone who comes into possession of a copy. This + License will therefore apply, along with any applicable section 7 + additional terms, to the whole of the work, and all its parts, + regardless of how they are packaged. This License gives no + permission to license the work in any other way, but it does not + invalidate such permission if you have separately received it. + + d) If the work has interactive user interfaces, each must display + Appropriate Legal Notices; however, if the Program has interactive + interfaces that do not display Appropriate Legal Notices, your + work need not make them do so. + + A compilation of a covered work with other separate and independent +works, which are not by their nature extensions of the covered work, +and which are not combined with it such as to form a larger program, +in or on a volume of a storage or distribution medium, is called an +"aggregate" if the compilation and its resulting copyright are not +used to limit the access or legal rights of the compilation's users +beyond what the individual works permit. Inclusion of a covered work +in an aggregate does not cause this License to apply to the other +parts of the aggregate. + + 6. Conveying Non-Source Forms. + + You may convey a covered work in object code form under the terms +of sections 4 and 5, provided that you also convey the +machine-readable Corresponding Source under the terms of this License, +in one of these ways: + + a) Convey the object code in, or embodied in, a physical product + (including a physical distribution medium), accompanied by the + Corresponding Source fixed on a durable physical medium + customarily used for software interchange. + + b) Convey the object code in, or embodied in, a physical product + (including a physical distribution medium), accompanied by a + written offer, valid for at least three years and valid for as + long as you offer spare parts or customer support for that product + model, to give anyone who possesses the object code either (1) a + copy of the Corresponding Source for all the software in the + product that is covered by this License, on a durable physical + medium customarily used for software interchange, for a price no + more than your reasonable cost of physically performing this + conveying of source, or (2) access to copy the + Corresponding Source from a network server at no charge. + + c) Convey individual copies of the object code with a copy of the + written offer to provide the Corresponding Source. This + alternative is allowed only occasionally and noncommercially, and + only if you received the object code with such an offer, in accord + with subsection 6b. + + d) Convey the object code by offering access from a designated + place (gratis or for a charge), and offer equivalent access to the + Corresponding Source in the same way through the same place at no + further charge. You need not require recipients to copy the + Corresponding Source along with the object code. If the place to + copy the object code is a network server, the Corresponding Source + may be on a different server (operated by you or a third party) + that supports equivalent copying facilities, provided you maintain + clear directions next to the object code saying where to find the + Corresponding Source. Regardless of what server hosts the + Corresponding Source, you remain obligated to ensure that it is + available for as long as needed to satisfy these requirements. + + e) Convey the object code using peer-to-peer transmission, provided + you inform other peers where the object code and Corresponding + Source of the work are being offered to the general public at no + charge under subsection 6d. + + A separable portion of the object code, whose source code is excluded +from the Corresponding Source as a System Library, need not be +included in conveying the object code work. + + A "User Product" is either (1) a "consumer product", which means any +tangible personal property which is normally used for personal, family, +or household purposes, or (2) anything designed or sold for incorporation +into a dwelling. In determining whether a product is a consumer product, +doubtful cases shall be resolved in favor of coverage. For a particular +product received by a particular user, "normally used" refers to a +typical or common use of that class of product, regardless of the status +of the particular user or of the way in which the particular user +actually uses, or expects or is expected to use, the product. A product +is a consumer product regardless of whether the product has substantial +commercial, industrial or non-consumer uses, unless such uses represent +the only significant mode of use of the product. + + "Installation Information" for a User Product means any methods, +procedures, authorization keys, or other information required to install +and execute modified versions of a covered work in that User Product from +a modified version of its Corresponding Source. The information must +suffice to ensure that the continued functioning of the modified object +code is in no case prevented or interfered with solely because +modification has been made. + + If you convey an object code work under this section in, or with, or +specifically for use in, a User Product, and the conveying occurs as +part of a transaction in which the right of possession and use of the +User Product is transferred to the recipient in perpetuity or for a +fixed term (regardless of how the transaction is characterized), the +Corresponding Source conveyed under this section must be accompanied +by the Installation Information. But this requirement does not apply +if neither you nor any third party retains the ability to install +modified object code on the User Product (for example, the work has +been installed in ROM). + + The requirement to provide Installation Information does not include a +requirement to continue to provide support service, warranty, or updates +for a work that has been modified or installed by the recipient, or for +the User Product in which it has been modified or installed. Access to a +network may be denied when the modification itself materially and +adversely affects the operation of the network or violates the rules and +protocols for communication across the network. + + Corresponding Source conveyed, and Installation Information provided, +in accord with this section must be in a format that is publicly +documented (and with an implementation available to the public in +source code form), and must require no special password or key for +unpacking, reading or copying. + + 7. Additional Terms. + + "Additional permissions" are terms that supplement the terms of this +License by making exceptions from one or more of its conditions. +Additional permissions that are applicable to the entire Program shall +be treated as though they were included in this License, to the extent +that they are valid under applicable law. If additional permissions +apply only to part of the Program, that part may be used separately +under those permissions, but the entire Program remains governed by +this License without regard to the additional permissions. + + When you convey a copy of a covered work, you may at your option +remove any additional permissions from that copy, or from any part of +it. (Additional permissions may be written to require their own +removal in certain cases when you modify the work.) You may place +additional permissions on material, added by you to a covered work, +for which you have or can give appropriate copyright permission. + + Notwithstanding any other provision of this License, for material you +add to a covered work, you may (if authorized by the copyright holders of +that material) supplement the terms of this License with terms: + + a) Disclaiming warranty or limiting liability differently from the + terms of sections 15 and 16 of this License; or + + b) Requiring preservation of specified reasonable legal notices or + author attributions in that material or in the Appropriate Legal + Notices displayed by works containing it; or + + c) Prohibiting misrepresentation of the origin of that material, or + requiring that modified versions of such material be marked in + reasonable ways as different from the original version; or + + d) Limiting the use for publicity purposes of names of licensors or + authors of the material; or + + e) Declining to grant rights under trademark law for use of some + trade names, trademarks, or service marks; or + + f) Requiring indemnification of licensors and authors of that + material by anyone who conveys the material (or modified versions of + it) with contractual assumptions of liability to the recipient, for + any liability that these contractual assumptions directly impose on + those licensors and authors. + + All other non-permissive additional terms are considered "further +restrictions" within the meaning of section 10. If the Program as you +received it, or any part of it, contains a notice stating that it is +governed by this License along with a term that is a further +restriction, you may remove that term. If a license document contains +a further restriction but permits relicensing or conveying under this +License, you may add to a covered work material governed by the terms +of that license document, provided that the further restriction does +not survive such relicensing or conveying. + + If you add terms to a covered work in accord with this section, you +must place, in the relevant source files, a statement of the +additional terms that apply to those files, or a notice indicating +where to find the applicable terms. + + Additional terms, permissive or non-permissive, may be stated in the +form of a separately written license, or stated as exceptions; +the above requirements apply either way. + + 8. Termination. + + You may not propagate or modify a covered work except as expressly +provided under this License. Any attempt otherwise to propagate or +modify it is void, and will automatically terminate your rights under +this License (including any patent licenses granted under the third +paragraph of section 11). + + However, if you cease all violation of this License, then your +license from a particular copyright holder is reinstated (a) +provisionally, unless and until the copyright holder explicitly and +finally terminates your license, and (b) permanently, if the copyright +holder fails to notify you of the violation by some reasonable means +prior to 60 days after the cessation. + + Moreover, your license from a particular copyright holder is +reinstated permanently if the copyright holder notifies you of the +violation by some reasonable means, this is the first time you have +received notice of violation of this License (for any work) from that +copyright holder, and you cure the violation prior to 30 days after +your receipt of the notice. + + Termination of your rights under this section does not terminate the +licenses of parties who have received copies or rights from you under +this License. If your rights have been terminated and not permanently +reinstated, you do not qualify to receive new licenses for the same +material under section 10. + + 9. Acceptance Not Required for Having Copies. + + You are not required to accept this License in order to receive or +run a copy of the Program. Ancillary propagation of a covered work +occurring solely as a consequence of using peer-to-peer transmission +to receive a copy likewise does not require acceptance. However, +nothing other than this License grants you permission to propagate or +modify any covered work. These actions infringe copyright if you do +not accept this License. Therefore, by modifying or propagating a +covered work, you indicate your acceptance of this License to do so. + + 10. Automatic Licensing of Downstream Recipients. + + Each time you convey a covered work, the recipient automatically +receives a license from the original licensors, to run, modify and +propagate that work, subject to this License. You are not responsible +for enforcing compliance by third parties with this License. + + An "entity transaction" is a transaction transferring control of an +organization, or substantially all assets of one, or subdividing an +organization, or merging organizations. If propagation of a covered +work results from an entity transaction, each party to that +transaction who receives a copy of the work also receives whatever +licenses to the work the party's predecessor in interest had or could +give under the previous paragraph, plus a right to possession of the +Corresponding Source of the work from the predecessor in interest, if +the predecessor has it or can get it with reasonable efforts. + + You may not impose any further restrictions on the exercise of the +rights granted or affirmed under this License. For example, you may +not impose a license fee, royalty, or other charge for exercise of +rights granted under this License, and you may not initiate litigation +(including a cross-claim or counterclaim in a lawsuit) alleging that +any patent claim is infringed by making, using, selling, offering for +sale, or importing the Program or any portion of it. + + 11. Patents. + + A "contributor" is a copyright holder who authorizes use under this +License of the Program or a work on which the Program is based. The +work thus licensed is called the contributor's "contributor version". + + A contributor's "essential patent claims" are all patent claims +owned or controlled by the contributor, whether already acquired or +hereafter acquired, that would be infringed by some manner, permitted +by this License, of making, using, or selling its contributor version, +but do not include claims that would be infringed only as a +consequence of further modification of the contributor version. For +purposes of this definition, "control" includes the right to grant +patent sublicenses in a manner consistent with the requirements of +this License. + + Each contributor grants you a non-exclusive, worldwide, royalty-free +patent license under the contributor's essential patent claims, to +make, use, sell, offer for sale, import and otherwise run, modify and +propagate the contents of its contributor version. + + In the following three paragraphs, a "patent license" is any express +agreement or commitment, however denominated, not to enforce a patent +(such as an express permission to practice a patent or covenant not to +sue for patent infringement). To "grant" such a patent license to a +party means to make such an agreement or commitment not to enforce a +patent against the party. + + If you convey a covered work, knowingly relying on a patent license, +and the Corresponding Source of the work is not available for anyone +to copy, free of charge and under the terms of this License, through a +publicly available network server or other readily accessible means, +then you must either (1) cause the Corresponding Source to be so +available, or (2) arrange to deprive yourself of the benefit of the +patent license for this particular work, or (3) arrange, in a manner +consistent with the requirements of this License, to extend the patent +license to downstream recipients. "Knowingly relying" means you have +actual knowledge that, but for the patent license, your conveying the +covered work in a country, or your recipient's use of the covered work +in a country, would infringe one or more identifiable patents in that +country that you have reason to believe are valid. + + If, pursuant to or in connection with a single transaction or +arrangement, you convey, or propagate by procuring conveyance of, a +covered work, and grant a patent license to some of the parties +receiving the covered work authorizing them to use, propagate, modify +or convey a specific copy of the covered work, then the patent license +you grant is automatically extended to all recipients of the covered +work and works based on it. + + A patent license is "discriminatory" if it does not include within +the scope of its coverage, prohibits the exercise of, or is +conditioned on the non-exercise of one or more of the rights that are +specifically granted under this License. You may not convey a covered +work if you are a party to an arrangement with a third party that is +in the business of distributing software, under which you make payment +to the third party based on the extent of your activity of conveying +the work, and under which the third party grants, to any of the +parties who would receive the covered work from you, a discriminatory +patent license (a) in connection with copies of the covered work +conveyed by you (or copies made from those copies), or (b) primarily +for and in connection with specific products or compilations that +contain the covered work, unless you entered into that arrangement, +or that patent license was granted, prior to 28 March 2007. + + Nothing in this License shall be construed as excluding or limiting +any implied license or other defenses to infringement that may +otherwise be available to you under applicable patent law. + + 12. No Surrender of Others' Freedom. + + If conditions are imposed on you (whether by court order, agreement or +otherwise) that contradict the conditions of this License, they do not +excuse you from the conditions of this License. If you cannot convey a +covered work so as to satisfy simultaneously your obligations under this +License and any other pertinent obligations, then as a consequence you may +not convey it at all. For example, if you agree to terms that obligate you +to collect a royalty for further conveying from those to whom you convey +the Program, the only way you could satisfy both those terms and this +License would be to refrain entirely from conveying the Program. + + 13. Use with the GNU Affero General Public License. + + Notwithstanding any other provision of this License, you have +permission to link or combine any covered work with a work licensed +under version 3 of the GNU Affero General Public License into a single +combined work, and to convey the resulting work. The terms of this +License will continue to apply to the part which is the covered work, +but the special requirements of the GNU Affero General Public License, +section 13, concerning interaction through a network will apply to the +combination as such. + + 14. Revised Versions of this License. + + The Free Software Foundation may publish revised and/or new versions of +the GNU General Public License from time to time. Such new versions will +be similar in spirit to the present version, but may differ in detail to +address new problems or concerns. + + Each version is given a distinguishing version number. If the +Program specifies that a certain numbered version of the GNU General +Public License "or any later version" applies to it, you have the +option of following the terms and conditions either of that numbered +version or of any later version published by the Free Software +Foundation. If the Program does not specify a version number of the +GNU General Public License, you may choose any version ever published +by the Free Software Foundation. + + If the Program specifies that a proxy can decide which future +versions of the GNU General Public License can be used, that proxy's +public statement of acceptance of a version permanently authorizes you +to choose that version for the Program. + + Later license versions may give you additional or different +permissions. However, no additional obligations are imposed on any +author or copyright holder as a result of your choosing to follow a +later version. + + 15. Disclaimer of Warranty. + + THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY +APPLICABLE LAW. EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT +HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY +OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO, +THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR +PURPOSE. THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM +IS WITH YOU. SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF +ALL NECESSARY SERVICING, REPAIR OR CORRECTION. + + 16. Limitation of Liability. + + IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING +WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS +THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY +GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE +USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF +DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD +PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS), +EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF +SUCH DAMAGES. + + 17. Interpretation of Sections 15 and 16. + + If the disclaimer of warranty and limitation of liability provided +above cannot be given local legal effect according to their terms, +reviewing courts shall apply local law that most closely approximates +an absolute waiver of all civil liability in connection with the +Program, unless a warranty or assumption of liability accompanies a +copy of the Program in return for a fee. + + END OF TERMS AND CONDITIONS + + How to Apply These Terms to Your New Programs + + If you develop a new program, and you want it to be of the greatest +possible use to the public, the best way to achieve this is to make it +free software which everyone can redistribute and change under these terms. + + To do so, attach the following notices to the program. It is safest +to attach them to the start of each source file to most effectively +state the exclusion of warranty; and each file should have at least +the "copyright" line and a pointer to where the full notice is found. + + + Copyright (C) + + This program is free software: you can redistribute it and/or modify + it under the terms of the GNU General Public License as published by + the Free Software Foundation, either version 3 of the License, or + (at your option) any later version. + + This program is distributed in the hope that it will be useful, + but WITHOUT ANY WARRANTY; without even the implied warranty of + MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the + GNU General Public License for more details. + + You should have received a copy of the GNU General Public License + along with this program. If not, see . + +Also add information on how to contact you by electronic and paper mail. + + If the program does terminal interaction, make it output a short +notice like this when it starts in an interactive mode: + + Copyright (C) + This program comes with ABSOLUTELY NO WARRANTY; for details type `show w'. + This is free software, and you are welcome to redistribute it + under certain conditions; type `show c' for details. + +The hypothetical commands `show w' and `show c' should show the appropriate +parts of the General Public License. Of course, your program's commands +might be different; for a GUI interface, you would use an "about box". + + You should also get your employer (if you work as a programmer) or school, +if any, to sign a "copyright disclaimer" for the program, if necessary. +For more information on this, and how to apply and follow the GNU GPL, see +. + + The GNU General Public License does not permit incorporating your program +into proprietary programs. If your program is a subroutine library, you +may consider it more useful to permit linking proprietary applications with +the library. If this is what you want to do, use the GNU Lesser General +Public License instead of this License. But first, please read +. diff --git a/README.md b/README.md new file mode 100644 index 0000000..36e512d --- /dev/null +++ b/README.md @@ -0,0 +1,123 @@ +# Deep Frank-Wolfe For Neural Network Optimization + +This repository contains the implementation of the paper [Deep Frank-Wolfe For Neural Network Optimization](https://openreview.net/forum?id=SyVU6s05K7) in pytorch. If you use this work for your research, please cite the paper: + +``` +@Article{berrada2018deep, + author = {Berrada, Leonard and Zisserman, Andrew and Kumar, M Pawan}, + title = {Deep Frank-Wolfe For Neural Network Optimization}, + journal = {Under review}, + year = {2018}, +} +``` + +The DFW algorithm is a first-order optimization algorithm for deep neural networks. To use it for your learning task, consider the two following requirements: +* the loss function has to be convex piecewise linear function (e.g. multi-class SVM [as implemented here](src/losses/hinge.py#L5), or l1 loss) +* the optimizer needs access to the value of the loss function of the current mini-batch [as shown here](src/epoch.py#L31) + +Beside these requirements, the optimizer can be used as plug-and-play, and its independent code is available in [src/optim/dfw.py](src/optim/dfw.py) + +## Requirements + +This code has been tested for pytorch 0.4.1 in python3. Detailed requirements are available in `requirements.txt`. + +## Reproducing the Results + +* To reproduce the CIFAR experiments: `VISION_DATA=[path/to/your/cifar/data] python scripts/reproduce_cifar.py` +* To reproduce the SNLI experiments: follow the [preparation instructions](https://github.com/lberrada/InferSent/tree/c4ded441cf701c256126c5283e4381abb8271792) and run `python scripts/reproduce_snli.py` + +Note that SGD benefits from a hand-designed learning rate schedule. In contrast, all the other optimizers (including DFW) automatically adapt their steps and rely on the tuning of the initial learning rate only. +On average, you should obtain similar results to the ones reported in the paper (there might be some variance on some instances of CIFAR experiments): + +### CIFAR-10: + + + +
Wide Residual Networks Densely Connected Networks
+ +| Optimizer | Test Accuracy (%) | +| --------- | :--------------: | +| Adagrad | 86.07 | +| Adam | 84.86 | +| AMSGrad | 86.08 | +| BPGrad | 88.62 | +| **DFW** | **90.18** | +| SGD | 90.08 | + + + +| Optimizer | Test Accuracy (%) | +| --------- | :--------------: | +| Adagrad | 87.32 | +| Adam | 88.44 | +| AMSGrad | 90.53 | +| **BPGrad**| **90.85** | +| DFW | 90.22 | +| **SGD** | **92.02** | + +
+ +### CIFAR-100: + + + +
Wide Residual Networks Densely Connected Networks
+ +| Optimizer | Test Accuracy (%) | +| --------- | :--------------: | +| Adagrad | 57.64 | +| Adam | 58.46 | +| AMSGrad | 60.73 | +| BPGrad | 60.31 | +| **DFW** | **67.83** | +| SGD | 66.78 | + + + +| Optimizer | Test Accuracy (%) | +| --------- | :--------------: | +| Adagrad | 56.47 | +| Adam | 64.61 | +| AMSGrad | 68.32 | +| BPGrad | 59.36 | +| **DFW** | **69.55** | +| **SGD** | **70.33** | + +
+ +### SNLI: + + + +
CE LossSVM Loss
+ +| Optimizer | Test Accuracy (%) | +| --------- | :--------------: | +| Adagrad | 83.8 | +| Adam | 84.5 | +| AMSGrad | 84.2 | +| BPGrad | 83.6 | +| DFW | - | +| SGD | 84.7 | +| SGD* | 84.5 | + + + +| Optimizer | Test Accuracy (%) | +| --------- | :--------------: | +| Adagrad | 84.6 | +| Adam | 85.0 | +| AMSGrad | 85.1 | +| BPGrad | 84.2 | +| **DFW** | **85.2** | +| **SGD** | **85.2** | +| SGD* | - | + +
+ +## Acknowledgments + +We use the following third-part implementations: +* [InferSent](https://github.com/facebookresearch/InferSent). +* [DenseNets](https://github.com/andreasveit/densenet-pytorch). +* [Wide ResNets](https://github.com/xternalz/WideResNet-pytorch). diff --git a/requirements.txt b/requirements.txt new file mode 100644 index 0000000..7d5c039 --- /dev/null +++ b/requirements.txt @@ -0,0 +1,8 @@ +torch==0.4.1 +numpy +tqdm +torchvision==0.2.1 +git+git://github.com/oval-group/logger@dev +visdom<=0.1.7; python_version < '3.0' +visdom<=0.1.8.4; python_version >= '3.0' +waitGPU==0.0.3 diff --git a/src/InferSent b/src/InferSent new file mode 160000 index 0000000..942a4f5 --- /dev/null +++ b/src/InferSent @@ -0,0 +1 @@ +Subproject commit 942a4f5269d9a51b1495bd82eae898aaa3e0265b diff --git a/src/cli.py b/src/cli.py new file mode 100644 index 0000000..fed7c63 --- /dev/null +++ b/src/cli.py @@ -0,0 +1,167 @@ +import os +import argparse + + +def parse_command(): + parser = argparse.ArgumentParser() + + _add_dataset_parser(parser) + _add_model_parser(parser) + _add_optimization_parser(parser) + _add_loss_parser(parser) + _add_misc_parser(parser) + + args = parser.parse_args() + filter_args(args) + + return args + + +def _add_dataset_parser(parser): + d_parser = parser.add_argument_group(title='Dataset parameters') + d_parser.add_argument('--dataset', + help='dataset') + d_parser.add_argument('--train-size', type=int, default=None, + help="training data size") + d_parser.add_argument('--val-size', type=int, default=None, + help="val data size") + d_parser.add_argument('--test-size', type=int, default=None, + help="test data size") + d_parser.add_argument('--no-data-augmentation', dest='data_aug', + action='store_false', help='no data augmentation') + d_parser.set_defaults(data_aug=True) + + +def _add_model_parser(parser): + m_parser = parser.add_argument_group(title='Model parameters') + m_parser.add_argument('--densenet', dest="densenet", action="store_true", + help="whether to use densenet on CIFAR") + m_parser.add_argument('--wrn', dest="wrn", action="store_true", + help="whether to use wide residual networks on CIFAR") + m_parser.add_argument('--depth', type=int, default=None, + help="depth of network on densenet / wide resnet") + m_parser.add_argument('--width', type=int, default=None, + help="width of network on wide resnet") + m_parser.add_argument('--growth', type=int, default=None, + help="growth rate of densenet") + m_parser.add_argument('--bottleneck', dest="bottleneck", action="store_true", + help="bottleneck on densenet") + m_parser.add_argument('--load-model', default=None, + help='data file with model') + m_parser.set_defaults(pretrained=False, wrn=False, densenet=False, bottleneck=True) + + +def _add_optimization_parser(parser): + o_parser = parser.add_argument_group(title='Training parameters') + o_parser.add_argument('--epochs', type=int, default=None, + help="number of epochs") + o_parser.add_argument('--batch-size', type=int, default=None, + help="batch size") + o_parser.add_argument('--eta', type=float, default=0.1, + help="initial eta / initial learning rate") + o_parser.add_argument('--momentum', type=float, default=0.9, + help="momentum value for SGD") + o_parser.add_argument('--opt', type=str, required=True, + help="optimizer to use") + o_parser.add_argument('--T', type=int, default=[-1], nargs='+', + help="number of epochs between proximal updates / lr decay") + o_parser.add_argument('--decay-factor', type=float, default=0.1, + help="decay factor for the learning rate / proximal term") + o_parser.add_argument('--load-opt', default=None, + help='data file with opt') + + +def _add_loss_parser(parser): + l_parser = parser.add_argument_group(title='Loss parameters') + l_parser.add_argument('--l2', type=float, default=0, + help="l2-regularization") + l_parser.add_argument('--loss', type=str, default='ce', choices=("svm", "ce"), + help="loss function to use ('svm' or 'ce')") + l_parser.add_argument('--smooth-svm', dest="smooth_svm", action="store_true", + help="smooth SVM") + l_parser.set_defaults(smooth_svm=False) + + +def _add_misc_parser(parser): + m_parser = parser.add_argument_group(title='Misc parameters') + m_parser.add_argument('--seed', type=int, default=None, + help="seed for pseudo-randomness") + m_parser.add_argument('--cuda', type=int, default=1, + help="use cuda") + m_parser.add_argument('--no-visdom', dest='visdom', action='store_false', + help='do not use visdom') + m_parser.add_argument('--server', type=str, default=None, + help="server for visdom") + m_parser.add_argument('--port', type=int, default=9014, + help="port for visdom") + m_parser.add_argument('--xp-name', type=str, default=None, + help="name of experiment") + m_parser.add_argument('--no-log', dest='log', action='store_false', + help='do not log results') + m_parser.add_argument('--debug', dest='debug', action='store_true', + help='debug mode') + m_parser.add_argument('--parallel-gpu', dest='parallel_gpu', action='store_true', + help="parallel gpu computation") + m_parser.add_argument('--no-tqdm', dest='tqdm', action='store_false', + help="use of tqdm progress bars") + m_parser.add_argument('--dump-every-epoch', dest='dump_epoch', action='store_true', + help="dump model every epoch") + m_parser.set_defaults(visdom=True, log=True, debug=False, parallel_gpu=False, + tqdm=True, dump_epoch=False) + + +def filter_args(args): + args.T = list(args.T) + + if args.debug: + args.log = args.visdom = False + args.xp_name = '../debug' + if not os.path.exists(args.xp_name): + os.makedirs(args.xp_name) + + if args.log: + # generate automatic experiment name if not provided + if args.xp_name is None: + arch = 'wrn' if args.wrn else 'dn' if args.densenet else 'nn' + args.xp_name = '../xp/{}-{}-{}'.format(args.dataset, arch, args.opt) + assert not os.path.exists(args.xp_name), \ + 'An experiment already exists at {}'.format(os.path.abspath(args.xp_name)) + os.makedirs(args.xp_name) + + if args.visdom: + if args.server is None: + if 'VISDOM_SERVER' in os.environ: + args.server = os.environ['VISDOM_SERVER'] + else: + args.visdom = False + print("Could not find a valid visdom server, de-activating visdom...") + + # default options for densenet + if args.densenet: + if not args.depth: + args.depth = 40 + if not args.growth: + args.growth = 40 + if not args.batch_size: + args.batch_size = 64 + + if args.epochs is None: + args.epochs = 300 + + # default options for wide residual network + if args.wrn: + if not args.depth: + args.depth = 40 + if not args.width: + args.width = 4 + if not args.batch_size: + args.batch_size = 128 + if args.epochs is None: + args.epochs = 200 + + if args.dataset == 'cifar10': + args.n_classes = 10 + elif args.dataset == 'cifar100': + args.n_classes = 100 + elif args.dataset == 'snli': + args.n_classes = 3 diff --git a/src/cuda.py b/src/cuda.py new file mode 100644 index 0000000..6232936 --- /dev/null +++ b/src/cuda.py @@ -0,0 +1,15 @@ +import os +try: + import waitGPU + ngpu = int(os.environ['NGPU']) if 'NGPU' in os.environ else 1 + waitGPU.wait(nproc=0, interval=10, ngpu=ngpu) +except ImportError: + print('Failed to import waitGPU --> no automatic scheduling on GPU') + pass +import torch # import torch *after* waitGPU.wait() + + +def set_cuda(args): + args.cuda = args.cuda and torch.cuda.is_available() + if args.cuda: + torch.zeros(1).cuda() # for quick initialization of process on device diff --git a/src/data/__init__.py b/src/data/__init__.py new file mode 100644 index 0000000..69ac546 --- /dev/null +++ b/src/data/__init__.py @@ -0,0 +1,26 @@ +from data.loaders import loaders_cifar, loaders_mnist, loaders_svhn + + +def get_data_loaders(args): + + print('Dataset: \t {}'.format(args.dataset.upper())) + + # remove values if None + for k in ('train_size', 'val_size', 'test_size'): + if args.__dict__[k] is None: + args.__dict__.pop(k) + + if args.dataset == 'mnist': + loader_train, loader_val, loader_test = loaders_mnist(**vars(args)) + elif 'cifar' in args.dataset: + loader_train, loader_val, loader_test = loaders_cifar(**vars(args)) + elif args.dataset == 'svhn': + loader_train, loader_val, loader_test = loaders_svhn(**vars(args)) + else: + raise NotImplementedError + + args.train_size = len(loader_train.dataset) + args.val_size = len(loader_val.dataset) + args.test_size = len(loader_test.dataset) + + return loader_train, loader_val, loader_test diff --git a/src/data/loaders.py b/src/data/loaders.py new file mode 100644 index 0000000..2344847 --- /dev/null +++ b/src/data/loaders.py @@ -0,0 +1,171 @@ +import os + +import torch.utils.data as data +import torchvision.datasets as datasets +import torchvision.transforms as transforms + +from .utils import random_subsets, Subset + + +def create_loaders(dataset_train, dataset_val, dataset_test, + train_size, val_size, test_size, batch_size, test_batch_size, + cuda, num_workers, split=True): + + kwargs = {'num_workers': num_workers, 'pin_memory': True} if cuda else {} + + if split: + train_indices, val_indices = random_subsets((train_size, val_size), + len(dataset_train), + seed=1234) + else: + train_size = train_size if train_size is not None else len(dataset_train) + train_indices, = random_subsets((train_size,), + len(dataset_train), + seed=1234) + val_size = val_size if val_size is not None else len(dataset_val) + val_indices, = random_subsets((val_size,), + len(dataset_val), + seed=1234) + + test_size = test_size if test_size is not None else len(dataset_test) + test_indices, = random_subsets((test_size,), + len(dataset_test), + seed=1234) + + dataset_train = Subset(dataset_train, train_indices) + dataset_val = Subset(dataset_val, val_indices) + dataset_test = Subset(dataset_test, test_indices) + + print('Dataset sizes: \t train: {} \t val: {} \t test: {}' + .format(len(dataset_train), len(dataset_val), len(dataset_test))) + print('Batch size: \t {}'.format(batch_size)) + + train_loader = data.DataLoader(dataset_train, + batch_size=batch_size, + shuffle=True, **kwargs) + + val_loader = data.DataLoader(dataset_val, + batch_size=test_batch_size, + shuffle=False, **kwargs) + + test_loader = data.DataLoader(dataset_test, + batch_size=test_batch_size, + shuffle=False, **kwargs) + + train_loader.tag = 'train' + val_loader.tag = 'val' + test_loader.tag = 'test' + + return train_loader, val_loader, test_loader + + +def loaders_mnist(dataset, batch_size=64, cuda=0, + train_size=50000, val_size=10000, test_size=10000, + test_batch_size=1000, augment=False, **kwargs): + + assert dataset == 'mnist' + root = '{}/{}'.format(os.environ['VISION_DATA'], dataset) + + # Data loading code + normalize = transforms.Normalize(mean=(0.1307,), + std=(0.3081,)) + + transform = transforms.Compose([transforms.ToTensor(), normalize]) + + # define two datasets in order to have different transforms + # on training and validation + dataset_train = datasets.MNIST(root=root, train=True, transform=transform) + dataset_val = datasets.MNIST(root=root, train=True, transform=transform) + dataset_test = datasets.MNIST(root=root, train=False, transform=transform) + + return create_loaders(dataset_train, dataset_val, + dataset_test, train_size, val_size, test_size, + batch_size=batch_size, + test_batch_size=test_batch_size, + cuda=cuda, num_workers=0) + + +def loaders_cifar(dataset, batch_size, cuda, + train_size=45000, augment=False, val_size=5000, test_size=10000, + test_batch_size=128, **kwargs): + + assert dataset in ('cifar10', 'cifar100') + # assert topk is None or topk == 1, "Top-k not wanted for CIFAR for now" + + root = '{}/{}'.format(os.environ['VISION_DATA'], dataset) + + # Data loading code + mean = [125.3, 123.0, 113.9] + std = [63.0, 62.1, 66.7] + normalize = transforms.Normalize(mean=[x / 255.0 for x in mean], + std=[x / 255.0 for x in std]) + + transform_test = transforms.Compose([ + transforms.ToTensor(), + normalize]) + + if augment: + transform_train = transforms.Compose([ + transforms.RandomCrop(32, padding=4), + transforms.RandomHorizontalFlip(), + transforms.ToTensor(), + normalize]) + else: + transform_train = transform_test + + # define two datasets in order to have different transforms + # on training and validation (no augmentation on validation) + dataset = datasets.CIFAR10 if dataset == 'cifar10' else datasets.CIFAR100 + dataset_train = dataset(root=root, train=True, + transform=transform_train) + dataset_val = dataset(root=root, train=True, + transform=transform_test) + dataset_test = dataset(root=root, train=False, + transform=transform_test) + + return create_loaders(dataset_train, dataset_val, + dataset_test, train_size, val_size, test_size, + batch_size, test_batch_size, cuda, num_workers=4) + + +def loaders_svhn(dataset, batch_size, cuda, + train_size=63257, augment=False, val_size=10000, test_size=26032, + test_batch_size=1000, **kwargs): + + assert dataset == 'svhn' + + root = '{}/{}'.format(os.environ['VISION_DATA'], dataset) + + # Data loading code + mean = [0.4380, 0.4440, 0.4730] + std = [0.1751, 0.1771, 0.1744] + + normalize = transforms.Normalize(mean=mean, + std=std) + + transform_test = transforms.Compose([ + transforms.ToTensor(), + normalize]) + + if augment: + transform_train = transforms.Compose([ + transforms.RandomCrop(32, padding=4), + transforms.RandomHorizontalFlip(), + transforms.ToTensor(), + normalize]) + else: + transform_train = transform_test + + # define two datasets in order to have different transforms + # on training and validation (no augmentation on validation) + dataset = datasets.SVHN + dataset_train = dataset(root=root, split='train', + transform=transform_train) + dataset_val = dataset(root=root, split='train', + transform=transform_test) + dataset_test = dataset(root=root, split='test', + transform=transform_test) + + return create_loaders(dataset_train, dataset_val, + dataset_test, train_size, val_size, test_size, + batch_size, test_batch_size, cuda, num_workers=4) diff --git a/src/data/utils.py b/src/data/utils.py new file mode 100644 index 0000000..281cba6 --- /dev/null +++ b/src/data/utils.py @@ -0,0 +1,61 @@ +import torch +import numpy as np +import torch.utils.data as data + + +class Subset(data.Dataset): + def __init__(self, dataset, indices=None): + """ + Subset of dataset given by indices. + """ + super(Subset, self).__init__() + self.dataset = dataset + self.indices = indices + + if self.indices is None: + self.n_samples = len(self.dataset) + else: + self.n_samples = len(self.indices) + assert self.n_samples >= 0 and \ + self.n_samples <= len(self.dataset), \ + "length of {} incompatible with dataset of size {}"\ + .format(self.n_samples, len(self.dataset)) + + def __getitem__(self, idx): + if torch.is_tensor(idx) and idx.dim(): + res = [self[iidx] for iidx in idx] + return torch.stack([x[0] for x in res]), torch.LongTensor([x[1] for x in res]) + if self.indices is None: + return self.dataset[idx] + else: + return self.dataset[self.indices[idx]] + + def __len__(self): + return self.n_samples + + +def random_subsets(subset_sizes, n_total, seed=None, replace=False): + """ + Return subsets of indices, with sizes given by the iterable + subset_sizes, drawn from {0, ..., n_total - 1} + Subsets may be distinct or not according to the replace option. + Optional seed for deterministic draw. + """ + # save current random state + state = np.random.get_state() + sum_sizes = sum(subset_sizes) + assert sum_sizes <= n_total + + np.random.seed(seed) + + total_subset = np.random.choice(n_total, size=sum_sizes, + replace=replace) + perm = np.random.permutation(total_subset) + res = [] + start = 0 + for size in subset_sizes: + res.append(perm[start: start + size]) + start += size + # restore initial random state + np.random.set_state(state) + return res diff --git a/src/epoch.py b/src/epoch.py new file mode 100644 index 0000000..50d257f --- /dev/null +++ b/src/epoch.py @@ -0,0 +1,89 @@ +import torch + +from tqdm import tqdm +from losses import set_smoothing_enabled +from utils import log_metrics, update_metrics, accuracy, regularization + + +def train(model, loss, optimizer, loader, xp, args): + + model.train() + + xp.Timer_Train.reset() + stats_dict = {} + + for x, y in tqdm(loader, disable=not args.tqdm, desc='Train Epoch', + leave=False, total=len(loader)): + (x, y) = (x.cuda(), y.cuda()) if args.cuda else (x, y) + + # forward pass + scores = model(x) + + # compute the loss function, possibly using smoothing + with set_smoothing_enabled(args.smooth_svm): + loss_value = loss(scores, y) + + # backward pass + optimizer.zero_grad() + loss_value.backward() + + # optimization step + optimizer.step(lambda: float(loss_value)) + + # monitoring + stats_dict['loss'] = float(loss(scores, y)) + stats_dict['acc'] = float(accuracy(scores, y)) + stats_dict['gamma'] = float(optimizer.gamma) + stats_dict['size'] = float(scores.size(0)) + update_metrics(xp, stats_dict) + + xp.Eta.update(optimizer.eta) + xp.Reg.update(regularization(model, args.l2)) + xp.Obj_Train.update(xp.Reg.value + xp.Loss_Train.value) + xp.Timer_Train.update() + + print('\nEpoch: [{0}] (Train) \t' + '({timer:.2f}s) \t' + 'Obj {obj:.3f}\t' + 'Loss {loss:.3f}\t' + 'Acc {acc:.2f}%\t' + .format(int(xp.Epoch.value), + timer=xp.Timer_Train.value, + acc=xp.Acc_Train.value, + obj=xp.Obj_Train.value, + loss=xp.Loss_Train.value)) + + log_metrics(xp) + + +@torch.autograd.no_grad() +def test(model, loader, xp, args): + model.eval() + + Acc = xp.get_metric(name='acc', tag=loader.tag) + Timer = xp.get_metric(name='timer', tag=loader.tag) + Acc.reset() + Timer.reset() + + for x, y in tqdm(loader, disable=not args.tqdm, + desc='{} Epoch'.format(loader.tag.title()), + leave=False, total=len(loader)): + (x, y) = (x.cuda(), y.cuda()) if args.cuda else (x, y) + scores = model(x) + acc = accuracy(scores, y) + Acc.update(acc, n=x.size(0)) + + Timer.update().log() + Acc.log() + print('Epoch: [{0}] ({tag})\t' + '({timer:.2f}s) \t' + 'Obj ----\t' + 'Loss ----\t' + 'Acc {acc:.2f}% \t' + .format(int(xp.Epoch.value), + tag=loader.tag.title(), + timer=Timer.value, + acc=Acc.value)) + + if loader.tag == 'val': + xp.Acc_Valbest.update(xp.Acc_Val.value).log() diff --git a/src/losses/__init__.py b/src/losses/__init__.py new file mode 100644 index 0000000..85fa3ca --- /dev/null +++ b/src/losses/__init__.py @@ -0,0 +1,20 @@ +import torch.nn as nn +from losses.hinge import MultiClassHingeLoss, set_smoothing_enabled + + +def get_loss(args): + if args.loss == 'svm': + loss_fn = MultiClassHingeLoss() + elif args.loss == 'ce': + loss_fn = nn.CrossEntropyLoss() + else: + raise ValueError + + print('L2 regularization: \t {}'.format(args.l2)) + print('\nLoss function:') + print(loss_fn) + + if args.cuda: + loss_fn = loss_fn.cuda() + + return loss_fn diff --git a/src/losses/hinge.py b/src/losses/hinge.py new file mode 100644 index 0000000..28573ad --- /dev/null +++ b/src/losses/hinge.py @@ -0,0 +1,75 @@ +import torch +import torch.nn as nn + + +class MultiClassHingeLoss(nn.Module): + r"""Creates a criterion that optimizes a multi-class classification hinge + loss (margin-based loss) between input `x` (a 2D mini-batch `Tensor`) and + output `y` (which is a 1D tensor of target class indices, + :math:`0 \leq y \leq \text{x.size}(1)`): + + This implements a Crammer & Singer formulation, which penalizes the maximal margin violation. + Note that `torch.nn.MultiMarginLoss` uses the Weston & Watkins formulation, + which penalizes the sum of the margin violations and performs significantly worse in our experience. + """ + + smooth = False + + def __init__(self): + super(MultiClassHingeLoss, self).__init__() + self.smooth = False + self._range = None + + def forward(self, x, y): + aug = self._augmented_scores(x, y) + xi = self._compute_xi(x, aug, y) + + loss = torch.sum(aug * xi) / x.size(0) + return loss + + def _augmented_scores(self, s, y): + if self._range is None: + delattr(self, '_range') + self.register_buffer('_range', torch.arange(s.size(1), device=s.device)[None, :]) + + delta = torch.ne(y[:, None], self._range).detach().float() + return s + delta - s.gather(1, y[:, None]) + + @torch.autograd.no_grad() + def _compute_xi(self, s, aug, y): + + # find argmax of augmented scores + _, y_star = torch.max(aug, 1) + # xi_max: one-hot encoding of maximal indices + xi_max = torch.eq(y_star[:, None], self._range).float() + + if MultiClassHingeLoss.smooth: + # find smooth argmax of scores + xi_smooth = nn.functional.softmax(s, dim=1) + # compute for each sample whether it has a positive contribution to the loss + losses = torch.sum(xi_smooth * aug, 1) + mask_smooth = torch.ge(losses, 0).float()[:, None] + # keep only smoothing for positive contributions + xi = mask_smooth * xi_smooth + (1 - mask_smooth) * xi_max + else: + xi = xi_max + + return xi + + def __repr__(self): + return 'MultiClassHingeLoss()' + + +class set_smoothing_enabled(object): + def __init__(self, mode): + """ + Context-manager similar to the torch.set_grad_enabled(mode). + Within the scope of the manager the MultiClassHingeLoss is smoothed (it is not otherwise). + """ + MultiClassHingeLoss.smooth = bool(mode) + + def __enter__(self): + pass + + def __exit__(self, exc_type, exc_val, exc_tb): + MultiClassHingeLoss.smooth = False diff --git a/src/main.py b/src/main.py new file mode 100644 index 0000000..ccdd08a --- /dev/null +++ b/src/main.py @@ -0,0 +1,42 @@ +# top-import for cuda device initialization +from cuda import set_cuda + +import logger + +from cli import parse_command +from losses import get_loss +from utils import get_xp, set_seed +from data import get_data_loaders +from models import get_model, load_best_model +from optim import get_optimizer, decay_optimizer +from epoch import train, test + + +def main(args): + + set_cuda(args) + set_seed(args) + + loader_train, loader_val, loader_test = get_data_loaders(args) + loss = get_loss(args) + model = get_model(args) + optimizer = get_optimizer(args, parameters=model.parameters()) + xp = get_xp(args, model, optimizer) + + for i in range(args.epochs): + xp.Epoch.update(1).log() + + train(model, loss, optimizer, loader_train, xp, args) + test(model, loader_val, xp, args) + + if (i + 1) in args.T: + decay_optimizer(optimizer, args.decay_factor) + + load_best_model(model, xp) + test(model, loader_test, xp, args) + + +if __name__ == '__main__': + args = parse_command() + with logger.stdout_to("{}/log.txt".format(args.xp_name)): + main(args) diff --git a/src/models/__init__.py b/src/models/__init__.py new file mode 100644 index 0000000..a9574ac --- /dev/null +++ b/src/models/__init__.py @@ -0,0 +1,51 @@ +import os +import torch +from .densenet import DenseNet3 +from .wide_resnet import WideResNet +from collections import OrderedDict + + +def get_model(args): + assert args.dataset in ('cifar10', 'cifar100') + + if args.densenet: + model = DenseNet3(args.depth, args.n_classes, args.growth, bottleneck=bool(args.bottleneck)) + elif args.wrn: + model = WideResNet(args.depth, args.n_classes, args.width) + else: + raise NotImplementedError + + if args.load_model: + state = torch.load(args.load_model)['model'] + new_state = OrderedDict() + for k in state: + # naming convention for data parallel + if 'module' in k: + v = state[k] + new_state[k.replace('module.', '')] = v + else: + new_state[k] = state[k] + model.load_state_dict(new_state) + print('Loaded model from {}'.format(args.load_model)) + + # Number of model parameters + args.nparams = sum([p.data.nelement() for p in model.parameters()]) + print('Number of model parameters: {}'.format(args.nparams)) + + if args.cuda: + if args.parallel_gpu: + model = torch.nn.DataParallel(model).cuda() + else: + model = model.cuda() + + return model + + +def load_best_model(model, xp): + best_model_file = '{}/best_model.pkl'.format(xp.name_and_dir) + if os.path.exists(best_model_file): + best_model_state = torch.load(best_model_file)['model'] + model.load_state_dict(best_model_state) + print('Loaded best model from {}'.format(best_model_file)) + else: + print('Could not find best model') diff --git a/src/models/densenet.py b/src/models/densenet.py new file mode 100644 index 0000000..e4a84d5 --- /dev/null +++ b/src/models/densenet.py @@ -0,0 +1,171 @@ +""" +Code from https://github.com/andreasveit/densenet-pytorch/blob/master/densenet.py + +BSD 3-Clause License + +Copyright (c) 2017, Andreas Veit +All rights reserved. + +Redistribution and use in source and binary forms, with or without +modification, are permitted provided that the following conditions are met: + +* Redistributions of source code must retain the above copyright notice, this + list of conditions and the following disclaimer. + +* Redistributions in binary form must reproduce the above copyright notice, + this list of conditions and the following disclaimer in the documentation + and/or other materials provided with the distribution. + +* Neither the name of the copyright holder nor the names of its + contributors may be used to endorse or promote products derived from + this software without specific prior written permission. + +THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" +AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE +IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE +DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE +FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL +DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR +SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER +CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, +OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE +OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. +""" + +import math +import torch +import torch.nn as nn +import torch.nn.functional as F + + +class BasicBlock(nn.Module): + def __init__(self, in_planes, out_planes, dropRate=0.0): + super(BasicBlock, self).__init__() + self.bn1 = nn.BatchNorm2d(in_planes) + self.relu = nn.ReLU(inplace=True) + self.conv1 = nn.Conv2d(in_planes, out_planes, kernel_size=3, stride=1, + padding=1, bias=False) + self.droprate = dropRate + + def forward(self, x): + out = self.conv1(self.relu(self.bn1(x))) + if self.droprate > 0: + out = F.dropout(out, p=self.droprate, training=self.training) + return torch.cat([x, out], 1) + + +class BottleneckBlock(nn.Module): + def __init__(self, in_planes, out_planes, dropRate=0.0): + super(BottleneckBlock, self).__init__() + inter_planes = out_planes * 4 + self.bn1 = nn.BatchNorm2d(in_planes) + self.relu = nn.ReLU(inplace=True) + self.conv1 = nn.Conv2d(in_planes, inter_planes, kernel_size=1, + stride=1, padding=0, bias=False) + self.bn2 = nn.BatchNorm2d(inter_planes) + self.conv2 = nn.Conv2d(inter_planes, out_planes, kernel_size=3, + stride=1, padding=1, bias=False) + self.droprate = dropRate + + def forward(self, x): + out = self.conv1(self.relu(self.bn1(x))) + if self.droprate > 0: + out = F.dropout(out, p=self.droprate, inplace=False, + training=self.training) + out = self.conv2(self.relu(self.bn2(out))) + if self.droprate > 0: + out = F.dropout(out, p=self.droprate, inplace=False, + training=self.training) + return torch.cat([x, out], 1) + + +class TransitionBlock(nn.Module): + def __init__(self, in_planes, out_planes, dropRate=0.0): + super(TransitionBlock, self).__init__() + self.bn1 = nn.BatchNorm2d(in_planes) + self.relu = nn.ReLU(inplace=True) + self.conv1 = nn.Conv2d(in_planes, out_planes, kernel_size=1, stride=1, + padding=0, bias=False) + self.droprate = dropRate + + def forward(self, x): + out = self.conv1(self.relu(self.bn1(x))) + if self.droprate > 0: + out = F.dropout(out, p=self.droprate, inplace=False, + training=self.training) + return F.avg_pool2d(out, 2) + + +class DenseBlock(nn.Module): + def __init__(self, nb_layers, in_planes, growth_rate, block, dropRate=0.0): + super(DenseBlock, self).__init__() + self.layer = self._make_layer(block, in_planes, growth_rate, + nb_layers, dropRate) + + def _make_layer(self, block, in_planes, growth_rate, nb_layers, dropRate): + layers = [] + for i in range(nb_layers): + layers.append(block(in_planes + i * growth_rate, + growth_rate, dropRate)) + return nn.Sequential(*layers) + + def forward(self, x): + return self.layer(x) + + +class DenseNet3(nn.Module): + def __init__(self, depth, num_classes, growth_rate=12, + reduction=0.5, bottleneck=True, dropRate=0.0): + super(DenseNet3, self).__init__() + in_planes = 2 * growth_rate + n = (depth - 4) / 3 + + if bottleneck is True: + n = n / 2 + block = BottleneckBlock + else: + block = BasicBlock + n = int(n) + # 1st conv before any dense block + self.conv1 = nn.Conv2d(3, in_planes, kernel_size=3, stride=1, + padding=1, bias=False) + # 1st block + self.block1 = DenseBlock(n, in_planes, growth_rate, block, dropRate) + in_planes = int(in_planes + n * growth_rate) + self.trans1 = TransitionBlock(in_planes, + int(math.floor(in_planes * reduction)), + dropRate=dropRate) + in_planes = int(math.floor(in_planes * reduction)) + # 2nd block + self.block2 = DenseBlock(n, in_planes, growth_rate, block, dropRate) + in_planes = int(in_planes + n * growth_rate) + self.trans2 = TransitionBlock(in_planes, + int(math.floor(in_planes * reduction)), + dropRate=dropRate) + in_planes = int(math.floor(in_planes * reduction)) + # 3rd block + self.block3 = DenseBlock(n, in_planes, growth_rate, block, dropRate) + in_planes = int(in_planes + n * growth_rate) + # global average pooling and classifier + self.bn1 = nn.BatchNorm2d(in_planes) + self.relu = nn.ReLU(inplace=True) + self.fc = nn.Linear(in_planes, num_classes, bias=False) + self.in_planes = in_planes + + for m in self.modules(): + if isinstance(m, nn.Conv2d): + n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels + m.weight.data.normal_(0, math.sqrt(2. / n)) + elif isinstance(m, nn.BatchNorm2d): + m.weight.data.fill_(1) + m.bias.data.zero_() + + def forward(self, x): + out = self.conv1(x) + out = self.trans1(self.block1(out)) + out = self.trans2(self.block2(out)) + out = self.block3(out) + out = self.relu(self.bn1(out)) + out = F.avg_pool2d(out, 8) + out = out.view(-1, self.in_planes) + return self.fc(out) diff --git a/src/models/wide_resnet.py b/src/models/wide_resnet.py new file mode 100644 index 0000000..34b35ec --- /dev/null +++ b/src/models/wide_resnet.py @@ -0,0 +1,95 @@ +""" +Code from https://github.com/xternalz/WideResNet-pytorch +""" + +import math +import torch +import torch.nn as nn +import torch.nn.functional as F + + +class BasicBlock(nn.Module): + def __init__(self, in_planes, out_planes, stride, dropRate=0.0): + super(BasicBlock, self).__init__() + self.bn1 = nn.BatchNorm2d(in_planes) + self.relu1 = nn.ReLU(inplace=True) + self.conv1 = nn.Conv2d(in_planes, out_planes, kernel_size=3, stride=stride, + padding=1, bias=False) + self.bn2 = nn.BatchNorm2d(out_planes) + self.relu2 = nn.ReLU(inplace=True) + self.conv2 = nn.Conv2d(out_planes, out_planes, kernel_size=3, stride=1, + padding=1, bias=False) + self.droprate = dropRate + self.equalInOut = (in_planes == out_planes) + self.convShortcut = (not self.equalInOut) and \ + nn.Conv2d(in_planes, out_planes, kernel_size=1, stride=stride, padding=0, bias=False) or None + + def forward(self, x): + if not self.equalInOut: + x = self.relu1(self.bn1(x)) + else: + out = self.relu1(self.bn1(x)) + out = self.relu2(self.bn2(self.conv1(out if self.equalInOut else x))) + if self.droprate > 0: + out = F.dropout(out, p=self.droprate, training=self.training) + out = self.conv2(out) + return torch.add(x if self.equalInOut else self.convShortcut(x), out) + + +class NetworkBlock(nn.Module): + def __init__(self, nb_layers, in_planes, out_planes, block, stride, dropRate=0.0): + super(NetworkBlock, self).__init__() + self.layer = self._make_layer(block, in_planes, out_planes, nb_layers, stride, dropRate) + + def _make_layer(self, block, in_planes, out_planes, nb_layers, stride, dropRate): + layers = [] + for i in range(nb_layers): + layers.append(block(i == 0 and in_planes or out_planes, out_planes, + i == 0 and stride or 1, dropRate)) + return nn.Sequential(*layers) + + def forward(self, x): + return self.layer(x) + + +class WideResNet(nn.Module): + def __init__(self, depth, num_classes, widen_factor=1, dropRate=0.0): + super(WideResNet, self).__init__() + nChannels = [16, 16 * widen_factor, 32 * widen_factor, 64 * widen_factor] + assert((depth - 4) % 6 == 0) + n = (depth - 4) // 6 + block = BasicBlock + # 1st conv before any network block + self.conv1 = nn.Conv2d(3, nChannels[0], kernel_size=3, stride=1, + padding=1, bias=False) + # 1st block + self.block1 = NetworkBlock(n, nChannels[0], nChannels[1], block, 1, dropRate) + # 2nd block + self.block2 = NetworkBlock(n, nChannels[1], nChannels[2], block, 2, dropRate) + # 3rd block + self.block3 = NetworkBlock(n, nChannels[2], nChannels[3], block, 2, dropRate) + # global average pooling and classifier + self.bn1 = nn.BatchNorm2d(nChannels[3]) + self.relu = nn.ReLU(inplace=True) + self.fc = nn.Linear(nChannels[3], num_classes) + self.nChannels = nChannels[3] + + for m in self.modules(): + if isinstance(m, nn.Conv2d): + n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels + m.weight.data.normal_(0, math.sqrt(2. / n)) + elif isinstance(m, nn.BatchNorm2d): + m.weight.data.fill_(1) + m.bias.data.zero_() + elif isinstance(m, nn.Linear): + m.bias.data.zero_() + + def forward(self, x): + out = self.conv1(x) + out = self.block1(out) + out = self.block2(out) + out = self.block3(out) + out = self.relu(self.bn1(out)) + out = F.avg_pool2d(out, 8) + out = out.view(-1, self.nChannels) + return self.fc(out) diff --git a/src/optim/__init__.py b/src/optim/__init__.py new file mode 100644 index 0000000..01ee17e --- /dev/null +++ b/src/optim/__init__.py @@ -0,0 +1,53 @@ +import torch.optim + +from .bpgrad import BPGrad +from .dfw import DFW + + +def get_optimizer(args, parameters): + """ + Available optimizers: + - SGD + - Adam + - Adagrad + - AMSGrad + - DFW + - BPGrad + """ + if args.opt == 'sgd': + optimizer = torch.optim.SGD(parameters, lr=args.eta, weight_decay=args.l2, + momentum=args.momentum, nesterov=bool(args.momentum)) + elif args.opt == "adam": + optimizer = torch.optim.Adam(parameters, lr=args.eta, weight_decay=args.l2) + elif args.opt == "adagrad": + optimizer = torch.optim.Adagrad(parameters, lr=args.eta, weight_decay=args.l2) + elif args.opt == "amsgrad": + optimizer = torch.optim.Adam(parameters, lr=args.eta, weight_decay=args.l2, amsgrad=True) + elif args.opt == 'dfw': + optimizer = DFW(parameters, eta=args.eta, momentum=args.momentum, weight_decay=args.l2) + elif args.opt == 'bpgrad': + optimizer = BPGrad(parameters, eta=args.eta, momentum=args.momentum, weight_decay=args.l2) + else: + raise ValueError(args.opt) + + print("Optimizer: \t {}".format(args.opt.upper())) + + optimizer.gamma = 1 + optimizer.eta = args.eta + + if args.load_opt: + state = torch.load(args.load_opt)['optimizer'] + optimizer.load_state_dict(state) + print('Loaded optimizer from {}'.format(args.load_opt)) + + return optimizer + + +def decay_optimizer(optimizer, decay_factor=0.1): + if isinstance(optimizer, torch.optim.SGD): + for param_group in optimizer.param_groups: + param_group['lr'] *= decay_factor + # update state + optimizer.eta = optimizer.param_groups[0]['lr'] + else: + raise ValueError diff --git a/src/optim/bpgrad.py b/src/optim/bpgrad.py new file mode 100644 index 0000000..d30df87 --- /dev/null +++ b/src/optim/bpgrad.py @@ -0,0 +1,86 @@ +import torch +import torch.optim as optim + +from torch.optim.optimizer import required + + +class BPGrad(optim.Optimizer): + r""" + Implements BPGrad: https://arxiv.org/abs/1711.06959. + + Args: + params (iterable): iterable of parameters to optimize or dicts defining + parameter groups + eta (float): initial learning rate + momentum (float, optional): momentum factor (default: 0) + weight_decay (float, optional): weight decay (L2 penalty) (default: 0) + eps (float, optional): small constant for numerical stability (default: 1e-5) + + Example: + >>> optimizer = BPGrad(model.parameters(), eta=1, momentum=0.9, weight_decay=1e-4) + >>> optimizer.zero_grad() + >>> loss_value = loss_fn(model(input), target) + >>> loss_value.backward() + >>> optimizer.step(lambda: float(loss_value)) + + .. note:: + In order to compute the step-size, it requires a closure at every step + that gives the current value of the objective function. + + Implementation notes: simplification of Algorithm 1 from https://arxiv.org/abs/1711.06959. + The authors recommend to use N=1 in practice. + This implies m=1 and therefore rho=0. + The update on `v` can thus be simplified to `v_{t+1} = mu v_t - (f(x_t) / (L * ||g_t||_2)) * g_t`, + where g_t is the (stochastic) gradient of f at x=x_t. + + For more details, see: + https://arxiv.org/abs/1711.06959. + """ + + def __init__(self, params, eta=required, momentum=0, weight_decay=0, eps=1e-5): + if eta is not required and eta <= 0.0: + raise ValueError("Invalid eta: {}".format(eta)) + if momentum < 0.0: + raise ValueError("Invalid momentum value: {}".format(momentum)) + if weight_decay < 0.0: + raise ValueError("Invalid weight_decay value: {}".format(weight_decay)) + + defaults = dict(eta=eta, momentum=momentum, weight_decay=weight_decay) + super(BPGrad, self).__init__(params, defaults) + self.eps = eps + + for group in self.param_groups: + group['L'] = 1. / group['eta'] + if group['momentum']: + for p in group['params']: + self.state[p]['v'] = torch.zeros_like(p.data, requires_grad=False) + + @torch.autograd.no_grad() + def step(self, closure): + + obj = float(closure()) + + for group in self.param_groups: + wd = group['weight_decay'] + if wd: + for p in group['params']: + obj += 0.5 * wd * p.data.norm() ** 2 + p.grad.data += wd * p.data + + grad_sqrd_norm = 0 + for group in self.param_groups: + for p in group['params']: + grad_sqrd_norm += p.grad.data.norm() ** 2 + + step_size = float(obj / (torch.sqrt(grad_sqrd_norm) + self.eps)) + + for group in self.param_groups: + L = group['L'] + mu = group['momentum'] + for p in group['params']: + v = self.state[p]['v'] + v *= mu + v -= step_size / L * p.grad.data + p.data += v + + self.gamma = step_size diff --git a/src/optim/dfw.py b/src/optim/dfw.py new file mode 100644 index 0000000..8dbd9ae --- /dev/null +++ b/src/optim/dfw.py @@ -0,0 +1,101 @@ +import torch +import torch.optim as optim + +from torch.optim.optimizer import required +from collections import defaultdict + + +class DFW(optim.Optimizer): + r""" + Implements Deep Frank Wolfe: https://openreview.net/forum?id=SyVU6s05K7. + Nesterov momentum is the *standard formula*, and differs + from pytorch NAG implementation. + + Args: + params (iterable): iterable of parameters to optimize or dicts defining + parameter groups + eta (float): initial learning rate + momentum (float, optional): momentum factor (default: 0) + weight_decay (float, optional): weight decay (L2 penalty) (default: 0) + eps (float, optional): small constant for numerical stability (default: 1e-5) + + Example: + >>> optimizer = DFW(model.parameters(), eta=1, momentum=0.9, weight_decay=1e-4) + >>> optimizer.zero_grad() + >>> loss_value = loss_fn(model(input), target) + >>> loss_value.backward() + >>> optimizer.step(lambda: float(loss_value)) + + .. note:: + This optimizer has been designed for convex piecewise linear loss functions only, + and should be used accordingly. + + In order to compute the step-size, it requires a closure at every step + that gives the current value of the loss function (without the regularization). + + For more details, see: + https://openreview.net/forum?id=SyVU6s05K7. + """ + + def __init__(self, params, eta=required, momentum=0, weight_decay=0, eps=1e-5): + if eta is not required and eta <= 0.0: + raise ValueError("Invalid eta: {}".format(eta)) + if momentum < 0.0: + raise ValueError("Invalid momentum value: {}".format(momentum)) + if weight_decay < 0.0: + raise ValueError("Invalid weight_decay value: {}".format(weight_decay)) + + defaults = dict(eta=eta, momentum=momentum, weight_decay=weight_decay) + super(DFW, self).__init__(params, defaults) + self.eps = eps + + for group in self.param_groups: + if group['momentum']: + for p in group['params']: + self.state[p]['momentum_buffer'] = torch.zeros_like(p.data, requires_grad=False) + + @torch.autograd.no_grad() + def step(self, closure): + loss = float(closure()) + + w_dict = defaultdict(dict) + for group in self.param_groups: + wd = group['weight_decay'] + for param in group['params']: + w_dict[param]['delta_t'] = param.grad.data + w_dict[param]['r_t'] = wd * param.data + + self._line_search(loss, w_dict) + + for group in self.param_groups: + eta = group['eta'] + mu = group['momentum'] + for param in group['params']: + state = self.state[param] + delta_t, r_t = w_dict[param]['delta_t'], w_dict[param]['r_t'] + + param.data -= eta * (r_t + self.gamma * delta_t) + + if mu: + z_t = state['momentum_buffer'] + z_t *= mu + z_t -= eta * self.gamma * (delta_t + r_t) + param.data += mu * z_t + + @torch.autograd.no_grad() + def _line_search(self, loss, w_dict): + """ + Computes the line search in closed form. + """ + + num = loss + denom = 0 + + for group in self.param_groups: + eta = group['eta'] + for param in group['params']: + delta_t, r_t = w_dict[param]['delta_t'], w_dict[param]['r_t'] + num -= eta * torch.sum(delta_t * r_t) + denom += eta * delta_t.norm() ** 2 + + self.gamma = float((num / (denom + self.eps)).clamp(min=0, max=1)) diff --git a/src/scripts/__init__.py b/src/scripts/__init__.py new file mode 100644 index 0000000..e69de29 diff --git a/src/scripts/reproduce_cifar.py b/src/scripts/reproduce_cifar.py new file mode 100644 index 0000000..ba16679 --- /dev/null +++ b/src/scripts/reproduce_cifar.py @@ -0,0 +1,82 @@ +from scheduling import launch + + +jobs = [ + + # SGD-CIFAR-10-WRN + """python main.py --dataset cifar10 --wrn --opt sgd --l2 5e-4 --eta 0.1 --T 60 120 160 --decay-factor 0.2 --no-tqdm""", + + # SGD-CIFAR-100-WRN + """python main.py --dataset cifar100 --wrn --opt sgd --l2 5e-4 --eta 0.1 --T 60 120 160 --decay-factor 0.2 --no-tqdm""", + + # SGD-CIFAR-10-DN + """python main.py --dataset cifar10 --densenet --opt sgd --l2 1e-4 --eta 0.1 --T 150 225 --decay-factor 0.1 --no-tqdm""", + + # SGD-CIFAR-100-DN + """python main.py --dataset cifar100 --densenet --opt sgd --l2 1e-4 --eta 0.1 --T 150 225 --decay-factor 0.1 --no-tqdm""", + + # DFW-CIFAR-10-WRN + """python main.py --dataset cifar10 --wrn --opt dfw --l2 1e-4 --eta 1. --loss svm --smooth --no-tqdm""", + + # DFW-CIFAR-100-WRN + """python main.py --dataset cifar100 --wrn --opt dfw --l2 1e-4 --eta 1. --loss svm --smooth --no-tqdm""", + + # DFW-CIFAR-10-DN + """python main.py --dataset cifar10 --densenet --opt dfw --l2 1e-4 --eta 0.1 --loss svm --smooth --no-tqdm""", + + # DFW-CIFAR-100-DN + """python main.py --dataset cifar100 --densenet --opt dfw --l2 1e-4 --eta 0.1 --loss svm --smooth --no-tqdm""", + + # ADAM-CIFAR-10-WRN + """python main.py --dataset cifar10 --wrn --opt adam --l2 1e-4 --eta 1e-3 --no-tqdm""", + + # ADAM-CIFAR-100-WRN + """python main.py --dataset cifar100 --wrn --opt adam --l2 1e-4 --eta 1e-3 --no-tqdm""", + + # ADAM-CIFAR-10-DN + """python main.py --dataset cifar10 --densenet --opt adam --l2 1e-4 --eta 1e-3 --no-tqdm""", + + # ADAM-CIFAR-100-DN + """python main.py --dataset cifar100 --densenet --opt adam --l2 1e-4 --eta 1e-3 --no-tqdm""", + + # ADAGRAD-CIFAR-10-WRN + """python main.py --dataset cifar10 --wrn --opt adagrad --l2 5e-4 --eta 1e-2 --no-tqdm""", + + # ADAGRAD-CIFAR-100-WRN + """python main.py --dataset cifar100 --wrn --opt adagrad --l2 1e-4 --eta 1e-2 --no-tqdm""", + + # ADAGRAD-CIFAR-10-DN + """python main.py --dataset cifar10 --densenet --opt adagrad --l2 1e-4 --eta 1e-2 --no-tqdm""", + + # ADAGRAD-CIFAR-100-DN + """python main.py --dataset cifar100 --densenet --opt adagrad --l2 1e-4 --eta 1e-2 --no-tqdm""", + + # AMSGRAD-CIFAR-10-WRN + """python main.py --dataset cifar10 --wrn --opt amsgrad --l2 1e-4 --eta 1e-3 --no-tqdm""", + + # AMSGRAD-CIFAR-100-WRN + """python main.py --dataset cifar100 --wrn --opt amsgrad --l2 1e-4 --eta 1e-3 --no-tqdm""", + + # AMSGRAD-CIFAR-10-DN + """python main.py --dataset cifar10 --densenet --opt amsgrad --l2 1e-4 --eta 1e-3 --no-tqdm""", + + # AMSGRAD-CIFAR-100-DN + """python main.py --dataset cifar100 --densenet --opt amsgrad --l2 1e-4 --eta 1e-3 --no-tqdm""", + + # BPGRAD-CIFAR-10-WRN + """python main.py --dataset cifar10 --wrn --opt bpgrad --l2 1e-4 --eta 0.1 --no-tqdm""", + + # BPGRAD-CIFAR-100-WRN + """python main.py --dataset cifar100 --wrn --opt bpgrad --l2 5e-4 --eta 0.1 --no-tqdm""", + + # BPGRAD-CIFAR-10-DN + """python main.py --dataset cifar10 --densenet --opt bpgrad --l2 1e-4 --eta 0.1 --no-tqdm""", + + # BPGRAD-CIFAR-10-DN + """python main.py --dataset cifar100 --densenet --opt bpgrad --l2 1e-4 --eta 0.1 --no-tqdm""", + +] + + +if __name__ == "__main__": + launch(jobs, interval=3) diff --git a/src/scripts/reproduce_snli.py b/src/scripts/reproduce_snli.py new file mode 100644 index 0000000..b66e492 --- /dev/null +++ b/src/scripts/reproduce_snli.py @@ -0,0 +1,29 @@ +import os + +from scheduling import launch + + +jobs = [ + # SGD-CE + "python train_nli.py --opt sgd --eta 1 --loss ce --no-tqdm", + + # SGD-SVM + "python train_nli.py --opt sgd --eta 0.1 --loss svm --no-tqdm", + + # ADAM-SVM + "python train_nli.py --opt adam --eta 1e-4 --loss svm --no-tqdm", + + # ADAM-CE + "python train_nli.py --opt adam --eta 1e-4 --loss ce --no-tqdm", + + # # DFW-SVM + "python train_nli.py --opt dfw --eta 1 --loss svm --no-tqdm", +] + + +if __name__ == "__main__": + # change current directory to InferSent + os.chdir('./InferSent/') + launch(jobs, interval=3) + # change current directory back to original + os.chdir('..') diff --git a/src/scripts/scheduling.py b/src/scripts/scheduling.py new file mode 100644 index 0000000..db56a5c --- /dev/null +++ b/src/scripts/scheduling.py @@ -0,0 +1,25 @@ +try: + import waitGPU +except ImportError: + print('Failed to import waitGPU --> no automatic scheduling on GPU') + waitGPU = None + pass +import subprocess +import time + + +def run_command(command, noprint=True): + if waitGPU is not None: + waitGPU.wait(nproc=0, interval=1, ngpu=1) + command = " ".join(command.split()) + if noprint: + command = "{} > /dev/null".format(command) + print(command) + subprocess.Popen(command, stderr=subprocess.STDOUT, stdout=None, shell=True) + + +def launch(jobs, interval): + for i, job in enumerate(jobs): + print("\nJob {} out of {}".format(i + 1, len(jobs))) + run_command(job) + time.sleep(interval) diff --git a/src/utils.py b/src/utils.py new file mode 100644 index 0000000..a4b8d8d --- /dev/null +++ b/src/utils.py @@ -0,0 +1,119 @@ +import os +import sys +import socket +import torch +import logger +import random +import numpy as np + + +def regularization(model, l2): + reg = 0.5 * l2 * sum([p.data.norm() ** 2 for p in model.parameters()]) if l2 else 0 + return reg + + +def set_seed(args, print_out=True): + if args.seed is None: + np.random.seed(None) + args.seed = np.random.randint(1e5) + if print_out: + print('Seed:\t {}'.format(args.seed)) + random.seed(args.seed) + np.random.seed(args.seed) + torch.manual_seed(args.seed) + if torch.cuda.is_available(): + torch.cuda.manual_seed(args.seed) + + +def get_xp(args, model, optimizer): + + # various useful information to store + args.command_line = 'python ' + ' '.join(sys.argv) + args.pid = os.getpid() + args.cwd = os.getcwd() + args.hostname = socket.gethostname() + + xp = logger.Experiment(args.xp_name, + use_visdom=args.visdom, + visdom_opts={'server': args.server, + 'port': args.port}, + time_indexing=False, xlabel='Epoch') + + xp.SumMetric(name='epoch', to_plot=False) + + xp.AvgMetric(name='acc', tag='train') + xp.AvgMetric(name='acc', tag='val') + xp.AvgMetric(name='acc', tag='test') + xp.BestMetric(name='acc', tag='valbest') + + xp.TimeMetric(name='timer', tag='train') + xp.TimeMetric(name='timer', tag='val') + xp.TimeMetric(name='timer', tag='test') + + xp.AvgMetric(name='loss', tag='train') + xp.AvgMetric(name='obj', tag='train') + xp.AvgMetric(name='reg') + + xp.log_config(vars(args)) + + xp.AvgMetric(name="gamma") + xp.SimpleMetric(name='eta') + + if args.log: + # log at each epoch + xp.Epoch.add_hook(lambda: xp.to_json('{}/results.json'.format(xp.name_and_dir))) + # log after final evaluation on test set + xp.Acc_Test.add_hook(lambda: xp.to_json('{}/results.json'.format(xp.name_and_dir))) + # save with different names at each epoch if needed + if args.dump_epoch: + filename = lambda: '{}-{}/model.pkl'.format(xp.name_and_dir, int(xp.Epoch.value)) + else: + filename = lambda: '{}/model.pkl'.format(xp.name_and_dir) + xp.Epoch.add_hook(lambda: save_state(model, optimizer, filename())) + + # save results and model for best validation performance + xp.Acc_Valbest.add_hook(lambda: xp.to_json('{}/best_results.json'.format(xp.name_and_dir))) + xp.Acc_Valbest.add_hook(lambda: save_state(model, optimizer, '{}/best_model.pkl'.format(xp.name_and_dir))) + + return xp + + +def save_state(model, optimizer, filename): + torch.save({'model': model.state_dict(), + 'optimizer': optimizer.state_dict()}, filename) + + +@torch.autograd.no_grad() +def accuracy(out, targets, topk=1): + if topk == 1: + _, pred = torch.max(out, 1) + acc = torch.mean(torch.eq(pred, targets).float()) + else: + _, pred = out.topk(topk, 1, True, True) + pred = pred.t() + correct = pred.eq(targets.view(1, -1).expand_as(pred)) + acc = correct[:topk].view(-1).float().sum(0) / out.size(0) + + return 100. * acc + + +def update_metrics(xp, state): + xp.Acc_Train.update(state['acc'], n=state['size']) + xp.Loss_Train.update(state['loss'], n=state['size']) + xp.Gamma.update(state['gamma'], n=state['size']) + + +def log_metrics(xp): + + # Average of accuracy and loss on training set + xp.Acc_Train.log_and_reset() + xp.Loss_Train.log_and_reset() + xp.Obj_Train.log_and_reset() + xp.Reg.log_and_reset() + + # timer of epoch + xp.Timer_Train.log_and_reset() + + # Log step-size + xp.Gamma.log_and_reset() + xp.Eta.log_and_reset()