Repository: WellyZhang/RAVEN Branch: master Commit: 77927ba3fe26 Files: 30 Total size: 192.3 KB Directory structure: gitextract_dccjdwzt/ ├── .gitignore ├── LICENSE ├── README.md ├── assets/ │ ├── README.md │ └── embedding.npy ├── requirements.txt └── src/ ├── dataset/ │ ├── AoT.py │ ├── Attribute.py │ ├── Rule.py │ ├── __init__.py │ ├── api.py │ ├── build_tree.py │ ├── const.py │ ├── constraints.py │ ├── main.py │ ├── rendering.py │ ├── sampling.py │ ├── serialize.py │ └── solver.py └── model/ ├── __init__.py ├── basic_model.py ├── cnn_lstm.py ├── cnn_mlp.py ├── const/ │ ├── __init__.py │ └── const.py ├── fc_tree_net.py ├── main.py ├── resnet18.py └── utility/ ├── __init__.py └── dataset_utility.py ================================================ FILE CONTENTS ================================================ ================================================ FILE: .gitignore ================================================ # Byte-compiled / optimized / DLL files __pycache__/ *.py[cod] *$py.class # C extensions *.so # Distribution / packaging .Python build/ develop-eggs/ dist/ downloads/ eggs/ .eggs/ lib/ lib64/ parts/ sdist/ var/ wheels/ *.egg-info/ .installed.cfg *.egg MANIFEST # PyInstaller # Usually these files are written by a python script from a template # before PyInstaller builds the exe, so as to inject date/other infos into it. *.manifest *.spec # Installer logs pip-log.txt pip-delete-this-directory.txt # Unit test / coverage reports htmlcov/ .tox/ .coverage .coverage.* .cache nosetests.xml coverage.xml *.cover .hypothesis/ .pytest_cache/ # Translations *.mo *.pot # Django stuff: *.log local_settings.py db.sqlite3 # 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 # Environments .env .venv env/ venv/ ENV/ env.bak/ venv.bak/ # Spyder project settings .spyderproject .spyproject # Rope project settings .ropeproject # mkdocs documentation /site # mypy .mypy_cache/ # vscode .vscode # experiments /experiments ================================================ FILE: LICENSE ================================================ 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 . ================================================ FILE: README.md ================================================ # RAVEN This repo contains code for our CVPR 2019 paper. [RAVEN: A Dataset for Relational and Analogical Visual rEasoNing](http://wellyzhang.github.io/attach/cvpr19zhang.pdf) Chi Zhang*, Feng Gao*, Baoxiong Jia, Yixin Zhu, Song-Chun Zhu *Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR)*, 2019 (* indicates equal contribution.) Dramatic progress has been witnessed in basic vision tasks involving low-level perception, such as object recognition, detection, and tracking. Unfortunately, there is still an enormous performance gap between artificial vision systems and human intelligence in terms of higher-level vision problems, especially ones involving reasoning. Earlier attempts in equipping machines with high-level reasoning have hovered around Visual Question Answering (VQA), one typical task associating vision and language understanding. In this work, we propose a new dataset, built in the context of Raven's Progressive Matrices (RPM) and aimed at lifting machine intelligence by associating vision with structural, relational, and analogical reasoning in a hierarchical representation. Unlike previous works in measuring abstract reasoning using RPM, we establish a semantic link between vision and reasoning by providing structure representation. This addition enables a new type of abstract reasoning by jointly operating on the structure representation. Machine reasoning ability using modern computer vision is evaluated in this newly proposed dataset. Additionally, we also provide human performance as a reference. Finally, we show consistent improvement across all models by incorporating a simple neural module that combines visual understanding and structure reasoning. ![framework](http://wellyzhang.github.io/img/in-post/RAVEN/process.jpg) # Dataset The dataset is generated using the attributed stochastic image grammar. An example is shown below. ![grammar](http://wellyzhang.github.io/img/in-post/RAVEN/prologue.jpg) The grammatical design makes the dataset flexible and extendable. In total, we come up with 7 different figural configurations. ![configurations](http://wellyzhang.github.io/img/in-post/RAVEN/peek_view.png) The dataset formatting document is in ```assets/README.md```. To download the dataset, please check [our project page](http://wellyzhang.github.io/project/raven.html#dataset). # Performance We show performance of models in the following table. For details, please check our [paper](http://wellyzhang.github.io/attach/cvpr19zhang.pdf). | Method | Acc | Center | 2x2Grid | 3x3Grid | L-R | U-D | O-IC | O-IG | | :--- | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | | LSTM | 13.07% | 13.19% | 14.13% | 13.69% | 12.84% | 12.35% | 12.15% | 12.99% | | WReN | 14.69% | 13.09% | 28.62% | 28.27% | 7.49% | 6.34% | 8.38% | 10.56% | | CNN | 36.97% | 33.58% | 30.30% | 33.53% | 39.43% | 41.26% | 43.20% | 37.54% | | ResNet | 53.43% | 52.82% | 41.86% | 44.29% | 58.77% | 60.16% | 63.19% | 53.12% | | LSTM+DRT | 13.96% | 14.29% | 15.08% | 14.09% | 13.79% | 13.24% | 13.99% | 13.29% | | WReN+DRT | 15.02% | 15.38% | 23.26% | 29.51% | 6.99% | 8.43% | 8.93% | 12.35% | | CNN+DRT | 39.42% | 37.30% | 30.06% | 34.57% | 45.49% | 45.54% | 45.93% | 37.54% | | ResNet+DRT | **59.56%** | **58.08%** | **46.53%** | **50.40%** | **65.82%** | **67.11%** | **69.09%** | **60.11%** | | Human | 84.41% | 95.45% | 81.82% | 79.55% | 86.36% | 81.81% | 86.36% | 81.81% | | Solver | 100% | 100% | 100% | 100% | 100% | 100% | 100% | 100% | # Dependencies **Important** * Python 2.7 * OpenCV * PyTorch * CUDA and cuDNN expected See ```requirements.txt``` for a full list of packages required. # Usage ## Dataset Generation Code to generate the dataset resides in the ```src/dataset``` folder. To generate a dataset, run ``` python src/dataset/main.py --num-samples --save-dir ``` Check the ```main.py``` file for a full list of arguments you can adjust. ## Benchmarking Code to benchmark the dataset resides in ```src/model```. To run the code, first put ```assets/embedding.npy``` in the dataset folder as specified in the ```src/model/utility/dataset_utility.py```. Then run ``` python src/model/main.py --model --path ``` You can check the ```main.py``` file for a full list of arguments. This repo only supports ```Resnet18_MLP```, ```CNN_MLP```, and ```CNN_LSTM```. For WReN, please check the implementation in [the WReN repo](https://github.com/Fen9/WReN). Note that for batch processing, we implement the DRT as a maximum tree of all possible tree structures and prune the branches during training based on an indicator. # Citation If you find the paper and/or the code helpful, please cite us. ``` @inproceedings{zhang2019raven, title={RAVEN: A Dataset for Relational and Analogical Visual rEasoNing}, author={Zhang, Chi and Gao, Feng and Jia, Baoxiong and Zhu, Yixin and Zhu, Song-Chun}, booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR)}, year={2019} } ``` # Acknowledgement We'd like to express our gratitude towards all the colleagues and anonymous reviewers for helping us improve the paper. The project is impossible to finish without the following open-source implementation. * [WReN](https://github.com/Fen9/WReN) ================================================ FILE: assets/README.md ================================================ # Dataset Format The dataset folder is organized as follows: ``` center_single/ RAVEN_0_train.npz RAVEN_0_train.xml ... RAVEN_6_val.npz RAVEN_6_val.xml ... RAVEN_8_test.npz RAVEN_8_test.xml ... distribute_four/ ... distribute_nine/ ... in_center_single_out_center_single/ ... in_distribute_four_out_center_single/ ... left_center_single_right_center_single/ ... up_center_single_down_center_single/ ... ``` Note that each npz file comes with an xml file. These 7 folders correspond to the 7 figure configurations in the paper. Specifically, * Center = center_single * 2x2Grid = distribute_four * 3x3Grid = distribute_nine * Left-Right = left_center_single_right_center_single * Up-Down = up_center_single_down_center_single * Out-InCenter = in_center_single_out_center_single * Out-InGrid = in_distribute_four_out_center_single ## Naming You might notice that the actual naming in this dataset is slightly different from what's reported in our paper. This is mostly due to the fact that things like **2x2** or **3x3** do not have corresponding word vectors. They are now **distribute_four** and **distribute_nine**. To make the paper concise, we also remove certain adjectives. **Center** was **Center_Single** and sometimes came with a component name. As described in the paper, embeddings for each of them are obtained from pre-trained GloVe vectors and held fixed during training. ## NPZ file Each npz file contains the following: * image: a (16, 160, 160) array where all 16 figures in each problem are stacked on the first dimension. Note that first 8 figures compose the problem matrix and the last 8 figures are choices. * target: the index of the correct answer in the answer set. Note that it starts from 0 and you should offset it by 8 if you want to retrieve it from the image array. * structure: the tree structure annotation for the problem. It's serialized into a sequence using pre-order traversal. * meta_matrix: similar to that in PGM. Detailed ordering could be found in ```src/dataset/const.py```. * meta_target: bitwise-or of meta_matrix on all rows. * meta_structure: it's similar to meta_matrix. Detailed ordering is in ```src/dataset/const.py```. ## XML file Each xml file contains the following: * Context panels and choice panels: each Panel could be further decomposed into Struct, Component, Layout, and Entity. * Each layer comes with its name and id if necessary. * Layout has its own attributes, whose values are indices into the value set (see also ```src/dataset/const.py```), except Position. Position is a list of slots entities could occupy, denoted by center and width/height. * Entity's attributes follow the same annotation. The bbox is retrieved from the Position array in its parent Layout and the real_bbox is the actual bounding box, denoted by center and width/height. The mask is encoded using the run-length encoding. To decode it, use the ```rle_decode``` function in ```src/dataset/api.py```. * Rules: rules are divided into groups, each of which applies to the corresponding component with the same id number. * ```attr``` could be ```Number/Position``` when the rule is ```Constant``` as these two attributes are deeply coupled. * When there is a rule on ```Number``` or ```Position```, we omit the rule on the other attribute, as it should be assumed **as is**, *i.e.*, following the rule on the other (could remain unchanged). * Therefore, each rule group has 4 rules. ================================================ FILE: requirements.txt ================================================ numpy scipy matplotlib pillow scikit-image opencv-contrib-python tqdm torch torchvision ================================================ FILE: src/dataset/AoT.py ================================================ # -*- coding: utf-8 -*- import copy import numpy as np from scipy.misc import comb from Attribute import Angle, Color, Number, Position, Size, Type, Uniformity from constraints import rule_constraint class AoTNode(object): """Superclass of AoT. """ levels_next = {"Root": "Structure", "Structure": "Component", "Component": "Layout", "Layout": "Entity"} def __init__(self, name, level, node_type, is_pg=False): self.name = name self.level = level self.node_type = node_type self.children = [] self.is_pg = is_pg def insert(self, node): """Used for public. Arguments: node(AoTNode): a node to insert """ assert isinstance(node, AoTNode) assert self.node_type != "leaf" assert node.level == self.levels_next[self.level] self.children.append(node) def _insert(self, node): """Used for private. Arguments: node(AoTNode): a node to insert """ assert isinstance(node, AoTNode) assert self.node_type != "leaf" assert node.level == self.levels_next[self.level] self.children.append(node) def _resample(self, change_number): """Resample the layout. If the number of entities change, resample also the position distribution; otherwise only resample each attribute for each entity. Arugments: change_number(bool): whether to the number has been reset """ assert self.is_pg if self.node_type == "and": for child in self.children: child._resample(change_number) else: self.children[0]._resample(change_number) def __repr__(self): return self.level + "." + self.name def __str__(self): return self.level + "." + self.name class Root(AoTNode): def __init__(self, name, is_pg=False): super(Root, self).__init__(name, level="Root", node_type="or", is_pg=is_pg) def sample(self): """The function returns a separate AoT that is correctly parsed. Note that a new node is needed so that modification does not alter settings in the original tree. Returns: new_node(Root): a newly instantiated node """ if self.is_pg: raise ValueError("Could not sample on a PG") new_node = Root(self.name, True) selected = np.random.choice(self.children) new_node.insert(selected._sample()) return new_node def resample(self, change_number=False): self._resample(change_number) def prune(self, rule_groups): """Prune the AoT such that all branches satisfy the constraints. Arguments: rule_groups(list of list of Rule): each list of Rule applies to a component Returns: new_node(Root): a newly instantiated node with branches all satisfying the constraints; None if no branches satisfy all the constraints """ new_node = Root(self.name) for structure in self.children: if len(structure.children) == len(rule_groups): new_child = structure._prune(rule_groups) if new_child is not None: new_node.insert(new_child) # during real execution, this should never happens if len(new_node.children) == 0: new_node = None return new_node def prepare(self): """This function prepares the AoT for rendering. Returns: structure.name(str): used for rendering structure entities(list of Entity): used for rendering each entity """ assert self.is_pg assert self.level == "Root" structure = self.children[0] components = [] for child in structure.children: components.append(child) entities = [] for component in components: for child in component.children[0].children: entities.append(child) return structure.name, entities def sample_new(self, component_idx, attr_name, min_level, max_level, root): """Sample a new configuration. This is used for generating answers. Arguments: component_idx(int): the component we will sample attr_name(str): name of the attribute to sample min_level(int): lower bound of value level for the attribute max_level(int): upper bound of value level for the attribute root(AoTNode): the answer AoT, used for storing previous value levels for each attribute """ assert self.is_pg self.children[0]._sample_new(component_idx, attr_name, min_level, max_level, root.children[0]) class Structure(AoTNode): def __init__(self, name, is_pg=False): super(Structure, self).__init__(name, level="Structure", node_type="and", is_pg=is_pg) def _sample(self): if self.is_pg: raise ValueError("Could not sample on a PG") new_node = Structure(self.name, True) for child in self.children: new_node.insert(child._sample()) return new_node def _prune(self, rule_groups): new_node = Structure(self.name) for i in range(len(self.children)): child = self.children[i] # if any of the components fails to satisfy the constraint # the structure could not be chosen new_child = child._prune(rule_groups[i]) if new_child is None: return None new_node.insert(new_child) return new_node def _sample_new(self, component_idx, attr_name, min_level, max_level, structure): self.children[component_idx]._sample_new(attr_name, min_level, max_level, structure.children[component_idx]) class Component(AoTNode): def __init__(self, name, is_pg=False): super(Component, self).__init__(name, level="Component", node_type="or", is_pg=is_pg) def _sample(self): if self.is_pg: raise ValueError("Could not sample on a PG") new_node = Component(self.name, True) selected = np.random.choice(self.children) new_node.insert(selected._sample()) return new_node def _prune(self, rule_group): new_node = Component(self.name) for child in self.children: new_child = child._update_constraint(rule_group) if new_child is not None: new_node.insert(new_child) if len(new_node.children) == 0: new_node = None return new_node def _sample_new(self, attr_name, min_level, max_level, component): self.children[0]._sample_new(attr_name, min_level, max_level, component.children[0]) class Layout(AoTNode): """Layout is the highest level of the hierarchy that has attributes (Number, Position and Uniformity). To copy a Layout, please use deepcopy such that newly instantiated and separated attributes are created. """ def __init__(self, name, layout_constraint, entity_constraint, orig_layout_constraint=None, orig_entity_constraint=None, sample_new_num_count=None, is_pg=False): super(Layout, self).__init__(name, level="Layout", node_type="and", is_pg=is_pg) self.layout_constraint = layout_constraint self.entity_constraint = entity_constraint self.number = Number(min_level=layout_constraint["Number"][0], max_level=layout_constraint["Number"][1]) self.position = Position(pos_type=layout_constraint["Position"][0], pos_list=layout_constraint["Position"][1]) self.uniformity = Uniformity(min_level=layout_constraint["Uni"][0], max_level=layout_constraint["Uni"][1]) self.number.sample() self.position.sample(self.number.get_value()) self.uniformity.sample() # store initial layout_constraint and entity_constraint for answer generation if orig_layout_constraint is None: self.orig_layout_constraint = copy.deepcopy(self.layout_constraint) else: self.orig_layout_constraint = orig_layout_constraint if orig_entity_constraint is None: self.orig_entity_constraint = copy.deepcopy(self.entity_constraint) else: self.orig_entity_constraint = orig_entity_constraint if sample_new_num_count is None: self.sample_new_num_count = dict() most_num = len(self.position.values) for i in range(layout_constraint["Number"][0], layout_constraint["Number"][1] + 1): self.sample_new_num_count[i] = [comb(most_num, i + 1), []] else: self.sample_new_num_count = sample_new_num_count def add_new(self, *bboxes): """Add new entities into this level. Arguments: *bboxes(tuple of bbox): bboxes of new entities """ name = self.number.get_value() uni = self.uniformity.get_value() for i in range(len(bboxes)): name += i bbox = bboxes[i] new_entity = copy.deepcopy(self.children[0]) new_entity.name = str(name) new_entity.bbox = bbox if not uni: new_entity.resample() self._insert(new_entity) def resample(self, change_number=False): self._resample(change_number) def _sample(self): """Though Layout is an "and" node, we do not enumerate all possible configurations, but rather we treat it as a sampling process such that different configurtions are sampled. After the sampling, the lower level Entities are instantiated. Returns: new_node(Layout): a separated node with independent attributes """ pos = self.position.get_value() new_node = copy.deepcopy(self) new_node.is_pg = True if self.uniformity.get_value(): node = Entity(name=str(0), bbox=pos[0], entity_constraint=self.entity_constraint) new_node._insert(node) for i in range(1, len(pos)): bbox = pos[i] node = copy.deepcopy(node) node.name = str(i) node.bbox = bbox new_node._insert(node) else: for i in range(len(pos)): bbox = pos[i] node = Entity(name=str(i), bbox=bbox, entity_constraint=self.entity_constraint) new_node._insert(node) return new_node def _resample(self, change_number): """Resample each attribute for every child. This function is called across rows. Arguments: change_number(bool): whether to resample a number """ if change_number: self.number.sample() del self.children[:] self.position.sample(self.number.get_value()) pos = self.position.get_value() if self.uniformity.get_value(): node = Entity(name=str(0), bbox=pos[0], entity_constraint=self.entity_constraint) self._insert(node) for i in range(1, len(pos)): bbox = pos[i] node = copy.deepcopy(node) node.name = str(i) node.bbox = bbox self._insert(node) else: for i in range(len(pos)): bbox = pos[i] node = Entity(name=str(i), bbox=bbox, entity_constraint=self.entity_constraint) self._insert(node) def _update_constraint(self, rule_group): """Update the constraint of the layout. If one constraint is not satisfied, return None such that this structure is disgarded. Arguments: rule_group(list of Rule): all rules to apply to this layout Returns: Layout(Layout): a new Layout node with independent attributes """ num_min = self.layout_constraint["Number"][0] num_max = self.layout_constraint["Number"][1] uni_min = self.layout_constraint["Uni"][0] uni_max = self.layout_constraint["Uni"][1] type_min = self.entity_constraint["Type"][0] type_max = self.entity_constraint["Type"][1] size_min = self.entity_constraint["Size"][0] size_max = self.entity_constraint["Size"][1] color_min = self.entity_constraint["Color"][0] color_max = self.entity_constraint["Color"][1] new_constraints = rule_constraint(rule_group, num_min, num_max, uni_min, uni_max, type_min, type_max, size_min, size_max, color_min, color_max) new_layout_constraint, new_entity_constraint = new_constraints new_num_min = new_layout_constraint["Number"][0] new_num_max = new_layout_constraint["Number"][1] if new_num_min > new_num_max: return None new_uni_min = new_layout_constraint["Uni"][0] new_uni_max = new_layout_constraint["Uni"][1] if new_uni_min > new_uni_max: return None new_type_min = new_entity_constraint["Type"][0] new_type_max = new_entity_constraint["Type"][1] if new_type_min > new_type_max: return None new_size_min = new_entity_constraint["Size"][0] new_size_max = new_entity_constraint["Size"][1] if new_size_min > new_size_max: return None new_color_min = new_entity_constraint["Color"][0] new_color_max = new_entity_constraint["Color"][1] if new_color_min > new_color_max: return None new_layout_constraint = copy.deepcopy(self.layout_constraint) new_layout_constraint["Number"][:] = [new_num_min, new_num_max] new_layout_constraint["Uni"][:] = [new_uni_min, new_uni_max] new_entity_constraint = copy.deepcopy(self.entity_constraint) new_entity_constraint["Type"][:] = [new_type_min, new_type_max] new_entity_constraint["Size"][:] = [new_size_min, new_size_max] new_entity_constraint["Color"][:] = [new_color_min, new_color_max] return Layout(self.name, new_layout_constraint, new_entity_constraint, self.orig_layout_constraint, self.orig_entity_constraint, self.sample_new_num_count) def reset_constraint(self, attr): attr_name = attr.lower() instance = getattr(self, attr_name) instance.min_level = self.layout_constraint[attr][0] instance.max_level = self.layout_constraint[attr][1] def _sample_new(self, attr_name, min_level, max_level, layout): if attr_name == "Number": while True: value_level = self.number.sample_new(min_level, max_level) if layout.sample_new_num_count[value_level][0] == 0: continue new_num = self.number.get_value(value_level) new_value_idx = self.position.sample_new(new_num) set_new_value_idx = set(new_value_idx) if set_new_value_idx not in layout.sample_new_num_count[value_level][1]: layout.sample_new_num_count[value_level][0] -= 1 layout.sample_new_num_count[value_level][1].append(set_new_value_idx) break self.number.set_value_level(value_level) self.position.set_value_idx(new_value_idx) pos = self.position.get_value() del self.children[:] for i in range(len(pos)): bbox = pos[i] node = Entity(name=str(i), bbox=bbox, entity_constraint=self.entity_constraint) self._insert(node) elif attr_name == "Position": new_value_idx = self.position.sample_new(self.number.get_value()) layout.position.previous_values.append(new_value_idx) self.position.set_value_idx(new_value_idx) pos = self.position.get_value() for i in range(len(pos)): bbox = pos[i] self.children[i].bbox = bbox elif attr_name == "Type": for index in range(len(self.children)): new_value_level = self.children[index].type.sample_new(min_level, max_level) self.children[index].type.set_value_level(new_value_level) layout.children[index].type.previous_values.append(new_value_level) elif attr_name == "Size": for index in range(len(self.children)): new_value_level = self.children[index].size.sample_new(min_level, max_level) self.children[index].size.set_value_level(new_value_level) layout.children[index].size.previous_values.append(new_value_level) elif attr_name == "Color": for index in range(len(self.children)): new_value_level = self.children[index].color.sample_new(min_level, max_level) self.children[index].color.set_value_level(new_value_level) layout.children[index].color.previous_values.append(new_value_level) else: raise ValueError("Unsupported operation") class Entity(AoTNode): def __init__(self, name, bbox, entity_constraint): super(Entity, self).__init__(name, level="Entity", node_type="leaf", is_pg=True) # Attributes # Sample each attribute such that the value lies in the admissible range # Otherwise, random sample self.entity_constraint = entity_constraint self.bbox = bbox self.type = Type(min_level=entity_constraint["Type"][0], max_level=entity_constraint["Type"][1]) self.type.sample() self.size = Size(min_level=entity_constraint["Size"][0], max_level=entity_constraint["Size"][1]) self.size.sample() self.color = Color(min_level=entity_constraint["Color"][0], max_level=entity_constraint["Color"][1]) self.color.sample() self.angle = Angle(min_level=entity_constraint["Angle"][0], max_level=entity_constraint["Angle"][1]) self.angle.sample() def reset_constraint(self, attr, min_level, max_level): attr_name = attr.lower() self.entity_constraint[attr][:] = [min_level, max_level] instance = getattr(self, attr_name) instance.min_level = min_level instance.max_level = max_level def resample(self): self.type.sample() self.size.sample() self.color.sample() self.angle.sample() ================================================ FILE: src/dataset/Attribute.py ================================================ # -*- coding: utf-8 -*- import numpy as np from const import (ANGLE_MAX, ANGLE_MIN, ANGLE_VALUES, COLOR_MAX, COLOR_MIN, COLOR_VALUES, NUM_MAX, NUM_MIN, NUM_VALUES, SIZE_MAX, SIZE_MIN, SIZE_VALUES, TYPE_MAX, TYPE_MIN, TYPE_VALUES, UNI_MAX, UNI_MIN, UNI_VALUES) class Attribute(object): """Super-class for all attributes. This should not be instantiated. In the sub-class, each attribute should have a pre-defined value set and a member to indicate the index in the value set. This design enables setting a value by modifying the index only. Also, each instance should come with value index boundaries, set as min_level and max_level. Boundaries are good when we want to set constraints on the value set. Before accessing the value, we should sample a value level by calling the sample function. """ def __init__(self, name): self.name = name self.level = "Attribute" # memory to store previous values self.previous_values = [] def sample(self): pass def get_value(self): pass def set_value(self): pass def __repr__(self): return self.level + "." + self.name def __str__(self): return self.level + "." + self.name class Number(Attribute): def __init__(self, min_level=NUM_MIN, max_level=NUM_MAX): super(Number, self).__init__("Number") self.value_level = 0 self.values = NUM_VALUES self.min_level = min_level self.max_level = max_level def sample(self, min_level=NUM_MIN, max_level=NUM_MAX): # min_level: min level index # max_level: max level index min_level = max(self.min_level, min_level) max_level = min(self.max_level, max_level) self.value_level = np.random.choice(range(min_level, max_level + 1)) def sample_new(self, min_level=None, max_level=None, previous_values=None): """Sample new values for generating the answer set. Returns: new_idx(int): a new value_level """ if min_level is None or max_level is None: values = range(self.min_level, self.max_level + 1) else: values = range(min_level, max_level + 1) if not previous_values: available = set(values) - set(self.previous_values) - set([self.value_level]) else: available = set(values) - set(previous_values) - set([self.value_level]) new_idx = np.random.choice(list(available)) return new_idx def get_value_level(self): return self.value_level def set_value_level(self, value_level): self.value_level = value_level def get_value(self, value_level=None): if value_level is None: value_level = self.value_level return self.values[value_level] class Type(Attribute): def __init__(self, min_level=TYPE_MIN, max_level=TYPE_MAX): super(Type, self).__init__("Type") self.value_level = 0 self.values = TYPE_VALUES self.min_level = min_level self.max_level = max_level def sample(self, min_level=TYPE_MIN, max_level=TYPE_MAX): min_level = max(self.min_level, min_level) max_level = min(self.max_level, max_level) self.value_level = np.random.choice(range(min_level, max_level + 1)) def sample_new(self, min_level=None, max_level=None, previous_values=None): if min_level is None or max_level is None: values = range(self.min_level, self.max_level + 1) else: values = range(min_level, max_level + 1) if not previous_values: available = set(values) - set(self.previous_values) - set([self.value_level]) else: available = set(values) - set(previous_values) - set([self.value_level]) new_idx = np.random.choice(list(available)) return new_idx def get_value_level(self): return self.value_level def set_value_level(self, value_level): self.value_level = value_level def get_value(self, value_level=None): if value_level is None: value_level = self.value_level return self.values[value_level] class Size(Attribute): def __init__(self, min_level=SIZE_MIN, max_level=SIZE_MAX): super(Size, self).__init__("Size") self.value_level = 3 self.values = SIZE_VALUES self.min_level = min_level self.max_level = max_level def sample(self, min_level=SIZE_MIN, max_level=SIZE_MAX): min_level = max(self.min_level, min_level) max_level = min(self.max_level, max_level) self.value_level = np.random.choice(range(min_level, max_level + 1)) def sample_new(self, min_level=None, max_level=None, previous_values=None): if min_level is None or max_level is None: values = range(self.min_level, self.max_level + 1) else: values = range(min_level, max_level + 1) if not previous_values: available = set(values) - set(self.previous_values) - set([self.value_level]) else: available = set(values) - set(previous_values) - set([self.value_level]) new_idx = np.random.choice(list(available)) return new_idx def get_value_level(self): return self.value_level def set_value_level(self, value_level): self.value_level = value_level def get_value(self, value_level=None): if value_level is None: value_level = self.value_level return self.values[value_level] class Color(Attribute): def __init__(self, min_level=COLOR_MIN, max_level=COLOR_MAX): super(Color, self).__init__("Color") self.value_level = 0 self.values = COLOR_VALUES self.min_level = min_level self.max_level = max_level def sample(self, min_level=COLOR_MIN, max_level=COLOR_MAX): min_level = max(self.min_level, min_level) max_level = min(self.max_level, max_level) self.value_level = np.random.choice(range(min_level, max_level + 1)) def sample_new(self, min_level=None, max_level=None, previous_values=None): if min_level is None or max_level is None: values = range(self.min_level, self.max_level + 1) else: values = range(min_level, max_level + 1) if not previous_values: available = set(values) - set(self.previous_values) - set([self.value_level]) else: available = set(values) - set(previous_values) - set([self.value_level]) new_idx = np.random.choice(list(available)) return new_idx def get_value_level(self): return self.value_level def set_value_level(self, value_level): self.value_level = value_level def get_value(self, value_level=None): if value_level is None: value_level = self.value_level return self.values[value_level] class Angle(Attribute): def __init__(self, min_level=ANGLE_MIN, max_level=ANGLE_MAX): super(Angle, self).__init__("Angle") self.value_level = 3 self.values = ANGLE_VALUES self.min_level = min_level self.max_level = max_level def sample(self, min_level=ANGLE_MIN, max_level=ANGLE_MAX): min_level = max(self.min_level, min_level) max_level = min(self.max_level, max_level) self.value_level = np.random.choice(range(min_level, max_level + 1)) def sample_new(self, min_level=None, max_level=None, previous_values=None): if min_level is None or max_level is None: values = range(self.min_level, self.max_level + 1) else: values = range(min_level, max_level + 1) if not previous_values: available = set(values) - set(self.previous_values) - set([self.value_level]) else: available = set(values) - set(previous_values) - set([self.value_level]) new_idx = np.random.choice(list(available)) return new_idx def get_value_level(self): return self.value_level def set_value_level(self, value_level): self.value_level = value_level def get_value(self, value_level=None): if value_level is None: value_level = self.value_level return self.values[value_level] class Uniformity(Attribute): def __init__(self, min_level=UNI_MIN, max_level=UNI_MAX): super(Uniformity, self).__init__("Uniformity") self.value_level = 0 self.values = UNI_VALUES self.min_level = min_level self.max_level = max_level def sample(self): self.value_level = np.random.choice(range(self.min_level, self.max_level + 1)) def sample_new(self): # Should not resample uniformity pass def set_value_level(self, value_level): self.value_level = value_level def get_value_level(self): return self.value_level def get_value(self, value_level=None): if value_level is None: value_level = self.value_level return self.values[value_level] class Position(Attribute): """Position is a special case. There are the planar position and the angular position. Planar position allows translation in the plane while angular Position performs roration around an axis penperdicular to the plane. """ def __init__(self, pos_type, pos_list): """Instantiate the Position attribute by passing a position type and a pre-defined position distribution on the plane. This attribute is strongly coupled with Number and hence value index boundaries are not needed. Arguments: pos_type(str): either "planar" or "angular pos_list(list of list of numbers): actual distribution on the plane """ super(Position, self).__init__("Position") # planar: [x_c, y_c, max_w, max_h] # angular: [x_c, y_c, max_w, max_h, x_r, y_r, omega] assert pos_type in ("planar", "angular") self.pos_type = pos_type self.values = pos_list self.value_idx = None def sample(self, num): """Sample multiple positions at the same time. Arguments: num(int): the number of positions to sample """ length = len(self.values) assert num <= length self.value_idx = np.random.choice(range(length), num, False) def sample_new(self, num, previous_values=None): # Here sample new relies on probability length = len(self.values) if not previous_values: constraints = self.previous_values else: constraints = previous_values while True: finished = True new_value_idx = np.random.choice(length, num, False) if set(new_value_idx) == set(self.value_idx): continue for previous_value in constraints: if set(new_value_idx) == set(previous_value): finished = False break if finished: break return new_value_idx def sample_add(self, num): """Sample additional number of positions. Arguments: num(int): the number of additional positions to sample Returns: ret(tuple of position): new positions to add to the layout """ ret = [] available = set(range(len(self.values))) - set(self.value_idx) idxes_2_add = np.random.choice(list(available), num, False) for index in idxes_2_add: self.value_idx = np.insert(self.value_idx, 0, index) ret.append(self.values[index]) return ret def get_value_idx(self): return self.value_idx def set_value_idx(self, value_idx): # Note that after sampling self.value_idx is a Numpy array self.value_idx = value_idx def get_value(self, value_idx=None): if value_idx is None: value_idx = self.value_idx ret = [] for idx in value_idx: ret.append(self.values[idx]) return ret def remove(self, bbox): # Note that after sampling self.value_idx is a Numpy array idx = self.values.index(bbox) np_idx = np.where(self.value_idx == idx)[0][0] self.value_idx = np.delete(self.value_idx, np_idx) ================================================ FILE: src/dataset/Rule.py ================================================ # -*- coding: utf-8 -*- import copy import numpy as np from const import COLOR_MAX, COLOR_MIN def Rule_Wrapper(name, attr, param, component_idx): ret = None if name == "Constant": ret = Constant(name, attr, param, component_idx) elif name == "Progression": ret = Progression(name, attr, param, component_idx) elif name == "Arithmetic": ret = Arithmetic(name, attr, param, component_idx) elif name == "Distribute_Three": ret = Distribute_Three(name, attr, param, component_idx) else: raise ValueError("Unsupported Rule") return ret class Rule(object): """General API for a rule. Priority order: Rule on Number/Position always comes first """ def __init__(self, name, attr, params, component_idx=0): """Instantiate a rule by its name, attribute, paramter list and the component it applies to. Each rule should be applied to all entities in a component. Arguments: name(str): pre-defined name of the rule attr(str): pre-defined name of the attribute params(list): a list of possible parameters for it to sample component_idx(int): the index of the component to apply the rule """ self.name = name self.attr = attr self.params = params self.component_idx = component_idx self.value = 0 self.sample() def sample(self): """Sample a parameter from the parameter list. """ if self.params is not None: self.value = np.random.choice(self.params) def apply_rule(self, aot, in_aot=None): """Apply the rule to a component in the AoT. Arguments: aot(AoTNode): an AoT for reference in_aot(AoTNode): an AoT to apply the rule Returns: second_aot(AoTNode): a modified AoT """ # Root -> Structure -> Component -> Layout -> Entity pass class Constant(Rule): """Unary operator. Nothing changes. """ def __init__(self, name, attr, param, component_idx): super(Constant, self).__init__(name, attr, param, component_idx) def apply_rule(self, aot, in_aot=None): if in_aot is None: in_aot = aot return copy.deepcopy(in_aot) class Progression(Rule): """Unary operator. Attribute difference on two consequetive Panels remains the same. """ def __init__(self, name, attr, param, component_idx): super(Progression, self).__init__(name, attr, param, component_idx) # Flag to trigger consistency of the attribute in the first column self.first_col = True def apply_rule(self, aot, in_aot=None): current_layout = aot.children[0].children[self.component_idx].children[0] if in_aot is None: in_aot = aot second_aot = copy.deepcopy(in_aot) second_layout = second_aot.children[0].children[self.component_idx].children[0] if self.attr == "Number": second_layout.number.set_value_level(second_layout.number.get_value_level() + self.value) second_layout.position.sample(second_layout.number.get_value()) pos = second_layout.position.get_value() del second_layout.children[:] for i in range(len(pos)): entity = copy.deepcopy(current_layout.children[0]) entity.name = str(i) entity.bbox = pos[i] if not current_layout.uniformity.get_value(): entity.resample() second_layout.insert(entity) elif self.attr == "Position": second_pos_idx = (second_layout.position.get_value_idx() + self.value) % len(second_layout.position.values) second_layout.position.set_value_idx(second_pos_idx) second_bbox = second_layout.position.get_value() for i in range(len(second_bbox)): second_layout.children[i].bbox = second_bbox[i] elif self.attr == "Type": old_value_level = current_layout.children[0].type.get_value_level() # enforce value consistency if self.first_col and not current_layout.uniformity.get_value(): for entity in current_layout.children: entity.type.set_value_level(old_value_level) for entity in second_layout.children: entity.type.set_value_level(old_value_level + self.value) elif self.attr == "Size": old_value_level = current_layout.children[0].size.get_value_level() # enforce value consistency if self.first_col and not current_layout.uniformity.get_value(): for entity in current_layout.children: entity.size.set_value_level(old_value_level) for entity in second_layout.children: entity.size.set_value_level(old_value_level + self.value) elif self.attr == "Color": old_value_level = current_layout.children[0].color.get_value_level() # enforce value consistency if self.first_col and not current_layout.uniformity.get_value(): for entity in current_layout.children: entity.color.set_value_level(old_value_level) for entity in second_layout.children: entity.color.set_value_level(old_value_level + self.value) else: raise ValueError("Unsupported attriubute") self.first_col = not self.first_col return second_aot class Arithmetic(Rule): """Binary operator. Basically: Panel_3 = Panel_1 + Panel_2. For Position: + means SET_UNION and - SET_DIFF. """ def __init__(self, name, attr, param, component_idx): super(Arithmetic, self).__init__(name, attr, param, component_idx) self.memory = [] self.color_count = 0 self.color_white_alarm = False def apply_rule(self, aot, in_aot=None): current_layout = aot.children[0].children[self.component_idx].children[0] if in_aot is None: in_aot = aot second_aot = copy.deepcopy(in_aot) second_layout = second_aot.children[0].children[self.component_idx].children[0] if self.attr == "Number": # the third col if len(self.memory) > 0: first_layout_number_level = self.memory.pop() if self.value > 0: total = first_layout_number_level + 1 + current_layout.number.get_value() else: total = first_layout_number_level + 1 - current_layout.number.get_value() second_layout.number.set_value_level(total - 1) # the second col else: old_value_level = current_layout.number.get_value_level() self.memory.append(old_value_level) if self.value > 0: num_max_level_orig = sum(current_layout.layout_constraint["Number"]) + 1 new_num_max_level = num_max_level_orig - old_value_level - 1 second_layout.layout_constraint["Number"][1] = new_num_max_level else: num_min_level_orig = (second_layout.layout_constraint["Number"][0] - 1) / 2 new_num_max_level = old_value_level - num_min_level_orig - 1 second_layout.layout_constraint["Number"][:] = [num_min_level_orig, new_num_max_level] second_layout.reset_constraint("Number") second_layout.number.sample() second_layout.position.sample(second_layout.number.get_value()) pos = second_layout.position.get_value() del second_layout.children[:] for i in range(len(pos)): entity = copy.deepcopy(current_layout.children[0]) entity.name = str(i) entity.bbox = pos[i] if not current_layout.uniformity.get_value(): entity.resample() second_layout.insert(entity) elif self.attr == "Position": # ADD is interpreted as SET_UNION; SUB is interpreted as SET_DIFF # the third col if len(self.memory) > 0: first_layout_value_idx = self.memory.pop() if self.value > 0: new_pos_idx = set(first_layout_value_idx) | set(current_layout.position.get_value_idx()) else: new_pos_idx = set(first_layout_value_idx) - set(current_layout.position.get_value_idx()) second_layout.number.set_value_level(len(new_pos_idx) - 1) second_layout.position.set_value_idx(np.array(list(new_pos_idx))) # the second col else: current_layout_value_idx = current_layout.position.get_value_idx() self.memory.append(current_layout_value_idx) while True: second_layout.number.sample() second_layout.position.sample(second_layout.number.get_value()) # if UNION, not a subset; otherwise not clearly a union if self.value > 0: if not (set(current_layout_value_idx) >= set(second_layout.position.get_value_idx())): break # if DIFF, not a subset; otherwise no entities left else: if not (set(current_layout_value_idx) <= set(second_layout.position.get_value_idx())): break pos = second_layout.position.get_value() del second_layout.children[:] for i in range(len(pos)): entity = copy.deepcopy(current_layout.children[0]) entity.name = str(i) entity.bbox = pos[i] if not current_layout.uniformity.get_value(): entity.resample() second_layout.insert(entity) elif self.attr == "Size": if len(self.memory) > 0: first_layout_size_level = self.memory.pop() if self.value > 0: new_size_value_level = first_layout_size_level + \ current_layout.children[0].size.get_value_level() + 1 else: new_size_value_level = first_layout_size_level - \ current_layout.children[0].size.get_value_level() - 1 for entity in second_layout.children: entity.size.set_value_level(new_size_value_level) else: # make sure of value consistency old_value_level = current_layout.children[0].size.get_value_level() self.memory.append(old_value_level) if not current_layout.uniformity.get_value(): for entity in current_layout.children: entity.size.set_value_level(old_value_level) if self.value > 0: size_max_level_orig = sum(current_layout.entity_constraint["Size"]) + 1 new_size_max_level = size_max_level_orig - old_value_level - 1 # deepcopy breaks the link of constraints between Layout and Entity # Need to reset each attribute second_layout.entity_constraint["Size"][1] = new_size_max_level else: size_min_level_orig = (current_layout.entity_constraint["Size"][0] - 1) / 2 new_size_max_level = old_value_level - size_min_level_orig - 1 second_layout.entity_constraint["Size"] = [size_min_level_orig, new_size_max_level] new_size_min_level, new_size_max_level = second_layout.entity_constraint["Size"] the_child = second_layout.children[0] the_child.reset_constraint("Size", new_size_min_level, new_size_max_level) the_child.size.sample() new_size_value_level = the_child.size.get_value_level() for idx in range(1, len(second_layout.children)): entity = second_layout.children[idx] entity.reset_constraint("Size", new_size_min_level, new_size_max_level) entity.size.set_value_level(new_size_value_level) elif self.attr == "Color": self.color_count += 1 if len(self.memory) > 0: first_layout_color_level = self.memory.pop() if self.value > 0: new_color_value_level = first_layout_color_level + \ current_layout.children[0].color.get_value_level() else: new_color_value_level = first_layout_color_level - \ current_layout.children[0].color.get_value_level() for entity in second_layout.children: entity.color.set_value_level(new_color_value_level) else: # Logic here: C_12 and C_22 could not be both 0, otherwise it's impossible to distinguish + and - # If C_12 == 0, we set an alarm # Under this alarm, if C_21 == MAX and ADD rule, then resample C_21 to ensure C_22 could be other than 0 # Similarly, if C_21 == 0 and SUB rule, then resample C_21 to ensure C_22 could be other than 0 # Finally, loop until C_22 is not 0 # make sure of value consistency old_value_level = current_layout.children[0].color.get_value_level() # the third time you apply this rule and find C_21 == MAX/0 if +/- reset_current_layout = False if self.color_count == 3 and self.color_white_alarm: if self.value > 0 and old_value_level == COLOR_MAX: old_value_level = current_layout.children[0].color.sample_new() reset_current_layout = True if self.value < 0 and old_value_level == COLOR_MIN: old_value_level = current_layout.children[0].color.sample_new() reset_current_layout = True self.memory.append(old_value_level) if reset_current_layout or not current_layout.uniformity.get_value(): for entity in current_layout.children: entity.color.set_value_level(old_value_level) if self.value > 0: color_max_level_orig = sum(current_layout.entity_constraint["Color"]) new_color_max_level = color_max_level_orig - old_value_level second_layout.entity_constraint["Color"][1] = new_color_max_level else: color_min_level_orig = second_layout.entity_constraint["Color"][0] / 2 new_color_max_level = old_value_level second_layout.entity_constraint["Color"][:] = [color_min_level_orig, new_color_max_level] new_color_min_level, new_color_max_level = second_layout.entity_constraint["Color"] the_child = second_layout.children[0] the_child.reset_constraint("Color", new_color_min_level, new_color_max_level) the_child.color.sample() new_color_value_level = the_child.color.get_value_level() # the first time you apply this rule and get C_12 == 0 # set the alarm if self.color_count == 1: self.color_white_alarm = (new_color_value_level == 0) if self.color_count == 3 and self.color_white_alarm and new_color_value_level == 0: new_color_value_level = the_child.color.sample_new() the_child.color.set_value_level(new_color_value_level) for idx in range(1, len(second_layout.children)): entity = second_layout.children[idx] entity.reset_constraint("Color", new_color_min_level, new_color_max_level) entity.color.set_value_level(new_color_value_level) else: raise ValueError("Unsupported attriubute") return second_aot class Distribute_Three(Rule): """Ternay operator. Three values across the columns form a fixed set. """ def __init__(self, name, attr, param, component_idx): super(Distribute_Three, self).__init__(name, attr, param, component_idx) self.value_levels = [] self.count = 0 def apply_rule(self, aot, in_aot=None): current_layout = aot.children[0].children[self.component_idx].children[0] if in_aot is None: in_aot = aot second_aot = copy.deepcopy(in_aot) second_layout = second_aot.children[0].children[self.component_idx].children[0] if self.attr == "Number": if self.count == 0: all_value_levels = range(current_layout.layout_constraint["Number"][0], current_layout.layout_constraint["Number"][1] + 1) current_value_level = current_layout.number.get_value_level() idx = all_value_levels.index(current_value_level) all_value_levels.pop(idx) three_value_levels = np.random.choice(all_value_levels, 2, False) three_value_levels = np.insert(three_value_levels, 0, current_value_level) self.value_levels.append(three_value_levels[[0, 1, 2]]) if np.random.uniform() >= 0.5: self.value_levels.append(three_value_levels[[1, 2, 0]]) self.value_levels.append(three_value_levels[[2, 0, 1]]) else: self.value_levels.append(three_value_levels[[2, 0, 1]]) self.value_levels.append(three_value_levels[[1, 2, 0]]) second_layout.number.set_value_level(self.value_levels[0][1]) else: row, col = divmod(self.count, 2) if col == 0: current_layout.number.set_value_level(self.value_levels[row][0]) current_layout.resample() second_aot = copy.deepcopy(aot) second_layout = second_aot.children[0].children[self.component_idx].children[0] second_layout.number.set_value_level(self.value_levels[row][1]) else: second_layout.number.set_value_level(self.value_levels[row][2]) second_layout.position.sample(second_layout.number.get_value()) pos = second_layout.position.get_value() del second_layout.children[:] for i in range(len(pos)): entity = copy.deepcopy(current_layout.children[0]) entity.name = str(i) entity.bbox = pos[i] if not current_layout.uniformity.get_value(): entity.resample() second_layout.insert(entity) self.count = (self.count + 1) % 6 elif self.attr == "Position": if self.count == 0: # sample new does not change value_level/value_idx num = current_layout.number.get_value() pos_0 = current_layout.position.get_value_idx() pos_1 = current_layout.position.sample_new(num) pos_2 = current_layout.position.sample_new(num, [pos_1]) three_value_levels = np.array([pos_0, pos_1, pos_2]) self.value_levels.append(three_value_levels[[0, 1, 2]]) if np.random.uniform() >= 0.5: self.value_levels.append(three_value_levels[[1, 2, 0]]) self.value_levels.append(three_value_levels[[2, 0, 1]]) else: self.value_levels.append(three_value_levels[[2, 0, 1]]) self.value_levels.append(three_value_levels[[1, 2, 0]]) second_layout.position.set_value_idx(self.value_levels[0][1]) else: row, col = divmod(self.count, 2) if col == 0: current_layout.number.set_value_level(len(self.value_levels[row][0]) - 1) current_layout.resample() current_layout.position.set_value_idx(self.value_levels[row][0]) pos = current_layout.position.get_value() for i in range(len(pos)): entity = current_layout.children[i] entity.bbox = pos[i] second_aot = copy.deepcopy(aot) second_layout = second_aot.children[0].children[self.component_idx].children[0] second_layout.position.set_value_idx(self.value_levels[row][1]) else: second_layout.position.set_value_idx(self.value_levels[row][2]) pos = second_layout.position.get_value() for i in range(len(pos)): entity = second_layout.children[i] entity.bbox = pos[i] self.count = (self.count + 1) % 6 elif self.attr == "Type": if self.count == 0: all_value_levels = range(current_layout.entity_constraint["Type"][0], current_layout.entity_constraint["Type"][1] + 1) # if np.random.uniform() >= 0.5 and 0 not in all_value_levels: # all_value_levels.insert(0, 0) three_value_levels = np.random.choice(all_value_levels, 3, False) np.random.shuffle(three_value_levels) self.value_levels.append(three_value_levels[[0, 1, 2]]) if np.random.uniform() >= 0.5: self.value_levels.append(three_value_levels[[1, 2, 0]]) self.value_levels.append(three_value_levels[[2, 0, 1]]) else: self.value_levels.append(three_value_levels[[2, 0, 1]]) self.value_levels.append(three_value_levels[[1, 2, 0]]) for entity in current_layout.children: entity.type.set_value_level(self.value_levels[0][0]) for entity in second_layout.children: entity.type.set_value_level(self.value_levels[0][1]) else: row, col = divmod(self.count, 2) if col == 0: value_level = self.value_levels[row][0] for entity in current_layout.children: entity.type.set_value_level(value_level) value_level = self.value_levels[row][1] for entity in second_layout.children: entity.type.set_value_level(value_level) else: value_level = self.value_levels[row][2] for entity in second_layout.children: entity.type.set_value_level(value_level) self.count = (self.count + 1) % 6 elif self.attr == "Size": if self.count == 0: all_value_levels = range(current_layout.entity_constraint["Size"][0], current_layout.entity_constraint["Size"][1] + 1) three_value_levels = np.random.choice(all_value_levels, 3, False) self.value_levels.append(three_value_levels[[0, 1, 2]]) if np.random.uniform() >= 0.5: self.value_levels.append(three_value_levels[[1, 2, 0]]) self.value_levels.append(three_value_levels[[2, 0, 1]]) else: self.value_levels.append(three_value_levels[[2, 0, 1]]) self.value_levels.append(three_value_levels[[1, 2, 0]]) for entity in current_layout.children: entity.size.set_value_level(self.value_levels[0][0]) for entity in second_layout.children: entity.size.set_value_level(self.value_levels[0][1]) else: row, col = divmod(self.count, 2) if col == 0: value_level = self.value_levels[row][0] for entity in current_layout.children: entity.size.set_value_level(value_level) value_level = self.value_levels[row][1] for entity in second_layout.children: entity.size.set_value_level(value_level) else: value_level = self.value_levels[row][2] for entity in second_layout.children: entity.size.set_value_level(value_level) self.count = (self.count + 1) % 6 elif self.attr == "Color": if self.count == 0: all_value_levels = range(current_layout.entity_constraint["Color"][0], current_layout.entity_constraint["Color"][1] + 1) three_value_levels = np.random.choice(all_value_levels, 3, False) self.value_levels.append(three_value_levels[[0, 1, 2]]) if np.random.uniform() >= 0.5: self.value_levels.append(three_value_levels[[1, 2, 0]]) self.value_levels.append(three_value_levels[[2, 0, 1]]) else: self.value_levels.append(three_value_levels[[2, 0, 1]]) self.value_levels.append(three_value_levels[[1, 2, 0]]) for entity in current_layout.children: entity.color.set_value_level(self.value_levels[0][0]) for entity in second_layout.children: entity.color.set_value_level(self.value_levels[0][1]) else: row, col = divmod(self.count, 2) if col == 0: value_level = self.value_levels[row][0] for entity in current_layout.children: entity.color.set_value_level(value_level) value_level = self.value_levels[row][1] for entity in second_layout.children: entity.color.set_value_level(value_level) else: value_level = self.value_levels[row][2] for entity in second_layout.children: entity.color.set_value_level(value_level) self.count = (self.count + 1) % 6 else: raise ValueError("Unsupported attriubute") return second_aot ================================================ FILE: src/dataset/__init__.py ================================================ """ RAVEN dataset generation code Author: Chi Zhang Data: 05/14/2019 Contact: chi.zhang@ucla.edu """ ================================================ FILE: src/dataset/api.py ================================================ # -*- coding: utf-8 -*- import xml.etree.ElementTree as ET import cv2 import numpy as np from const import DEFAULT_WIDTH, IMAGE_SIZE from rendering import render_entity class Bunch: """Dummy class""" def __init__(self, **kwds): self.__dict__.update(kwds) def get_real_bbox(entity_bbox, entity_type, entity_size, entity_angle): assert entity_type != "none" center = (int(entity_bbox[1] * IMAGE_SIZE), int(entity_bbox[0] * IMAGE_SIZE)) M = cv2.getRotationMatrix2D(center, entity_angle, 1) unit = min(entity_bbox[2], entity_bbox[3]) * IMAGE_SIZE / 2 delta = DEFAULT_WIDTH * 1.5 / IMAGE_SIZE if entity_type == "circle": radius = unit * entity_size real_bbox = [center[1] * 1.0 / IMAGE_SIZE, center[0] * 1.0 / IMAGE_SIZE, 2 * radius / IMAGE_SIZE + delta, 2 * radius / IMAGE_SIZE + delta] else: if entity_type == "triangle": dl = int(unit * entity_size) homo_pts = np.array([[center[0], center[1] - dl, 1], [center[0] + int(dl / 2.0 * np.sqrt(3)), center[1] + int(dl / 2.0), 1], [center[0] - int(dl / 2.0 * np.sqrt(3)), center[1] + int(dl / 2.0), 1]], np.int32) if entity_type == "square": dl = int(unit / 2 * np.sqrt(2) * entity_size) homo_pts = np.array([[center[0] - dl, center[1] - dl, 1], [center[0] - dl, center[1] + dl, 1], [center[0] + dl, center[1] + dl, 1], [center[0] + dl, center[1] - dl, 1]], np.int32) if entity_type == "pentagon": dl = int(unit * entity_size) homo_pts = np.array([[center[0], center[1] - dl, 1], [center[0] - int(dl * np.cos(np.pi / 10)), center[1] - int(dl * np.sin(np.pi / 10)), 1], [center[0] - int(dl * np.sin(np.pi / 5)), center[1] + int(dl * np.cos(np.pi / 5)), 1], [center[0] + int(dl * np.sin(np.pi / 5)), center[1] + int(dl * np.cos(np.pi / 5)), 1], [center[0] + int(dl * np.cos(np.pi / 10)), center[1] - int(dl * np.sin(np.pi / 10)), 1]], np.int32) if entity_type == "hexagon": dl = int(unit * entity_size) homo_pts = np.array([[center[0], center[1] - dl, 1], [center[0] - int(dl / 2.0 * np.sqrt(3)), center[1] - int(dl / 2.0), 1], [center[0] - int(dl / 2.0 * np.sqrt(3)), center[1] + int(dl / 2.0), 1], [center[0], center[1] + dl, 1], [center[0] + int(dl / 2.0 * np.sqrt(3)), center[1] + int(dl / 2.0), 1], [center[0] + int(dl / 2.0 * np.sqrt(3)), center[1] - int(dl / 2.0), 1]], np.int32) after_pts = np.dot(M, homo_pts.T) min_x = min(after_pts[1, :]) / IMAGE_SIZE max_x = max(after_pts[1, :]) / IMAGE_SIZE min_y = min(after_pts[0, :]) / IMAGE_SIZE max_y = max(after_pts[0, :]) / IMAGE_SIZE real_bbox = [(min_x + max_x) / 2, (min_y + max_y) / 2, max_x - min_x + delta, max_y - min_y + delta] return list(np.round(real_bbox, 4)) def get_mask(entity_bbox, entity_type, entity_size, entity_angle): dummy_entity = Bunch() dummy_entity.bbox = entity_bbox dummy_entity.type = Bunch(get_value=lambda : entity_type) dummy_entity.size = Bunch(get_value=lambda : entity_size) dummy_entity.color = Bunch(get_value=lambda : 0) dummy_entity.angle = Bunch(get_value=lambda : entity_angle) mask = render_entity(dummy_entity) / 255 return mask # ref: https://www.kaggle.com/stainsby/fast-tested-rle # ref: https://www.kaggle.com/paulorzp/run-length-encode-and-decode def rle_encode(img): ''' img: numpy array, 1 - mask, 0 - background Returns run length as string formated ''' pixels = img.flatten() pixels = np.concatenate([[0], pixels, [0]]) runs = np.where(pixels[1:] != pixels[:-1])[0] + 1 runs[1::2] -= runs[::2] return "[" + ",".join(str(x) for x in runs) + "]" def rle_decode(mask_rle, shape): ''' mask_rle: run-length as string formated (start length) shape: (height,width) of array to return Returns numpy array, 1 - mask, 0 - background ''' s = mask_rle[1:-1].split(",") starts, lengths = [np.asarray(x, dtype=int) for x in (s[0:][::2], s[1:][::2])] starts -= 1 ends = starts + lengths img = np.zeros(shape[0] * shape[1], dtype=np.uint8) for lo, hi in zip(starts, ends): img[lo:hi] = 1 return img.reshape(shape) ================================================ FILE: src/dataset/build_tree.py ================================================ # -*- coding: utf-8 -*- from AoT import Component, Layout, Root, Structure from constraints import (gen_entity_constraint, gen_layout_constraint, rule_constraint) def build_center_single(): # Build AoT here root = Root("Scene") # Singleton struct struct = Structure("Singleton") # Singleton comp comp = Component("Grid") # Center_Single layout entity_constraint = gen_entity_constraint(type_min=1) layout_constraint = gen_layout_constraint("planar", [(0.5, 0.5, 1, 1)], num_min=0, num_max=0) layout = Layout("Center_Single", layout_constraint, entity_constraint) comp.insert(layout) struct.insert(comp) root.insert(struct) return root def build_distribute_four(): # Build AoT here root = Root("Scene") # Singleton struct struct = Structure("Singleton") # Singleton comp comp = Component("Grid") # Distribute_Four entity_constraint = gen_entity_constraint(type_min=1) layout_constraint = gen_layout_constraint("planar", [(0.25, 0.25, 0.5, 0.5), (0.25, 0.75, 0.5, 0.5), (0.75, 0.25, 0.5, 0.5), (0.75, 0.75, 0.5, 0.5)], num_min=0, num_max=3) layout = Layout("Distribute_Four", layout_constraint, entity_constraint) comp.insert(layout) struct.insert(comp) root.insert(struct) return root def build_distribute_nine(): # Build AoT here root = Root("Scene") # Singleton struct struct = Structure("Singleton") # Singleton comp comp = Component("Grid") # Distribute_Nine entity_constraint = gen_entity_constraint(type_min=1) layout_constraint = gen_layout_constraint("planar", [(0.16, 0.16, 0.33, 0.33), (0.16, 0.5, 0.33, 0.33), (0.16, 0.83, 0.33, 0.33), (0.5, 0.16, 0.33, 0.33), (0.5, 0.5, 0.33, 0.33), (0.5, 0.83, 0.33, 0.33), (0.83, 0.16, 0.33, 0.33), (0.83, 0.5, 0.33, 0.33), (0.83, 0.83, 0.33, 0.33)], num_min=0, num_max=8) layout = Layout("Distribute_Nine", layout_constraint, entity_constraint) comp.insert(layout) struct.insert(comp) root.insert(struct) return root def build_left_center_single_right_center_single(): # Build AoT here root = Root("Scene") # Left-Right Structure struct = Structure("Left_Right") # Left Component comp_left = Component("Left") # Left_Center_Single entity_constraint = gen_entity_constraint(type_min=1) layout_constraint = gen_layout_constraint("planar", [(0.5, 0.25, 0.5, 0.5)], num_min=0, num_max=0) layout = Layout("Left_Center_Single", layout_constraint, entity_constraint) comp_left.insert(layout) # Right Component comp_right = Component("Right") # Right_Center_Single entity_constraint = gen_entity_constraint(type_min=1) layout_constraint = gen_layout_constraint("planar", [(0.5, 0.75, 0.5, 0.5)], num_min=0, num_max=0) layout = Layout("Right_Center_Single", layout_constraint, entity_constraint) comp_right.insert(layout) struct.insert(comp_left) struct.insert(comp_right) root.insert(struct) return root def build_up_center_single_down_center_single(): # Build AoT here root = Root("Scene") # Up-Down Structure struct = Structure("Up_Down") # Left Component comp_up = Component("Up") # Up_Center_Single entity_constraint = gen_entity_constraint(type_min=1) layout_constraint = gen_layout_constraint("planar", [(0.25, 0.5, 0.5, 0.5)], num_min=0, num_max=0) layout = Layout("Up_Center_Single", layout_constraint, entity_constraint) comp_up.insert(layout) # Down Component comp_down = Component("Down") # Down_Center_Single entity_constraint = gen_entity_constraint(type_min=1) layout_constraint = gen_layout_constraint("planar", [(0.75, 0.5, 0.5, 0.5)], num_min=0, num_max=0) layout = Layout("Down_Center_Single", layout_constraint, entity_constraint) comp_down.insert(layout) struct.insert(comp_up) struct.insert(comp_down) root.insert(struct) return root def build_in_center_single_out_center_single(): # Build AoT here root = Root("Scene") # In-Out Structure struct = Structure("Out_In") # Out Component comp_out = Component("Out") # Out_One entity_constraint = gen_entity_constraint(type_min=1, size_min=3, color_max=0) layout_constraint = gen_layout_constraint("planar", [(0.5, 0.5, 1, 1)], num_min=0, num_max=0) layout = Layout("Out_Center_Single", layout_constraint, entity_constraint) comp_out.insert(layout) # In Component comp_in = Component("In") # In_Center_Single entity_constraint = gen_entity_constraint(type_min=1) layout_constraint = gen_layout_constraint("planar", [(0.5, 0.5, 0.33, 0.33)], num_min=0, num_max=0) layout = Layout("In_Center_Single", layout_constraint, entity_constraint) comp_in.insert(layout) struct.insert(comp_out) struct.insert(comp_in) root.insert(struct) return root def build_in_distribute_four_out_center_single(): # Build AoT here root = Root("Scene") # In-Out Structure struct = Structure("Out_In") # Out Component comp_out = Component("Out") # Out_One entity_constraint = gen_entity_constraint(type_min=1, size_min=3, color_max=0) layout_constraint = gen_layout_constraint("planar", [(0.5, 0.5, 1, 1)], num_min=0, num_max=0) layout = Layout("Out_Center_Single", layout_constraint, entity_constraint) comp_out.insert(layout) # In Component comp_in = Component("In") # In_Four entity_constraint = gen_entity_constraint(type_min=1, size_min=2) layout_constraint = gen_layout_constraint("planar", [(0.42, 0.42, 0.15, 0.15), (0.42, 0.58, 0.15, 0.15), (0.58, 0.42, 0.15, 0.15), (0.58, 0.58, 0.15, 0.15)], num_min=0, num_max=3) layout = Layout("In_Distribute_Four", layout_constraint, entity_constraint) comp_in.insert(layout) struct.insert(comp_out) struct.insert(comp_in) root.insert(struct) return root ================================================ FILE: src/dataset/const.py ================================================ # -*- coding: utf-8 -*- # Maximum number of components in a RPM MAX_COMPONENTS = 2 # Canvas parameters IMAGE_SIZE = 160 CENTER = (IMAGE_SIZE / 2, IMAGE_SIZE / 2) DEFAULT_RADIUS = IMAGE_SIZE / 4 DEFAULT_WIDTH = 2 # Attribute parameters # Number NUM_VALUES = [1, 2, 3, 4, 5, 6, 7, 8, 9] NUM_MIN = 0 NUM_MAX = len(NUM_VALUES) - 1 # Uniformity UNI_VALUES = [False, False, False, True] UNI_MIN = 0 UNI_MAX = len(UNI_VALUES) - 1 # Type TYPE_VALUES = ["none", "triangle", "square", "pentagon", "hexagon", "circle"] TYPE_MIN = 0 TYPE_MAX = len(TYPE_VALUES) - 1 # Size SIZE_VALUES = [0.4, 0.5, 0.6, 0.7, 0.8, 0.9] SIZE_MIN = 0 SIZE_MAX = len(SIZE_VALUES) - 1 # Color COLOR_VALUES = [255, 224, 196, 168, 140, 112, 84, 56, 28, 0] COLOR_MIN = 0 COLOR_MAX = len(COLOR_VALUES) - 1 # Angle: self-rotation ANGLE_VALUES = [-135, -90, -45, 0, 45, 90, 135, 180] ANGLE_MIN = 0 ANGLE_MAX = len(ANGLE_VALUES) - 1 META_TARGET_FORMAT = ["Constant", "Progression", "Arithmetic", "Distribute_Three", "Number", "Position", "Type", "Size", "Color"] META_STRUCTURE_FORMAT = ["Singleton", "Left_Right", "Up_Down", "Out_In", "Left", "Right", "Up", "Down", "Out", "In", "Grid", "Center_Single", "Distribute_Four", "Distribute_Nine", "Left_Center_Single", "Right_Center_Single", "Up_Center_Single", "Down_Center_Single", "Out_Center_Single", "In_Center_Single", "In_Distribute_Four"] # Rule, Attr, Param # The design encodes rule priority order: Number/Position always comes first # Number and Position could not both be sampled # Progression on Number: Number on each Panel +1/2 or -1/2 # Progression on Position: Entities on each Panel roll over the layout # Arithmetic on Number: Numeber on the third Panel = Number on first +/- Number on second (1 for + and -1 for -) # Arithmetic on Position: 1 for SET_UNION and -1 for SET_DIFF # Distribute_Three on Number: Three numbers through each row # Distribute_Three on Position: Three positions (same number) through each row # Constant on Number/Position: Nothing changes # Progression on Type: Type progression defined as the number of edges on each entity (Triangle, Square, Pentagon, Hexagon, Circle) # Distribute_Three on Type: Three types through each row # Constant on Type: Nothing changes # Progression on Size: Size on each entity +1/2 or -1/2 # Arithmetic on Size: Size on the third Panel = Size on the first +/- Size on the second (1 for + and -1 for -) # Distribute_Three on Size: Three sizes through each row # Constant on Size: Nothing changes # Progression on Color: Color +1/2 or -1/2 # Arithmetic on Color: Color on the third Panel = Color on the first +/- Color on the second (1 for + and -1 for -) # Distribute_Three on Color: Three colors through each row # Constant on Color: Nothing changes # Note that all rules on Type, Size and Color enforce value consistency in a panel RULE_ATTR = [[["Progression", "Number", [-2, -1, 1, 2]], ["Progression", "Position", [-2, -1, 1, 2]], ["Arithmetic", "Number", [1, -1]], ["Arithmetic", "Position", [1, -1]], ["Distribute_Three", "Number", None], ["Distribute_Three", "Position", None], ["Constant", "Number/Position", None]], [["Progression", "Type", [-2, -1, 1, 2]], ["Distribute_Three", "Type", None], ["Constant", "Type", None]], [["Progression", "Size", [-2, -1, 1, 2]], ["Arithmetic", "Size", [1, -1]], ["Distribute_Three", "Size", None], ["Constant", "Size", None]], [["Progression", "Color", [-2, -1, 1, 2]], ["Arithmetic", "Color", [1, -1]], ["Distribute_Three", "Color", None], ["Constant", "Color", None]]] ================================================ FILE: src/dataset/constraints.py ================================================ # -*- coding: utf-8 -*- from const import (ANGLE_MAX, ANGLE_MIN, COLOR_MAX, COLOR_MIN, NUM_MAX, NUM_MIN, SIZE_MAX, SIZE_MIN, TYPE_MAX, TYPE_MIN, UNI_MAX, UNI_MIN) def gen_layout_constraint(pos_type, pos_list, num_min=NUM_MIN, num_max=NUM_MAX, uni_min=UNI_MIN, uni_max=UNI_MAX): constraint = {"Number": [num_min, num_max], "Position": [pos_type, pos_list[:]], "Uni": [uni_min, uni_max]} return constraint def gen_entity_constraint(type_min=TYPE_MIN, type_max=TYPE_MAX, size_min=SIZE_MIN, size_max=SIZE_MAX, color_min=COLOR_MIN, color_max=COLOR_MAX, angle_min=ANGLE_MIN, angle_max=ANGLE_MAX): constraint = {"Type": [type_min, type_max], "Size": [size_min, size_max], "Color": [color_min, color_max], "Angle": [angle_min, angle_max]} return constraint def rule_constraint(rule_list, num_min, num_max, uni_min, uni_max, type_min, type_max, size_min, size_max, color_min, color_max): """Generate constraints given the rules and the original constraints from layout and entity. Note that each attribute has at most one rule applied on it. Arguments: rule_list(ordered list of Rule): all rules applied to this layout others (int): boundary levels for each attribute in a layout; note that num_max + 1 == len(layout.position.values) Returns: layout_constraint(dict): a new layout constraint entity_constraint(dict): a new entity constraint """ assert len(rule_list) > 0 for rule in rule_list: if rule.name == "Progression": # rule.value: add/sub how many levels if rule.attr == "Number": if rule.value > 0: num_max = num_max - rule.value * 2 else: num_min = num_min - rule.value * 2 if rule.attr == "Position": # Progression here means moving in Layout slots in order abs_value = abs(rule.value) num_max = num_max - abs_value * 2 if rule.attr == "Type": if rule.value > 0: type_max = type_max - rule.value * 2 else: type_min = type_min - rule.value * 2 if rule.attr == "Size": if rule.value > 0: size_max = size_max - rule.value * 2 else: size_min = size_min - rule.value * 2 if rule.attr == "Color": if rule.value > 0: color_max = color_max - rule.value * 2 else: color_min = color_min - rule.value * 2 if rule.name == "Arithmetic": # rule.value > 0 if add col_0 + col_1 # rule.value < 0 if sub col_0 - col_1 if rule.attr == "Number": if rule.value > 0: num_max = num_max - num_min - 1 else: num_min = 2 * num_min + 1 if rule.attr == "Position": # SET_UNION # at least two position configurations if rule.value > 0: num_max = num_max - 1 # num_min makes sure of overlap # at least two configurations # SET_DIFF else: num_min = (num_max + 2) / 2 - 1 num_max = num_max - 1 if rule.attr == "Size": if rule.value > 0: size_max = size_max - size_min - 1 else: size_min = 2 * size_min + 1 if rule.attr == "Color": # at least two different colors if color_max - color_min < 1: color_max = color_min - 1 else: if rule.value > 0: color_max = color_max - color_min if rule.value < 0: color_min = 2 * color_min if rule.name == "Distribute_Three": # if less than 3 values, invalidate it if rule.attr == "Number": if num_max - num_min + 1 < 3: num_max = num_min - 1 if rule.attr == "Position": # max number allowed in the layout should be >= 3 if num_max + 1 < 3: num_max = num_min - 1 # num_max + 1 == len(layout.position.values) # C_{num_max + 1}^{num_value} >= 3 # C_{num_max + 1} = num_max + 1 >= 3 # hence only need to constrain num_max: num_max = num_max - 1 # Check Yang Hui’s Triangle (Pascal's Triangle): https://www.varsitytutors.com/hotmath/hotmath_help/topics/yang-huis-triangle else: num_max = num_max - 1 if rule.attr == "Type": if type_max - type_min + 1 < 3: type_max = type_min - 1 if rule.attr == "Size": if size_max - size_min + 1 < 3: size_max = size_min - 1 if rule.attr == "Color": if color_max - color_min + 1 < 3: color_max = color_min - 1 return gen_layout_constraint(None, [], num_min, num_max, uni_min, uni_max), \ gen_entity_constraint(type_min, type_max, size_min, size_max, color_min, color_max) ================================================ FILE: src/dataset/main.py ================================================ # -*- coding: utf-8 -*- import argparse import copy import os import random import sys import numpy as np from tqdm import trange from build_tree import (build_center_single, build_distribute_four, build_distribute_nine, build_in_center_single_out_center_single, build_in_distribute_four_out_center_single, build_left_center_single_right_center_single, build_up_center_single_down_center_single) from const import IMAGE_SIZE, RULE_ATTR from rendering import (generate_matrix, generate_matrix_answer, imsave, imshow, render_panel) from Rule import Rule_Wrapper from sampling import sample_attr, sample_attr_avail, sample_rules from serialize import dom_problem, serialize_aot, serialize_rules from solver import solve def merge_component(dst_aot, src_aot, component_idx): src_component = src_aot.children[0].children[component_idx] dst_aot.children[0].children[component_idx] = src_component def fuse(args, all_configs): random.seed(args.seed) np.random.seed(args.seed) acc = 0 for k in trange(args.num_samples * len(all_configs)): if k < args.num_samples * (1 - args.val - args.test): set_name = "train" elif k < args.num_samples * (1 - args.test): set_name = "val" else: set_name = "test" tree_name = random.choice(all_configs.keys()) root = all_configs[tree_name] while True: rule_groups = sample_rules() new_root = root.prune(rule_groups) if new_root is not None: break start_node = new_root.sample() row_1_1 = copy.deepcopy(start_node) for l in range(len(rule_groups)): rule_group = rule_groups[l] rule_num_pos = rule_group[0] row_1_2 = rule_num_pos.apply_rule(row_1_1) row_1_3 = rule_num_pos.apply_rule(row_1_2) for i in range(1, len(rule_group)): rule = rule_group[i] row_1_2 = rule.apply_rule(row_1_1, row_1_2) for i in range(1, len(rule_group)): rule = rule_group[i] row_1_3 = rule.apply_rule(row_1_2, row_1_3) if l == 0: to_merge = [row_1_1, row_1_2, row_1_3] else: merge_component(to_merge[1], row_1_2, l) merge_component(to_merge[2], row_1_3, l) row_1_1, row_1_2, row_1_3 = to_merge row_2_1 = copy.deepcopy(start_node) row_2_1.resample(True) for l in range(len(rule_groups)): rule_group = rule_groups[l] rule_num_pos = rule_group[0] row_2_2 = rule_num_pos.apply_rule(row_2_1) row_2_3 = rule_num_pos.apply_rule(row_2_2) for i in range(1, len(rule_group)): rule = rule_group[i] row_2_2 = rule.apply_rule(row_2_1, row_2_2) for i in range(1, len(rule_group)): rule = rule_group[i] row_2_3 = rule.apply_rule(row_2_2, row_2_3) if l == 0: to_merge = [row_2_1, row_2_2, row_2_3] else: merge_component(to_merge[1], row_2_2, l) merge_component(to_merge[2], row_2_3, l) row_2_1, row_2_2, row_2_3 = to_merge row_3_1 = copy.deepcopy(start_node) row_3_1.resample(True) for l in range(len(rule_groups)): rule_group = rule_groups[l] rule_num_pos = rule_group[0] row_3_2 = rule_num_pos.apply_rule(row_3_1) row_3_3 = rule_num_pos.apply_rule(row_3_2) for i in range(1, len(rule_group)): rule = rule_group[i] row_3_2 = rule.apply_rule(row_3_1, row_3_2) for i in range(1, len(rule_group)): rule = rule_group[i] row_3_3 = rule.apply_rule(row_3_2, row_3_3) if l == 0: to_merge = [row_3_1, row_3_2, row_3_3] else: merge_component(to_merge[1], row_3_2, l) merge_component(to_merge[2], row_3_3, l) row_3_1, row_3_2, row_3_3 = to_merge imgs = [render_panel(row_1_1), render_panel(row_1_2), render_panel(row_1_3), render_panel(row_2_1), render_panel(row_2_2), render_panel(row_2_3), render_panel(row_3_1), render_panel(row_3_2), np.zeros((IMAGE_SIZE, IMAGE_SIZE), np.uint8)] context = [row_1_1, row_1_2, row_1_3, row_2_1, row_2_2, row_2_3, row_3_1, row_3_2] modifiable_attr = sample_attr_avail(rule_groups, row_3_3) answer_AoT = copy.deepcopy(row_3_3) candidates = [answer_AoT] for j in range(7): component_idx, attr_name, min_level, max_level = sample_attr(modifiable_attr) answer_j = copy.deepcopy(answer_AoT) answer_j.sample_new(component_idx, attr_name, min_level, max_level, answer_AoT) candidates.append(answer_j) random.shuffle(candidates) answers = [] for candidate in candidates: answers.append(render_panel(candidate)) # imsave(generate_matrix_answer(imgs + answers), "./experiments/fuse/{}.jpg".format(k)) image = imgs[0:8] + answers target = candidates.index(answer_AoT) predicted = solve(rule_groups, context, candidates) meta_matrix, meta_target = serialize_rules(rule_groups) structure, meta_structure = serialize_aot(start_node) np.savez("{}/RAVEN_{}_{}.npz".format(args.save_dir, k, set_name), image=image, target=target, predict=predicted, meta_matrix=meta_matrix, meta_target=meta_target, structure=structure, meta_structure=meta_structure) with open("{}/RAVEN_{}_{}.xml".format(args.save_dir, k, set_name), "w") as f: dom = dom_problem(context + candidates, rule_groups) f.write(dom) if target == predicted: acc += 1 print "Accuracy: {}".format(float(acc) / (args.num_samples * len(all_configs))) def separate(args, all_configs): random.seed(args.seed) np.random.seed(args.seed) for key in all_configs.keys(): acc = 0 for k in trange(args.num_samples): count_num = k % 10 if count_num < (10 - args.val - args.test): set_name = "train" elif count_num < (10 - args.test): set_name = "val" else: set_name = "test" root = all_configs[key] while True: rule_groups = sample_rules() new_root = root.prune(rule_groups) if new_root is not None: break start_node = new_root.sample() row_1_1 = copy.deepcopy(start_node) for l in range(len(rule_groups)): rule_group = rule_groups[l] rule_num_pos = rule_group[0] row_1_2 = rule_num_pos.apply_rule(row_1_1) row_1_3 = rule_num_pos.apply_rule(row_1_2) for i in range(1, len(rule_group)): rule = rule_group[i] row_1_2 = rule.apply_rule(row_1_1, row_1_2) for i in range(1, len(rule_group)): rule = rule_group[i] row_1_3 = rule.apply_rule(row_1_2, row_1_3) if l == 0: to_merge = [row_1_1, row_1_2, row_1_3] else: merge_component(to_merge[1], row_1_2, l) merge_component(to_merge[2], row_1_3, l) row_1_1, row_1_2, row_1_3 = to_merge row_2_1 = copy.deepcopy(start_node) row_2_1.resample(True) for l in range(len(rule_groups)): rule_group = rule_groups[l] rule_num_pos = rule_group[0] row_2_2 = rule_num_pos.apply_rule(row_2_1) row_2_3 = rule_num_pos.apply_rule(row_2_2) for i in range(1, len(rule_group)): rule = rule_group[i] row_2_2 = rule.apply_rule(row_2_1, row_2_2) for i in range(1, len(rule_group)): rule = rule_group[i] row_2_3 = rule.apply_rule(row_2_2, row_2_3) if l == 0: to_merge = [row_2_1, row_2_2, row_2_3] else: merge_component(to_merge[1], row_2_2, l) merge_component(to_merge[2], row_2_3, l) row_2_1, row_2_2, row_2_3 = to_merge row_3_1 = copy.deepcopy(start_node) row_3_1.resample(True) for l in range(len(rule_groups)): rule_group = rule_groups[l] rule_num_pos = rule_group[0] row_3_2 = rule_num_pos.apply_rule(row_3_1) row_3_3 = rule_num_pos.apply_rule(row_3_2) for i in range(1, len(rule_group)): rule = rule_group[i] row_3_2 = rule.apply_rule(row_3_1, row_3_2) for i in range(1, len(rule_group)): rule = rule_group[i] row_3_3 = rule.apply_rule(row_3_2, row_3_3) if l == 0: to_merge = [row_3_1, row_3_2, row_3_3] else: merge_component(to_merge[1], row_3_2, l) merge_component(to_merge[2], row_3_3, l) row_3_1, row_3_2, row_3_3 = to_merge imgs = [render_panel(row_1_1), render_panel(row_1_2), render_panel(row_1_3), render_panel(row_2_1), render_panel(row_2_2), render_panel(row_2_3), render_panel(row_3_1), render_panel(row_3_2), np.zeros((IMAGE_SIZE, IMAGE_SIZE), np.uint8)] context = [row_1_1, row_1_2, row_1_3, row_2_1, row_2_2, row_2_3, row_3_1, row_3_2] modifiable_attr = sample_attr_avail(rule_groups, row_3_3) answer_AoT = copy.deepcopy(row_3_3) candidates = [answer_AoT] for j in range(7): component_idx, attr_name, min_level, max_level = sample_attr(modifiable_attr) answer_j = copy.deepcopy(answer_AoT) answer_j.sample_new(component_idx, attr_name, min_level, max_level, answer_AoT) candidates.append(answer_j) random.shuffle(candidates) answers = [] for candidate in candidates: answers.append(render_panel(candidate)) # imsave(generate_matrix_answer(imgs + answers), "./experiments/{}/{}.jpg".format(key, k)) image = imgs[0:8] + answers target = candidates.index(answer_AoT) predicted = solve(rule_groups, context, candidates) meta_matrix, meta_target = serialize_rules(rule_groups) structure, meta_structure = serialize_aot(start_node) np.savez("{}/{}/RAVEN_{}_{}.npz".format(args.save_dir, key, k, set_name), image=image, target=target, predict=predicted, meta_matrix=meta_matrix, meta_target=meta_target, structure=structure, meta_structure=meta_structure) with open("{}/{}/RAVEN_{}_{}.xml".format(args.save_dir, key, k, set_name), "w") as f: dom = dom_problem(context + candidates, rule_groups) f.write(dom) if target == predicted: acc += 1 print "Accuracy of {}: {}".format(key, float(acc) / args.num_samples) def main(): main_arg_parser = argparse.ArgumentParser(description="parser for RAVEN") main_arg_parser.add_argument("--num-samples", type=int, default=20000, help="number of samples for each component configuration") main_arg_parser.add_argument("--save-dir", type=str, default="~/Datasets/", help="path to folder where the generated dataset will be saved.") main_arg_parser.add_argument("--seed", type=int, default=1234, help="random seed for dataset generation") main_arg_parser.add_argument("--fuse", type=int, default=0, help="whether to fuse different configurations") main_arg_parser.add_argument("--val", type=float, default=2, help="the proportion of the size of validation set") main_arg_parser.add_argument("--test", type=float, default=2, help="the proportion of the size of test set") args = main_arg_parser.parse_args() all_configs = {"center_single": build_center_single(), "distribute_four": build_distribute_four(), "distribute_nine": build_distribute_nine(), "left_center_single_right_center_single": build_left_center_single_right_center_single(), "up_center_single_down_center_single": build_up_center_single_down_center_single(), "in_center_single_out_center_single": build_in_center_single_out_center_single(), "in_distribute_four_out_center_single": build_in_distribute_four_out_center_single()} if not os.path.exists(args.save_dir): os.mkdir(args.save_dir) if args.fuse: if not os.path.exists(os.path.join(args.save_dir, "fuse")): os.mkdir(os.path.join(args.save_dir, "fuse")) fuse(args, all_configs) else: for key in all_configs.keys(): if not os.path.exists(os.path.join(args.save_dir, key)): os.mkdir(os.path.join(args.save_dir, key)) separate(args, all_configs) if __name__ == "__main__": main() ================================================ FILE: src/dataset/rendering.py ================================================ # -*- coding: utf-8 -*- import cv2 import numpy as np from PIL import Image from AoT import Root from const import CENTER, DEFAULT_WIDTH, IMAGE_SIZE def imshow(array): image = Image.fromarray(array) image.show() def imsave(array, filepath): image = Image.fromarray(array) image.save(filepath) def generate_matrix(array_list): # row-major array_list assert len(array_list) <= 9 img_grid = np.zeros((IMAGE_SIZE * 3, IMAGE_SIZE * 3), np.uint8) for idx in range(len(array_list)): i, j = divmod(idx, 3) img_grid[i * IMAGE_SIZE:(i + 1) * IMAGE_SIZE, j * IMAGE_SIZE:(j + 1) * IMAGE_SIZE] = array_list[idx] # draw grid for x in [0.33, 0.67]: img_grid[int(x * IMAGE_SIZE * 3) - 1:int(x * IMAGE_SIZE * 3) + 1, :] = 0 for y in [0.33, 0.67]: img_grid[:, int(y * IMAGE_SIZE * 3) - 1:int(y * IMAGE_SIZE * 3) + 1] = 0 return img_grid def generate_answers(array_list): assert len(array_list) <= 8 img_grid = np.zeros((IMAGE_SIZE * 2, IMAGE_SIZE * 4), np.uint8) for idx in range(len(array_list)): i, j = divmod(idx, 4) img_grid[i * IMAGE_SIZE:(i + 1) * IMAGE_SIZE, j * IMAGE_SIZE:(j + 1) * IMAGE_SIZE] = array_list[idx] # draw grid for x in [0.5]: img_grid[int(x * IMAGE_SIZE * 2) - 1:int(x * IMAGE_SIZE * 2) + 1, :] = 0 for y in [0.25, 0.5, 0.75]: img_grid[:, int(y * IMAGE_SIZE * 4) - 1:int(y * IMAGE_SIZE * 4) + 1] = 0 return img_grid def generate_matrix_answer(array_list): # row-major array_list assert len(array_list) <= 18 img_grid = np.zeros((IMAGE_SIZE * 6, IMAGE_SIZE * 3), np.uint8) for idx in range(len(array_list)): i, j = divmod(idx, 3) img_grid[i * IMAGE_SIZE:(i + 1) * IMAGE_SIZE, j * IMAGE_SIZE:(j + 1) * IMAGE_SIZE] = array_list[idx] # draw grid for x in [0.33, 0.67, 1.00, 1.33, 1.67]: img_grid[int(x * IMAGE_SIZE * 3), :] = 0 for y in [0.33, 0.67]: img_grid[:, int(y * IMAGE_SIZE * 3)] = 0 return img_grid def merge_matrix_answer(matrix, answer): matrix_image = generate_matrix(matrix) answer_image = generate_answers(answer) img_grid = np.ones((IMAGE_SIZE * 5 + 20, IMAGE_SIZE * 4), np.uint8) * 255 img_grid[:IMAGE_SIZE * 3, int(0.5 * IMAGE_SIZE):int(3.5 * IMAGE_SIZE)] = matrix_image img_grid[-(IMAGE_SIZE * 2):, :] = answer_image return img_grid def render_panel(root): # Decompose the panel into a structure and its entities assert isinstance(root, Root) canvas = np.ones((IMAGE_SIZE, IMAGE_SIZE), np.uint8) * 255 structure, entities = root.prepare() structure_img = render_structure(structure) background = np.zeros((IMAGE_SIZE, IMAGE_SIZE), np.uint8) # note left components entities are in the lower layer for entity in entities: entity_img = render_entity(entity) background = layer_add(background, entity_img) background = layer_add(background, structure_img) return canvas - background def render_structure(structure_name): ret = None if structure_name == "Left_Right": ret = np.zeros((IMAGE_SIZE, IMAGE_SIZE), np.uint8) ret[:, int(0.5 * IMAGE_SIZE)] = 255.0 elif structure_name == "Up_Down": ret = np.zeros((IMAGE_SIZE, IMAGE_SIZE), np.uint8) ret[int(0.5 * IMAGE_SIZE), :] = 255.0 else: ret = np.zeros((IMAGE_SIZE, IMAGE_SIZE), np.uint8) return ret def render_entity(entity): entity_bbox = entity.bbox entity_type = entity.type.get_value() entity_size = entity.size.get_value() entity_color = entity.color.get_value() entity_angle = entity.angle.get_value() img = np.zeros((IMAGE_SIZE, IMAGE_SIZE), np.uint8) # planar position: [x, y, w, h] # angular position: [x, y, w, h, x_c, y_c, omega] # center: (columns, rows) center = (int(entity_bbox[1] * IMAGE_SIZE), int(entity_bbox[0] * IMAGE_SIZE)) if entity_type == "triangle": unit = min(entity_bbox[2], entity_bbox[3]) * IMAGE_SIZE / 2 dl = int(unit * entity_size) pts = np.array([[center[0], center[1] - dl], [center[0] + int(dl / 2.0 * np.sqrt(3)), center[1] + int(dl / 2.0)], [center[0] - int(dl / 2.0 * np.sqrt(3)), center[1] + int(dl / 2.0)]], np.int32) pts = pts.reshape((-1, 1, 2)) color = 255 - entity_color width = DEFAULT_WIDTH draw_triangle(img, pts, color, width) elif entity_type == "square": unit = min(entity_bbox[2], entity_bbox[3]) * IMAGE_SIZE / 2 dl = int(unit / 2 * np.sqrt(2) * entity_size) pt1 = (center[0] - dl, center[1] - dl) pt2 = (center[0] + dl, center[1] + dl) color = 255 - entity_color width = DEFAULT_WIDTH draw_square(img, pt1, pt2, color, width) elif entity_type == "pentagon": unit = min(entity_bbox[2], entity_bbox[3]) * IMAGE_SIZE / 2 dl = int(unit * entity_size) pts = np.array([[center[0], center[1] - dl], [center[0] - int(dl * np.cos(np.pi / 10)), center[1] - int(dl * np.sin(np.pi / 10))], [center[0] - int(dl * np.sin(np.pi / 5)), center[1] + int(dl * np.cos(np.pi / 5))], [center[0] + int(dl * np.sin(np.pi / 5)), center[1] + int(dl * np.cos(np.pi / 5))], [center[0] + int(dl * np.cos(np.pi / 10)), center[1] - int(dl * np.sin(np.pi / 10))]], np.int32) pts = pts.reshape((-1, 1, 2)) color = 255 - entity_color width = DEFAULT_WIDTH draw_pentagon(img, pts, color, width) elif entity_type == "hexagon": unit = min(entity_bbox[2], entity_bbox[3]) * IMAGE_SIZE / 2 dl = int(unit * entity_size) pts = np.array([[center[0], center[1] - dl], [center[0] - int(dl / 2.0 * np.sqrt(3)), center[1] - int(dl / 2.0)], [center[0] - int(dl / 2.0 * np.sqrt(3)), center[1] + int(dl / 2.0)], [center[0], center[1] + dl], [center[0] + int(dl / 2.0 * np.sqrt(3)), center[1] + int(dl / 2.0)], [center[0] + int(dl / 2.0 * np.sqrt(3)), center[1] - int(dl / 2.0)]], np.int32) pts = pts.reshape((-1, 1, 2)) color = 255 - entity_color width = DEFAULT_WIDTH draw_hexagon(img, pts, color, width) elif entity_type == "circle": # Minus because of the way we show the image. See: render_panel's return color = 255 - entity_color unit = min(entity_bbox[2], entity_bbox[3]) * IMAGE_SIZE / 2 radius = int(unit * entity_size) width = DEFAULT_WIDTH draw_circle(img, center, radius, color, width) elif entity_type == "none": pass # angular if len(entity_bbox) > 4: # [x, y, w, h, x_c, y_c, omega] entity_angle = entity_bbox[6] center = (int(entity_bbox[5] * IMAGE_SIZE), int(entity_bbox[4] * IMAGE_SIZE)) img = rotate(img, entity_angle, center=center) # planar else: img = rotate(img, entity_angle, center=center) # img = shift(img, *entity_position) return img def shift(img, dx, dy): M = np.array([[1, 0, dx], [0, 1, dy]], np.float32) img = cv2.warpAffine(img, M, (IMAGE_SIZE, IMAGE_SIZE), flags=cv2.INTER_LINEAR) return img def rotate(img, angle, center=CENTER): M = cv2.getRotationMatrix2D(center, angle, 1) img = cv2.warpAffine(img, M, (IMAGE_SIZE, IMAGE_SIZE), flags=cv2.INTER_LINEAR) return img def scale(img, tx, ty, center=CENTER): M = np.array([[tx, 0, center[0] * (1 - tx)], [0, ty, center[1] * (1 - ty)]], np.float32) img = cv2.warpAffine(img, M, (IMAGE_SIZE, IMAGE_SIZE), flags=cv2.INTER_LINEAR) return img def layer_add(lower_layer_np, higher_layer_np): # higher_layer_np is superimposed on lower_layer_np # new_np = lower_layer_np.copy() # lower_layer_np is modified lower_layer_np[higher_layer_np > 0] = 0 return lower_layer_np + higher_layer_np # Draw primitives def draw_triangle(img, pts, color, width): # if filled if color != 0: # fill the interior cv2.fillConvexPoly(img, pts, color) # draw the edge cv2.polylines(img, [pts], True, 255, width) # if not filled else: cv2.polylines(img, [pts], True, 255, width) def draw_square(img, pt1, pt2, color, width): # if filled if color != 0: # fill the interior cv2.rectangle(img, pt1, pt2, color, -1) # draw the edge cv2.rectangle(img, pt1, pt2, 255, width) # if not filled else: cv2.rectangle(img, pt1, pt2, 255, width) def draw_pentagon(img, pts, color, width): # if filled if color != 0: # fill the interior cv2.fillConvexPoly(img, pts, color) # draw the edge cv2.polylines(img, [pts], True, 255, width) # if not filled else: cv2.polylines(img, [pts], True, 255, width) def draw_hexagon(img, pts, color, width): # if filled if color != 0: # fill the interior cv2.fillConvexPoly(img, pts, color) # draw the edge cv2.polylines(img, [pts], True, 255, width) # if not filled else: cv2.polylines(img, [pts], True, 255, width) def draw_circle(img, center, radius, color, width): # if filled if color != 0: # fill the interior cv2.circle(img, center, radius, color, -1) # draw the edge cv2.circle(img, center, radius, 255, width) # if not filled else: cv2.circle(img, center, radius, 255, width) ================================================ FILE: src/dataset/sampling.py ================================================ # -*- coding: utf-8 -*- import numpy as np from scipy.misc import comb from const import MAX_COMPONENTS, RULE_ATTR from Rule import Rule_Wrapper def sample_rules(): """First sample # components; for each component, sample a rule on each attribute. """ num_components = np.random.randint(1, MAX_COMPONENTS + 1) all_rules = [] for i in range(num_components): all_rules_component = [] for j in range(len(RULE_ATTR)): idx = np.random.choice(len(RULE_ATTR[j])) name_attr_param = RULE_ATTR[j][idx] all_rules_component.append(Rule_Wrapper(name_attr_param[0], name_attr_param[1], name_attr_param[2], component_idx=i)) all_rules.append(all_rules_component) return all_rules # pay attention to Position Arithmetic, new entities (resample) def sample_attr_avail(rule_groups, row_3_3): """Sample available attributes whose values could be modified. Arguments: rule_groups(list of list of Rule): a list of rules to apply to the component row_3_3(AoTNode): the answer AoT Returns: ret(list of list): [component_idx, attr, available_times, constraints] """ ret = [] for i in range(len(rule_groups)): rule_group = rule_groups[i] start_node_layout = row_3_3.children[0].children[i].children[0] row_3_3_layout = row_3_3.children[0].children[i].children[0] uni = row_3_3_layout.uniformity.get_value() # Number/Position # If Rule on Number: Only change Number # If Rule on Position: Both Number and Position could be changed rule = rule_group[0] num = row_3_3_layout.number.get_value() most_num = len(start_node_layout.position.values) if rule.attr == "Number": num_times = 0 min_level = start_node_layout.orig_layout_constraint["Number"][0] max_level = start_node_layout.orig_layout_constraint["Number"][1] for k in range(min_level, max_level + 1): if k + 1 != num: num_times += comb(most_num, k + 1) if num_times > 0: ret.append([i, "Number", num_times, min_level, max_level]) # Constant or on Position else: num_times = 0 min_level = start_node_layout.orig_layout_constraint["Number"][0] max_level = start_node_layout.orig_layout_constraint["Number"][1] for k in range(min_level, max_level + 1): if k + 1 != num: num_times += comb(most_num, k + 1) if num_times > 0: ret.append([i, "Number", num_times, min_level, max_level]) pos_times = comb(most_num, row_3_3_layout.number.get_value()) pos_times -= 1 if pos_times > 0: ret.append([i, "Position", pos_times, None, None]) # Type, Size, Color for j in range(1, len(rule_group)): rule = rule_group[j] rule_attr = rule.attr min_level = start_node_layout.orig_entity_constraint[rule_attr][0] max_level = start_node_layout.orig_entity_constraint[rule_attr][1] if rule.name == "Constant": if uni or rule_group[0].name == "Constant" or \ (rule_group[0].attr == "Position" and (rule_group[0].name == "Progression" or rule_group[0].name == "Distribute_Three")): times = max_level - min_level + 1 times = times - 1 if times > 0: ret.append([i, rule_attr, times, min_level, max_level]) else: times = max_level - min_level + 1 times = times - 1 if times > 0: ret.append([i, rule_attr, times, min_level, max_level]) return ret def sample_attr(attrs_list): """Given the attr_avail list, sample one attribute to modify the value. If the available times becomes zero, delete it. Arguments: attrs_list(list of list): a flat component of available attributes to change the values; consisting of different component indexes """ attr_idx = np.random.choice(len(attrs_list)) component_idx, attr_name, _, min_level, max_level = attrs_list[attr_idx] attrs_list[attr_idx][2] -= 1 if attrs_list[attr_idx][2] == 0: del attrs_list[attr_idx] return component_idx, attr_name, min_level, max_level ================================================ FILE: src/dataset/serialize.py ================================================ # -*- coding: utf-8 -*- import json import xml.etree.ElementTree as ET import numpy as np from const import META_STRUCTURE_FORMAT from api import get_real_bbox, get_mask, rle_encode def n_tree_serialize(aot): assert aot.is_pg ret = "" if aot.level == "Layout": return aot.name + "./" else: ret += aot.name + "." for child in aot.children: x = n_tree_serialize(child) ret += x ret += "." ret += "/" return ret def serialize_aot(aot): """Meta Structure format META_STRUCTURE_FORMAT provided by const.py """ n_tree = n_tree_serialize(aot) meta_structure = np.zeros(len(META_STRUCTURE_FORMAT), np.uint8) split = n_tree.split(".") for node in split: try: node_index = META_STRUCTURE_FORMAT.index(node) meta_structure[node_index] = 1 except ValueError: continue return split, meta_structure def serialize_rules(rule_groups): """Meta matrix format ["Constant", "Progression", "Arithmetic", "Distribute_Three", "Number", "Position", "Type", "Size", "Color"] """ meta_matrix = np.zeros((8, 9), np.uint8) counter = 0 for rule_group in rule_groups: for rule in rule_group: if rule.name == "Constant": meta_matrix[counter, 0] = 1 elif rule.name == "Progression": meta_matrix[counter, 1] = 1 elif rule.name == "Arithmetic": meta_matrix[counter, 2] = 1 else: meta_matrix[counter, 3] = 1 if rule.attr == "Number/Position": meta_matrix[counter, 4] = 1 meta_matrix[counter, 5] = 1 elif rule.attr == "Number": meta_matrix[counter, 4] = 1 elif rule.attr == "Position": meta_matrix[counter, 5] = 1 elif rule.attr == "Type": meta_matrix[counter, 6] = 1 elif rule.attr == "Size": meta_matrix[counter, 7] = 1 else: meta_matrix[counter, 8] = 1 counter += 1 return meta_matrix, np.bitwise_or.reduce(meta_matrix) def dom_problem(instances, rule_groups): data = ET.Element("Data") panels = ET.SubElement(data, "Panels") for i in range(len(instances)): panel = instances[i] panel_i = ET.SubElement(panels, "Panel") struct = panel.children[0] struct_i = ET.SubElement(panel_i, "Struct") struct_i.set("name", struct.name) for j in range(len(struct.children)): component = struct.children[j] component_j = ET.SubElement(struct_i, "Component") component_j.set("id", str(j)) component_j.set("name", component.name) layout = component.children[0] layout_k = ET.SubElement(component_j, "Layout") layout_k.set("name", layout.name) layout_k.set("Number", str(layout.number.get_value_level())) layout_k.set("Position", json.dumps(layout.position.values)) layout_k.set("Uniformity", str(layout.uniformity.get_value_level())) for l in range(len(layout.children)): entity = layout.children[l] entity_l = ET.SubElement(layout_k, "Entity") entity_bbox = entity.bbox entity_type = entity.type.get_value() entity_size = entity.size.get_value() entity_angle = entity.angle.get_value() entity_l.set("bbox", json.dumps(entity_bbox)) entity_l.set("real_bbox", json.dumps(get_real_bbox(entity_bbox, entity_type, entity_size, entity_angle))) entity_l.set("mask", rle_encode(get_mask(entity_bbox, entity_type, entity_size, entity_angle))) entity_l.set("Type", str(entity.type.get_value_level())) entity_l.set("Size", str(entity.size.get_value_level())) entity_l.set("Color", str(entity.color.get_value_level())) entity_l.set("Angle", str(entity.angle.get_value_level())) rules = ET.SubElement(data, "Rules") for i in range(len(rule_groups)): rule_group = rule_groups[i] rule_group_i = ET.SubElement(rules, "Rule_Group") rule_group_i.set("id", str(i)) for rule in rule_group: rule_j = ET.SubElement(rule_group_i, "Rule") rule_j.set("name", rule.name) rule_j.set("attr", rule.attr) return ET.tostring(data) ================================================ FILE: src/dataset/solver.py ================================================ # -*- coding: utf-8 -*- import numpy as np def solve(rule_groups, context, candidates): """Search-based Heuristic Solver. Arguments: rule_groups(list of list of Rule): rules that apply to each component context(list of AoTNode): a list of context AoTs in a row-major order; should be of length 8 candidates(list of AoTNode): a list of candidate answer AoTs; should be of length 8 Returns: ans(int): index of the correct answer in the candidates """ satisfied = [0] * len(candidates) for i in range(len(candidates)): candidate = candidates[i] # note that rule.component_idx should be the same as j for j in range(len(rule_groups)): rule_group = rule_groups[j] rule_num_pos = rule_group[0] satisfied[i] += check_num_pos(rule_num_pos, context, candidate) regenerate = False if rule_num_pos.attr == "Number" or rule_num_pos.name == "Arithmetic": regenerate = True rule_type = rule_group[1] satisfied[i] += check_entity(rule_type, context, candidate, "Type", regenerate) rule_size = rule_group[2] satisfied[i] += check_entity(rule_size, context, candidate, "Size", regenerate) rule_color = rule_group[3] satisfied[i] += check_entity(rule_color, context, candidate, "Color", regenerate) satisfied = np.array(satisfied) answer_set = np.where(satisfied == max(satisfied))[0] return np.random.choice(answer_set) def check_num_pos(rule_num_pos, context, candidate): """Check whether Rule on layout attribute is satisfied. Arguments: rule_num_pos(Rule): the rule to check context(list of AoTNode): the 8 context figures candidate(AoTNode): the candidate AoT Returns: ret(int): 0 if failure, 1 if success """ ret = 0 component_idx = rule_num_pos.component_idx row_3_1_layout = context[6].children[0].children[component_idx].children[0] row_3_2_layout = context[7].children[0].children[component_idx].children[0] candidate_layout = candidate.children[0].children[component_idx].children[0] if rule_num_pos.name == "Constant": set_row_3_1_pos = set(row_3_1_layout.position.get_value_idx()) set_row_3_2_pos = set(row_3_2_layout.position.get_value_idx()) set_candidate_pos = set(candidate_layout.position.get_value_idx()) # note that set equal only when len(Number) equal and content equal if set_candidate_pos == set_row_3_1_pos and set_candidate_pos == set_row_3_2_pos: ret = 1 elif rule_num_pos.name == "Progression": if rule_num_pos.attr == "Number": row_3_1_num = row_3_1_layout.number.get_value_level() row_3_2_num = row_3_2_layout.number.get_value_level() candidate_num = candidate_layout.number.get_value_level() if row_3_2_num * 2 == row_3_1_num + candidate_num: ret = 1 else: row_3_1_pos = row_3_1_layout.position.get_value_idx() row_3_2_pos = row_3_2_layout.position.get_value_idx() candidate_pos = candidate_layout.position.get_value_idx() most_num = len(candidate_layout.position.values) diff = rule_num_pos.value if (set((row_3_1_pos + diff) % most_num) == set(row_3_2_pos)) and \ (set((row_3_2_pos + diff) % most_num) == set(candidate_pos)): ret = 1 elif rule_num_pos.name == "Arithmetic": mode = rule_num_pos.value if rule_num_pos.attr == "Number": row_3_1_num = row_3_1_layout.number.get_value() row_3_2_num = row_3_2_layout.number.get_value() candidate_num = candidate_layout.number.get_value() if mode > 0 and (candidate_num == row_3_1_num + row_3_2_num): ret = 1 if mode < 0 and (candidate_num == row_3_1_num - row_3_2_num): ret = 1 else: row_3_1_pos = row_3_1_layout.position.get_value_idx() row_3_2_pos = row_3_2_layout.position.get_value_idx() candidate_pos = candidate_layout.position.get_value_idx() if mode > 0 and (set(candidate_pos) == set(row_3_1_pos) | set(row_3_2_pos)): ret = 1 if mode < 0 and (set(candidate_pos) == set(row_3_1_pos) - set(row_3_2_pos)): ret = 1 else: three_values = rule_num_pos.value_levels[2] if rule_num_pos.attr == "Number": row_3_1_num = row_3_1_layout.number.get_value_level() row_3_2_num = row_3_2_layout.number.get_value_level() candidate_num = candidate_layout.number.get_value_level() if row_3_1_num == three_values[0] and \ row_3_2_num == three_values[1] and \ candidate_num == three_values[2]: ret = 1 else: row_3_1_pos = row_3_1_layout.position.get_value_idx() row_3_2_pos = row_3_2_layout.position.get_value_idx() candidate_pos = candidate_layout.position.get_value_idx() if set(row_3_1_pos) == set(three_values[0]) and \ set(row_3_2_pos) == set(three_values[1]) and \ set(candidate_pos) == set(three_values[2]): ret = 1 return ret def check_consistency(candidate, attr, component_idx): candidate_layout = candidate.children[0].children[component_idx].children[0] entity_0 = candidate_layout.children[0] attr_name = attr.lower() entity_0_value = getattr(entity_0, attr_name).get_value_level() for i in range(1, len(candidate_layout.children)): entity_i = candidate_layout.children[i] entity_i_value = getattr(entity_i, attr_name).get_value_level() if entity_i_value != entity_0_value: return False return True def check_entity(rule, context, candidate, attr, regenerate): """Check whether Rule on entity attribute is satisfied. Arguments: rule(Rule): the rule to check context(list of AoTNode): the 8 context figures candidate(AoTNode): the candidate AoT attr(str): attribute name Returns: ret(int): 0 if failure, 1 if success """ ret = 0 component_idx = rule.component_idx row_3_1_layout = context[6].children[0].children[component_idx].children[0] row_3_2_layout = context[7].children[0].children[component_idx].children[0] candidate_layout = candidate.children[0].children[component_idx].children[0] uni = candidate_layout.uniformity.get_value() attr_name = attr.lower() if rule.name == "Constant": if uni: if check_consistency(candidate, attr, component_idx): if getattr(candidate_layout.children[0], attr_name).get_value_level() == \ getattr(row_3_2_layout.children[0], attr_name).get_value_level(): ret = 1 else: row_3_1_num = row_3_1_layout.number.get_value_level() row_3_2_num = row_3_2_layout.number.get_value_level() candidate_num = candidate_layout.number.get_value_level() if (row_3_1_num == row_3_2_num) and (row_3_2_num == candidate_num): if regenerate: ret = 1 else: flag = True for i in range(len(candidate_layout.children)): if not (getattr(candidate_layout.children[i], attr_name).get_value_level() == getattr(row_3_2_layout.children[i], attr_name).get_value_level()): flag = False break if flag: ret = 1 else: ret = 1 elif rule.name == "Progression": if check_consistency(candidate, attr, component_idx): row_3_1_value = getattr(row_3_1_layout.children[0], attr_name).get_value_level() row_3_2_value = getattr(row_3_2_layout.children[0], attr_name).get_value_level() candidate_value = getattr(candidate_layout.children[0], attr_name).get_value_level() if row_3_2_value * 2 == row_3_1_value + candidate_value: ret = 1 elif rule.name == "Arithmetic": if check_consistency(candidate, attr, component_idx): row_3_1_value = getattr(row_3_1_layout.children[0], attr_name).get_value_level() row_3_2_value = getattr(row_3_2_layout.children[0], attr_name).get_value_level() candidate_value = getattr(candidate_layout.children[0], attr_name).get_value_level() if rule.value > 0: if attr == "Color": if candidate_value == row_3_1_value + row_3_2_value: ret = 1 else: if candidate_value == row_3_1_value + row_3_2_value + 1: ret = 1 if rule.value < 0: if attr == "Color": if candidate_value == row_3_1_value - row_3_2_value: ret = 1 else: if candidate_value == row_3_1_value - row_3_2_value - 1: ret = 1 else: if check_consistency(candidate, attr, component_idx): row_3_1_value = getattr(row_3_1_layout.children[0], attr_name).get_value_level() row_3_2_value = getattr(row_3_2_layout.children[0], attr_name).get_value_level() candidate_value = getattr(candidate_layout.children[0], attr_name).get_value_level() three_values = rule.value_levels[2] if row_3_1_value == three_values[0] and \ row_3_2_value == three_values[1] and \ candidate_value == three_values[2]: ret = 1 return ret ================================================ FILE: src/model/__init__.py ================================================ """ RAVEN benchmarking code Author: Chi Zhang Data: 05/14/2019 Contact: chi.zhang@ucla.edu """ ================================================ FILE: src/model/basic_model.py ================================================ import torch import torch.nn as nn import torch.nn.functional as F class BasicModel(nn.Module): def __init__(self, args): super(BasicModel, self).__init__() self.name = args.model def load_model(self, path, epoch): state_dict = torch.load(path+'{}_epoch_{}.pth'.format(self.name, epoch))['state_dict'] self.load_state_dict(state_dict) def save_model(self, path, epoch, acc, loss): torch.save({'state_dict': self.state_dict(), 'acc': acc, 'loss': loss}, path+'{}_epoch_{}.pth'.format(self.name, epoch)) def compute_loss(self, output, target, meta_target, meta_structure): pass def train_(self, image, target, meta_target, meta_structure, embedding, indicator): self.optimizer.zero_grad() output = self(image, embedding, indicator) loss = self.compute_loss(output, target, meta_target, meta_structure) loss.backward() self.optimizer.step() pred = output[0].data.max(1)[1] correct = pred.eq(target.data).cpu().sum().numpy() accuracy = correct * 100.0 / target.size()[0] return loss.item(), accuracy def validate_(self, image, target, meta_target, meta_structure, embedding, indicator): with torch.no_grad(): output = self(image, embedding, indicator) loss = self.compute_loss(output, target, meta_target, meta_structure) pred = output[0].data.max(1)[1] correct = pred.eq(target.data).cpu().sum().numpy() accuracy = correct * 100.0 / target.size()[0] return loss.item(), accuracy def test_(self, image, target, meta_target, meta_structure, embedding, indicator): with torch.no_grad(): output = self(image, embedding, indicator) pred = output[0].data.max(1)[1] correct = pred.eq(target.data).cpu().sum().numpy() accuracy = correct * 100.0 / target.size()[0] return accuracy ================================================ FILE: src/model/cnn_lstm.py ================================================ import numpy as np import torch import torch.nn as nn import torch.optim as optim import torch.nn.functional as F from basic_model import BasicModel from fc_tree_net import FCTreeNet class conv_module(nn.Module): def __init__(self): super(conv_module, self).__init__() self.conv1 = nn.Conv2d(1, 16, kernel_size=3, stride=2) self.batch_norm1 = nn.BatchNorm2d(16) self.relu1 = nn.ReLU() self.conv2 = nn.Conv2d(16, 16, kernel_size=3, stride=2) self.batch_norm2 = nn.BatchNorm2d(16) self.relu2 = nn.ReLU() self.conv3 = nn.Conv2d(16, 16, kernel_size=3, stride=2) self.batch_norm3 = nn.BatchNorm2d(16) self.relu3 = nn.ReLU() self.conv4 = nn.Conv2d(16, 16, kernel_size=3, stride=2) self.batch_norm4 = nn.BatchNorm2d(16) self.relu4 = nn.ReLU() def forward(self, x): x = self.conv1(x) x = self.relu1(self.batch_norm1(x)) x = self.conv2(x) x = self.relu2(self.batch_norm2(x)) x = self.conv3(x) x = self.relu3(self.batch_norm3(x)) x = self.conv4(x) x = self.relu4(self.batch_norm4(x)) return x.view(-1, 16, 16*4*4) class lstm_module(nn.Module): def __init__(self): super(lstm_module, self).__init__() self.lstm = nn.LSTM(input_size=16*4*4, hidden_size=96, num_layers=1) self.dropout = nn.Dropout(0.5) self.fc = nn.Linear(96, 8) def forward(self, x): x = x.permute(1, 0, 2) hidden, _ = self.lstm(x) score = self.fc(hidden[-1, :, :]) return score class CNN_LSTM(BasicModel): def __init__(self, args): super(CNN_LSTM, self).__init__(args) self.conv = conv_module() self.lstm = lstm_module() self.fc_tree_net = FCTreeNet(in_dim=300, img_dim=256) self.optimizer = optim.Adam(self.parameters(), lr=args.lr, betas=(args.beta1, args.beta2), eps=args.epsilon) def compute_loss(self, output, target, meta_target, meta_structure): pred = output[0] loss = F.cross_entropy(pred, target) return loss def forward(self, x, embedding, indicator): alpha = 1.0 features = self.conv(x.view(-1, 1, 80, 80)) features_tree = self.fc_tree_net(features, embedding, indicator) features_tree = features_tree.view(-1, 16, 256) final_features = features + alpha * features_tree score = self.lstm(final_features) return score, None ================================================ FILE: src/model/cnn_mlp.py ================================================ import numpy as np import torch import torch.nn as nn import torch.optim as optim import torch.nn.functional as F from basic_model import BasicModel from fc_tree_net import FCTreeNet class conv_module(nn.Module): def __init__(self): super(conv_module, self).__init__() self.conv1 = nn.Conv2d(16, 32, kernel_size=3, stride=2) self.batch_norm1 = nn.BatchNorm2d(32) self.relu1 = nn.ReLU() self.conv2 = nn.Conv2d(32, 32, kernel_size=3, stride=2) self.batch_norm2 = nn.BatchNorm2d(32) self.relu2 = nn.ReLU() self.conv3 = nn.Conv2d(32, 32, kernel_size=3, stride=2) self.batch_norm3 = nn.BatchNorm2d(32) self.relu3 = nn.ReLU() self.conv4 = nn.Conv2d(32, 32, kernel_size=3, stride=2) self.batch_norm4 = nn.BatchNorm2d(32) self.relu4 = nn.ReLU() def forward(self, x): x = self.conv1(x) x = self.relu1(self.batch_norm1(x)) x = self.conv2(x) x = self.relu2(self.batch_norm2(x)) x = self.conv3(x) x = self.relu3(self.batch_norm3(x)) x = self.conv4(x) x = self.relu4(self.batch_norm4(x)) return x.view(-1, 32*4*4) class mlp_module(nn.Module): def __init__(self): super(mlp_module, self).__init__() self.fc1 = nn.Linear(32*4*4, 512) self.relu1 = nn.ReLU() self.fc2 = nn.Linear(512, 8) self.dropout = nn.Dropout(0.5) def forward(self, x): x = self.relu1(self.fc1(x)) x = self.dropout(x) x = self.fc2(x) return x class CNN_MLP(BasicModel): def __init__(self, args): super(CNN_MLP, self).__init__(args) self.conv = conv_module() self.mlp = mlp_module() self.fc_tree_net = FCTreeNet(in_dim=300, img_dim=512) self.optimizer = optim.Adam(self.parameters(), lr=args.lr, betas=(args.beta1, args.beta2), eps=args.epsilon) def compute_loss(self, output, target, meta_target, meta_structure): pred = output[0] loss = F.cross_entropy(pred, target) return loss def forward(self, x, embedding, indicator): alpha = 1.0 features = self.conv(x.view(-1, 16, 80, 80)) features_tree = features.view(-1, 1, 512) features_tree = self.fc_tree_net(features_tree, embedding, indicator) final_features = features + alpha * features_tree score = self.mlp(final_features) return score, None ================================================ FILE: src/model/const/__init__.py ================================================ from const import * ================================================ FILE: src/model/const/const.py ================================================ # -*- coding: utf-8 -*- # Maximum number of components in a RPM MAX_COMPONENTS = 2 # Canvas parameters IMAGE_SIZE = 160 CENTER = (IMAGE_SIZE / 2, IMAGE_SIZE / 2) DEFAULT_RADIUS = IMAGE_SIZE / 4 DEFAULT_WIDTH = 2 # Attribute parameters # Number NUM_VALUES = [1, 2, 3, 4, 5, 6, 7, 8, 9] NUM_MIN = 0 NUM_MAX = len(NUM_VALUES) - 1 # Uniformity UNI_VALUES = [False, False, False, True] UNI_MIN = 0 UNI_MAX = len(UNI_VALUES) - 1 # Type TYPE_VALUES = ["none", "triangle", "square", "pentagon", "hexagon", "circle"] TYPE_MIN = 0 TYPE_MAX = len(TYPE_VALUES) - 1 # Size SIZE_VALUES = [0.4, 0.5, 0.6, 0.7, 0.8, 0.9] SIZE_MIN = 0 SIZE_MAX = len(SIZE_VALUES) - 1 # Color COLOR_VALUES = [255, 224, 196, 168, 140, 112, 84, 56, 28, 0] COLOR_MIN = 0 COLOR_MAX = len(COLOR_VALUES) - 1 # Angle: self-rotation ANGLE_VALUES = [-135, -90, -45, 0, 45, 90, 135, 180] ANGLE_MIN = 0 ANGLE_MAX = len(ANGLE_VALUES) - 1 META_STRUCTURE_FORMAT = ["Singleton", "Left_Right", "Up_Down", "Out_In", "Left", "Right", \ "Up","Down", "Out", "In", "Grid", "Center_Single", "Distribute_Four", "Distribute_Nine", "Left_Center_Single", "Right_Center_Single", \ "Up_Center_Single", "Down_Center_Single", "Out_Center_Single", "In_Center_Single", "In_Distribute_Four"] # Rule, Attr, Param # The design encodes rule priority order: Number/Position always comes first # Number and Position could not both be sampled # Progression on Number: Number on each Panel +1/2 or -1/2 # Progression on Position: Entities on each Panel roll over the layout # Arithmetic on Number: Numeber on the third Panel = Number on first +/- Number on second (1 for + and -1 for -) # Arithmetic on Position: 1 for SET_UNION and -1 for SET_DIFF # Distribute_Three on Number: Three numbers through each row # Distribute_Three on Position: Three positions (same number) through each row # Constant on Number/Position: Nothing changes # Progression on Type: Type progression defined as the number of edges on each entity (Triangle, Square, Pentagon, Hexagon, Circle) # Distribute_Three on Type: Three types through each row # Constant on Type: Nothing changes # Progression on Size: Size on each entity +1/2 or -1/2 # Arithmetic on Size: Size on the third Panel = Size on the first +/- Size on the second (1 for + and -1 for -) # Distribute_Three on Size: Three sizes through each row # Constant on Size: Nothing changes # Progression on Color: Color +1/2 or -1/2 # Arithmetic on Color: Color on the third Panel = Color on the first +/- Color on the second (1 for + and -1 for -) # Distribute_Three on Color: Three colors through each row # Constant on Color: Nothing changes # Note that all rules on Type, Size and Color enforce value consistency in a panel RULE_ATTR = [[["Progression", "Number", [-2, -1, 1, 2]], ["Progression", "Position", [-2, -1, 1, 2]], ["Arithmetic", "Number", [1, -1]], ["Arithmetic", "Position", [1, -1]], ["Distribute_Three", "Number", None], ["Distribute_Three", "Position", None], ["Constant", "Number/Position", None]], [["Progression", "Type", [-2, -1, 1, 2]], ["Distribute_Three", "Type", None], ["Constant", "Type", None]], [["Progression", "Size", [-2, -1, 1, 2]], ["Arithmetic", "Size", [1, -1]], ["Distribute_Three", "Size", None], ["Constant", "Size", None]], [["Progression", "Color", [-2, -1, 1, 2]], ["Arithmetic", "Color", [1, -1]], ["Distribute_Three", "Color", None], ["Constant", "Color", None]]] ================================================ FILE: src/model/fc_tree_net.py ================================================ import torch import torch.nn as nn import torch.nn.functional as F import numpy as np class FCTreeNet(torch.nn.Module): def __init__(self, in_dim=300, img_dim=256, use_cuda=True): ''' initialization for TreeNet model, basically a ChildSumLSTM model with non-linear activation embedding for different nodes in the AoG. Shared weigths for all LSTM cells. :param in_dim: input feature dimension for word embedding (from string to vector space) :param img_dim: dimension of the input image feature, should be (panel_pair_number * img_feature_dim (e.g. 512 or 256)) ''' super(FCTreeNet, self).__init__() self.in_dim = in_dim self.img_dim = img_dim self.fc = nn.Linear(self.in_dim, self.in_dim) self.leaf = nn.Linear(self.in_dim + self.img_dim, self.img_dim) self.middle = nn.Linear(self.in_dim + self.img_dim, self.img_dim) self.merge = nn.Linear(self.in_dim + self.img_dim, self.img_dim) self.root = nn.Linear(self.in_dim + self.img_dim, self.img_dim) self.relu = nn.ReLU() def forward(self, image_feature, input, indicator): ''' Forward funciton for TreeNet model :param input: input should be (batch_size * 6 * input_word_embedding_dimension), got from the embedding vector :param indicator: indicating whether the input is of structure with branches (batch_size * 1) :param image_feature: input dictionary for each node, primarily feature, for example (batch_size * 16 (panel_pair_number) * feature_dim (output from CNN)) :return: ''' # image_feature = image_feature.view(-1, 16, image_feature.size(2)) input = self.fc(input.view(-1, input.size(-1))) input = input.view(-1, 6, input.size(-1)) input = input.unsqueeze(1).repeat(1, image_feature.size(1), 1, 1) indicator = indicator.unsqueeze(1).repeat(1, image_feature.size(1), 1).view(-1, 1) leaf_left = input[:, :, 3, :].view(-1, input.size(-1)) # (batch_size * panel_pair_num) * input_word_embedding_dimension leaf_right = input[:, :, 5, :].view(-1, input.size(-1)) inter_left = input[:, :, 2, :].view(-1, input.size(-1)) inter_right = input[:, :, 4, :].view(-1, input.size(-1)) merge = input[:, :, 1, :].view(-1, input.size(-1)) root = input[:, :, 0, :].view(-1, input.size(-1)) # concating image_feature and word_embeddings for leaf node inputs leaf_left = torch.cat((leaf_left, image_feature.view(-1, image_feature.size(-1))), dim=-1) leaf_right = torch.cat((leaf_right, image_feature.view(-1, image_feature.size(-1))), dim=-1) out_leaf_left = self.leaf(leaf_left) out_leaf_right = self.leaf(leaf_right) out_leaf_left = self.relu(out_leaf_left) out_leaf_right = self.relu(out_leaf_right) out_left = self.middle(torch.cat((inter_left, out_leaf_left), dim=-1)) out_right = self.middle(torch.cat((inter_right, out_leaf_right), dim=-1)) out_left = self.relu(out_left) out_right = self.relu(out_right) out_right = torch.mul(out_right, indicator) merge_input = torch.cat((merge, out_left + out_right), dim=-1) out_merge = self.merge(merge_input) out_merge = self.relu(out_merge) out_root = self.root(torch.cat((root, out_merge), dim=-1)) out_root = self.relu(out_root) # size ((batch_size * panel_pair) * feature_dim) return out_root ================================================ FILE: src/model/main.py ================================================ import os import numpy as np import argparse import torch import torch.nn as nn import torch.optim as optim from torch.utils.data import Dataset, DataLoader from torchvision import transforms, utils from utility import dataset, ToTensor from cnn_mlp import CNN_MLP from cnn_lstm import CNN_LSTM from resnet18 import Resnet18_MLP parser = argparse.ArgumentParser(description='our_model') parser.add_argument('--model', type=str, default='Resnet18_MLP') parser.add_argument('--epochs', type=int, default=200) parser.add_argument('--batch_size', type=int, default=32) parser.add_argument('--seed', type=int, default=12345) parser.add_argument('--device', type=int, default=0) parser.add_argument('--load_workers', type=int, default=16) parser.add_argument('--resume', type=bool, default=False) parser.add_argument('--path', type=str, default='/home/chizhang/Datasets/RAVEN-10000/') parser.add_argument('--save', type=str, default='./experiments/checkpoint/') parser.add_argument('--img_size', type=int, default=224) parser.add_argument('--lr', type=float, default=1e-4) parser.add_argument('--beta1', type=float, default=0.9) parser.add_argument('--beta2', type=float, default=0.999) parser.add_argument('--epsilon', type=float, default=1e-8) parser.add_argument('--meta_alpha', type=float, default=0.0) parser.add_argument('--meta_beta', type=float, default=0.0) args = parser.parse_args() args.cuda = torch.cuda.is_available() torch.cuda.set_device(args.device) if args.cuda: torch.cuda.manual_seed(args.seed) if not os.path.exists(args.save): os.makedirs(args.save) train = dataset(args.path, "train", args.img_size, transform=transforms.Compose([ToTensor()])) valid = dataset(args.path, "val", args.img_size, transform=transforms.Compose([ToTensor()])) test = dataset(args.path, "test", args.img_size, transform=transforms.Compose([ToTensor()])) trainloader = DataLoader(train, batch_size=args.batch_size, shuffle=True, num_workers=16) validloader = DataLoader(valid, batch_size=args.batch_size, shuffle=False, num_workers=16) testloader = DataLoader(test, batch_size=args.batch_size, shuffle=False, num_workers=16) if args.model == "CNN_MLP": model = CNN_MLP(args) elif args.model == "CNN_LSTM": model = CNN_LSTM(args) elif args.model == "Resnet18_MLP": model = Resnet18_MLP(args) if args.resume: model.load_model(args.save, 0) print('Loaded model') if args.cuda: model = model.cuda() def train(epoch): model.train() train_loss = 0 accuracy = 0 loss_all = 0.0 acc_all = 0.0 counter = 0 for batch_idx, (image, target, meta_target, meta_structure, embedding, indicator) in enumerate(trainloader): counter += 1 if args.cuda: image = image.cuda() target = target.cuda() meta_target = meta_target.cuda() meta_structure = meta_structure.cuda() embedding = embedding.cuda() indicator = indicator.cuda() loss, acc = model.train_(image, target, meta_target, meta_structure, embedding, indicator) print('Train: Epoch:{}, Batch:{}, Loss:{:.6f}, Acc:{:.4f}.'.format(epoch, batch_idx, loss, acc)) loss_all += loss acc_all += acc if counter > 0: print("Avg Training Loss: {:.6f}".format(loss_all/float(counter))) def validate(epoch): model.eval() val_loss = 0 accuracy = 0 loss_all = 0.0 acc_all = 0.0 counter = 0 for batch_idx, (image, target, meta_target, meta_structure, embedding, indicator) in enumerate(validloader): counter += 1 if args.cuda: image = image.cuda() target = target.cuda() meta_target = meta_target.cuda() meta_structure = meta_structure.cuda() embedding = embedding.cuda() indicator = indicator.cuda() loss, acc = model.validate_(image, target, meta_target, meta_structure, embedding, indicator) # print('Validate: Epoch:{}, Batch:{}, Loss:{:.6f}, Acc:{:.4f}.'.format(epoch, batch_idx, loss, acc)) loss_all += loss acc_all += acc if counter > 0: print("Total Validation Loss: {:.6f}, Acc: {:.4f}".format(loss_all/float(counter), acc_all/float(counter))) return loss_all/float(counter), acc_all/float(counter) def test(epoch): model.eval() accuracy = 0 acc_all = 0.0 counter = 0 for batch_idx, (image, target, meta_target, meta_structure, embedding, indicator) in enumerate(testloader): counter += 1 if args.cuda: image = image.cuda() target = target.cuda() meta_target = meta_target.cuda() meta_structure = meta_structure.cuda() embedding = embedding.cuda() indicator = indicator.cuda() acc = model.test_(image, target, meta_target, meta_structure, embedding, indicator) # print('Test: Epoch:{}, Batch:{}, Acc:{:.4f}.'.format(epoch, batch_idx, acc)) acc_all += acc if counter > 0: print("Total Testing Acc: {:.4f}".format(acc_all / float(counter))) return acc_all/float(counter) def main(): for epoch in range(0, args.epochs): train(epoch) avg_loss, avg_acc = validate(epoch) test(epoch) model.save_model(args.save, epoch, avg_acc, avg_loss) if __name__ == '__main__': main() ================================================ FILE: src/model/resnet18.py ================================================ import numpy as np import torch import torch.nn as nn import torch.optim as optim import torch.nn.functional as F import torchvision.models as models from basic_model import BasicModel from fc_tree_net import FCTreeNet class identity(nn.Module): def __init__(self): super(identity, self).__init__() def forward(self, x): return x class mlp_module(nn.Module): def __init__(self): super(mlp_module, self).__init__() self.fc1 = nn.Linear(512, 512) self.relu1 = nn.ReLU() self.fc2 = nn.Linear(512, 8+9+21) self.dropout = nn.Dropout(0.5) def forward(self, x): x = self.relu1(self.fc1(x)) x = self.dropout(x) x = self.fc2(x) return x class Resnet18_MLP(BasicModel): def __init__(self, args): super(Resnet18_MLP, self).__init__(args) self.resnet18 = models.resnet18(pretrained=False) self.resnet18.conv1 = nn.Conv2d(16, 64, kernel_size=7, stride=2, padding=3, bias=False) self.resnet18.fc = identity() self.mlp = mlp_module() self.fc_tree_net = FCTreeNet(in_dim=300, img_dim=512) self.optimizer = optim.Adam(self.parameters(), lr=args.lr, betas=(args.beta1, args.beta2), eps=args.epsilon) self.meta_alpha = args.meta_alpha self.meta_beta = args.meta_beta def compute_loss(self, output, target, meta_target, meta_structure): pred, meta_target_pred, meta_struct_pred = output[0], output[1], output[2] target_loss = F.cross_entropy(pred, target) meta_target_pred = torch.chunk(meta_target_pred, chunks=9, dim=1) meta_target = torch.chunk(meta_target, chunks=9, dim=1) meta_target_loss = 0. for idx in range(0, 9): meta_target_loss += F.binary_cross_entropy(F.sigmoid(meta_target_pred[idx]), meta_target[idx]) meta_struct_pred = torch.chunk(meta_struct_pred, chunks=21, dim=1) meta_structure = torch.chunk(meta_structure, chunks=21, dim=1) meta_struct_loss = 0. for idx in range(0, 21): meta_struct_loss += F.binary_cross_entropy(F.sigmoid(meta_struct_pred[idx]), meta_structure[idx]) loss = target_loss + self.meta_alpha*meta_struct_loss/21. + self.meta_beta*meta_target_loss/9. return loss def forward(self, x, embedding, indicator): alpha = 1.0 features = self.resnet18(x.view(-1, 16, 224, 224)) features_tree = features.view(-1, 1, 512) features_tree = self.fc_tree_net(features_tree, embedding, indicator) final_features = features + alpha * features_tree output = self.mlp(final_features) pred = output[:,0:8] meta_target_pred = output[:,8:17] meta_struct_pred = output[:,17:38] return pred, meta_target_pred, meta_struct_pred ================================================ FILE: src/model/utility/__init__.py ================================================ from dataset_utility import * ================================================ FILE: src/model/utility/dataset_utility.py ================================================ import os import glob import numpy as np from scipy import misc import torch from torch.utils.data import Dataset, DataLoader from torchvision import transforms, utils class ToTensor(object): def __call__(self, sample): return torch.tensor(sample, dtype=torch.float32) class dataset(Dataset): def __init__(self, root_dir, dataset_type, img_size, transform=None, shuffle=False): self.root_dir = root_dir self.transform = transform self.file_names = [f for f in glob.glob(os.path.join(root_dir, "*", "*.npz")) \ if dataset_type in f] self.img_size = img_size self.embeddings = np.load(os.path.join(root_dir, 'embedding.npy'), allow_pickle=True) self.shuffle = shuffle def __len__(self): return len(self.file_names) def __getitem__(self, idx): data_path = self.file_names[idx] data = np.load(data_path) image = data["image"].reshape(16, 160, 160) target = data["target"] structure = data["structure"] meta_target = data["meta_target"] meta_structure = data["meta_structure"] if self.shuffle: context = image[:8, :, :] choices = image[8:, :, :] indices = range(8) np.random.shuffle(indices) new_target = indices.index(target) new_choices = choices[indices, :, :] image = np.concatenate((context, new_choices)) target = new_target resize_image = [] for idx in range(0, 16): resize_image.append(misc.imresize(image[idx,:,:], (self.img_size, self.img_size))) resize_image = np.stack(resize_image) # image = resize(image, (16, 128, 128)) # meta_matrix = data["mata_matrix"] embedding = torch.zeros((6, 300), dtype=torch.float) indicator = torch.zeros(1, dtype=torch.float) element_idx = 0 for element in structure: if element != '/': embedding[element_idx, :] = torch.tensor(self.embeddings.item().get(element), dtype=torch.float) element_idx += 1 if element_idx == 6: indicator[0] = 1. # if meta_target.dtype == np.int8: # meta_target = meta_target.astype(np.uint8) # if meta_structure.dtype == np.int8: # meta_structure = meta_structure.astype(np.uint8) del data if self.transform: resize_image = self.transform(resize_image) # meta_matrix = self.transform(meta_matrix) target = torch.tensor(target, dtype=torch.long) meta_target = self.transform(meta_target) meta_structure = self.transform(meta_structure) # meta_target = torch.tensor(meta_target, dtype=torch.long) return resize_image, target, meta_target, meta_structure, embedding, indicator