[ { "path": ".gitignore", "content": "data/\nmdr.egg*/\napex/\nmodels/\nlogs/\n.DS_Store\n*.pyc\n*.swp\n" }, { "path": "CODE_OF_CONDUCT.md", "content": "# Code of Conduct\n\n## Our Pledge\n\nIn the interest of fostering an open and welcoming environment, we as\ncontributors and maintainers pledge to make participation in our project and\nour community a harassment-free experience for everyone, regardless of age, body\nsize, disability, ethnicity, sex characteristics, gender identity and expression,\nlevel of experience, education, socio-economic status, nationality, personal\nappearance, race, religion, or sexual identity and orientation.\n\n## Our Standards\n\nExamples of behavior that contributes to creating a positive environment\ninclude:\n\n* Using welcoming and inclusive language\n* Being respectful of differing viewpoints and experiences\n* Gracefully accepting constructive criticism\n* Focusing on what is best for the community\n* Showing empathy towards other community members\n\nExamples of unacceptable behavior by participants include:\n\n* The use of sexualized language or imagery and unwelcome sexual attention or\nadvances\n* Trolling, insulting/derogatory comments, and personal or political attacks\n* Public or private harassment\n* Publishing others' private information, such as a physical or electronic\naddress, without explicit permission\n* Other conduct which could reasonably be considered inappropriate in a\nprofessional setting\n\n## Our Responsibilities\n\nProject maintainers are responsible for clarifying the standards of acceptable\nbehavior and are expected to take appropriate and fair corrective action in\nresponse to any instances of unacceptable behavior.\n\nProject maintainers have the right and responsibility to remove, edit, or\nreject comments, commits, code, wiki edits, issues, and other contributions\nthat are not aligned to this Code of Conduct, or to ban temporarily or\npermanently any contributor for other behaviors that they deem inappropriate,\nthreatening, offensive, or harmful.\n\n## Scope\n\nThis Code of Conduct applies within all project spaces, and it also applies when\nan individual is representing the project or its community in public spaces.\nExamples of representing a project or community include using an official\nproject e-mail address, posting via an official social media account, or acting\nas an appointed representative at an online or offline event. Representation of\na project may be further defined and clarified by project maintainers.\n\n## Enforcement\n\nInstances of abusive, harassing, or otherwise unacceptable behavior may be\nreported by contacting the project team at . All\ncomplaints will be reviewed and investigated and will result in a response that\nis deemed necessary and appropriate to the circumstances. The project team is\nobligated to maintain confidentiality with regard to the reporter of an incident.\nFurther details of specific enforcement policies may be posted separately.\n\nProject maintainers who do not follow or enforce the Code of Conduct in good\nfaith may face temporary or permanent repercussions as determined by other\nmembers of the project's leadership.\n\n## Attribution\n\nThis Code of Conduct is adapted from the [Contributor Covenant][homepage], version 1.4,\navailable at https://www.contributor-covenant.org/version/1/4/code-of-conduct.html\n\n[homepage]: https://www.contributor-covenant.org\n\nFor answers to common questions about this code of conduct, see\nhttps://www.contributor-covenant.org/faq\n" }, { "path": "CONTRIBUTING.md", "content": "# Contributing to multihop_dense_retrieval\nWe want to make contributing to this project as easy and transparent as\npossible.\n\n## Pull Requests\nWe actively welcome your pull requests.\n\n1. Fork the repo and create your branch from `master`.\n2. If you've added code that should be tested, add tests.\n3. If you've changed APIs, update the documentation.\n4. Ensure the test suite passes.\n5. Make sure your code lints.\n6. If you haven't already, complete the Contributor License Agreement (\"CLA\").\n\n## Contributor License Agreement (\"CLA\")\nIn order to accept your pull request, we need you to submit a CLA. You only need\nto do this once to work on any of Facebook's open source projects.\n\nComplete your CLA here: \n\n## Issues\nWe use GitHub issues to track public bugs. Please ensure your description is\nclear and has sufficient instructions to be able to reproduce the issue.\n\nFacebook has a [bounty program](https://www.facebook.com/whitehat/) for the safe\ndisclosure of security bugs. In those cases, please go through the process\noutlined on that page and do not file a public issue.\n\n## License\nBy contributing to multihop_dense_retrieval, you agree that your contributions will be licensed\nunder the LICENSE file in the root directory of this source tree." }, { "path": "LICENSE", "content": "Attribution-NonCommercial 4.0 International\n\n=======================================================================\n\nCreative Commons Corporation (\"Creative Commons\") is not a law firm and\ndoes not provide legal services or legal advice. Distribution of\nCreative Commons public licenses does not create a lawyer-client or\nother relationship. Creative Commons makes its licenses and related\ninformation available on an \"as-is\" basis. Creative Commons gives no\nwarranties regarding its licenses, any material licensed under their\nterms and conditions, or any related information. Creative Commons\ndisclaims all liability for damages resulting from their use to the\nfullest extent possible.\n\nUsing Creative Commons Public Licenses\n\nCreative Commons public licenses provide a standard set of terms and\nconditions that creators and other rights holders may use to share\noriginal works of authorship and other material subject to copyright\nand certain other rights specified in the public license below. The\nfollowing considerations are for informational purposes only, are not\nexhaustive, and do not form part of our licenses.\n\n Considerations for licensors: Our public licenses are\n intended for use by those authorized to give the public\n permission to use material in ways otherwise restricted by\n copyright and certain other rights. Our licenses are\n irrevocable. Licensors should read and understand the terms\n and conditions of the license they choose before applying it.\n Licensors should also secure all rights necessary before\n applying our licenses so that the public can reuse the\n material as expected. Licensors should clearly mark any\n material not subject to the license. This includes other CC-\n licensed material, or material used under an exception or\n limitation to copyright. More considerations for licensors:\n\twiki.creativecommons.org/Considerations_for_licensors\n\n Considerations for the public: By using one of our public\n licenses, a licensor grants the public permission to use the\n licensed material under specified terms and conditions. If\n the licensor's permission is not necessary for any reason--for\n example, because of any applicable exception or limitation to\n copyright--then that use is not regulated by the license. Our\n licenses grant only permissions under copyright and certain\n other rights that a licensor has authority to grant. Use of\n the licensed material may still be restricted for other\n reasons, including because others have copyright or other\n rights in the material. A licensor may make special requests,\n such as asking that all changes be marked or described.\n Although not required by our licenses, you are encouraged to\n respect those requests where reasonable. More_considerations\n for the public:\n\twiki.creativecommons.org/Considerations_for_licensees\n\n=======================================================================\n\nCreative Commons Attribution-NonCommercial 4.0 International Public\nLicense\n\nBy exercising the Licensed Rights (defined below), You accept and agree\nto be bound by the terms and conditions of this Creative Commons\nAttribution-NonCommercial 4.0 International Public License (\"Public\nLicense\"). To the extent this Public License may be interpreted as a\ncontract, You are granted the Licensed Rights in consideration of Your\nacceptance of these terms and conditions, and the Licensor grants You\nsuch rights in consideration of benefits the Licensor receives from\nmaking the Licensed Material available under these terms and\nconditions.\n\nSection 1 -- Definitions.\n\n a. Adapted Material means material subject to Copyright and Similar\n Rights that is derived from or based upon the Licensed Material\n and in which the Licensed Material is translated, altered,\n arranged, transformed, or otherwise modified in a manner requiring\n permission under the Copyright and Similar Rights held by the\n Licensor. For purposes of this Public License, where the Licensed\n Material is a musical work, performance, or sound recording,\n Adapted Material is always produced where the Licensed Material is\n synched in timed relation with a moving image.\n\n b. Adapter's License means the license You apply to Your Copyright\n and Similar Rights in Your contributions to Adapted Material in\n accordance with the terms and conditions of this Public License.\n\n c. Copyright and Similar Rights means copyright and/or similar rights\n closely related to copyright including, without limitation,\n performance, broadcast, sound recording, and Sui Generis Database\n Rights, without regard to how the rights are labeled or\n categorized. For purposes of this Public License, the rights\n specified in Section 2(b)(1)-(2) are not Copyright and Similar\n Rights.\n d. Effective Technological Measures means those measures that, in the\n absence of proper authority, may not be circumvented under laws\n fulfilling obligations under Article 11 of the WIPO Copyright\n Treaty adopted on December 20, 1996, and/or similar international\n agreements.\n\n e. Exceptions and Limitations means fair use, fair dealing, and/or\n any other exception or limitation to Copyright and Similar Rights\n that applies to Your use of the Licensed Material.\n\n f. Licensed Material means the artistic or literary work, database,\n or other material to which the Licensor applied this Public\n License.\n\n g. Licensed Rights means the rights granted to You subject to the\n terms and conditions of this Public License, which are limited to\n all Copyright and Similar Rights that apply to Your use of the\n Licensed Material and that the Licensor has authority to license.\n\n h. Licensor means the individual(s) or entity(ies) granting rights\n under this Public License.\n\n i. NonCommercial means not primarily intended for or directed towards\n commercial advantage or monetary compensation. For purposes of\n this Public License, the exchange of the Licensed Material for\n other material subject to Copyright and Similar Rights by digital\n file-sharing or similar means is NonCommercial provided there is\n no payment of monetary compensation in connection with the\n exchange.\n\n j. Share means to provide material to the public by any means or\n process that requires permission under the Licensed Rights, such\n as reproduction, public display, public performance, distribution,\n dissemination, communication, or importation, and to make material\n available to the public including in ways that members of the\n public may access the material from a place and at a time\n individually chosen by them.\n\n k. Sui Generis Database Rights means rights other than copyright\n resulting from Directive 96/9/EC of the European Parliament and of\n the Council of 11 March 1996 on the legal protection of databases,\n as amended and/or succeeded, as well as other essentially\n equivalent rights anywhere in the world.\n\n l. You means the individual or entity exercising the Licensed Rights\n under this Public License. Your has a corresponding meaning.\n\nSection 2 -- Scope.\n\n a. License grant.\n\n 1. Subject to the terms and conditions of this Public License,\n the Licensor hereby grants You a worldwide, royalty-free,\n non-sublicensable, non-exclusive, irrevocable license to\n exercise the Licensed Rights in the Licensed Material to:\n\n a. reproduce and Share the Licensed Material, in whole or\n in part, for NonCommercial purposes only; and\n\n b. produce, reproduce, and Share Adapted Material for\n NonCommercial purposes only.\n\n 2. Exceptions and Limitations. For the avoidance of doubt, where\n Exceptions and Limitations apply to Your use, this Public\n License does not apply, and You do not need to comply with\n its terms and conditions.\n\n 3. Term. The term of this Public License is specified in Section\n 6(a).\n\n 4. Media and formats; technical modifications allowed. The\n Licensor authorizes You to exercise the Licensed Rights in\n all media and formats whether now known or hereafter created,\n and to make technical modifications necessary to do so. The\n Licensor waives and/or agrees not to assert any right or\n authority to forbid You from making technical modifications\n necessary to exercise the Licensed Rights, including\n technical modifications necessary to circumvent Effective\n Technological Measures. For purposes of this Public License,\n simply making modifications authorized by this Section 2(a)\n (4) never produces Adapted Material.\n\n 5. Downstream recipients.\n\n a. Offer from the Licensor -- Licensed Material. Every\n recipient of the Licensed Material automatically\n receives an offer from the Licensor to exercise the\n Licensed Rights under the terms and conditions of this\n Public License.\n\n b. No downstream restrictions. You may not offer or impose\n any additional or different terms or conditions on, or\n apply any Effective Technological Measures to, the\n Licensed Material if doing so restricts exercise of the\n Licensed Rights by any recipient of the Licensed\n Material.\n\n 6. No endorsement. Nothing in this Public License constitutes or\n may be construed as permission to assert or imply that You\n are, or that Your use of the Licensed Material is, connected\n with, or sponsored, endorsed, or granted official status by,\n the Licensor or others designated to receive attribution as\n provided in Section 3(a)(1)(A)(i).\n\n b. Other rights.\n\n 1. Moral rights, such as the right of integrity, are not\n licensed under this Public License, nor are publicity,\n privacy, and/or other similar personality rights; however, to\n the extent possible, the Licensor waives and/or agrees not to\n assert any such rights held by the Licensor to the limited\n extent necessary to allow You to exercise the Licensed\n Rights, but not otherwise.\n\n 2. Patent and trademark rights are not licensed under this\n Public License.\n\n 3. To the extent possible, the Licensor waives any right to\n collect royalties from You for the exercise of the Licensed\n Rights, whether directly or through a collecting society\n under any voluntary or waivable statutory or compulsory\n licensing scheme. In all other cases the Licensor expressly\n reserves any right to collect such royalties, including when\n the Licensed Material is used other than for NonCommercial\n purposes.\n\nSection 3 -- License Conditions.\n\nYour exercise of the Licensed Rights is expressly made subject to the\nfollowing conditions.\n\n a. Attribution.\n\n 1. If You Share the Licensed Material (including in modified\n form), You must:\n\n a. retain the following if it is supplied by the Licensor\n with the Licensed Material:\n\n i. identification of the creator(s) of the Licensed\n Material and any others designated to receive\n attribution, in any reasonable manner requested by\n the Licensor (including by pseudonym if\n designated);\n\n ii. a copyright notice;\n\n iii. a notice that refers to this Public License;\n\n iv. a notice that refers to the disclaimer of\n warranties;\n\n v. a URI or hyperlink to the Licensed Material to the\n extent reasonably practicable;\n\n b. indicate if You modified the Licensed Material and\n retain an indication of any previous modifications; and\n\n c. indicate the Licensed Material is licensed under this\n Public License, and include the text of, or the URI or\n hyperlink to, this Public License.\n\n 2. You may satisfy the conditions in Section 3(a)(1) in any\n reasonable manner based on the medium, means, and context in\n which You Share the Licensed Material. For example, it may be\n reasonable to satisfy the conditions by providing a URI or\n hyperlink to a resource that includes the required\n information.\n\n 3. If requested by the Licensor, You must remove any of the\n information required by Section 3(a)(1)(A) to the extent\n reasonably practicable.\n\n 4. If You Share Adapted Material You produce, the Adapter's\n License You apply must not prevent recipients of the Adapted\n Material from complying with this Public License.\n\nSection 4 -- Sui Generis Database Rights.\n\nWhere the Licensed Rights include Sui Generis Database Rights that\napply to Your use of the Licensed Material:\n\n a. for the avoidance of doubt, Section 2(a)(1) grants You the right\n to extract, reuse, reproduce, and Share all or a substantial\n portion of the contents of the database for NonCommercial purposes\n only;\n\n b. if You include all or a substantial portion of the database\n contents in a database in which You have Sui Generis Database\n Rights, then the database in which You have Sui Generis Database\n Rights (but not its individual contents) is Adapted Material; and\n\n c. You must comply with the conditions in Section 3(a) if You Share\n all or a substantial portion of the contents of the database.\n\nFor the avoidance of doubt, this Section 4 supplements and does not\nreplace Your obligations under this Public License where the Licensed\nRights include other Copyright and Similar Rights.\n\nSection 5 -- Disclaimer of Warranties and Limitation of Liability.\n\n a. UNLESS OTHERWISE SEPARATELY UNDERTAKEN BY THE LICENSOR, TO THE\n EXTENT POSSIBLE, THE LICENSOR OFFERS THE LICENSED MATERIAL AS-IS\n AND AS-AVAILABLE, AND MAKES NO REPRESENTATIONS OR WARRANTIES OF\n ANY KIND CONCERNING THE LICENSED MATERIAL, WHETHER EXPRESS,\n IMPLIED, STATUTORY, OR OTHER. THIS INCLUDES, WITHOUT LIMITATION,\n WARRANTIES OF TITLE, MERCHANTABILITY, FITNESS FOR A PARTICULAR\n PURPOSE, NON-INFRINGEMENT, ABSENCE OF LATENT OR OTHER DEFECTS,\n ACCURACY, OR THE PRESENCE OR ABSENCE OF ERRORS, WHETHER OR NOT\n KNOWN OR DISCOVERABLE. WHERE DISCLAIMERS OF WARRANTIES ARE NOT\n ALLOWED IN FULL OR IN PART, THIS DISCLAIMER MAY NOT APPLY TO YOU.\n\n b. TO THE EXTENT POSSIBLE, IN NO EVENT WILL THE LICENSOR BE LIABLE\n TO YOU ON ANY LEGAL THEORY (INCLUDING, WITHOUT LIMITATION,\n NEGLIGENCE) OR OTHERWISE FOR ANY DIRECT, SPECIAL, INDIRECT,\n INCIDENTAL, CONSEQUENTIAL, PUNITIVE, EXEMPLARY, OR OTHER LOSSES,\n COSTS, EXPENSES, OR DAMAGES ARISING OUT OF THIS PUBLIC LICENSE OR\n USE OF THE LICENSED MATERIAL, EVEN IF THE LICENSOR HAS BEEN\n ADVISED OF THE POSSIBILITY OF SUCH LOSSES, COSTS, EXPENSES, OR\n DAMAGES. WHERE A LIMITATION OF LIABILITY IS NOT ALLOWED IN FULL OR\n IN PART, THIS LIMITATION MAY NOT APPLY TO YOU.\n\n c. The disclaimer of warranties and limitation of liability provided\n above shall be interpreted in a manner that, to the extent\n possible, most closely approximates an absolute disclaimer and\n waiver of all liability.\n\nSection 6 -- Term and Termination.\n\n a. This Public License applies for the term of the Copyright and\n Similar Rights licensed here. However, if You fail to comply with\n this Public License, then Your rights under this Public License\n terminate automatically.\n\n b. Where Your right to use the Licensed Material has terminated under\n Section 6(a), it reinstates:\n\n 1. automatically as of the date the violation is cured, provided\n it is cured within 30 days of Your discovery of the\n violation; or\n\n 2. upon express reinstatement by the Licensor.\n\n For the avoidance of doubt, this Section 6(b) does not affect any\n right the Licensor may have to seek remedies for Your violations\n of this Public License.\n\n c. For the avoidance of doubt, the Licensor may also offer the\n Licensed Material under separate terms or conditions or stop\n distributing the Licensed Material at any time; however, doing so\n will not terminate this Public License.\n\n d. Sections 1, 5, 6, 7, and 8 survive termination of this Public\n License.\n\nSection 7 -- Other Terms and Conditions.\n\n a. The Licensor shall not be bound by any additional or different\n terms or conditions communicated by You unless expressly agreed.\n\n b. Any arrangements, understandings, or agreements regarding the\n Licensed Material not stated herein are separate from and\n independent of the terms and conditions of this Public License.\n\nSection 8 -- Interpretation.\n\n a. For the avoidance of doubt, this Public License does not, and\n shall not be interpreted to, reduce, limit, restrict, or impose\n conditions on any use of the Licensed Material that could lawfully\n be made without permission under this Public License.\n\n b. To the extent possible, if any provision of this Public License is\n deemed unenforceable, it shall be automatically reformed to the\n minimum extent necessary to make it enforceable. If the provision\n cannot be reformed, it shall be severed from this Public License\n without affecting the enforceability of the remaining terms and\n conditions.\n\n c. No term or condition of this Public License will be waived and no\n failure to comply consented to unless expressly agreed to by the\n Licensor.\n\n d. Nothing in this Public License constitutes or may be interpreted\n as a limitation upon, or waiver of, any privileges and immunities\n that apply to the Licensor or You, including from the legal\n processes of any jurisdiction or authority.\n\n=======================================================================\n\nCreative Commons is not a party to its public\nlicenses. Notwithstanding, Creative Commons may elect to apply one of\nits public licenses to material it publishes and in those instances\nwill be considered the “Licensor.” The text of the Creative Commons\npublic licenses is dedicated to the public domain under the CC0 Public\nDomain Dedication. Except for the limited purpose of indicating that\nmaterial is shared under a Creative Commons public license or as\notherwise permitted by the Creative Commons policies published at\ncreativecommons.org/policies, Creative Commons does not authorize the\nuse of the trademark \"Creative Commons\" or any other trademark or logo\nof Creative Commons without its prior written consent including,\nwithout limitation, in connection with any unauthorized modifications\nto any of its public licenses or any other arrangements,\nunderstandings, or agreements concerning use of licensed material. For\nthe avoidance of doubt, this paragraph does not form part of the\npublic licenses.\n\nCreative Commons may be contacted at creativecommons.org." }, { "path": "README.md", "content": "\n\n\n\n# [

Multi-Hop Dense Text Retrieval (`MDR`)

](#p-aligncentermulti-hop-dense-text-retrieval-mdrp)\n\n**\\*\\*\\*\\*\\* Update 3/4/2021: Adding simple demo code based on [streamlit](https://streamlit.io/) \\*\\*\\*\\*\\***\n\n`MDR` is a simple and generalized dense retrieval method which recursively retrieves supporting text passages for answering complex open-domain questions. The repo provides code and pretrained retrieval models that produce **state-of-the-art** retrieval performance on two multi-hop QA datasets (the [HotpotQA](https://hotpotqa.github.io) dataset and the multi-hop subset of the [FEVER fact extraction and verification dataset](https://fever.ai)). \n\nMore details about our approach are described in our ICLR paper [Answering Complex Open-Domain Questions with Multi-Hop Dense Retrieval](https://arxiv.org/abs/2009.12756)\n\n

\n\n- [Use the trained models](#use-the-trained-models)\n - [Evaluating retrieval](#evaluating-retrieval)\n - [Evaluating QA](#evaluating-qa)\n - [Demo](#end-to-end-demo)\n- [Train models from scratch](#train-models-from-scratch)\n - [Retriever training](#retriever-training)\n - [Encoding the corpus for retrieval](#encoding-the-corpus-for-retrieval)\n - [ELECTRA QA model training](#electra-qa-model-training)\n\n\n## Use the trained models\n\n1. Set up the environment\n```bash\nconda create --name MDR python=3.6\nconda activate MDR\ngit clone git@github.com:facebookresearch/multihop_dense_retrieval.git\ncd multihop_dense_retrieval \nbash setup.sh\n```\n\n2. Download the necessary data files and pretrained retrieval models\nSimplified data files with **quesitons** and ground-truth **supporting passages**:\n\n```\n# save pretrained models to models/ and all processed hotpotQA into data/ \n# models will take about 2GB, and data will take 20GB since the pre-trained wikipedia index are included.\nbash ./scripts/download_hotpot.sh\n```\n\n\n### Evaluating retrieval\nEvalauting direct retrieval performance (The printed statistics might not adhere to the metric names defined in the paper. \n* **PR**: whether one of the supporting passages is included in all retrieved passages; \n* **P-EM**: whether **both** supporting passages are included in all retrieval passages; \n* **Path Recall**: whether any of the topk retrieved chain extract match the ground-truth supporting passages.) and saving topk retrieved passage chains for downstream QA. \n\nHere's an example evaluating the top1 ranked passage chains:\n\n```\npython scripts/eval/eval_mhop_retrieval.py \\\n data/hotpot/hotpot_qas_val.json \\\n data/hotpot_index/wiki_index.npy \\\n data/hotpot_index/wiki_id2doc.json \\\n models/q_encoder.pt \\\n --batch-size 100 \\\n --beam-size 1 \\\n --topk 1 \\\n --shared-encoder \\\n --model-name roberta-base \\\n --gpu \\\n --save-path ${SAVE_RETRIEVAL_FOR_QA}\n```\nSevaral important options includes \n* `--beam-size-n`: beam size at each hop; \n* `--topk`: topk passage chains from beam search \n* `--gpu`: move the dense index to GPU, resulting in much faster search\n\n\nExpected results (Top1):\n\n```\nEvaluating 7405 samples...\n\tAvg PR: 0.8428089128966915\n\tAvg P-EM: 0.6592842673869007\n\tAvg 1-Recall: 0.7906819716407832\n\tPath Recall: 0.6592842673869007\ncomparison Questions num: 1487\n\tAvg PR: 0.9932750504371217\n\tAvg P-EM: 0.9482178883658372\n\tAvg 1-Recall: 0.9643577673167452\n\tPath Recall: 0.9482178883658372\nbridge Questions num: 5918\n\tAvg PR: 0.805001689760054\n\tAvg P-EM: 0.5866846907739101\n\tAvg 1-Recall: 0.7470429199053734\n\tPath Recall: 0.5866846907739101\n```\n\n**Note:** For more efficient retrieval on CPU, check out the `--hnsw` option in `scripts/eval/eval_mhop_retrieval.py`. \n\n### Evaluating QA\nThe best answer extraction model is based on the pretrained [ELECTRA](https://arxiv.org/abs/2003.10555), outperforming the **BERT-large-whole-word-masking** by ~2 points answer EM/F1. We construct the training data with the pretrained MDR retriever and always include the ground-truth passage chain if the MDR failed. Each training question is paired with the groundtruth SP passage chain and also 5 (hyperparameter) retrieved chains which do not match the groundtruth. \n\nAs the HotpotQA task requires evaluating the prediction of supporting sentences, we do sentence segmetation on the MDR retrieval result before feeding into the answer extraction models. Follow the script [scripts/add_sp_label.sh](scripts/add_sp_label.sh) to annotate the retrieved chains for train/val data. Supposing we got the top100 retrieved results in `data/hotpot/dev_retrieval_top100_sp.json`: \n\n```\npython scripts/train_qa.py \\\n --do_predict \\\n --predict_batch_size 200 \\\n --model_name google/electra-large-discriminator \\\n --fp16 \\\n --predict_file data/hotpot/dev_retrieval_top100_sp.json \\\n --max_seq_len 512 \\\n --max_q_len 64 \\\n --init_checkpoint models/qa_electra.pt \\\n --sp-pred \\\n --max_ans_len 30 \\\n --save-prediction hotpot_val_top100.json\n```\n\nExpected results:\n\n```\n01/21/2021 17:01:49 - INFO - __main__ - evaluated 7405 questions...\n01/21/2021 17:01:49 - INFO - __main__ - chain ranking em: 0.8113436866981769\n01/21/2021 17:01:50 - INFO - __main__ - .......Using combination factor 0.8......\n01/21/2021 17:01:50 - INFO - __main__ - answer em: 0.6233625928426739, count: 7405\n01/21/2021 17:01:50 - INFO - __main__ - answer f1: 0.7504594111976622, count: 7405\n01/21/2021 17:01:50 - INFO - __main__ - sp em: 0.5654287643484133, count: 7405\n01/21/2021 17:01:50 - INFO - __main__ - sp f1: 0.7942837708469039, count: 7405\n01/21/2021 17:01:50 - INFO - __main__ - joint em: 0.42052667116812964, count: 7405\n01/21/2021 17:01:50 - INFO - __main__ - joint f1: 0.6631669237532106, count: 7405\n01/21/2021 17:01:50 - INFO - __main__ - Best joint F1 from combination 0.7504594111976622\n01/21/2021 17:01:51 - INFO - __main__ - test performance {'em': 0.6233625928426739, 'f1': 0.7504594111976622, 'joint_em': 0.42052667116812964, 'joint_f1': 0.6631669237532106, 'sp_em': 0.5654287643484133, 'sp_f1': 0.7942837708469039}\n```\n\n## End to end Demo\nA simple demo code using our pretrained models.\n```\nstreamlit run scripts/demo.py\n```\n\n

\n\n\n## Train models from scratch\nOur experiments are mostly run on 8 GPUs, however, we observed similar performance when using a smaller performance. \n\n### Retriever training\n\n```\nCUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python scripts/train_mhop.py \\\n --do_train \\\n --prefix ${RUN_ID} \\\n --predict_batch_size 3000 \\\n --model_name roberta-base \\\n --train_batch_size 150 \\\n --learning_rate 2e-5 \\\n --fp16 \\\n --train_file ${TRAIN_DATA_PATH} \\\n --predict_file ${DEV_DATA_PATH} \\\n --seed 16 \\\n --eval-period -1 \\\n --max_c_len 300 \\\n --max_q_len 70 \\\n --max_q_sp_len 350 \\\n --shared-encoder \\\n --warmup-ratio 0.1\n```\nProcessed train/validation data for retrieval training:\n* `${TRAIN_DATA_PATH}`: data/hotpot/hotpot_train_with_neg_v0.json\n* `${DEV_DATA_PATH}`: data/hotpot/hotpot_dev_with_neg_v0.json\n\n### Finetune the question encoder with frozen memory bank\nThis step happens after the previous training stage and reuses the checkpoint\npoint.\n\n\n```\nCUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python train_momentum.py \\\n --do_train \\\n --prefix {RUN_ID} \\\n --predict_batch_size 3000 \\\n --model_name roberta-base \\\n --train_batch_size 150 \\\n --learning_rate 1e-5 \\\n --fp16 \\\n --train_file {TRAIN_DATA_PATH} \\\n --predict_file {DEV_DATA_PATH} \\\n --seed 16 \\\n --eval-period -1 \\\n --max_c_len 300 \\\n --max_q_len 70 \\\n --max_q_sp_len 350 \\\n --momentum \\\n --k 76800 \\\n --m 0.999 \\\n --temperature 1 \\\n --init-retriever {CHECKPOINT_PT}\n```\n\n## Encoding the corpus for retrieval\n```\nCUDA_VISIBLE_DEVICES=0,1,2,3 python scripts/encode_corpus.py \\\n --do_predict \\\n --predict_batch_size 1000 \\\n --model_name roberta-base \\\n --predict_file ${CORPUS_PATH} \\\n --init_checkpoint ${MODEL_CHECKPOINT} \\\n --embed_save_path ${SAVE_PATH} \\\n --fp16 \\\n --max_c_len 300 \\\n --num_workers 20 \n```\n* `${CORPUS_PATH}`: each line of this file should be an json encoded object ({\"title\": str, \"text\": str}). For HotpotQA, check the authors' [guide](https://hotpotqa.github.io/wiki-readme.html) to get the processed Wikipedia corpus (abstract only).\n* `${SAVE_PATH}`: path to save the numpy vectors and ID2DOC lookup table.\n\n### ELECTRA QA model training\n\nThe ELECTRA-based QA model is sensitive to the learning rate schedule and adding a 10% warmup stage is necessary to achieve good answer extraction performance in our experiments:\n```\nCUDA_VISIBLE_DEVICES=0 python train_qa.py \\\n --do_train \\\n --prefix electra_large_debug_sn \\\n --predict_batch_size 1024 \\\n --model_name google/electra-large-discriminator \\\n --train_batch_size 12 \\\n --learning_rate 5e-5 \\\n --train_file ${QA_TRAIN_DATA} \\\n --predict_file ${QA_DEV_DATA} \\\n --seed 42 \\\n --eval-period 250 \\\n --max_seq_len 512 \\\n --max_q_len 64 \\\n --gradient_accumulation_steps 8 \\\n --neg-num 5 \\\n --fp16 \\\n --use-adam \\\n --warmup-ratio 0.1 \\\n --sp-weight 0.05 \\\n --sp-pred\n```\n\nProcessed (ran [scripts/add_sp_label.sh](scripts/add_sp_label.sh)) train/validata data for QA training.\n* `${QA_TRAIN_DATA}`: data/hotpot/train_retrieval_b100_k100_sp.json\n* `${QA_DEV_DATA}`: data/hotpot/dev_retrieval_b50_k50_sp.json\n\n## Cite\n```\n@article{xiong2020answering,\n title={Answering Complex Open-Domain Questions with Multi-Hop Dense Retrieval},\n author={Xiong, Wenhan and Li, Xiang Lorraine and Iyer, Srinivasan and Du, Jingfei and Lewis, Patrick and Wang, William Yang and Mehdad, Yashar and Yih, Wen-tau and Riedel, Sebastian and Kiela, Douwe and O{\\u{g}}uz, Barlas},\n journal={International Conference on Learning Representations},\n year={2021}\n}\n```\n\n## License\nCC-BY-NC 4.0\n" }, { "path": "mdr/__init__.py", "content": "# Copyright (c) Facebook, Inc. and its affiliates.\n# All rights reserved.\n#\n# This source code is licensed under the license found in the \n# LICENSE file in the root directory of this source tree.\n\n\nfrom . import qa\nfrom . import retrieval" }, { "path": "mdr/qa/__init__.py", "content": "# Copyright (c) Facebook, Inc. and its affiliates.\n# All rights reserved.\n#\n# This source code is licensed under the license found in the \n# LICENSE file in the root directory of this source tree.\n\n\n" }, { "path": "mdr/qa/basic_tokenizer.py", "content": "# Copyright (c) Facebook, Inc. and its affiliates.\n# All rights reserved.\n#\n# This source code is licensed under the license found in the \n# LICENSE file in the root directory of this source tree.\n#!/usr/bin/env python3\n# Copyright 2017-present, Facebook, Inc.\n# All rights reserved.\n#\n# This source code is licensed under the license found in the\n# LICENSE file in the root directory of this source tree.\n\"\"\"Base tokenizer/tokens classes and utilities.\"\"\"\n\nimport copy\n\nclass Tokens(object):\n \"\"\"A class to represent a list of tokenized text.\"\"\"\n TEXT = 0\n TEXT_WS = 1\n SPAN = 2\n POS = 3\n LEMMA = 4\n NER = 5\n\n def __init__(self, data, annotators, opts=None):\n self.data = data\n self.annotators = annotators\n self.opts = opts or {}\n\n def __len__(self):\n \"\"\"The number of tokens.\"\"\"\n return len(self.data)\n\n def slice(self, i=None, j=None):\n \"\"\"Return a view of the list of tokens from [i, j).\"\"\"\n new_tokens = copy.copy(self)\n new_tokens.data = self.data[i: j]\n return new_tokens\n\n def untokenize(self):\n \"\"\"Returns the original text (with whitespace reinserted).\"\"\"\n return ''.join([t[self.TEXT_WS] for t in self.data]).strip()\n\n def words(self, uncased=False):\n \"\"\"Returns a list of the text of each token\n\n Args:\n uncased: lower cases text\n \"\"\"\n if uncased:\n return [t[self.TEXT].lower() for t in self.data]\n else:\n return [t[self.TEXT] for t in self.data]\n\n def offsets(self):\n \"\"\"Returns a list of [start, end) character offsets of each token.\"\"\"\n return [t[self.SPAN] for t in self.data]\n\n def pos(self):\n \"\"\"Returns a list of part-of-speech tags of each token.\n Returns None if this annotation was not included.\n \"\"\"\n if 'pos' not in self.annotators:\n return None\n return [t[self.POS] for t in self.data]\n\n def lemmas(self):\n \"\"\"Returns a list of the lemmatized text of each token.\n Returns None if this annotation was not included.\n \"\"\"\n if 'lemma' not in self.annotators:\n return None\n return [t[self.LEMMA] for t in self.data]\n\n def entities(self):\n \"\"\"Returns a list of named-entity-recognition tags of each token.\n Returns None if this annotation was not included.\n \"\"\"\n if 'ner' not in self.annotators:\n return None\n return [t[self.NER] for t in self.data]\n\n def ngrams(self, n=1, uncased=False, filter_fn=None, as_strings=True):\n \"\"\"Returns a list of all ngrams from length 1 to n.\n\n Args:\n n: upper limit of ngram length\n uncased: lower cases text\n filter_fn: user function that takes in an ngram list and returns\n True or False to keep or not keep the ngram\n as_string: return the ngram as a string vs list\n \"\"\"\n def _skip(gram):\n if not filter_fn:\n return False\n return filter_fn(gram)\n\n words = self.words(uncased)\n ngrams = [(s, e + 1)\n for s in range(len(words))\n for e in range(s, min(s + n, len(words)))\n if not _skip(words[s:e + 1])]\n\n # Concatenate into strings\n if as_strings:\n ngrams = ['{}'.format(' '.join(words[s:e])) for (s, e) in ngrams]\n\n return ngrams\n\n def entity_groups(self):\n \"\"\"Group consecutive entity tokens with the same NER tag.\"\"\"\n entities = self.entities()\n if not entities:\n return None\n non_ent = self.opts.get('non_ent', 'O')\n groups = []\n idx = 0\n while idx < len(entities):\n ner_tag = entities[idx]\n # Check for entity tag\n if ner_tag != non_ent:\n # Chomp the sequence\n start = idx\n while (idx < len(entities) and entities[idx] == ner_tag):\n idx += 1\n groups.append((self.slice(start, idx).untokenize(), ner_tag))\n else:\n idx += 1\n return groups\n\n\nclass Tokenizer(object):\n \"\"\"Base tokenizer class.\n Tokenizers implement tokenize, which should return a Tokens class.\n \"\"\"\n\n def tokenize(self, text):\n raise NotImplementedError\n\n def shutdown(self):\n pass\n\n def __del__(self):\n self.shutdown()\n\n\nimport regex\nimport logging\n\nlogger = logging.getLogger(__name__)\n\n\nclass RegexpTokenizer(Tokenizer):\n DIGIT = r'\\p{Nd}+([:\\.\\,]\\p{Nd}+)*'\n TITLE = (r'(dr|esq|hon|jr|mr|mrs|ms|prof|rev|sr|st|rt|messrs|mmes|msgr)'\n r'\\.(?=\\p{Z})')\n ABBRV = r'([\\p{L}]\\.){2,}(?=\\p{Z}|$)'\n ALPHA_NUM = r'[\\p{L}\\p{N}\\p{M}]++'\n HYPHEN = r'{A}([-\\u058A\\u2010\\u2011]{A})+'.format(A=ALPHA_NUM)\n NEGATION = r\"((?!n't)[\\p{L}\\p{N}\\p{M}])++(?=n't)|n't\"\n CONTRACTION1 = r\"can(?=not\\b)\"\n CONTRACTION2 = r\"'([tsdm]|re|ll|ve)\\b\"\n START_DQUOTE = r'(?<=[\\p{Z}\\(\\[{<]|^)(``|[\"\\u0093\\u201C\\u00AB])(?!\\p{Z})'\n START_SQUOTE = r'(?<=[\\p{Z}\\(\\[{<]|^)[\\'\\u0091\\u2018\\u201B\\u2039](?!\\p{Z})'\n END_DQUOTE = r'(?%s)|(?P%s)|(?P<abbr>%s)|(?P<neg>%s)|(?P<hyph>%s)|'\n '(?P<contr1>%s)|(?P<alphanum>%s)|(?P<contr2>%s)|(?P<sdquote>%s)|'\n '(?P<edquote>%s)|(?P<ssquote>%s)|(?P<esquote>%s)|(?P<dash>%s)|'\n '(?<ellipses>%s)|(?P<punct>%s)|(?P<nonws>%s)' %\n (self.DIGIT, self.TITLE, self.ABBRV, self.NEGATION, self.HYPHEN,\n self.CONTRACTION1, self.ALPHA_NUM, self.CONTRACTION2,\n self.START_DQUOTE, self.END_DQUOTE, self.START_SQUOTE,\n self.END_SQUOTE, self.DASH, self.ELLIPSES, self.PUNCT,\n self.NON_WS),\n flags=regex.IGNORECASE + regex.UNICODE + regex.MULTILINE\n )\n if len(kwargs.get('annotators', {})) > 0:\n logger.warning('%s only tokenizes! Skipping annotators: %s' %\n (type(self).__name__, kwargs.get('annotators')))\n self.annotators = set()\n self.substitutions = kwargs.get('substitutions', True)\n\n def tokenize(self, text):\n data = []\n matches = [m for m in self._regexp.finditer(text)]\n for i in range(len(matches)):\n # Get text\n token = matches[i].group()\n\n # Make normalizations for special token types\n if self.substitutions:\n groups = matches[i].groupdict()\n if groups['sdquote']:\n token = \"``\"\n elif groups['edquote']:\n token = \"''\"\n elif groups['ssquote']:\n token = \"`\"\n elif groups['esquote']:\n token = \"'\"\n elif groups['dash']:\n token = '--'\n elif groups['ellipses']:\n token = '...'\n\n # Get whitespace\n span = matches[i].span()\n start_ws = span[0]\n if i + 1 < len(matches):\n end_ws = matches[i + 1].span()[0]\n else:\n end_ws = span[1]\n\n # Format data\n data.append((\n token,\n text[start_ws: end_ws],\n span,\n ))\n return Tokens(data, self.annotators)\n\n\nclass SimpleTokenizer(Tokenizer):\n ALPHA_NUM = r'[\\p{L}\\p{N}\\p{M}]+'\n NON_WS = r'[^\\p{Z}\\p{C}]'\n\n def __init__(self, **kwargs):\n \"\"\"\n Args:\n annotators: None or empty set (only tokenizes).\n \"\"\"\n self._regexp = regex.compile(\n '(%s)|(%s)' % (self.ALPHA_NUM, self.NON_WS),\n flags=regex.IGNORECASE + regex.UNICODE + regex.MULTILINE\n )\n if len(kwargs.get('annotators', {})) > 0:\n logger.warning('%s only tokenizes! Skipping annotators: %s' %\n (type(self).__name__, kwargs.get('annotators')))\n self.annotators = set()\n\n def tokenize(self, text):\n data = []\n matches = [m for m in self._regexp.finditer(text)]\n for i in range(len(matches)):\n # Get text\n token = matches[i].group()\n\n # Get whitespace\n span = matches[i].span()\n start_ws = span[0]\n if i + 1 < len(matches):\n end_ws = matches[i + 1].span()[0]\n else:\n end_ws = span[1]\n\n # Format data\n data.append((\n token,\n text[start_ws: end_ws],\n span,\n ))\n return Tokens(data, self.annotators)\n\n\n\n" }, { "path": "mdr/qa/config.py", "content": "# Copyright (c) Facebook, Inc. and its affiliates.\n# All rights reserved.\n#\n# This source code is licensed under the license found in the \n# LICENSE file in the root directory of this source tree.\nimport argparse\nfrom ast import parse\n\nfrom typing import NamedTuple\n\nfrom torch.nn import parallel\nclass ClusterConfig(NamedTuple):\n dist_backend: str\n dist_url: str\n\ndef common_args():\n parser = argparse.ArgumentParser()\n\n # task\n parser.add_argument(\"--train_file\", type=str,\n default=\"../data/nq-with-neg-train.txt\")\n parser.add_argument(\"--predict_file\", type=str,\n default=\"../data/nq-with-neg-dev.txt\")\n parser.add_argument(\"--num_workers\", default=10, type=int)\n parser.add_argument(\"--do_train\", default=False,\n action='store_true', help=\"Whether to run training.\")\n parser.add_argument(\"--do_predict\", default=False,\n action='store_true', help=\"Whether to run eval on the dev set.\")\n parser.add_argument(\"--do_test\", default=False, action=\"store_true\", help=\"for final test submission\")\n\n # model\n parser.add_argument(\"--model_name\",\n default=\"bert-base-uncased\", type=str)\n parser.add_argument(\"--init_checkpoint\", type=str,\n help=\"Initial checkpoint (usually from a pre-trained BERT model).\",\n default=\"\")\n parser.add_argument(\"--max_seq_len\", default=512, type=int,\n help=\"The maximum total input sequence length after WordPiece tokenization. Sequences \"\n \"longer than this will be truncated, and sequences shorter than this will be padded.\")\n parser.add_argument(\"--max_q_len\", default=64, type=int)\n parser.add_argument(\"--max_ans_len\", default=35, type=int)\n parser.add_argument('--fp16', action='store_true')\n parser.add_argument('--fp16_opt_level', type=str, default='O1',\n help=\"For fp16: Apex AMP optimization level selected in ['O0', 'O1', 'O2', and 'O3'].\"\n \"See details at https://nvidia.github.io/apex/amp.html\")\n parser.add_argument(\"--no_cuda\", default=False, action='store_true',\n help=\"Whether not to use CUDA when available\")\n parser.add_argument(\"--local_rank\", type=int, default=-1,\n help=\"local_rank for distributed training on gpus\")\n parser.add_argument(\"--predict_batch_size\", default=256,\n type=int, help=\"Total batch size for predictions.\")\n parser.add_argument(\"--save-prediction\", default=\"\", type=str)\n\n parser.add_argument(\"--sp-pred\", action=\"store_true\", help=\"whether to predict sentence sp\")\n return parser\n\ndef train_args():\n parser = common_args()\n # optimization\n parser.add_argument('--prefix', type=str, default=\"eval\")\n parser.add_argument(\"--weight_decay\", default=0.0, type=float,\n help=\"Weight decay if we apply some.\")\n parser.add_argument(\"--output_dir\", default=\"./logs\", type=str,\n help=\"The output directory where the model checkpoints will be written.\")\n parser.add_argument(\"--train_batch_size\", default=128,\n type=int, help=\"Total batch size for training.\")\n parser.add_argument(\"--num_q_per_gpu\", default=1) \n parser.add_argument(\"--learning_rate\", default=1e-5,\n type=float, help=\"The initial learning rate for Adam.\")\n parser.add_argument(\"--num_train_epochs\", default=5, type=float,\n help=\"Total number of training epochs to perform.\")\n parser.add_argument('--seed', type=int, default=3,\n help=\"random seed for initialization\")\n parser.add_argument('--gradient_accumulation_steps', type=int, default=1,\n help=\"Number of updates steps to accumualte before performing a backward/update pass.\")\n parser.add_argument('--eval-period', type=int, default=2500)\n parser.add_argument(\"--max_grad_norm\", default=2.0, type=float, help=\"Max gradient norm.\")\n parser.add_argument(\"--adam_epsilon\", default=1e-8, type=float, help=\"Epsilon for Adam optimizer.\")\n parser.add_argument(\"--neg-num\", type=int, default=9, help=\"how many neg/distant passage chains to use\")\n parser.add_argument(\"--shared-norm\", action=\"store_true\")\n parser.add_argument(\"--qa-drop\", default=0, type=float)\n parser.add_argument(\"--rank-drop\", default=0, type=float)\n parser.add_argument(\"--sp-drop\", default=0, type=float)\n parser.add_argument(\"--final-metric\", default=\"joint_f1\")\n parser.add_argument(\"--use-adam\", action=\"store_true\", help=\"use adam or adamW\")\n parser.add_argument(\"--warmup-ratio\", default=0, type=float, help=\"Linear warmup over warmup_steps.\")\n parser.add_argument(\"--sp-weight\", default=0, type=float, help=\"weight of the sp loss\")\n return parser.parse_args()\n" }, { "path": "mdr/qa/data_utils.py", "content": "# Copyright (c) Facebook, Inc. and its affiliates.\n# All rights reserved.\n#\n# This source code is licensed under the license found in the \n# LICENSE file in the root directory of this source tree.\n\nimport json\nfrom tqdm import tqdm\nimport numpy as np\n\ndef explore(path):\n train = json.load(open(path))\n\n neg_counts = []\n for item in train:\n tfidf_neg = item[\"tfidf_neg\"]\n linked_neg = item[\"linked_neg\"]\n neg_counts.append(len(tfidf_neg + linked_neg))\n \n import pdb; pdb.set_trace()\n return \n\ndef load_corpus(corpus_path=\"/private/home/xwhan/data/hotpot/tfidf/abstracts.txt\"):\n content = [json.loads(l) for l in open(corpus_path).readlines()]\n title2doc = {item[\"title\"]:item[\"text\"] for item in content}\n\nif __name__ == \"__main__\":\n explore(\"/private/home/xwhan/data/hotpot/hotpot_rerank_train_2_neg_types.json\")" }, { "path": "mdr/qa/hotpot_evaluate_v1.py", "content": "# Copyright (c) Facebook, Inc. and its affiliates.\n# All rights reserved.\n#\n# This source code is licensed under the license found in the \n# LICENSE file in the root directory of this source tree.\nimport sys\nimport ujson as json\nimport re\nimport string\nfrom collections import Counter\nimport pickle\n\ndef normalize_answer(s):\n\n def remove_articles(text):\n return re.sub(r'\\b(a|an|the)\\b', ' ', text)\n\n def white_space_fix(text):\n return ' '.join(text.split())\n\n def remove_punc(text):\n exclude = set(string.punctuation)\n return ''.join(ch for ch in text if ch not in exclude)\n\n def lower(text):\n return text.lower()\n\n return white_space_fix(remove_articles(remove_punc(lower(s))))\n\n\ndef f1_score(prediction, ground_truth):\n normalized_prediction = normalize_answer(prediction)\n normalized_ground_truth = normalize_answer(ground_truth)\n\n ZERO_METRIC = (0, 0, 0)\n\n if normalized_prediction in ['yes', 'no', 'noanswer'] and normalized_prediction != normalized_ground_truth:\n return ZERO_METRIC\n if normalized_ground_truth in ['yes', 'no', 'noanswer'] and normalized_prediction != normalized_ground_truth:\n return ZERO_METRIC\n\n prediction_tokens = normalized_prediction.split()\n ground_truth_tokens = normalized_ground_truth.split()\n common = Counter(prediction_tokens) & Counter(ground_truth_tokens)\n num_same = sum(common.values())\n if num_same == 0:\n return ZERO_METRIC\n precision = 1.0 * num_same / len(prediction_tokens)\n recall = 1.0 * num_same / len(ground_truth_tokens)\n f1 = (2 * precision * recall) / (precision + recall)\n return f1, precision, recall\n\n\ndef exact_match_score(prediction, ground_truth):\n return (normalize_answer(prediction) == normalize_answer(ground_truth))\n\ndef update_answer(metrics, prediction, gold):\n em = exact_match_score(prediction, gold)\n f1, prec, recall = f1_score(prediction, gold)\n metrics['em'] += float(em)\n metrics['f1'] += f1\n metrics['prec'] += prec\n metrics['recall'] += recall\n return em, prec, recall\n\ndef update_sp(metrics, prediction, gold):\n cur_sp_pred = set(map(tuple, prediction))\n gold_sp_pred = set(map(tuple, gold))\n tp, fp, fn = 0, 0, 0\n for e in cur_sp_pred:\n if e in gold_sp_pred:\n tp += 1\n else:\n fp += 1\n for e in gold_sp_pred:\n if e not in cur_sp_pred:\n fn += 1\n prec = 1.0 * tp / (tp + fp) if tp + fp > 0 else 0.0\n recall = 1.0 * tp / (tp + fn) if tp + fn > 0 else 0.0\n f1 = 2 * prec * recall / (prec + recall) if prec + recall > 0 else 0.0\n em = 1.0 if fp + fn == 0 else 0.0\n metrics['sp_em'] += em\n metrics['sp_f1'] += f1\n metrics['sp_prec'] += prec\n metrics['sp_recall'] += recall\n return em, prec, recall\n\ndef eval(prediction_file, gold_file):\n with open(prediction_file) as f:\n prediction = json.load(f)\n with open(gold_file) as f:\n gold = json.load(f)\n\n metrics = {'em': 0, 'f1': 0, 'prec': 0, 'recall': 0,\n 'sp_em': 0, 'sp_f1': 0, 'sp_prec': 0, 'sp_recall': 0,\n 'joint_em': 0, 'joint_f1': 0, 'joint_prec': 0, 'joint_recall': 0}\n for dp in gold:\n cur_id = dp['_id']\n can_eval_joint = True\n if cur_id not in prediction['answer']:\n print('missing answer {}'.format(cur_id))\n can_eval_joint = False\n else:\n em, prec, recall = update_answer(\n metrics, prediction['answer'][cur_id], dp['answer'])\n if cur_id not in prediction['sp']:\n print('missing sp fact {}'.format(cur_id))\n can_eval_joint = False\n else:\n sp_em, sp_prec, sp_recall = update_sp(\n metrics, prediction['sp'][cur_id], dp['supporting_facts'])\n\n if can_eval_joint:\n joint_prec = prec * sp_prec\n joint_recall = recall * sp_recall\n if joint_prec + joint_recall > 0:\n joint_f1 = 2 * joint_prec * joint_recall / (joint_prec + joint_recall)\n else:\n joint_f1 = 0.\n joint_em = em * sp_em\n\n metrics['joint_em'] += joint_em\n metrics['joint_f1'] += joint_f1\n metrics['joint_prec'] += joint_prec\n metrics['joint_recall'] += joint_recall\n\n N = len(gold)\n for k in metrics.keys():\n metrics[k] /= N\n\n print(metrics)\n\nif __name__ == '__main__':\n eval(sys.argv[1], sys.argv[2])\n\n" }, { "path": "mdr/qa/qa_dataset.py", "content": "# Copyright (c) Facebook, Inc. and its affiliates.\n# All rights reserved.\n#\n# This source code is licensed under the license found in the \n# LICENSE file in the root directory of this source tree.\nimport collections\nimport json\nimport random\n\nimport torch\nfrom torch.utils.data import Dataset, Sampler\nfrom tqdm import tqdm\n\nfrom .basic_tokenizer import SimpleTokenizer\nfrom .utils import (find_ans_span_with_char_offsets, match_answer_span, para_has_answer, _is_whitespace)\n\ndef collate_tokens(values, pad_idx, eos_idx=None, left_pad=False, move_eos_to_beginning=False):\n \"\"\"Convert a list of 1d tensors into a padded 2d tensor.\"\"\"\n if len(values[0].size()) > 1:\n values = [v.view(-1) for v in values]\n size = max(v.size(0) for v in values)\n res = values[0].new(len(values), size).fill_(pad_idx)\n\n def copy_tensor(src, dst):\n assert dst.numel() == src.numel()\n if move_eos_to_beginning:\n assert src[-1] == eos_idx\n dst[0] = eos_idx\n dst[1:] = src[:-1]\n else:\n dst.copy_(src)\n\n for i, v in enumerate(values):\n copy_tensor(v, res[i][size - len(v):] if left_pad else res[i][:len(v)])\n return res\n\n\ndef prepare(item, tokenizer, special_toks=[\"[SEP]\", \"[unused1]\", \"[unused2]\"]):\n \"\"\"\n tokenize the passages chains, add sentence start markers for SP sentence identification\n \"\"\"\n def _process_p(para):\n \"\"\"\n handle each para\n \"\"\"\n title, sents = para[\"title\"].strip(), para[\"sents\"]\n # return \"[unused1] \" + title + \" [unused1] \" + text # mark title\n # return title + \" \" + text\n pre_sents = []\n for idx, sent in enumerate(sents):\n pre_sents.append(\"[unused1] \" + sent.strip())\n return title + \" \" + \" \".join(pre_sents)\n # return \" \".join(pre_sents)\n # mark passage boundary\n contexts = []\n for para in item[\"passages\"]:\n contexts.append(_process_p(para))\n context = \" [SEP] \".join(contexts)\n\n doc_tokens = []\n char_to_word_offset = []\n prev_is_whitespace = True\n\n context = \"yes no [SEP] \" + context\n\n for c in context:\n if _is_whitespace(c):\n prev_is_whitespace = True\n else:\n if prev_is_whitespace:\n doc_tokens.append(c)\n else:\n doc_tokens[-1] += c\n prev_is_whitespace = False\n char_to_word_offset.append(len(doc_tokens) - 1)\n\n sent_starts = []\n orig_to_tok_index = []\n tok_to_orig_index = []\n all_doc_tokens = []\n for (i, token) in enumerate(doc_tokens):\n orig_to_tok_index.append(len(all_doc_tokens))\n\n if token in special_toks:\n if token == \"[unused1]\":\n sent_starts.append(len(all_doc_tokens))\n\n sub_tokens = [token]\n else:\n sub_tokens = tokenizer.tokenize(token)\n\n for sub_token in sub_tokens:\n tok_to_orig_index.append(i)\n all_doc_tokens.append(sub_token)\n\n item[\"context_processed\"] = {\n \"doc_tokens\": doc_tokens,\n \"char_to_word_offset\": char_to_word_offset,\n \"orig_to_tok_index\": orig_to_tok_index,\n \"tok_to_orig_index\": tok_to_orig_index,\n \"all_doc_tokens\": all_doc_tokens,\n \"context\": context,\n \"sent_starts\": sent_starts\n }\n\n return item\n\nclass QAEvalDataset(Dataset):\n\n def __init__(self,\n tokenizer,\n retrievel_results,\n max_seq_len,\n max_q_len,\n ):\n\n retriever_outputs = retrievel_results\n self.tokenizer = tokenizer\n self.max_seq_len = max_seq_len\n self.max_q_len = max_q_len\n self.data = []\n\n for item in retriever_outputs:\n if item[\"question\"].endswith(\"?\"):\n item[\"question\"] = item[\"question\"][:-1]\n\n # for validation, add target predictions\n sp_titles = None\n gold_answer = item.get(\"answer\", [])\n sp_gold = []\n\n for chain in item[\"candidate_chains\"]:\n chain_titles = [_[\"title\"] for _ in chain]\n\n if sp_titles:\n label = int(set(chain_titles) == sp_titles)\n else:\n label = -1\n self.data.append({\n \"question\": item[\"question\"],\n \"passages\": chain,\n \"label\": label,\n \"qid\": item[\"_id\"],\n \"gold_answer\": gold_answer,\n \"sp_gold\": sp_gold\n })\n\n print(f\"Total instances size {len(self.data)}\")\n\n def __len__(self):\n return len(self.data)\n\n def __getitem__(self, index):\n item = prepare(self.data[index], self.tokenizer) \n context_ann = item[\"context_processed\"]\n q_toks = self.tokenizer.tokenize(item[\"question\"])[:self.max_q_len]\n para_offset = len(q_toks) + 2 # cls and seq\n item[\"wp_tokens\"] = context_ann[\"all_doc_tokens\"]\n assert item[\"wp_tokens\"][0] == \"yes\" and item[\"wp_tokens\"][1] == \"no\"\n item[\"para_offset\"] = para_offset\n max_toks_for_doc = self.max_seq_len - para_offset - 1\n if len(item[\"wp_tokens\"]) > max_toks_for_doc:\n item[\"wp_tokens\"] = item[\"wp_tokens\"][:max_toks_for_doc]\n item[\"encodings\"] = self.tokenizer.encode_plus(q_toks, text_pair=item[\"wp_tokens\"], max_length=self.max_seq_len, return_tensors=\"pt\", is_pretokenized=True)\n\n item[\"paragraph_mask\"] = torch.zeros(item[\"encodings\"][\"input_ids\"].size()).view(-1)\n item[\"paragraph_mask\"][para_offset:-1] = 1\n \n item[\"doc_tokens\"] = context_ann[\"doc_tokens\"]\n item[\"tok_to_orig_index\"] = context_ann[\"tok_to_orig_index\"]\n\n # filter sentence offsets exceeding max sequence length\n sent_labels, sent_offsets = [], []\n for idx, s in enumerate(item[\"context_processed\"][\"sent_starts\"]):\n if s >= len(item[\"wp_tokens\"]):\n break\n if \"sp_sent_labels\" in item:\n sent_labels.append(item[\"sp_sent_labels\"][idx])\n sent_offsets.append(s + para_offset)\n assert item[\"encodings\"][\"input_ids\"].view(-1)[s+para_offset] == self.tokenizer.convert_tokens_to_ids(\"[unused1]\")\n\n # supporting fact label\n item[\"sent_offsets\"] = sent_offsets\n item[\"sent_offsets\"] = torch.LongTensor(item[\"sent_offsets\"])\n item[\"label\"] = torch.LongTensor([item[\"label\"]])\n return item\n\nclass QADataset(Dataset):\n\n def __init__(self,\n tokenizer,\n data_path,\n max_seq_len,\n max_q_len,\n train=False,\n no_sent_label=False\n ):\n\n retriever_outputs = [json.loads(l) for l in tqdm(open(data_path).readlines())]\n self.tokenizer = tokenizer\n self.max_seq_len = max_seq_len\n self.max_q_len = max_q_len\n self.train = train\n self.no_sent_label = no_sent_label\n self.simple_tok = SimpleTokenizer()\n self.data = []\n\n if train:\n self.qid2gold = collections.defaultdict(list) # idx \n self.qid2neg = collections.defaultdict(list)\n for item in retriever_outputs:\n if item[\"question\"].endswith(\"?\"):\n item[\"question\"] = item[\"question\"][:-1]\n\n sp_sent_labels = []\n sp_gold = []\n if not self.no_sent_label:\n for sp in item[\"sp\"]:\n for _ in sp[\"sp_sent_ids\"]:\n sp_gold.append([sp[\"title\"], _])\n for idx in range(len(sp[\"sents\"])):\n sp_sent_labels.append(int(idx in sp[\"sp_sent_ids\"]))\n\n question_type = item[\"type\"]\n self.data.append({\n \"question\": item[\"question\"],\n \"passages\": item[\"sp\"], \n \"label\": 1,\n \"qid\": item[\"_id\"],\n \"gold_answer\": item[\"answer\"],\n \"sp_sent_labels\": sp_sent_labels,\n \"ans_covered\": 1, # includes partial chains.\n \"sp_gold\": sp_gold\n })\n self.qid2gold[item[\"_id\"]].append(len(self.data) - 1)\n\n sp_titles = set([_[\"title\"] for _ in item[\"sp\"]])\n if question_type == \"bridge\":\n ans_titles = set([p[\"title\"] for p in item[\"sp\"] if para_has_answer(item[\"answer\"], \"\".join(p[\"sents\"]), self.simple_tok)])\n else:\n ans_titles = set()\n # top ranked negative chains\n ds_count = 0 # track how many distant supervised chain to use\n ds_limit = 5\n for chain in item[\"candidate_chains\"]:\n chain_titles = [_[\"title\"] for _ in chain]\n if set(chain_titles) == sp_titles:\n continue\n if question_type == \"bridge\":\n answer_covered = int(len(set(chain_titles) & ans_titles) > 0)\n ds_count += answer_covered\n else:\n answer_covered = 0\n self.data.append({\n \"question\": item[\"question\"],\n \"passages\": chain,\n \"label\": 0,\n \"qid\": item[\"_id\"],\n \"gold_answer\": item[\"answer\"],\n \"ans_covered\": answer_covered,\n \"sp_gold\": sp_gold\n })\n self.qid2neg[item[\"_id\"]].append(len(self.data) - 1)\n else:\n for item in retriever_outputs:\n if item[\"question\"].endswith(\"?\"):\n item[\"question\"] = item[\"question\"][:-1]\n\n # for validation, add target predictions\n sp_titles = set([_[\"title\"] for _ in item[\"sp\"]]) if \"sp\" in item else None\n gold_answer = item.get(\"answer\", [])\n sp_gold = []\n if \"sp\" in item:\n for sp in item[\"sp\"]:\n for _ in sp[\"sp_sent_ids\"]:\n sp_gold.append([sp[\"title\"], _])\n\n chain_seen = set()\n for chain in item[\"candidate_chains\"]:\n chain_titles = [_[\"title\"] for _ in chain]\n\n # title_set = frozenset(chain_titles)\n # if len(title_set) == 0 or title_set in chain_seen:\n # continue\n # chain_seen.add(title_set)\n\n if sp_titles:\n label = int(set(chain_titles) == sp_titles)\n else:\n label = -1\n self.data.append({\n \"question\": item[\"question\"],\n \"passages\": chain,\n \"label\": label,\n \"qid\": item[\"_id\"],\n \"gold_answer\": gold_answer,\n \"sp_gold\": sp_gold\n })\n\n print(f\"Data size {len(self.data)}\")\n\n def __len__(self):\n return len(self.data)\n\n def __getitem__(self, index):\n item = prepare(self.data[index], self.tokenizer) \n context_ann = item[\"context_processed\"]\n q_toks = self.tokenizer.tokenize(item[\"question\"])[:self.max_q_len]\n para_offset = len(q_toks) + 2 # cls and seq\n item[\"wp_tokens\"] = context_ann[\"all_doc_tokens\"]\n assert item[\"wp_tokens\"][0] == \"yes\" and item[\"wp_tokens\"][1] == \"no\"\n item[\"para_offset\"] = para_offset\n max_toks_for_doc = self.max_seq_len - para_offset - 1\n if len(item[\"wp_tokens\"]) > max_toks_for_doc:\n item[\"wp_tokens\"] = item[\"wp_tokens\"][:max_toks_for_doc]\n item[\"encodings\"] = self.tokenizer.encode_plus(q_toks, text_pair=item[\"wp_tokens\"], max_length=self.max_seq_len, return_tensors=\"pt\", is_pretokenized=True)\n\n item[\"paragraph_mask\"] = torch.zeros(item[\"encodings\"][\"input_ids\"].size()).view(-1)\n item[\"paragraph_mask\"][para_offset:-1] = 1\n \n if self.train:\n # if item[\"label\"] == 1:\n if item[\"ans_covered\"]:\n if item[\"gold_answer\"][0] == \"yes\":\n # ans_type = 0\n starts, ends= [para_offset], [para_offset]\n elif item[\"gold_answer\"][0] == \"no\":\n # ans_type = 1\n starts, ends= [para_offset + 1], [para_offset + 1]\n else:\n # ans_type = 2\n matched_spans = match_answer_span(context_ann[\"context\"], item[\"gold_answer\"], self.simple_tok)\n ans_starts, ans_ends= [], []\n for span in matched_spans:\n char_starts = [i for i in range(len(context_ann[\"context\"])) if context_ann[\"context\"].startswith(span, i)]\n if len(char_starts) > 0:\n char_ends = [start + len(span) - 1 for start in char_starts]\n answer = {\"text\": span, \"char_spans\": list(zip(char_starts, char_ends))}\n ans_spans = find_ans_span_with_char_offsets(\n answer, context_ann[\"char_to_word_offset\"], context_ann[\"doc_tokens\"], context_ann[\"all_doc_tokens\"], context_ann[\"orig_to_tok_index\"], self.tokenizer)\n for s, e in ans_spans:\n ans_starts.append(s)\n ans_ends.append(e)\n starts, ends = [], []\n for s, e in zip(ans_starts, ans_ends):\n if s >= len(item[\"wp_tokens\"]):\n continue\n else:\n s = min(s, len(item[\"wp_tokens\"]) - 1) + para_offset\n e = min(e, len(item[\"wp_tokens\"]) - 1) + para_offset\n starts.append(s)\n ends.append(e)\n if len(starts) == 0:\n starts, ends = [-1], [-1] \n else:\n starts, ends= [-1], [-1]\n # ans_type = -1\n \n item[\"starts\"] = torch.LongTensor(starts)\n item[\"ends\"] = torch.LongTensor(ends)\n # item[\"ans_type\"] = torch.LongTensor([ans_type])\n\n if item[\"label\"]:\n assert len(item[\"sp_sent_labels\"]) == len(item[\"context_processed\"][\"sent_starts\"])\n else:\n # # for answer extraction\n item[\"doc_tokens\"] = context_ann[\"doc_tokens\"]\n item[\"tok_to_orig_index\"] = context_ann[\"tok_to_orig_index\"]\n\n # filter sentence offsets exceeding max sequence length\n sent_labels, sent_offsets = [], []\n for idx, s in enumerate(item[\"context_processed\"][\"sent_starts\"]):\n if s >= len(item[\"wp_tokens\"]):\n break\n if \"sp_sent_labels\" in item:\n sent_labels.append(item[\"sp_sent_labels\"][idx])\n sent_offsets.append(s + para_offset)\n assert item[\"encodings\"][\"input_ids\"].view(-1)[s+para_offset] == self.tokenizer.convert_tokens_to_ids(\"[unused1]\")\n\n # supporting fact label\n item[\"sent_offsets\"] = sent_offsets\n item[\"sent_offsets\"] = torch.LongTensor(item[\"sent_offsets\"])\n if self.train:\n item[\"sent_labels\"] = sent_labels if len(sent_labels) != 0 else [0] * len(sent_offsets)\n item[\"sent_labels\"] = torch.LongTensor(item[\"sent_labels\"])\n item[\"ans_covered\"] = torch.LongTensor([item[\"ans_covered\"]])\n\n item[\"label\"] = torch.LongTensor([item[\"label\"]])\n return item\n\nclass MhopSampler(Sampler):\n \"\"\"\n Shuffle QA pairs not context, make sure data within the batch are from the same QA pair\n \"\"\"\n\n def __init__(self, data_source, num_neg=9, n_gpu=8):\n # for each QA pair, sample negative paragraphs\n self.qid2gold = data_source.qid2gold\n self.qid2neg = data_source.qid2neg\n self.neg_num = num_neg\n self.n_gpu = n_gpu\n self.all_qids = list(self.qid2gold.keys())\n assert len(self.qid2gold) == len(self.qid2neg)\n\n self.q_num_per_epoch = len(self.qid2gold) - len(self.qid2gold) % self.n_gpu\n self._num_samples = self.q_num_per_epoch * (self.neg_num + 1)\n\n def __len__(self):\n return self._num_samples\n\n def __iter__(self):\n sample_indice = []\n random.shuffle(self.all_qids)\n \n # when use shared-normalization, passages for each question should be on the same GPU\n qids_to_use = self.all_qids[:self.q_num_per_epoch]\n for qid in qids_to_use:\n neg_samples = self.qid2neg[qid]\n random.shuffle(neg_samples)\n sample_indice += self.qid2gold[qid]\n sample_indice += neg_samples[:self.neg_num]\n return iter(sample_indice)\n\ndef qa_collate(samples, pad_id=0):\n if len(samples) == 0:\n return {}\n\n batch = {\n 'input_ids': collate_tokens([s[\"encodings\"]['input_ids'] for s in samples], pad_id),\n 'attention_mask': collate_tokens([s[\"encodings\"]['attention_mask'] for s in samples], 0),\n 'paragraph_mask': collate_tokens([s['paragraph_mask'] for s in samples], 0),\n 'label': collate_tokens([s[\"label\"] for s in samples], -1),\n \"sent_offsets\": collate_tokens([s[\"sent_offsets\"] for s in samples], 0),\n }\n\n # training labels\n if \"starts\" in samples[0]:\n batch[\"starts\"] = collate_tokens([s['starts'] for s in samples], -1)\n batch[\"ends\"] = collate_tokens([s['ends'] for s in samples], -1)\n # batch[\"ans_types\"] = collate_tokens([s['ans_type'] for s in samples], -1)\n batch[\"sent_labels\"] = collate_tokens([s['sent_labels'] for s in samples], 0)\n batch[\"ans_covered\"] = collate_tokens([s['ans_covered'] for s in samples], 0)\n\n # roberta does not use token_type_ids\n if \"token_type_ids\" in samples[0][\"encodings\"]:\n batch[\"token_type_ids\"] = collate_tokens([s[\"encodings\"]['token_type_ids']for s in samples], 0)\n\n batched = {\n \"qids\": [s[\"qid\"] for s in samples],\n \"passages\": [s[\"passages\"] for s in samples],\n \"gold_answer\": [s[\"gold_answer\"] for s in samples],\n \"sp_gold\": [s[\"sp_gold\"] for s in samples],\n \"para_offsets\": [s[\"para_offset\"] for s in samples],\n \"net_inputs\": batch,\n }\n\n # for answer extraction\n if \"doc_tokens\" in samples[0]:\n batched[\"doc_tokens\"] = [s[\"doc_tokens\"] for s in samples]\n batched[\"tok_to_orig_index\"] = [s[\"tok_to_orig_index\"] for s in samples]\n batched[\"wp_tokens\"] = [s[\"wp_tokens\"] for s in samples]\n\n return batched\n" }, { "path": "mdr/qa/qa_model.py", "content": "# Copyright (c) Facebook, Inc. and its affiliates.\n# All rights reserved.\n#\n# This source code is licensed under the license found in the \n# LICENSE file in the root directory of this source tree.\n\nfrom transformers import AutoModel, BertModel\nimport torch.nn as nn\nfrom torch.nn import CrossEntropyLoss\nimport torch\nimport torch.nn.functional as F\n\nclass BertPooler(nn.Module):\n def __init__(self, config):\n super().__init__()\n self.dense = nn.Linear(config.hidden_size, config.hidden_size)\n self.activation = nn.Tanh()\n\n def forward(self, hidden_states):\n # We \"pool\" the model by simply taking the hidden state corresponding\n # to the first token.\n first_token_tensor = hidden_states[:, 0]\n pooled_output = self.dense(first_token_tensor)\n pooled_output = self.activation(pooled_output)\n return pooled_output\n\nclass QAModel(nn.Module):\n\n def __init__(self,\n config,\n args\n ):\n super().__init__()\n self.model_name = args.model_name\n self.sp_weight = args.sp_weight\n self.sp_pred = args.sp_pred\n self.encoder = AutoModel.from_pretrained(args.model_name)\n\n if \"electra\" in args.model_name:\n self.pooler = BertPooler(config)\n\n self.qa_outputs = nn.Linear(config.hidden_size, 2)\n self.rank = nn.Linear(config.hidden_size, 1) # noan\n\n if self.sp_pred:\n self.sp = nn.Linear(config.hidden_size, 1)\n self.loss_fct = CrossEntropyLoss(ignore_index=-1, reduction=\"none\")\n\n def forward(self, batch):\n\n outputs = self.encoder(batch['input_ids'], batch['attention_mask'], batch.get('token_type_ids', None))\n\n if \"electra\" in self.model_name:\n sequence_output = outputs[0]\n pooled_output = self.pooler(sequence_output)\n else:\n sequence_output, pooled_output = outputs[0], outputs[1]\n\n logits = self.qa_outputs(sequence_output)\n outs = [o.squeeze(-1) for o in logits.split(1, dim=-1)]\n outs = [o.float().masked_fill(batch[\"paragraph_mask\"].ne(1), float(\"-inf\")).type_as(o) for o in outs]\n\n start_logits, end_logits = outs[0], outs[1]\n rank_score = self.rank(pooled_output)\n\n if self.sp_pred:\n gather_index = batch[\"sent_offsets\"].unsqueeze(2).expand(-1, -1, sequence_output.size()[-1])\n sent_marker_rep = torch.gather(sequence_output, 1, gather_index)\n sp_score = self.sp(sent_marker_rep).squeeze(2)\n else:\n sp_score = None\n\n if self.training:\n\n rank_target = batch[\"label\"]\n if self.sp_pred:\n sp_loss = F.binary_cross_entropy_with_logits(sp_score, batch[\"sent_labels\"].float(), reduction=\"none\")\n sp_loss = (sp_loss * batch[\"sent_offsets\"]) * batch[\"label\"]\n sp_loss = sp_loss.sum()\n\n start_positions, end_positions = batch[\"starts\"], batch[\"ends\"]\n\n rank_loss = F.binary_cross_entropy_with_logits(rank_score, rank_target.float(), reduction=\"sum\")\n\n start_losses = [self.loss_fct(start_logits, starts) for starts in torch.unbind(start_positions, dim=1)]\n end_losses = [self.loss_fct(end_logits, ends) for ends in torch.unbind(end_positions, dim=1)]\n loss_tensor = torch.cat([t.unsqueeze(1) for t in start_losses], dim=1) + torch.cat([t.unsqueeze(1) for t in end_losses], dim=1)\n log_prob = - loss_tensor\n log_prob = log_prob.float().masked_fill(log_prob == 0, float('-inf')).type_as(log_prob)\n marginal_probs = torch.sum(torch.exp(log_prob), dim=1)\n m_prob = [marginal_probs[idx] for idx in marginal_probs.nonzero()]\n if len(m_prob) == 0:\n span_loss = self.loss_fct(start_logits, start_logits.new_zeros(\n start_logits.size(0)).long()-1).sum()\n else:\n span_loss = - torch.log(torch.cat(m_prob)).sum()\n\n if self.sp_pred:\n loss = rank_loss + span_loss + sp_loss * self.sp_weight\n else:\n loss = rank_loss + span_loss\n return loss.unsqueeze(0)\n\n return {\n 'start_logits': start_logits,\n 'end_logits': end_logits,\n 'rank_score': rank_score,\n \"sp_score\": sp_score\n }\n" }, { "path": "mdr/qa/qa_trainer.py", "content": "# Copyright (c) Facebook, Inc. and its affiliates.\n# All rights reserved.\n#\n# This source code is licensed under the license found in the \n# LICENSE file in the root directory of this source tree.\nimport argparse\nimport json\nimport os\nimport os.path as osp\nimport random\nfrom functools import partial\nfrom pathlib import Path\nfrom typing import NamedTuple, Optional\nimport collections\nfrom torch.optim import lr_scheduler\nfrom tqdm import tqdm\n\nimport apex\nimport attr\nimport numpy as np\nimport submitit\nimport torch\nimport torch.distributed\nimport torch.nn as nn\nimport torch.optim as optim\nfrom apex import amp\nfrom torch.utils.tensorboard import SummaryWriter\nfrom transformers import (AdamW, AutoConfig, AutoTokenizer,\n get_linear_schedule_with_warmup)\n\nfrom config import ClusterConfig\nfrom hotpot_evaluate_v1 import exact_match_score, f1_score, update_sp\nfrom qa_model import QAModel\nfrom reranking_datasets import RankingDataset, rank_collate, MhopSampler\nfrom utils import AverageMeter, move_to_cuda, get_final_text\n\napex.amp.register_half_function(torch, 'einsum')\n\n@attr.s(auto_attribs=True)\nclass TrainerState:\n \"\"\"\n Contains the state of the Trainer.\n It can be saved to checkpoint the training and loaded to resume it.\n \"\"\"\n\n epoch: int\n model: nn.Module\n optimizer: optim.Optimizer\n lr_scheduler: torch.optim.lr_scheduler._LRScheduler\n global_step: int\n\n def save(self, filename: str) -> None:\n data = attr.asdict(self)\n # store only the state dict\n data[\"model\"] = self.model.state_dict()\n data[\"optimizer\"] = self.optimizer.state_dict()\n data[\"lr_scheduler\"] = self.lr_scheduler.state_dict()\n torch.save(data, filename)\n\n @classmethod\n def load(cls, filename: str, default: \"TrainerState\", gpu: int) -> \"TrainerState\":\n data = torch.load(filename, map_location=lambda storage, loc: storage.cuda(gpu))\n # We need this default to load the state dict\n model = default.model\n model.load_state_dict(data[\"model\"])\n data[\"model\"] = model\n\n optimizer = default.optimizer\n optimizer.load_state_dict(data[\"optimizer\"])\n data[\"optimizer\"] = optimizer\n\n lr_scheduler = default.lr_scheduler\n lr_scheduler.load_state_dict(data[\"lr_scheduler\"])\n data[\"lr_scheduler\"] = lr_scheduler\n\n return cls(**data)\n\nclass Trainer:\n def __init__(self, train_cfg: NamedTuple, cluster_cfg: ClusterConfig) -> None:\n self._train_cfg = train_cfg\n self._cluster_cfg = cluster_cfg\n\n def __call__(self) -> Optional[float]:\n \"\"\"\n Called by submitit for each task.\n :return: The master task return the final accuracy of the model.\n \"\"\"\n self._setup_process_group()\n self._init_state()\n final_acc = self._train()\n return final_acc\n\n def log(self, log_data: dict):\n job_env = submitit.JobEnvironment()\n # z = {**vars(self._train_cfg), **log_data}\n save_dir = Path(self._train_cfg.output_dir)\n os.makedirs(save_dir, exist_ok=True)\n with open(save_dir / 'log.txt', 'a') as f:\n f.write(json.dumps(log_data) + '\\n')\n\n def checkpoint(self, rm_init=True) -> submitit.helpers.DelayedSubmission:\n # will be called by submitit in case of preemption\n job_env = submitit.JobEnvironment()\n save_dir = osp.join(self._train_cfg.output_dir, str(job_env.job_id))\n os.makedirs(save_dir, exist_ok=True)\n self._state.save(osp.join(save_dir, \"checkpoint.pth\"))\n\n # Trick here: when the job will be requeue, we will use the same init file\n # but it must not exist when we initialize the process group\n # so we delete it, but only when this method is called by submitit for requeue\n if rm_init and osp.exists(self._cluster_cfg.dist_url[7:]):\n os.remove(self._cluster_cfg.dist_url[7:]) # remove file:// at the beginning\n # This allow to remove any non-pickable part of the Trainer instance.\n empty_trainer = Trainer(self._train_cfg, self._cluster_cfg)\n return submitit.helpers.DelayedSubmission(empty_trainer)\n\n def _setup_process_group(self) -> None:\n job_env = submitit.JobEnvironment()\n torch.cuda.set_device(job_env.local_rank)\n torch.distributed.init_process_group(\n backend=self._cluster_cfg.dist_backend,\n init_method=self._cluster_cfg.dist_url,\n world_size=job_env.num_tasks,\n rank=job_env.global_rank,\n )\n print(f\"Process group: {job_env.num_tasks} tasks, rank: {job_env.global_rank}\")\n\n def _init_state(self) -> None:\n \"\"\"\n Initialize the state and load it from an existing checkpoint if any\n \"\"\"\n job_env = submitit.JobEnvironment()\n\n if job_env.global_rank == 0:\n # config_path = Path(args.save_folder) / str(job_env.job_id) / 'config.json'\n os.makedirs(self._train_cfg.output_dir, exist_ok=True)\n config_path = Path(self._train_cfg.output_dir) / 'config.json'\n with open(config_path, \"w\") as g:\n g.write(json.dumps(self._train_cfg._asdict()))\n\n print(f\"Setting random seed {self._train_cfg.seed}\", flush=True)\n random.seed(self._train_cfg.seed)\n np.random.seed(self._train_cfg.seed)\n torch.manual_seed(self._train_cfg.seed)\n\n print(\"Create data loaders\", flush=True)\n tokenizer = AutoTokenizer.from_pretrained(self._train_cfg.model_name)\n collate_fc = partial(rank_collate, pad_id=tokenizer.pad_token_id)\n train_set = RankingDataset(tokenizer, self._train_cfg.train_file, self._train_cfg.max_seq_len, self._train_cfg.max_q_len, train=True)\n\n train_sampler = MhopSampler(train_set, num_neg=self._train_cfg.neg_num)\n\n batch_size_per_gpu = (1 + self._train_cfg.neg_num) * self._train_cfg.num_q_per_gpu\n n_gpu = torch.cuda.device_count()\n print(f\"Number of GPUs: {n_gpu}\", flush=True)\n print(f\"Batch size per node: {batch_size_per_gpu * n_gpu}\", flush=True)\n\n self._train_loader = torch.utils.data.DataLoader(train_set, batch_size=batch_size_per_gpu * n_gpu, num_workers=self._train_cfg.num_workers, collate_fn=collate_fc, sampler=train_sampler)\n test_set = RankingDataset(tokenizer, self._train_cfg.predict_file, self._train_cfg.max_seq_len, self._train_cfg.max_q_len)\n self._test_loader = torch.utils.data.DataLoader(\n test_set,\n batch_size=self._train_cfg.predict_batch_size,\n num_workers=self._train_cfg.num_workers, collate_fn=collate_fc\n )\n\n print(\"Create model\", flush=True)\n print(f\"Local rank {job_env.local_rank}\", flush=True)\n bert_config = AutoConfig.from_pretrained(self._train_cfg.model_name)\n model = QAModel(bert_config, self._train_cfg)\n model.cuda(job_env.local_rank)\n\n no_decay = ['bias', 'LayerNorm.weight']\n optimizer_parameters = [\n {'params': [p for n, p in model.named_parameters() if not any(\n nd in n for nd in no_decay)], 'weight_decay': self._train_cfg.weight_decay},\n {'params': [p for n, p in model.named_parameters() if any(\n nd in n for nd in no_decay)], 'weight_decay': 0.0}\n ]\n\n if self._train_cfg.use_adam:\n optimizer = optim.Adam(optimizer_parameters, lr=self._train_cfg.learning_rate)\n else:\n optimizer = AdamW(optimizer_parameters, lr=self._train_cfg.learning_rate)\n # lr_scheduler = optim.lr_scheduler.ReduceLROnPlateau(optimizer, mode='max', factor=0.5, patience=2)\n\n if self._train_cfg.fp16:\n model, optimizer = amp.initialize(\n model, optimizer, opt_level=self._train_cfg.fp16_opt_level)\n\n t_total = len(self._train_loader) // self._train_cfg.gradient_accumulation_steps * self._train_cfg.num_train_epochs\n warmup_steps = t_total * self._train_cfg.warmup_ratio\n lr_scheduler = get_linear_schedule_with_warmup(\n optimizer, num_warmup_steps=warmup_steps, num_training_steps=t_total\n )\n\n model = torch.nn.DataParallel(model)\n self._state = TrainerState(\n epoch=0, model=model, optimizer=optimizer, lr_scheduler=lr_scheduler, global_step=0\n )\n self.tb_logger = SummaryWriter(self._train_cfg.output_dir.replace(\"logs\", \"tflogs\"))\n\n checkpoint_fn = osp.join(self._train_cfg.output_dir, str(job_env.job_id), \"checkpoint.pth\")\n # checkpoint_fn = osp.join(self._train_cfg.output_dir, \"checkpoint.pth\")\n if os.path.isfile(checkpoint_fn):\n print(f\"Load existing checkpoint from {checkpoint_fn}\", flush=True)\n self._state = TrainerState.load(\n checkpoint_fn, default=self._state, gpu=job_env.local_rank)\n\n def _train(self) -> Optional[float]:\n job_env = submitit.JobEnvironment()\n batch_step = 0 # forward batch count\n best_metric = 0\n train_loss_meter = AverageMeter()\n print(f\"Start training\", flush=True)\n # Start from the loaded epoch\n start_epoch = self._state.epoch\n global_step = self._state.global_step\n for epoch in range(start_epoch, self._train_cfg.num_train_epochs):\n print(f\"Start epoch {epoch}\", flush=True)\n self._state.model.train()\n self._state.epoch = epoch\n\n for batch in self._train_loader:\n batch_step += 1\n batch_inputs = move_to_cuda(batch[\"net_inputs\"])\n loss = self._state.model(batch_inputs)\n if torch.cuda.device_count() > 1:\n loss = loss.mean()\n if self._train_cfg.gradient_accumulation_steps > 1:\n loss = loss / self._train_cfg.gradient_accumulation_steps\n if self._train_cfg.fp16:\n with amp.scale_loss(loss, self._state.optimizer) as scaled_loss:\n scaled_loss.backward()\n else:\n loss.backward()\n train_loss_meter.update(loss.item())\n if (batch_step + 1) % self._train_cfg.gradient_accumulation_steps == 0:\n if self._train_cfg.fp16:\n torch.nn.utils.clip_grad_norm_(\n amp.master_params(self._state.optimizer), self._train_cfg.max_grad_norm)\n else:\n torch.nn.utils.clip_grad_norm_(\n self._state.model.parameters(), self._train_cfg.max_grad_norm)\n self._state.optimizer.step()\n self._state.lr_scheduler.step()\n self._state.model.zero_grad()\n global_step += 1\n self._state.global_step = global_step\n\n self.tb_logger.add_scalar('batch_train_loss',\n loss.item(), global_step)\n self.tb_logger.add_scalar('smoothed_train_loss',\n train_loss_meter.avg, global_step)\n if job_env.global_rank == 0:\n if self._train_cfg.eval_period != -1 and global_step % self._train_cfg.eval_period == 0:\n metrics = self._eval()\n for k, v in metrics.items():\n self.tb_logger.add_scalar(k, v*100, global_step)\n score = metrics[self._train_cfg.final_metric]\n if best_metric < score:\n print(\"Saving model with best %s %.2f -> em %.2f\" % (self._train_cfg.final_metric, best_metric*100, score*100), flush=True)\n torch.save(self._state.model.state_dict(), os.path.join(self._train_cfg.output_dir, f\"checkpoint_best.pt\"))\n best_metric = score\n # Checkpoint only on the master\n if job_env.global_rank == 0:\n self.checkpoint(rm_init=False)\n metrics = self._eval()\n for k, v in metrics.items():\n self.tb_logger.add_scalar(k, v*100, global_step)\n score = metrics[self._train_cfg.final_metric]\n if best_metric < score:\n print(\"Saving model with best %s %.2f -> em %.2f\" % (self._train_cfg.final_metric, best_metric*100, score*100), flush=True)\n torch.save(self._state.model.state_dict(), os.path.join(self._train_cfg.output_dir, f\"checkpoint_best.pt\"))\n best_metric = score\n self.log({\n \"best_score\": best_metric,\n \"curr_score\": score,\n \"smoothed_loss\": train_loss_meter.avg,\n \"epoch\": epoch\n })\n return best_metric\n\n def _eval(self) -> dict:\n print(\"Start evaluation of the model\", flush=True)\n job_env = submitit.JobEnvironment()\n args = self._train_cfg\n eval_dataloader = self._test_loader\n model = self._state.model\n model.eval()\n id2result = collections.defaultdict(list)\n id2answer = collections.defaultdict(list)\n id2gold = {}\n id2goldsp = {}\n for batch in tqdm(eval_dataloader):\n batch_to_feed = move_to_cuda(batch[\"net_inputs\"])\n batch_qids = batch[\"qids\"]\n batch_labels = batch[\"net_inputs\"][\"label\"].view(-1).tolist()\n with torch.no_grad():\n outputs = model(batch_to_feed)\n scores = outputs[\"rank_score\"]\n scores = scores.view(-1).tolist()\n sp_scores = outputs[\"sp_score\"]\n sp_scores = sp_scores.float().masked_fill(batch_to_feed[\"sent_offsets\"].eq(0), float(\"-inf\")).type_as(sp_scores)\n batch_sp_scores = sp_scores.sigmoid()\n # ans_type_predicted = torch.argmax(outputs[\"ans_type_logits\"], dim=1).view(-1).tolist()\n outs = [outputs[\"start_logits\"], outputs[\"end_logits\"]]\n for qid, label, score in zip(batch_qids, batch_labels, scores):\n id2result[qid].append((label, score))\n\n # answer prediction\n span_scores = outs[0][:, :, None] + outs[1][:, None]\n max_seq_len = span_scores.size(1)\n span_mask = np.tril(np.triu(np.ones((max_seq_len, max_seq_len)), 0), args.max_ans_len)\n span_mask = span_scores.data.new(max_seq_len, max_seq_len).copy_(torch.from_numpy(span_mask))\n span_scores_masked = span_scores.float().masked_fill((1 - span_mask[None].expand_as(span_scores)).bool(), -1e10).type_as(span_scores)\n start_position = span_scores_masked.max(dim=2)[0].max(dim=1)[1]\n end_position = span_scores_masked.max(dim=2)[1].gather(\n 1, start_position.unsqueeze(1)).squeeze(1)\n answer_scores = span_scores_masked.max(dim=2)[0].max(dim=1)[0].tolist()\n para_offset = batch['para_offsets']\n start_position_ = list(\n np.array(start_position.tolist()) - np.array(para_offset))\n end_position_ = list(\n np.array(end_position.tolist()) - np.array(para_offset)) \n \n for idx, qid in enumerate(batch_qids):\n id2gold[qid] = batch[\"gold_answer\"][idx]\n id2goldsp[qid] = batch[\"sp_gold\"][idx]\n rank_score = scores[idx]\n sp_score = batch_sp_scores[idx].tolist()\n start = start_position_[idx]\n end = end_position_[idx]\n span_score = answer_scores[idx]\n\n tok_to_orig_index = batch['tok_to_orig_index'][idx]\n doc_tokens = batch['doc_tokens'][idx]\n wp_tokens = batch['wp_tokens'][idx]\n orig_doc_start = tok_to_orig_index[start]\n orig_doc_end = tok_to_orig_index[end]\n orig_tokens = doc_tokens[orig_doc_start:(orig_doc_end + 1)]\n tok_tokens = wp_tokens[start:end+1]\n tok_text = \" \".join(tok_tokens)\n tok_text = tok_text.replace(\" ##\", \"\")\n tok_text = tok_text.replace(\"##\", \"\")\n tok_text = tok_text.strip()\n tok_text = \" \".join(tok_text.split())\n orig_text = \" \".join(orig_tokens)\n pred_str = get_final_text(tok_text, orig_text, do_lower_case=True, verbose_logging=False)\n\n pred_sp = []\n passages = batch[\"passages\"][idx]\n for passage, sent_offset in zip(passages, [0, len(passages[0][\"sents\"])]):\n for idx, _ in enumerate(passage[\"sents\"]):\n try:\n if sp_score[idx + sent_offset] > 0.5:\n pred_sp.append([passage[\"title\"], idx])\n except:\n continue\n id2answer[qid].append((pred_str.strip(), rank_score, span_score, pred_sp))\n\n acc = []\n for qid, res in id2result.items():\n res.sort(key=lambda x: x[1], reverse=True)\n acc.append(res[0][0] == 1)\n print(f\"evaluated {len(id2result)} questions...\", flush=True)\n print(f'chain ranking em: {np.mean(acc)}', flush=True)\n\n best_em, best_f1, best_joint_em, best_joint_f1 = 0, 0, 0, 0\n lambdas = [0, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1]\n for lambda_ in lambdas:\n ems, f1s = [], []\n sp_ems, sp_f1s = [], []\n joint_ems, joint_f1s = [], []\n for qid, res in id2result.items():\n ans_res = id2answer[qid]\n ans_res.sort(key=lambda x: lambda_ * x[1] + (1 - lambda_) * x[2], reverse=True)\n top_pred = ans_res[0][0]\n ems.append(exact_match_score(top_pred, id2gold[qid][0]))\n f1, prec, recall = f1_score(top_pred, id2gold[qid][0])\n f1s.append(f1)\n\n top_pred_sp = ans_res[0][3]\n metrics = {'sp_em': 0, 'sp_f1': 0, 'sp_prec': 0, 'sp_recall': 0}\n update_sp(metrics, top_pred_sp, id2goldsp[qid])\n sp_ems.append(metrics['sp_em'])\n sp_f1s.append(metrics['sp_f1'])\n\n # joint metrics\n joint_prec = prec * metrics[\"sp_prec\"]\n joint_recall = recall * metrics[\"sp_recall\"]\n if joint_prec + joint_recall > 0:\n joint_f1 = 2 * joint_prec * joint_recall / (joint_prec + joint_recall)\n else:\n joint_f1 = 0\n joint_em = ems[-1] * sp_ems[-1]\n joint_ems.append(joint_em)\n joint_f1s.append(joint_f1)\n\n if best_joint_f1 < np.mean(joint_f1s):\n best_joint_f1 = np.mean(joint_f1s)\n best_joint_em = np.mean(joint_ems)\n best_f1 = np.mean(f1s)\n best_em = np.mean(ems)\n\n print(f\".......Using combination factor {lambda_}......\", flush=True)\n print(f'answer em: {np.mean(ems)}, count: {len(ems)}', flush=True)\n print(f'answer f1: {np.mean(f1s)}, count: {len(f1s)}', flush=True)\n print(f'sp em: {np.mean(sp_ems)}, count: {len(sp_ems)}', flush=True)\n print(f'sp f1: {np.mean(sp_f1s)}, count: {len(sp_f1s)}', flush=True)\n print(f'joint em: {np.mean(joint_ems)}, count: {len(joint_ems)}', flush=True)\n print(f'joint f1: {np.mean(joint_f1s)}, count: {len(joint_f1s)}', flush=True)\n print(f\"Best joint EM/F1 from combination {best_em}/{best_f1}\", flush=True)\n\n model.train()\n return {\"em\": best_em, \"f1\": best_f1, \"joint_em\": best_joint_em, \"joint_f1\": best_joint_f1}\n\n" }, { "path": "mdr/qa/train.md", "content": "\n\nCUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python train_qa.py \\\n --do_train \\\n --prefix qa_wwm_bert_title_mark_eval_debug \\\n --predict_batch_size 512 \\\n --model_name bert-large-uncased-whole-word-masking \\\n --train_batch_size 80 \\\n --learning_rate 3e-5 \\\n --fp16 \\\n --train_file /private/home/xwhan/data/hotpot/dense_train_b10_top20_outputs.json \\\n --predict_file /private/home/xwhan/data/hotpot/dense_val_outputs.json \\\n --seed 3 \\\n --eval-period 10 \\\n --max_seq_len 512 \\\n --max_q_len 100 \\\n --gradient_accumulation_steps 8 \\\n --neg-num 4\n\n\n# spanbert debug, fp16 does not work\nCUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python train_qa.py \\\n --do_train \\\n --prefix ranked_spanbert_debug \\\n --predict_batch_size 1024 \\\n --model_name spanbert \\\n --train_batch_size 48 \\\n --learning_rate 3e-5 \\\n --train_file /private/home/xwhan/data/hotpot/dense_train_b10_top20_outputs_sents.json \\\n --predict_file /private/home/xwhan/data/hotpot/dense_val_outputs_sents.json \\\n --seed 3 \\\n --eval-period 500 \\\n --max_seq_len 512 \\\n --max_q_len 64 \\\n --gradient_accumulation_steps 8 \\\n --neg-num 5 \\\n --use-adam\n\n# test electra\nCUDA_VISIBLE_DEVICES=0 python train_qa.py \\\n --do_train \\\n --prefix electra_large_debug_sn \\\n --predict_batch_size 1024 \\\n --model_name google/electra-large-discriminator \\\n --train_batch_size 12 \\\n --learning_rate 5e-5 \\\n --train_file /private/home/xwhan/data/hotpot/dense_train_b100_k100_sents.json \\\n --predict_file /private/home/xwhan/data/hotpot/dense_val_b30_k30_roberta_sents.json \\\n --seed 42 \\\n --eval-period 250 \\\n --max_seq_len 512 \\\n --max_q_len 64 \\\n --gradient_accumulation_steps 8 \\\n --neg-num 11 \\\n --fp16 \\\n --use-adam \\\n --warmup-ratio 0.1 \\\n --sp-weight 0.05 \\\n --sp-pred \\\n --shared-norm\n\n\n# QA evaluation\nCUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python train_qa.py \\\n --do_predict \\\n --predict_batch_size 2000 \\\n --model_name google/electra-large-discriminator \\\n --fp16 \\\n --predict_file /private/home/xwhan/data/hotpot/dense_val_b100_k100_roberta_best_sents.json \\\n --max_seq_len 512 \\\n --max_q_len 64 \\\n --init_checkpoint qa/logs/08-10-2020/electra_val_top30-epoch7-lr5e-05-seed42-rdrop0-qadrop0-decay0-qpergpu2-aggstep8-clip2-evalper250-evalbsize1024-negnum5-warmup0.1-adamTrue-spweight0.025/checkpoint_best.pt \\\n --sp-pred \\\n --max_ans_len 30 \\\n --save-prediction hotpot_val_top100.json\n\n# QA evaluation with wwm\n\nCUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python train_qa.py \\\n --do_predict \\\n --predict_batch_size 1024 \\\n --model_name bert-large-uncased-whole-word-masking \\\n --fp16 \\\n --predict_file /private/home/xwhan/data/hotpot/dense_hotpot_val_b250_k250_roberta_best_sents.json \\\n --max_seq_len 512 \\\n --max_q_len 64 \\\n --init_checkpoint qa/logs/08-17-2020/wwm_val_top50-epoch7-lr5e-05-seed42-rdrop0-qadrop0-decay0-qpergpu2-aggstep8-clip2-evalper250-evalbsize1024-negnum5-warmup0.2-adamTrue-spweight0.025-snFalse/checkpoint_best.pt \\\n --sp-pred \\\n --max_ans_len 30 \\\n --save-prediction hotpot_val_wwm_top250.json\n\n\nCUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python train_qa.py \\\n --do_predict \\\n --predict_batch_size 1024 \\\n --model_name google/electra-large-discriminator \\\n --fp16 \\\n --predict_file /private/home/xwhan/data/hotpot/dense_val_b50_k50_roberta_best_sents.json \\\n --max_seq_len 512 \\\n --max_q_len 64 \\\n --init_checkpoint qa/logs/08-10-2020/electra_val_top30-epoch7-lr5e-05-seed42-rdrop0-qadrop0-decay0-qpergpu2-aggstep8-clip2-evalper250-evalbsize1024-negnum5-warmup0.1-adamTrue-spweight0.025/checkpoint_best.pt \\\n --sp-pred \\\n --max_ans_len 30 \\\n --save-prediction hotpot_val_b5_k5.json \\\n\nsrun --gres=gpu:8 --partition learnfair --time=48:00:00 --mem 500G --constraint volta32gb --cpus-per-task 80 --pty /bin/bash -l\n" }, { "path": "mdr/qa/train_ranker.py", "content": "# Copyright (c) Facebook, Inc. and its affiliates.\n# All rights reserved.\n#\n# This source code is licensed under the license found in the \n# LICENSE file in the root directory of this source tree.\nimport collections\nimport json\nimport logging\nimport os\nimport random\nfrom datetime import date\nfrom functools import partial\nimport copy\n\nimport numpy as np\nimport torch\nimport torch.nn.functional as F\nfrom torch.utils.data import DataLoader\nfrom torch.utils.tensorboard import SummaryWriter\nfrom torch.optim import Adam\nfrom tqdm import tqdm\nfrom transformers import (AdamW, AutoConfig, AutoTokenizer,\n get_linear_schedule_with_warmup)\n\nfrom config import train_args\nfrom reranking_datasets import RankingDataset, rank_collate\nfrom reranking_model import RankModel\nfrom utils import AverageMeter, convert_to_half, move_to_cuda\n\ndef load_saved(model, path):\n state_dict = torch.load(path)\n def filter(x): return x[7:] if x.startswith('module.') else x\n state_dict = {filter(k): v for (k, v) in state_dict.items()}\n model.load_state_dict(state_dict)\n return model\n\ndef main():\n args = train_args()\n if args.fp16:\n import apex\n apex.amp.register_half_function(torch, 'einsum')\n date_curr = date.today().strftime(\"%m-%d-%Y\")\n model_name = f\"{args.prefix}-seed{args.seed}-bsz{args.train_batch_size}-fp16{args.fp16}-lr{args.learning_rate}-decay{args.weight_decay}\"\n args.output_dir = os.path.join(args.output_dir, date_curr, model_name)\n tb_logger = SummaryWriter(os.path.join(args.output_dir.replace(\"logs\",\"tflogs\")))\n\n if os.path.exists(args.output_dir) and os.listdir(args.output_dir):\n print(\n f\"output directory {args.output_dir} already exists and is not empty.\")\n if not os.path.exists(args.output_dir):\n os.makedirs(args.output_dir, exist_ok=True)\n\n logging.basicConfig(format='%(asctime)s - %(levelname)s - %(name)s - %(message)s', datefmt='%m/%d/%Y %H:%M:%S',\n level=logging.INFO,\n handlers=[logging.FileHandler(os.path.join(args.output_dir, \"log.txt\")),\n logging.StreamHandler()])\n logger = logging.getLogger(__name__)\n logger.info(args)\n\n if args.local_rank == -1 or args.no_cuda:\n device = torch.device(\n \"cuda\" if torch.cuda.is_available() and not args.no_cuda else \"cpu\")\n n_gpu = torch.cuda.device_count()\n else:\n device = torch.device(\"cuda\", args.local_rank)\n n_gpu = 1\n # Initializes the distributed backend which will take care of sychronizing nodes/GPUs\n torch.distributed.init_process_group(backend='nccl')\n logger.info(\"device %s n_gpu %d distributed training %r\",\n device, n_gpu, bool(args.local_rank != -1))\n\n if args.accumulate_gradients < 1:\n raise ValueError(\"Invalid accumulate_gradients parameter: {}, should be >= 1\".format(\n args.accumulate_gradients))\n\n args.train_batch_size = int(\n args.train_batch_size / args.accumulate_gradients)\n random.seed(args.seed)\n np.random.seed(args.seed)\n torch.manual_seed(args.seed)\n if n_gpu > 0:\n torch.cuda.manual_seed_all(args.seed)\n\n bert_config = AutoConfig.from_pretrained(args.model_name)\n model = RankModel(bert_config, args)\n tokenizer = AutoTokenizer.from_pretrained(args.model_name)\n collate_fc = partial(rank_collate, pad_id=tokenizer.pad_token_id)\n if args.do_train and args.max_seq_len > bert_config.max_position_embeddings:\n raise ValueError(\n \"Cannot use sequence length %d because the BERT model \"\n \"was only trained up to sequence length %d\" %\n (args.max_seq_len, bert_config.max_position_embeddings))\n\n eval_dataset = RankingDataset(\n tokenizer, args.predict_file, args.max_seq_len, args.max_q_len)\n eval_dataloader = DataLoader(\n eval_dataset, batch_size=args.predict_batch_size, collate_fn=collate_fc, pin_memory=True, num_workers=args.num_workers)\n logger.info(f\"Num of dev batches: {len(eval_dataloader)}\")\n\n if args.init_checkpoint != \"\":\n model = load_saved(model, args.init_checkpoint)\n\n model.to(device)\n print(f\"number of trainable parameters: {sum(p.numel() for p in model.parameters() if p.requires_grad)}\")\n\n if args.do_train:\n no_decay = ['bias', 'LayerNorm.weight']\n optimizer_parameters = [\n {'params': [p for n, p in model.named_parameters() if not any(\n nd in n for nd in no_decay)], 'weight_decay': args.weight_decay},\n {'params': [p for n, p in model.named_parameters() if any(\n nd in n for nd in no_decay)], 'weight_decay': 0.0}\n ]\n optimizer = AdamW(optimizer_parameters,\n lr=args.learning_rate, eps=args.adam_epsilon)\n\n if args.fp16:\n from apex import amp\n model, optimizer = amp.initialize(\n model, optimizer, opt_level=args.fp16_opt_level)\n else:\n if args.fp16:\n from apex import amp\n model = amp.initialize(model, opt_level=args.fp16_opt_level)\n\n if args.local_rank != -1:\n model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[args.local_rank],\n output_device=args.local_rank)\n elif n_gpu > 1:\n model = torch.nn.DataParallel(model)\n\n if args.do_train:\n global_step = 0 # gradient update step\n batch_step = 0 # forward batch count\n best_acc = 0\n train_loss_meter = AverageMeter()\n model.train()\n train_dataset = RankingDataset(tokenizer, args.train_file, args.max_seq_len, args.max_q_len, train=True)\n train_dataloader = DataLoader(train_dataset, batch_size=args.train_batch_size, pin_memory=True, collate_fn=collate_fc, num_workers=args.num_workers, shuffle=True)\n\n logger.info('Start training....')\n for epoch in range(int(args.num_train_epochs)):\n for batch in tqdm(train_dataloader):\n batch_step += 1\n batch_inputs = move_to_cuda(batch[\"net_inputs\"])\n loss = model(batch_inputs)\n\n if n_gpu > 1:\n loss = loss.mean()\n\n if args.gradient_accumulation_steps > 1:\n loss = loss / args.gradient_accumulation_steps\n\n if args.fp16:\n with amp.scale_loss(loss, optimizer) as scaled_loss:\n scaled_loss.backward()\n else:\n loss.backward()\n\n train_loss_meter.update(loss.item())\n tb_logger.add_scalar('batch_train_loss',\n loss.item(), global_step)\n tb_logger.add_scalar('smoothed_train_loss',\n train_loss_meter.avg, global_step)\n \n if (batch_step + 1) % args.gradient_accumulation_steps == 0:\n if args.fp16:\n torch.nn.utils.clip_grad_norm_(\n amp.master_params(optimizer), args.max_grad_norm)\n else:\n torch.nn.utils.clip_grad_norm_(\n model.parameters(), args.max_grad_norm)\n optimizer.step() # We have accumulated enought gradients\n model.zero_grad()\n global_step += 1\n\n if args.eval_period != -1 and global_step % args.eval_period == 0:\n acc = predict(args, model, eval_dataloader,\n device, logger)\n logger.info(\"Step %d Train loss %.2f acc %.2f on epoch=%d\" % (global_step, train_loss_meter.avg, acc*100, epoch))\n\n # save most recent model\n torch.save(model.state_dict(), os.path.join(\n args.output_dir, f\"checkpoint_last.pt\"))\n\n if best_acc < acc:\n logger.info(\"Saving model with best acc %.2f -> acc %.2f on epoch=%d\" %\n (best_acc*100, acc*100, epoch))\n torch.save(model.state_dict(), os.path.join(\n args.output_dir, f\"checkpoint_best.pt\"))\n model = model.to(device)\n best_acc = acc\n\n acc = predict(args, model, eval_dataloader, device, logger)\n logger.info(\"Step %d Train loss %.2f acc %.2f on epoch=%d\" % (\n global_step, train_loss_meter.avg, acc*100, epoch))\n tb_logger.add_scalar('dev_acc', acc*100, epoch)\n torch.save(model.state_dict(), os.path.join(args.output_dir, f\"checkpoint_last.pt\"))\n\n if best_acc < acc:\n logger.info(\"Saving model with best acc %.2f -> acc %.2f on epoch=%d\" % (best_acc*100, acc*100, epoch))\n torch.save(model.state_dict(), os.path.join(\n args.output_dir, f\"checkpoint_best.pt\"))\n best_acc = acc\n\n logger.info(\"Training finished!\")\n\n elif args.do_predict:\n acc = predict(args, model, eval_dataloader, device, logger)\n logger.info(f\"test performance {acc}\")\n\ndef predict(args, model, eval_dataloader, device, logger):\n model.eval()\n id2result = collections.defaultdict(list)\n for batch in tqdm(eval_dataloader):\n batch_to_feed = move_to_cuda(batch[\"net_inputs\"])\n batch_qids = batch[\"qids\"]\n batch_labels = batch[\"net_inputs\"][\"label\"].view(-1).tolist()\n with torch.no_grad():\n scores = model(batch_to_feed)\n scores = scores.view(-1).tolist()\n for qid, label, score in zip(batch_qids, batch_labels, scores):\n id2result[qid].append((label, score))\n\n acc = []\n top_pred = {}\n for qid, res in id2result.items():\n res.sort(key=lambda x: x[1], reverse=True)\n acc.append(res[0][0] == 1)\n logger.info(f\"evaluated {len(id2result)} questions...\")\n logger.info(f'acc: {np.mean(acc)}')\n model.train()\n return np.mean(acc)\n\n\nif __name__ == \"__main__\":\n main()\n" }, { "path": "mdr/qa/utils.py", "content": "# Copyright (c) Facebook, Inc. and its affiliates.\n# All rights reserved.\n#\n# This source code is licensed under the license found in the \n# LICENSE file in the root directory of this source tree.\nimport torch\nimport sqlite3\nimport unicodedata\nimport collections\nimport logging\nimport re\n\ndef set_global_logging_level(level=logging.ERROR, prefices=[\"\"]):\n \"\"\"\n Override logging levels of different modules based on their name as a prefix.\n It needs to be invoked after the modules have been loaded so that their loggers have been initialized.\n\n Args:\n - level: desired level. e.g. logging.INFO. Optional. Default is logging.ERROR\n - prefices: list of one or more str prefices to match (e.g. [\"transformers\", \"torch\"]). Optional.\n Default is `[\"\"]` to match all active loggers.\n The match is a case-sensitive `module_name.startswith(prefix)`\n \"\"\"\n prefix_re = re.compile(fr'^(?:{ \"|\".join(prefices) })')\n for name in logging.root.manager.loggerDict:\n if re.match(prefix_re, name):\n logging.getLogger(name).setLevel(level)\n\ndef load_saved(model, path, exact=True):\n try:\n state_dict = torch.load(path)\n except:\n state_dict = torch.load(path, map_location=torch.device('cpu'))\n\n def filter(x): return x[7:] if x.startswith('module.') else x\n if exact:\n state_dict = {filter(k): v for (k, v) in state_dict.items()}\n else:\n state_dict = {filter(k): v for (\n k, v) in state_dict.items() if filter(k) in model.state_dict()}\n model.load_state_dict(state_dict)\n return model\n\n\ndef move_to_cuda(sample):\n if len(sample) == 0:\n return {}\n\n def _move_to_cuda(maybe_tensor):\n if torch.is_tensor(maybe_tensor):\n return maybe_tensor.cuda()\n elif isinstance(maybe_tensor, dict):\n return {\n key: _move_to_cuda(value)\n for key, value in maybe_tensor.items()\n }\n elif isinstance(maybe_tensor, list):\n return [_move_to_cuda(x) for x in maybe_tensor]\n else:\n return maybe_tensor\n\n return _move_to_cuda(sample)\n\n\ndef convert_to_half(sample):\n if len(sample) == 0:\n return {}\n\n def _convert_to_half(maybe_floatTensor):\n if torch.is_tensor(maybe_floatTensor) and maybe_floatTensor.type() == \"torch.FloatTensor\":\n return maybe_floatTensor.half()\n elif isinstance(maybe_floatTensor, dict):\n return {\n key: _convert_to_half(value)\n for key, value in maybe_floatTensor.items()\n }\n elif isinstance(maybe_floatTensor, list):\n return [_convert_to_half(x) for x in maybe_floatTensor]\n else:\n return maybe_floatTensor\n\n return _convert_to_half(sample)\n\n\nclass AverageMeter(object):\n \"\"\"Computes and stores the average and current value\"\"\"\n\n def __init__(self):\n self.reset()\n\n def reset(self):\n self.val = 0\n self.avg = 0\n self.sum = 0\n self.count = 0\n\n def update(self, val, n=1):\n self.val = val\n self.sum += val * n\n self.count += n\n self.avg = self.sum / self.count\n\n\ndef normalize(text):\n \"\"\"Resolve different type of unicode encodings.\"\"\"\n return unicodedata.normalize('NFD', text)\n\n\ndef para_has_answer(answer, para, tokenizer):\n text = normalize(para)\n tokens = tokenizer.tokenize(text)\n text = tokens.words(uncased=True)\n assert len(text) == len(tokens)\n for single_answer in answer:\n single_answer = normalize(single_answer)\n single_answer = tokenizer.tokenize(single_answer)\n single_answer = single_answer.words(uncased=True)\n for i in range(0, len(text) - len(single_answer) + 1):\n if single_answer == text[i: i + len(single_answer)]:\n return True\n return False\n\n\ndef match_answer_span(p, answer, tokenizer, match=\"string\"):\n # p has been normalized\n if match == 'string':\n tokens = tokenizer.tokenize(p)\n text = tokens.words(uncased=True)\n matched = set()\n for single_answer in answer:\n single_answer = normalize(single_answer)\n single_answer = tokenizer.tokenize(single_answer)\n single_answer = single_answer.words(uncased=True)\n for i in range(0, len(text) - len(single_answer) + 1):\n if single_answer == text[i: i + len(single_answer)]:\n matched.add(tokens.slice(\n i, i + len(single_answer)).untokenize())\n return list(matched)\n elif match == 'regex':\n # Answer is a regex\n single_answer = normalize(answer[0])\n return regex_match(p, single_answer)\n\n\ndef _is_whitespace(char):\n \"\"\"Checks whether `chars` is a whitespace character.\"\"\"\n # \\t, \\n, and \\r are technically contorl characters but we treat them\n # as whitespace since they are generally considered as such.\n if char == \" \" or char == \"\\t\" or char == \"\\n\" or char == \"\\r\":\n return True\n cat = unicodedata.category(char)\n if cat == \"Zs\":\n return True\n return False\n\n\n\n\n\ndef _improve_answer_span(doc_tokens, input_start, input_end, tokenizer,\n orig_answer_text):\n tok_answer_text = \" \".join(tokenizer.tokenize(orig_answer_text))\n\n for new_start in range(input_start, input_end + 1):\n for new_end in range(input_end, new_start - 1, -1):\n text_span = \" \".join(doc_tokens[new_start:(new_end + 1)])\n if text_span == tok_answer_text:\n return (new_start, new_end)\n\n return (input_start, input_end)\n\n\ndef whitespace_tokenize(text):\n \"\"\"Runs basic whitespace cleaning and splitting on a peice of text.\"\"\"\n text = text.strip()\n if not text:\n return []\n tokens = text.split()\n return tokens\n\n\ndef find_ans_span_with_char_offsets(detected_ans, char_to_word_offset, doc_tokens, all_doc_tokens, orig_to_tok_index, tokenizer):\n # could return mutiple spans for an answer string\n ans_text = detected_ans[\"text\"]\n char_spans = detected_ans[\"char_spans\"]\n ans_subtok_spans = []\n for char_start, char_end in char_spans:\n tok_start = char_to_word_offset[char_start]\n # char_end points to the last char of the answer, not one after\n tok_end = char_to_word_offset[char_end]\n sub_tok_start = orig_to_tok_index[tok_start]\n\n if tok_end < len(doc_tokens) - 1:\n sub_tok_end = orig_to_tok_index[tok_end + 1] - 1\n else:\n sub_tok_end = len(all_doc_tokens) - 1\n\n actual_text = \" \".join(doc_tokens[tok_start:(tok_end + 1)])\n cleaned_answer_text = \" \".join(whitespace_tokenize(ans_text))\n if actual_text.find(cleaned_answer_text) == -1:\n print(\"Could not find answer: '{}' vs. '{}'\".format(\n actual_text, cleaned_answer_text))\n\n (sub_tok_start, sub_tok_end) = _improve_answer_span(\n all_doc_tokens, sub_tok_start, sub_tok_end, tokenizer, ans_text)\n ans_subtok_spans.append((sub_tok_start, sub_tok_end))\n\n return ans_subtok_spans\n\nimport six\n\ndef convert_to_unicode(text):\n \"\"\"Converts `text` to Unicode (if it's not already), assuming utf-8 input.\"\"\"\n if six.PY3:\n if isinstance(text, str):\n return text\n elif isinstance(text, bytes):\n return text.decode(\"utf-8\", \"ignore\")\n else:\n raise ValueError(\"Unsupported string type: %s\" % (type(text)))\n elif six.PY2:\n if isinstance(text, str):\n return text.decode(\"utf-8\", \"ignore\")\n elif isinstance(text, unicode):\n return text\n else:\n raise ValueError(\"Unsupported string type: %s\" % (type(text)))\n else:\n raise ValueError(\"Not running on Python2 or Python 3?\")\n\n\ndef _is_control(char):\n \"\"\"Checks whether `chars` is a control character.\"\"\"\n # These are technically control characters but we count them as whitespace\n # characters.\n if char == \"\\t\" or char == \"\\n\" or char == \"\\r\":\n return False\n cat = unicodedata.category(char)\n if cat.startswith(\"C\"):\n return True\n return False\n\ndef _is_punctuation(char):\n \"\"\"Checks whether `chars` is a punctuation character.\"\"\"\n cp = ord(char)\n # We treat all non-letter/number ASCII as punctuation.\n # Characters such as \"^\", \"$\", and \"`\" are not in the Unicode\n # Punctuation class but we treat them as punctuation anyways, for\n # consistency.\n if ((cp >= 33 and cp <= 47) or (cp >= 58 and cp <= 64) or\n (cp >= 91 and cp <= 96) or (cp >= 123 and cp <= 126)):\n return True\n cat = unicodedata.category(char)\n if cat.startswith(\"P\"):\n return True\n return False\n\n\nclass BasicTokenizer(object):\n \"\"\"Runs basic tokenization (punctuation splitting, lower casing, etc.).\"\"\"\n\n def __init__(self, do_lower_case=True):\n \"\"\"Constructs a BasicTokenizer.\n Args:\n do_lower_case: Whether to lower case the input.\n \"\"\"\n self.do_lower_case = do_lower_case\n\n def tokenize(self, text):\n \"\"\"Tokenizes a piece of text.\"\"\"\n text = convert_to_unicode(text)\n text = self._clean_text(text)\n orig_tokens = whitespace_tokenize(text)\n split_tokens = []\n for token in orig_tokens:\n if self.do_lower_case:\n token = token.lower()\n token = self._run_strip_accents(token)\n split_tokens.extend(self._run_split_on_punc(token))\n\n output_tokens = whitespace_tokenize(\" \".join(split_tokens))\n return output_tokens\n\n def _run_strip_accents(self, text):\n \"\"\"Strips accents from a piece of text.\"\"\"\n text = unicodedata.normalize(\"NFD\", text)\n output = []\n for char in text:\n cat = unicodedata.category(char)\n if cat == \"Mn\":\n continue\n output.append(char)\n return \"\".join(output)\n\n def _run_split_on_punc(self, text):\n \"\"\"Splits punctuation on a piece of text.\"\"\"\n chars = list(text)\n i = 0\n start_new_word = True\n output = []\n while i < len(chars):\n char = chars[i]\n if _is_punctuation(char):\n output.append([char])\n start_new_word = True\n else:\n if start_new_word:\n output.append([])\n start_new_word = False\n output[-1].append(char)\n i += 1\n\n return [\"\".join(x) for x in output]\n\n def _clean_text(self, text):\n \"\"\"Performs invalid character removal and whitespace cleanup on text.\"\"\"\n output = []\n for char in text:\n cp = ord(char)\n if cp == 0 or cp == 0xfffd or _is_control(char):\n continue\n if _is_whitespace(char):\n output.append(\" \")\n else:\n output.append(char)\n return \"\".join(output)\n\n\ndef get_final_text(pred_text, orig_text, do_lower_case=False, verbose_logging=True):\n \"\"\"Project the tokenized prediction back to the original text.\"\"\"\n def _strip_spaces(text):\n ns_chars = []\n ns_to_s_map = collections.OrderedDict()\n for (i, c) in enumerate(text):\n if c == \" \":\n continue\n ns_to_s_map[len(ns_chars)] = i\n ns_chars.append(c)\n ns_text = \"\".join(ns_chars)\n return (ns_text, ns_to_s_map)\n\n # We first tokenize `orig_text`, strip whitespace from the result\n # and `pred_text`, and check if they are the same length. If they are\n # NOT the same length, the heuristic has failed. If they are the same\n # length, we assume the characters are one-to-one aligned.\n tokenizer = BasicTokenizer(do_lower_case=do_lower_case)\n\n tok_text = \" \".join(tokenizer.tokenize(orig_text))\n\n start_position = tok_text.find(pred_text)\n if start_position == -1:\n if verbose_logging:\n print(\n \"Unable to find text: '%s' in '%s'\" % (pred_text, orig_text))\n return orig_text\n end_position = start_position + len(pred_text) - 1\n\n (orig_ns_text, orig_ns_to_s_map) = _strip_spaces(orig_text)\n (tok_ns_text, tok_ns_to_s_map) = _strip_spaces(tok_text)\n\n if len(orig_ns_text) != len(tok_ns_text):\n if verbose_logging:\n print(\"Length not equal after stripping spaces: '%s' vs '%s'\",\n orig_ns_text, tok_ns_text)\n return orig_text\n\n # We then project the characters in `pred_text` back to `orig_text` using\n # the character-to-character alignment.\n tok_s_to_ns_map = {}\n for (i, tok_index) in six.iteritems(tok_ns_to_s_map):\n tok_s_to_ns_map[tok_index] = i\n\n orig_start_position = None\n if start_position in tok_s_to_ns_map:\n ns_start_position = tok_s_to_ns_map[start_position]\n if ns_start_position in orig_ns_to_s_map:\n orig_start_position = orig_ns_to_s_map[ns_start_position]\n\n if orig_start_position is None:\n if verbose_logging:\n print(\"Couldn't map start position\")\n return orig_text\n\n orig_end_position = None\n if end_position in tok_s_to_ns_map:\n ns_end_position = tok_s_to_ns_map[end_position]\n if ns_end_position in orig_ns_to_s_map:\n orig_end_position = orig_ns_to_s_map[ns_end_position]\n\n if orig_end_position is None:\n if verbose_logging:\n print(\"Couldn't map end position\")\n return orig_text\n\n output_text = orig_text[orig_start_position:(orig_end_position + 1)]\n return output_text\n" }, { "path": "mdr/retrieval/__init__.py", "content": "# Copyright (c) Facebook, Inc. and its affiliates.\n# All rights reserved.\n#\n# This source code is licensed under the license found in the \n# LICENSE file in the root directory of this source tree.\n#!/usr/bin/env python\n# Copyright 2017-present, Facebook, Inc.\n# All rights reserved.\n#\n# This source code is licensed under the license found in the\n# LICENSE file in the root directory of this source tree.\n\nfrom . import data\nfrom . import models\nfrom . import utils\n" }, { "path": "mdr/retrieval/config.py", "content": "# Copyright (c) Facebook, Inc. and its affiliates.\n# All rights reserved.\n#\n# This source code is licensed under the license found in the \n# LICENSE file in the root directory of this source tree.\nimport argparse\nfrom ast import parse\nfrom typing import NamedTuple\n\nclass ClusterConfig(NamedTuple):\n dist_backend: str\n dist_url: str\n\ndef common_args():\n parser = argparse.ArgumentParser()\n\n # task\n parser.add_argument(\"--train_file\", type=str,\n default=\"../data/nq-with-neg-train.txt\")\n parser.add_argument(\"--predict_file\", type=str,\n default=\"../data/nq-with-neg-dev.txt\")\n parser.add_argument(\"--num_workers\", default=30, type=int)\n parser.add_argument(\"--do_train\", default=False,\n action='store_true', help=\"Whether to run training.\")\n parser.add_argument(\"--do_predict\", default=False,\n action='store_true', help=\"Whether to run eval on the dev set.\")\n\n # model\n parser.add_argument(\"--model_name\",\n default=\"bert-base-uncased\", type=str)\n parser.add_argument(\"--init_checkpoint\", type=str,\n help=\"Initial checkpoint (usually from a pre-trained BERT model).\",\n default=\"\")\n parser.add_argument(\"--max_c_len\", default=512, type=int,\n help=\"The maximum total input sequence length after WordPiece tokenization. Sequences \"\n \"longer than this will be truncated, and sequences shorter than this will be padded.\")\n parser.add_argument(\"--max_q_len\", default=50, type=int,\n help=\"The maximum number of tokens for the question. Questions longer than this will \"\n \"be truncated to this length.\")\n parser.add_argument('--fp16', action='store_true')\n parser.add_argument('--fp16_opt_level', type=str, default='O1',\n help=\"For fp16: Apex AMP optimization level selected in ['O0', 'O1', 'O2', and 'O3'].\"\n \"See details at https://nvidia.github.io/apex/amp.html\")\n parser.add_argument(\"--no_cuda\", default=False, action='store_true',\n help=\"Whether not to use CUDA when available\")\n parser.add_argument(\"--local_rank\", type=int, default=-1,\n help=\"local_rank for distributed training on gpus\")\n parser.add_argument(\"--max_q_sp_len\", default=50, type=int)\n parser.add_argument(\"--sent-level\", action=\"store_true\")\n parser.add_argument(\"--rnn-retriever\", action=\"store_true\")\n parser.add_argument(\"--predict_batch_size\", default=512,\n type=int, help=\"Total batch size for predictions.\")\n parser.add_argument(\"--shared-encoder\", action=\"store_true\")\n\n # multi vector scheme\n parser.add_argument(\"--multi-vector\", type=int, default=1)\n parser.add_argument(\"--scheme\", type=str, help=\"how to get the multivector, layerwise or tokenwise\", default=\"none\")\n\n # momentum\n parser.add_argument(\"--momentum\", action=\"store_true\")\n parser.add_argument(\"--init-retriever\", type=str, default=\"\")\n parser.add_argument(\"--k\", type=int, default=38400, help=\"memory bank size\")\n parser.add_argument(\"--m\", type=float, default=0.999, help=\"momentum\")\n\n\n # NQ multihop trial\n parser.add_argument(\"--nq-multi\", action=\"store_true\", help=\"train the NQ retrieval model to recover from error cases\")\n\n return parser\n\ndef train_args():\n parser = common_args()\n # optimization\n parser.add_argument('--prefix', type=str, default=\"eval\")\n parser.add_argument(\"--weight_decay\", default=0.0, type=float,\n help=\"Weight decay if we apply some.\")\n parser.add_argument(\"--temperature\", default=1, type=float)\n parser.add_argument(\"--output_dir\", default=\"./logs\", type=str,\n help=\"The output directory where the model checkpoints will be written.\")\n parser.add_argument(\"--train_batch_size\", default=128,\n type=int, help=\"Total batch size for training.\")\n parser.add_argument(\"--learning_rate\", default=1e-5,\n type=float, help=\"The initial learning rate for Adam.\")\n parser.add_argument(\"--adam_epsilon\", default=1e-8, type=float,\n help=\"Epsilon for Adam optimizer.\")\n parser.add_argument(\"--num_train_epochs\", default=50, type=float,\n help=\"Total number of training epochs to perform.\")\n parser.add_argument(\"--save_checkpoints_steps\", default=20000, type=int,\n help=\"How often to save the model checkpoint.\")\n parser.add_argument(\"--iterations_per_loop\", default=1000, type=int,\n help=\"How many steps to make in each estimator call.\")\n parser.add_argument(\"--accumulate_gradients\", type=int, default=1,\n help=\"Number of steps to accumulate gradient on (divide the batch_size and accumulate)\")\n parser.add_argument('--seed', type=int, default=3,\n help=\"random seed for initialization\")\n parser.add_argument('--gradient_accumulation_steps', type=int, default=1,\n help=\"Number of updates steps to accumualte before performing a backward/update pass.\")\n parser.add_argument('--eval-period', type=int, default=2500)\n parser.add_argument(\"--max_grad_norm\", default=2.0, type=float, help=\"Max gradient norm.\")\n parser.add_argument(\"--stop-drop\", default=0, type=float)\n parser.add_argument(\"--use-adam\", action=\"store_true\")\n parser.add_argument(\"--warmup-ratio\", default=0, type=float, help=\"Linear warmup over warmup_steps.\")\n\n\n return parser.parse_args()\n\ndef encode_args():\n parser = common_args()\n parser.add_argument('--embed_save_path', type=str, default=\"\")\n parser.add_argument('--is_query_embed', action=\"store_true\")\n args = parser.parse_args()\n return args\n" }, { "path": "mdr/retrieval/criterions.py", "content": "# Copyright (c) Facebook, Inc. and its affiliates.\n# All rights reserved.\n#\n# This source code is licensed under the license found in the \n# LICENSE file in the root directory of this source tree.\nimport torch\nfrom torch.nn import CrossEntropyLoss\nimport torch.nn.functional as F\n\n# def loss_single(model, batch, momentum=False):\n# outputs = model(batch)\n# q = outputs['q']\n# c = outputs['c']\n# neg_c = outputs['neg_c']\n# product_in_batch = torch.mm(q, c.t())\n# product_neg = (q * neg_c).sum(-1).unsqueeze(1)\n# product = torch.cat([product_in_batch, product_neg], dim=-1)\n\n# if momentum:\n# queue_c = model.module.encode_queue_ctx()\n# product_queue = torch.mm(q, queue_c.t())\n# product = torch.cat([product, product_queue], dim=-1)\n# model.module.dequeue_and_enqueue(batch)\n\n# target = torch.arange(product.size(0)).to(product.device)\n# loss = F.cross_entropy(product, target)\n# return loss\n\n\n# \"\"\"\n# multi-hop retrieval for NQ, train the model to recover from\n# \"\"\"\n# def loss_nq_mhop(model, batch, momentum=False):\n# outputs = model(batch)\n# product_in_batch = torch.mm(outputs['q'], outputs['c'].t())\n# product_neg = (outputs['q'] * outputs['neg']).sum(-1).unsqueeze(1)\n# # product_neg1 = (outputs['q'] * outputs['dense_neg1']).sum(-1).unsqueeze(1)\n# # product_neg2 = (outputs['q'] * outputs['dense_neg2']).sum(-1).unsqueeze(1)\n# scores1 = torch.cat([product_in_batch, product_neg], dim=-1)\n\n# product_in_batch_from_error = torch.mm(outputs[\"q_neg1\"], outputs['c'].t())\n# dense_neg = torch.cat([outputs[\"dense_neg1\"].unsqueeze(1), outputs[\"dense_neg2\"].unsqueeze(1)], dim=1)\n# product_neg_from_error = torch.bmm(outputs[\"q_neg1\"].unsqueeze(1), dense_neg.transpose(1,2)).squeeze(1)\n# scores2 = torch.cat([product_in_batch_from_error, product_neg_from_error], dim=-1)\n# if momentum:\n# queue_neg_scores_1 = torch.mm(outputs['q'], model.module.queue.clone().detach().t())\n# queue_neg_scores_2 = torch.mm(outputs[\"q_neg1\"], model.module.queue.clone().detach().t())\n# scores1 = torch.cat([scores1, queue_neg_scores_1], dim=1)\n# scores2 = torch.cat([scores2, queue_neg_scores_2], dim=1)\n# model.module.dequeue_and_enqueue(outputs[\"c\"].detach())\n# # model.module.momentum_update_key_encoder()\n\n# target = torch.arange(scores1.size(0)).to(scores1.device)\n# loss = F.cross_entropy(scores1, target) + F.cross_entropy(scores2, target)\n# # loss = F.cross_entropy(scores1, target)\n# return loss\n\n# def eval_nq_mhop(model, batch):\n# outputs = model(batch)\n# product_in_batch = torch.mm(outputs['q'], outputs['c'].t())\n# product_neg = (outputs['q'] * outputs['neg']).sum(-1).unsqueeze(1)\n# # product_neg1 = (outputs['q'] * outputs['dense_neg1']).sum(-1).unsqueeze(1)\n# # product_neg2 = (outputs['q'] * outputs['dense_neg2']).sum(-1).unsqueeze(1)\n# scores1 = torch.cat([product_in_batch, product_neg], dim=-1)\n\n# product_in_batch_from_error = torch.mm(outputs[\"q_neg1\"], outputs['c'].t())\n# dense_neg = torch.cat([outputs[\"dense_neg1\"].unsqueeze(1), outputs[\"dense_neg2\"].unsqueeze(1)], dim=1)\n# product_neg_from_error = torch.bmm(outputs[\"q_neg1\"].unsqueeze(1), dense_neg.transpose(1,2)).squeeze(1)\n# scores2 = torch.cat([product_in_batch_from_error, product_neg_from_error], dim=-1)\n\n# target = torch.arange(scores1.size(0)).to(scores1.device)\n\n# rrs, rrs_2hop = [], []\n# ranked = scores1.argsort(dim=1, descending=True)\n# ranked_2hop = scores2.argsort(dim=1, descending=True)\n# idx2rank = ranked.argsort(dim=1)\n# for idx, t in enumerate(target.tolist()):\n# rrs.append(1 / (idx2rank[idx][t].item() +1))\n# idx2rank2hop = ranked_2hop.argsort(dim=1)\n# for idx, t in enumerate(target.tolist()):\n# rrs_2hop.append(1 / (idx2rank2hop[idx][t].item() +1))\n# return rrs, rrs_2hop\n\n\n\n# def eval_vanilla(outputs):\n# \"\"\"\n# view the two sp passages as the same, no multi-hop modeling;\n# select the passages from all passages in the batch\n# \"\"\"\n# rrs = []\n# q = outputs['q']\n# c1 = outputs['c1'] \n# c2 = outputs['c2']\n# c = torch.cat([c1.unsqueeze(1), c2.unsqueeze(1)], dim=1) # B x 2 x D\n# c = c.view(-1, q.size(-1)) # 2B x D\n# product_in_batch = torch.mm(q, c.t()) # Bx2B\n# neg_c = outputs['neg_c']\n# product_neg = (q * neg_c).sum(-1).unsqueeze(1)\n# product = torch.cat([product_in_batch, product_neg], dim=-1) \n# target = torch.arange(product.size(0)).to(product.device).unsqueeze(1)\n# target = torch.cat([target*2, target*2+1], dim=1)\n# ranked = product.argsort(dim=1, descending=True)\n# # MRR\n# idx2rank = ranked.argsort(dim=1)\n# for idx, t in enumerate(target):\n# correct_idx = t.tolist()\n# for _ in correct_idx:\n# rrs.append(1 / (idx2rank[idx][_].item() + 1))\n# return rrs\n\n\n\ndef mhop_loss(model, batch, args):\n\n outputs = model(batch)\n loss_fct = CrossEntropyLoss(ignore_index=-1)\n\n all_ctx = torch.cat([outputs['c1'], outputs['c2']], dim=0)\n neg_ctx = torch.cat([outputs[\"neg_1\"].unsqueeze(1), outputs[\"neg_2\"].unsqueeze(1)], dim=1) # B x 2 x M x h\n \n scores_1_hop = torch.mm(outputs[\"q\"], all_ctx.t())\n neg_scores_1 = torch.bmm(outputs[\"q\"].unsqueeze(1), neg_ctx.transpose(1,2)).squeeze(1)\n scores_2_hop = torch.mm(outputs[\"q_sp1\"], all_ctx.t())\n neg_scores_2 = torch.bmm(outputs[\"q_sp1\"].unsqueeze(1), neg_ctx.transpose(1,2)).squeeze(1)\n\n # mask the 1st hop\n bsize = outputs[\"q\"].size(0)\n scores_1_mask = torch.cat([torch.zeros(bsize, bsize), torch.eye(bsize)], dim=1).to(outputs[\"q\"].device)\n scores_1_hop = scores_1_hop.float().masked_fill(scores_1_mask.bool(), float('-inf')).type_as(scores_1_hop)\n scores_1_hop = torch.cat([scores_1_hop, neg_scores_1], dim=1)\n scores_2_hop = torch.cat([scores_2_hop, neg_scores_2], dim=1)\n\n if args.momentum:\n queue_neg_scores_1 = torch.mm(outputs[\"q\"], model.module.queue.clone().detach().t())\n queue_neg_scores_2 = torch.mm(outputs[\"q_sp1\"], model.module.queue.clone().detach().t())\n\n # queue_neg_scores_1 = queue_neg_scores_1 / args.temperature\n # queue_neg_scores_2 = queue_neg_scores_2 / args.temperature \n\n scores_1_hop = torch.cat([scores_1_hop, queue_neg_scores_1], dim=1)\n scores_2_hop = torch.cat([scores_2_hop, queue_neg_scores_2], dim=1)\n model.module.dequeue_and_enqueue(all_ctx.detach())\n # model.module.momentum_update_key_encoder()\n\n target_1_hop = torch.arange(outputs[\"q\"].size(0)).to(outputs[\"q\"].device)\n target_2_hop = torch.arange(outputs[\"q\"].size(0)).to(outputs[\"q\"].device) + outputs[\"q\"].size(0)\n\n retrieve_loss = loss_fct(scores_1_hop, target_1_hop) + loss_fct(scores_2_hop, target_2_hop)\n\n return retrieve_loss\n\ndef mhop_eval(outputs, args):\n all_ctx = torch.cat([outputs['c1'], outputs['c2']], dim=0)\n neg_ctx = torch.cat([outputs[\"neg_1\"].unsqueeze(1), outputs[\"neg_2\"].unsqueeze(1)], dim=1)\n\n\n scores_1_hop = torch.mm(outputs[\"q\"], all_ctx.t())\n neg_scores_1 = torch.bmm(outputs[\"q\"].unsqueeze(1), neg_ctx.transpose(1,2)).squeeze(1)\n scores_2_hop = torch.mm(outputs[\"q_sp1\"], all_ctx.t())\n neg_scores_2 = torch.bmm(outputs[\"q_sp1\"].unsqueeze(1), neg_ctx.transpose(1,2)).squeeze(1)\n\n\n bsize = outputs[\"q\"].size(0)\n scores_1_mask = torch.cat([torch.zeros(bsize, bsize), torch.eye(bsize)], dim=1).to(outputs[\"q\"].device)\n scores_1_hop = scores_1_hop.float().masked_fill(scores_1_mask.bool(), float('-inf')).type_as(scores_1_hop)\n scores_1_hop = torch.cat([scores_1_hop, neg_scores_1], dim=1)\n scores_2_hop = torch.cat([scores_2_hop, neg_scores_2], dim=1)\n target_1_hop = torch.arange(outputs[\"q\"].size(0)).to(outputs[\"q\"].device)\n target_2_hop = torch.arange(outputs[\"q\"].size(0)).to(outputs[\"q\"].device) + outputs[\"q\"].size(0)\n\n ranked_1_hop = scores_1_hop.argsort(dim=1, descending=True)\n ranked_2_hop = scores_2_hop.argsort(dim=1, descending=True)\n idx2ranked_1 = ranked_1_hop.argsort(dim=1)\n idx2ranked_2 = ranked_2_hop.argsort(dim=1)\n rrs_1, rrs_2 = [], []\n for t, idx2ranked in zip(target_1_hop, idx2ranked_1):\n rrs_1.append(1 / (idx2ranked[t].item() + 1))\n for t, idx2ranked in zip(target_2_hop, idx2ranked_2):\n rrs_2.append(1 / (idx2ranked[t].item() + 1))\n \n return {\"rrs_1\": rrs_1, \"rrs_2\": rrs_2}\n\n\ndef unified_loss(model, batch, args):\n\n outputs = model(batch)\n all_ctx = torch.cat([outputs['c1'], outputs['c2']], dim=0)\n neg_ctx = torch.cat([outputs[\"neg_1\"].unsqueeze(1), outputs[\"neg_2\"].unsqueeze(1)], dim=1)\n scores_1_hop = torch.mm(outputs[\"q\"], all_ctx.t())\n neg_scores_1 = torch.bmm(outputs[\"q\"].unsqueeze(1), neg_ctx.transpose(1,2)).squeeze(1)\n scores_2_hop = torch.mm(outputs[\"q_sp1\"], all_ctx.t())\n neg_scores_2 = torch.bmm(outputs[\"q_sp1\"].unsqueeze(1), neg_ctx.transpose(1,2)).squeeze(1)\n\n # mask for 1st hop\n bsize = outputs[\"q\"].size(0)\n scores_1_mask = torch.cat([torch.zeros(bsize, bsize), torch.eye(bsize)], dim=1).to(outputs[\"q\"].device)\n scores_1_hop = scores_1_hop.float().masked_fill(scores_1_mask.bool(), float('-inf')).type_as(scores_1_hop)\n scores_1_hop = torch.cat([scores_1_hop, neg_scores_1], dim=1)\n scores_2_hop = torch.cat([scores_2_hop, neg_scores_2], dim=1)\n\n stop_loss = F.cross_entropy(outputs[\"stop_logits\"], batch[\"stop_targets\"].view(-1), reduction=\"sum\")\n\n target_1_hop = torch.arange(outputs[\"q\"].size(0)).to(outputs[\"q\"].device)\n target_2_hop = torch.arange(outputs[\"q\"].size(0)).to(outputs[\"q\"].device) + outputs[\"q\"].size(0)\n\n\n retrieve_loss = F.cross_entropy(scores_1_hop, target_1_hop, reduction=\"sum\") + (F.cross_entropy(scores_2_hop, target_2_hop, reduction=\"none\") * batch[\"stop_targets\"].view(-1)).sum()\n\n return retrieve_loss + stop_loss\n\ndef unified_eval(outputs, batch):\n all_ctx = torch.cat([outputs['c1'], outputs['c2']], dim=0)\n neg_ctx = torch.cat([outputs[\"neg_1\"].unsqueeze(1), outputs[\"neg_2\"].unsqueeze(1)], dim=1)\n scores_1_hop = torch.mm(outputs[\"q\"], all_ctx.t())\n scores_2_hop = torch.mm(outputs[\"q_sp1\"], all_ctx.t())\n neg_scores_1 = torch.bmm(outputs[\"q\"].unsqueeze(1), neg_ctx.transpose(1,2)).squeeze(1)\n neg_scores_2 = torch.bmm(outputs[\"q_sp1\"].unsqueeze(1), neg_ctx.transpose(1,2)).squeeze(1)\n bsize = outputs[\"q\"].size(0)\n scores_1_mask = torch.cat([torch.zeros(bsize, bsize), torch.eye(bsize)], dim=1).to(outputs[\"q\"].device)\n scores_1_hop = scores_1_hop.float().masked_fill(scores_1_mask.bool(), float('-inf')).type_as(scores_1_hop)\n scores_1_hop = torch.cat([scores_1_hop, neg_scores_1], dim=1)\n scores_2_hop = torch.cat([scores_2_hop, neg_scores_2], dim=1)\n target_1_hop = torch.arange(outputs[\"q\"].size(0)).to(outputs[\"q\"].device)\n target_2_hop = torch.arange(outputs[\"q\"].size(0)).to(outputs[\"q\"].device) + outputs[\"q\"].size(0)\n\n # stop accuracy\n stop_pred = outputs[\"stop_logits\"].argmax(dim=1)\n stop_targets = batch[\"stop_targets\"].view(-1)\n stop_acc = (stop_pred == stop_targets).float().tolist()\n\n ranked_1_hop = scores_1_hop.argsort(dim=1, descending=True)\n ranked_2_hop = scores_2_hop.argsort(dim=1, descending=True)\n idx2ranked_1 = ranked_1_hop.argsort(dim=1)\n idx2ranked_2 = ranked_2_hop.argsort(dim=1)\n\n rrs_1_mhop, rrs_2_mhop, rrs_nq = [], [], []\n for t1, idx2ranked1, t2, idx2ranked2, stop in zip(target_1_hop, idx2ranked_1, target_2_hop, idx2ranked_2, stop_targets):\n if stop: # \n rrs_1_mhop.append(1 / (idx2ranked1[t1].item() + 1))\n rrs_2_mhop.append(1 / (idx2ranked2[t2].item() + 1))\n else:\n rrs_nq.append(1 / (idx2ranked1[t1].item() + 1))\n\n return {\n \"stop_acc\": stop_acc, \n \"rrs_1_mhop\": rrs_1_mhop,\n \"rrs_2_mhop\": rrs_2_mhop,\n \"rrs_nq\": rrs_nq\n }\n" }, { "path": "mdr/retrieval/decomposed_analysis.py", "content": "# Copyright (c) Facebook, Inc. and its affiliates.\n# All rights reserved.\n#\n# This source code is licensed under the license found in the \n# LICENSE file in the root directory of this source tree.\nimport json\n\n\ndef decomposed_errors():\n top1_pred = [json.loads(l) for l in open(\"/private/home/xwhan/data/hotpot/dense_val_b1_top1.json\").readlines()]\n analysis_folder = \"/private/home/xwhan/data/hotpot/analysis\"\n\n start_errors, bridge_errors, failed = [], [], []\n correct = []\n for item in top1_pred:\n pred_titles = [_[0] for _ in item[\"candidate_chains\"][0]]\n gold_titles = [_[0] for _ in item[\"sp\"]]\n if set(pred_titles) == set(gold_titles):\n if item[\"type\"] == \"bridge\":\n correct.append(item)\n continue\n if item[\"type\"] == \"bridge\":\n start_title = None\n for t in gold_titles:\n if t != item[\"bridge\"]:\n start_title = t\n assert start_title is not None\n if item[\"bridge\"] in pred_titles and start_title not in pred_titles:\n start_errors.append(item)\n elif item[\"bridge\"] not in pred_titles and start_title in pred_titles:\n bridge_errors.append(item)\n else:\n failed.append(item)\n\n with open(analysis_folder + \"/correct.json\", \"w\") as g:\n for _ in correct:\n _[\"predicted\"] = _.pop(\"candidate_chains\")[0]\n g.write(json.dumps(_) + \"\\n\")\n\n with open(analysis_folder + \"/start_errors.json\", \"w\") as g:\n for _ in start_errors:\n _[\"predicted\"] = _.pop(\"candidate_chains\")[0]\n g.write(json.dumps(_) + \"\\n\")\n\n with open(analysis_folder + \"/bridge_errors.json\", \"w\") as g:\n for _ in bridge_errors:\n _[\"predicted\"] = _.pop(\"candidate_chains\")[0]\n g.write(json.dumps(_) + \"\\n\")\n\n with open(analysis_folder + \"/total_errors.json\", \"w\") as g:\n for _ in failed:\n _[\"predicted\"] = _.pop(\"candidate_chains\")[0]\n g.write(json.dumps(_) + \"\\n\")\n\n\n print(len(correct))\n print(len(start_errors))\n print(len(bridge_errors))\n print(len(failed))\n\nimport random\ndef collect_gold_decomposition():\n \"\"\"\n interactively collect\n \"\"\"\n dev_qdmr = [json.loads(l) for l in open(\"/private/home/xwhan/data/QDMR/dev.json\").readlines()]\n bridge_dev = [_ for _ in dev_qdmr if _[\"type\"] == \"bridge\"]\n\n random.shuffle(bridge_dev)\n idx = 0\n samples_to_inspect = []\n while True:\n print(f\"\\n-----{len(samples_to_inspect)} samples collected so far-----\")\n sample = bridge_dev[idx]\n idx += 1\n print(f\"Original Q: {sample['q']}\")\n print(f\"Decomposed Q: {sample['q_decom']}\")\n print(f\"Supporting Passages: {sample['sp']}\")\n subq1 = input(\"Type SUB Q1:\")\n if subq1 == \"bad\":\n continue\n elif subq1 == \"stop\":\n break\n subq2 = input(\"Type SUB Q2:\")\n samples_to_inspect.append({\n \"id\": sample[\"id\"],\n \"sp\": sample[\"sp\"],\n \"orig_q\": sample['q'],\n \"subQ_1\": subq1,\n \"subQ_2\": subq2\n })\n\n print(f\"{len(samples_to_inspect)} samples collected in total..\")\n\n with open(\"/private/home/xwhan/data/QDMR/inspect.json\", \"w\") as g:\n for _ in samples_to_inspect:\n g.write(json.dumps(_) + \"\\n\")\n\ndef qdmr_utils():\n \"\"\"\n change file format for decomposed and end-to-end retrieval\n \"\"\"\n qdmr_data = [json.loads(l) for l in open(\"/private/home/xwhan/data/QDMR/inspect.json\").readlines()]\n\n mhop_data, decomposed_data = [], []\n for idx, item in enumerate(qdmr_data):\n if idx in [65,66,67]:\n continue\n sp = [_[\"title\"] for _ in item[\"sp\"]]\n question = item[\"orig_q\"]\n mhop_data.append({\n \"question\": question,\n \"sp\": sp,\n \"type\": \"bridge\",\n \"_id\": item[\"id\"]\n })\n decomposed_data.append(item)\n\n # with open(\"/private/home/xwhan/data/QDMR/qdmr_decomposed.json\", \"w\") as g:\n # for item in decomposed_data:\n # g.write(json.dumps(item) + \"\\n\")\n\n with open(\"/private/home/xwhan/data/QDMR/qdmr_e2e.json\", \"w\") as g:\n for item in mhop_data:\n g.write(json.dumps(item) + \"\\n\")\n\n\ndef analyze_results():\n decomposed_results = [json.loads(l) for l in open(\"/private/home/xwhan/data/QDMR/qdmr_decomposed_results.json\")]\n e2e_results = [json.loads(l) for l in open(\"/private/home/xwhan/data/QDMR/qdmr_e2e_results.json\")]\n better = 0\n worse = 0\n both = 0\n for res1, res2 in zip(decomposed_results, e2e_results):\n sp_titles = set([_[0] for _ in res1[\"sp\"]])\n\n res1_top1 = [_[0] for _ in res1[\"candidate_chains\"][0]]\n res2_top1 = [_[0] for _ in res2[\"candidate_chains\"][0]]\n\n assert res1[\"_id\"] == res2[\"_id\"]\n\n question = res1[\"question\"]\n q_pairs = res1[\"q_pairs\"]\n\n if set(res2_top1) == sp_titles and set(res1_top1) != sp_titles:\n # print(sp_titles)\n # import pdb; pdb.set_trace()\n better += 1\n elif set(res2_top1) != sp_titles and set(res1_top1) == sp_titles:\n worse += 1\n elif set(res2_top1) == sp_titles and set(res1_top1) == sp_titles:\n both += 1\n\n print(both)\n print(better)\n print(worse)\n print(len(decomposed_results))\n\nif __name__ == \"__main__\":\n # collect_gold_decomposition()\n # qdmr_utils()\n\n analyze_results()\n" }, { "path": "mdr/retrieval/fever.ipynb", "content": "{\n \"cells\": [\n {\n \"cell_type\": \"code\",\n \"execution_count\": 90,\n \"metadata\": {},\n \"outputs\": [\n {\n \"name\": \"stdout\",\n \"output_type\": \"stream\",\n \"text\": [\n \"1059\\n\",\n \"12273\\n\"\n ]\n }\n ],\n \"source\": [\n \"import json\\n\",\n \"import numpy as np\\n\",\n \"import random\\n\",\n \"\\n\",\n \"fever_path = \\\"/private/home/xwhan/data/fever/retrieval/\\\"\\n\",\n \"\\n\",\n \"dev = [json.loads(l) for l in open(fever_path + \\\"dev.txt\\\").readlines()]\\n\",\n \"multi_dev = []\\n\",\n \"single_dev = []\\n\",\n \"all_evidence_lens = [] # for multi evidence\\n\",\n \"random.shuffle(dev)\\n\",\n \"all_claim_lens = []\\n\",\n \"for item in dev:\\n\",\n \" evidence_lens = []\\n\",\n \" all_claim_lens.append(len(item[\\\"claim\\\"].split()))\\n\",\n \" \\n\",\n \" for chain in item[\\\"evidence\\\"]:\\n\",\n \" if len(chain) > 1:\\n\",\n \"# evidence_lens.append(len(chain))\\n\",\n \" chain_titles = set([p[\\\"title\\\"] for p in chain])\\n\",\n \" evidence_lens.append(len(chain_titles)) \\n\",\n \"# print(item[\\\"claim\\\"])\\n\",\n \"# print(chain)\\n\",\n \"# assert False\\n\",\n \" else:\\n\",\n \" evidence_lens.append(1)\\n\",\n \" multi_count = np.sum([int(c > 1) for c in evidence_lens])\\n\",\n \" \\n\",\n \" if multi_count == len(evidence_lens):\\n\",\n \" multi_dev.append(item)\\n\",\n \" all_evidence_lens += evidence_lens\\n\",\n \" else:\\n\",\n \" single_dev.append(item)\\n\",\n \" \\n\",\n \"print(len(multi_dev))\\n\",\n \"print(len(single_dev))\\n\",\n \"with open(\\\"/private/home/xwhan/data/fever/retrieval/dev_multi_evidence_compact.txt\\\", \\\"w\\\") as g:\\n\",\n \" for l in multi_dev:\\n\",\n \" g.write(json.dumps(l) + \\\"\\\\n\\\")\\n\",\n \"# with open(\\\"/private/home/xwhan/data/fever/retrieval/dev_single_evidence.txt\\\", \\\"w\\\") as g:\\n\",\n \"# for l in single_dev:\\n\",\n \"# g.write(json.dumps(l) + \\\"\\\\n\\\")\\n\",\n \"# with open(\\\"/private/home/xwhan/data/fever/retrieval/dev_all.txt\\\", \\\"w\\\") as g:\\n\",\n \"# for l in single_dev + multi_dev:\\n\",\n \"# g.write(json.dumps(l) + \\\"\\\\n\\\")\"\n ]\n },\n {\n \"cell_type\": \"code\",\n \"execution_count\": 81,\n \"metadata\": {},\n \"outputs\": [\n {\n \"name\": \"stdout\",\n \"output_type\": \"stream\",\n \"text\": [\n \"1741\\n\",\n \"2.0\\n\",\n \"0.5835726593911545 0.5835726593911545 0.5835726593911545\\n\"\n ]\n }\n ],\n \"source\": [\n \"# baseline retrieval for single/multihop subsets\\n\",\n \"\\n\",\n \"import unicodedata\\n\",\n \"def normalize(text):\\n\",\n \" \\\"\\\"\\\"Resolve different type of unicode encodings.\\\"\\\"\\\"\\n\",\n \" return unicodedata.normalize('NFD', text)\\n\",\n \"\\n\",\n \"el_results = [json.loads(l) for l in open(\\\"/private/home/xwhan/data/fever/retrieval/dev.ensembles.s10.jsonl\\\").readlines()]\\n\",\n \"id2el_docs = {_[\\\"id\\\"]:_[\\\"predicted_pages\\\"] for _ in el_results}\\n\",\n \"\\n\",\n \"dense_multi_results = [json.loads(l) for l in open(\\\"/private/home/xwhan/data/fever/retrieval/dense_fever_b1_20_k20.json\\\").readlines()] \\n\",\n \"\\n\",\n \"single_gold = {_[\\\"id\\\"]:_ for _ in single_dev}\\n\",\n \"multi_gold = {_[\\\"id\\\"]:_ for _ in multi_dev}\\n\",\n \"all_gold = {_[\\\"id\\\"]:_ for _ in multi_dev + single_dev}\\n\",\n \"\\n\",\n \"subset = multi_gold\\n\",\n \"precs, recalls = [], []\\n\",\n \"doc_count = []\\n\",\n \"dense_docs = []\\n\",\n \"for item in dense_multi_results:\\n\",\n \" if item[\\\"id\\\"] in subset:\\n\",\n \"# pred = set(item[\\\"predicted_pages\\\"])\\n\",\n \" retrieved_chains = item[\\\"candidate_chains\\\"] \\n\",\n \" pred = []\\n\",\n \" for chain in retrieved_chains:\\n\",\n \" for p in chain:\\n\",\n \" if normalize(p[0]) not in pred:\\n\",\n \" pred.append(normalize(p[0]))\\n\",\n \" pred = pred[:2]\\n\",\n \"# pred = [_[\\\"title\\\"] for _ in item[\\\"topk\\\"][:1]]\\n\",\n \" \\n\",\n \" pred = set(pred)\\n\",\n \"# el_pred = id2el_docs[item[\\\"id\\\"]]\\n\",\n \"# el_count = 0\\n\",\n \"# for title in el_pred:\\n\",\n \"# if title not in pred:\\n\",\n \"# pred.add(title)\\n\",\n \"# el_count +=1\\n\",\n \"# if el_count == 2:\\n\",\n \"# break\\n\",\n \" pred = list(pred)\\n\",\n \" \\n\",\n \" dense_docs.append({\\n\",\n \" \\\"claim\\\": item[\\\"claim\\\"],\\n\",\n \" \\\"id\\\": item[\\\"id\\\"],\\n\",\n \" \\\"predicted_pages\\\": list(pred)\\n\",\n \" })\\n\",\n \" \\n\",\n \" doc_count.append(len(pred))\\n\",\n \" \\n\",\n \" gold_docs = set()\\n\",\n \" recall = 0\\n\",\n \" for chain in subset[item[\\\"id\\\"]][\\\"evidence\\\"]:\\n\",\n \" chain_titles = set([normalize(p[\\\"title\\\"]) for p in chain])\\n\",\n \" for t in chain_titles: gold_docs.add(t)\\n\",\n \" chain_covered = [int(t in pred) for t in chain_titles]\\n\",\n \" if np.sum(chain_covered) == len(chain_titles):\\n\",\n \" recall = 1\\n\",\n \" break\\n\",\n \" \\n\",\n \" if len(gold_docs) > 0:\\n\",\n \" if len(pred) == 0:\\n\",\n \" prec = 0\\n\",\n \" else:\\n\",\n \" prec = np.mean([int(doc in gold_docs) for doc in pred])\\n\",\n \" \\n\",\n \" precs.append(prec)\\n\",\n \" recalls.append(recall)\\n\",\n \" \\n\",\n \"print(len(precs))\\n\",\n \"print(np.mean(doc_count))\\n\",\n \"pr, rec = np.mean(precs), np.mean(recalls)\\n\",\n \"print(pr, rec, 2.0 * pr * rec / (pr + rec))\\n\",\n \"\\n\",\n \"# with open(\\\"/private/home/xwhan/data/fever/retrieval/dense_wiki_pages_top2.jsonl\\\", \\\"w\\\") as g:\\n\",\n \"# for _ in dense_docs:\\n\",\n \"# g.write(json.dumps(_) + \\\"\\\\n\\\")\"\n ]\n },\n {\n \"cell_type\": \"code\",\n \"execution_count\": 95,\n \"metadata\": {},\n \"outputs\": [\n {\n \"name\": \"stdout\",\n \"output_type\": \"stream\",\n \"text\": [\n \"1741\\n\",\n \"2.9959793222286044\\n\",\n \"0.6764680633362424 1.0 0.8070157471297632\\n\"\n ]\n }\n ],\n \"source\": [\n \"# inspect the FEVER results\\n\",\n \"dev_all_results = [json.loads(l) for l in open(\\\"/private/home/xwhan/code/Transformer-XH/data/fever_dev_graph.json\\\").readlines()]\\n\",\n \"subset = multi_gold\\n\",\n \"precs, recalls = [], []\\n\",\n \"doc_count = []\\n\",\n \"\\n\",\n \"for item in dev_all_results:\\n\",\n \" if item[\\\"qid\\\"] in subset:\\n\",\n \" \\n\",\n \" pred = [_[\\\"name\\\"] for _ in item[\\\"node\\\"]]\\n\",\n \" pred = set(pred)\\n\",\n \" pred = list(pred)\\n\",\n \" \\n\",\n \" doc_count.append(len(pred))\\n\",\n \" \\n\",\n \" gold_docs = set()\\n\",\n \" recall = 0\\n\",\n \" for chain in subset[item[\\\"qid\\\"]][\\\"evidence\\\"]:\\n\",\n \" chain_titles = set([normalize(p[\\\"title\\\"]) for p in chain])\\n\",\n \" for t in chain_titles: gold_docs.add(t)\\n\",\n \" chain_covered = [int(t in pred) for t in chain_titles]\\n\",\n \" if np.sum(chain_covered) == len(chain_titles):\\n\",\n \" recall = 1\\n\",\n \" break\\n\",\n \" \\n\",\n \" if len(gold_docs) > 0:\\n\",\n \" if len(pred) == 0:\\n\",\n \" prec = 0\\n\",\n \" else:\\n\",\n \" prec = np.mean([int(doc in gold_docs) for doc in pred])\\n\",\n \" \\n\",\n \" precs.append(prec)\\n\",\n \" recalls.append(recall)\\n\",\n \" \\n\",\n \"print(len(precs))\\n\",\n \"print(np.mean(doc_count))\\n\",\n \"pr, rec = np.mean(precs), np.mean(recalls)\\n\",\n \"print(pr, rec, 2.0 * pr * rec / (pr + rec))\"\n ]\n },\n {\n \"cell_type\": \"code\",\n \"execution_count\": 40,\n \"metadata\": {},\n \"outputs\": [],\n \"source\": [\n \"# passage retrieval evaluation \\n\",\n \"def fever_retrieval_eval(results, topk=5):\\n\",\n \" \\n\",\n \" precs, recalls = [], []\\n\",\n \" for item in results:\\n\",\n \" gold = item[\\\"correct_normalized\\\"]\\n\",\n \" pred = item[\\\"bm25_topk\\\"][:topk]\\n\",\n \" \\n\",\n \" if len(gold) > 0:\\n\",\n \" prec = np.mean([int(doc in gold) for doc in pred])\\n\",\n \" else:\\n\",\n \" prec = 1\\n\",\n \" recall = 0\\n\",\n \" for chain in item[\\\"evidence\\\"]:\\n\",\n \" chain_titles = set([normalize(p[\\\"title\\\"]) for p in chain])\\n\",\n \" chain_covered = [int(t in pred) for t in chain_titles]\\n\",\n \" if np.sum(chain_covered) == len(chain_titles):\\n\",\n \" recall = 1\\n\",\n \" break\\n\",\n \" precs.append(prec)\\n\",\n \" recalls.append(recall)\\n\",\n \" pr, rec = np.mean(precs), np.mean(recalls)\\n\",\n \" return pr, rec, 2.0 * pr * rec / (pr + rec)\"\n ]\n },\n {\n \"cell_type\": \"code\",\n \"execution_count\": 42,\n \"metadata\": {},\n \"outputs\": [\n {\n \"name\": \"stdout\",\n \"output_type\": \"stream\",\n \"text\": [\n \"(0.12268811028144745, 0.5020103388856979, 0.19718537768662206)\\n\"\n ]\n }\n ],\n \"source\": [\n \"tfidf_results = [json.loads(l) for l in open(\\\"/private/home/xwhan/data/fever/retrieval/multi_dev_tfidf.txt\\\").readlines()]\\n\",\n \"print(fever_retrieval_eval(tfidf_results, 10))\"\n ]\n },\n {\n \"cell_type\": \"code\",\n \"execution_count\": 234,\n \"metadata\": {},\n \"outputs\": [],\n \"source\": []\n },\n {\n \"cell_type\": \"code\",\n \"execution_count\": 244,\n \"metadata\": {},\n \"outputs\": [\n {\n \"name\": \"stdout\",\n \"output_type\": \"stream\",\n \"text\": [\n \"1741\\n\",\n \"1741\\n\",\n \"(0.6223243346735593, 0.46927053417576103, 0.5350675077296432)\\n\"\n ]\n }\n ],\n \"source\": [\n \"# phrase_matching_results = [json.loads(l) for l in open(\\\"/private/home/xwhan/data/fever/retrieval/all_dev.json\\\").readlines()]\\n\",\n \"# phrase_matching_results = [json.loads(l.decode('utf-8').strip('\\\\r\\\\n')) for l in open(\\\"/private/home/xwhan/data/fever/retrieval/dev.ensembles.s10.jsonl\\\").readlines()]\\n\",\n \"phrase_matching_results = [json.loads(l) for l in open(\\\"/private/home/xwhan/code/Transformer-XH/data/fever_dev_graph.json\\\").readlines()]\\n\",\n \"# for _ in phrase_matching_results:\\n\",\n \"# _[\\\"id\\\"] = _[\\\"qid\\\"]\\n\",\n \"\\n\",\n \"phrase_matching_results = [_ for _ in phrase_matching_results if _[\\\"id\\\"] in multihop_ids]\\n\",\n \"# phrase_matching_results = [json.loads(l) for l in open(\\\"/private/home/xwhan/data/fever/retrieval/esim_mhop_dev.json\\\").readlines()]\\n\",\n \"\\n\",\n \"# json.dump(phrase_matching_results, open(\\\"/private/home/xwhan/data/fever/retrieval/dev_el_wiki_pages.jsonl\\\", \\\"w\\\"))\\n\",\n \"\\n\",\n \"print(len(phrase_matching_results))\\n\",\n \"\\n\",\n \"tfidf_results = [json.loads(l) for l in open(\\\"/private/home/xwhan/data/fever/retrieval/multi_dev_tfidf.txt\\\").readlines()]\\n\",\n \"print(len(tfidf_results))\\n\",\n \"pred_lens = []\\n\",\n \"def fever_retrieval_eval_phrase(tfidf_results, phrase_results, topk=5):\\n\",\n \" id2gold = {_[\\\"id\\\"]:_[\\\"correct_normalized\\\"] for _ in tfidf_results}\\n\",\n \" id2gold_evidence = {_[\\\"id\\\"]:_[\\\"evidence\\\"] for _ in tfidf_results}\\n\",\n \" precs, recalls = [], []\\n\",\n \" for item in phrase_results:\\n\",\n \" gold = id2gold[item[\\\"id\\\"]]\\n\",\n \"# print(gold)\\n\",\n \" retrieved_evidence = item[\\\"evidence\\\"] \\n\",\n \" pred = []\\n\",\n \" for e in retrieved_evidence:\\n\",\n \" pred.append(normalize(e[0]))\\n\",\n \" \\n\",\n \"# pred = item[\\\"predicted_pages\\\"] + item[\\\"wiki_results\\\"]\\n\",\n \"# pred = item[\\\"wiki_results\\\"]\\n\",\n \"# pred = item[\\\"predicted_pages\\\"]\\n\",\n \" \\n\",\n \"# pred = []\\n\",\n \"# for n in item[\\\"node\\\"]:\\n\",\n \"# pred.append(n[\\\"name\\\"])\\n\",\n \" \\n\",\n \" pred = list(set(pred))\\n\",\n \" pred_lens.append(len(pred))\\n\",\n \" \\n\",\n \" if len(gold) > 0:\\n\",\n \" if len(pred) == 0:\\n\",\n \" prec = 0\\n\",\n \" else:\\n\",\n \" prec = np.mean([int(doc in gold) for doc in pred])\\n\",\n \" else:\\n\",\n \" prec = 1\\n\",\n \" recall = 0\\n\",\n \" for chain in id2gold_evidence[item[\\\"id\\\"]]:\\n\",\n \" chain_titles = set([normalize(p[\\\"title\\\"]) for p in chain])\\n\",\n \" chain_covered = [int(t in pred) for t in chain_titles]\\n\",\n \" if np.sum(chain_covered) == len(chain_titles):\\n\",\n \" recall = 1\\n\",\n \" break\\n\",\n \" precs.append(prec)\\n\",\n \" recalls.append(recall)\\n\",\n \" \\n\",\n \" pr, rec = np.mean(precs), np.mean(recalls)\\n\",\n \" return pr, rec, 2.0 * pr * rec / (pr + rec)\\n\",\n \"print(fever_retrieval_eval_phrase(tfidf_results, phrase_matching_results, topk=5))\"\n ]\n },\n {\n \"cell_type\": \"code\",\n \"execution_count\": 181,\n \"metadata\": {},\n \"outputs\": [\n {\n \"data\": {\n \"text/plain\": [\n \"2.1039632395175185\"\n ]\n },\n \"execution_count\": 181,\n \"metadata\": {},\n \"output_type\": \"execute_result\"\n }\n ],\n \"source\": [\n \"np.mean(pred_lens)\"\n ]\n },\n {\n \"cell_type\": \"code\",\n \"execution_count\": 153,\n \"metadata\": {},\n \"outputs\": [\n {\n \"data\": {\n \"text/plain\": [\n \"0.12426242624262426\"\n ]\n },\n \"execution_count\": 153,\n \"metadata\": {},\n \"output_type\": \"execute_result\"\n }\n ],\n \"source\": [\n \"test_with_all = [json.loads(l) for l in open(\\\"/private/home/xwhan/data/fever/retrieval/all_dev.json\\\").readlines()]\\n\",\n \"title_count = []\\n\",\n \"for item in test_with_all:\\n\",\n \" titles = set()\\n\",\n \" for e in item[\\\"evidence\\\"]:\\n\",\n \" titles.add(e[0])\\n\",\n \" title_count.append(len(titles))\\n\",\n \"np.sum(np.array(title_count) > 7) / len(title_count)\"\n ]\n },\n {\n \"cell_type\": \"code\",\n \"execution_count\": 154,\n \"metadata\": {},\n \"outputs\": [\n {\n \"data\": {\n \"text/plain\": [\n \"0.11321132113211321\"\n ]\n },\n \"execution_count\": 154,\n \"metadata\": {},\n \"output_type\": \"execute_result\"\n }\n ],\n \"source\": [\n \"test_with_all = [json.loads(l) for l in open(\\\"/private/home/xwhan/data/fever/retrieval/all_test.json\\\").readlines()]\\n\",\n \"title_count = []\\n\",\n \"for item in test_with_all:\\n\",\n \" titles = set()\\n\",\n \" for e in item[\\\"evidence\\\"]:\\n\",\n \" titles.add(e[0])\\n\",\n \" title_count.append(len(titles))\\n\",\n \"np.sum(np.array(title_count) > 7) / len(title_count)\"\n ]\n },\n {\n \"cell_type\": \"code\",\n \"execution_count\": 91,\n \"metadata\": {},\n \"outputs\": [\n {\n \"name\": \"stdout\",\n \"output_type\": \"stream\",\n \"text\": [\n \"1960\\n\",\n \"1960\\n\"\n ]\n }\n ],\n \"source\": [\n \"# build dense final prediction for evaluation\\n\",\n \"final_pred = [json.loads(l) for l in open(\\\"/private/home/xwhan/code/KernelGAT/kgat/output/dense_bert_dev_mhop_top4.json\\\").readlines()]\\n\",\n \"final_retrieval = [json.loads(l) for l in open(\\\"/private/home/xwhan/code/KernelGAT/data/bert_dense_top4_mhop_sents.json\\\").readlines()]\\n\",\n \"all_dev_gold = [json.loads(l) for l in open(\\\"/private/home/xwhan/data/fever/shared_task_dev.jsonl\\\").readlines()]\\n\",\n \"id2gold = {_[\\\"id\\\"]:_ for _ in all_dev_gold}\\n\",\n \"\\n\",\n \"print(len(final_retrieval))\\n\",\n \"final = []\\n\",\n \"for pred, retrieval in zip(final_pred, final_retrieval):\\n\",\n \" assert pred[\\\"id\\\"] == retrieval[\\\"id\\\"]\\n\",\n \" final.append({\\n\",\n \" \\\"id\\\": pred[\\\"id\\\"],\\n\",\n \" \\\"label\\\": id2gold[pred[\\\"id\\\"]][\\\"label\\\"],\\n\",\n \" \\\"evidence\\\": id2gold[pred[\\\"id\\\"]][\\\"evidence\\\"],\\n\",\n \" \\\"predicted_label\\\": pred[\\\"predicted_label\\\"],\\n\",\n \" \\\"predicted_evidence\\\": [[normalize(e[0]), int(e[1])] for e in retrieval[\\\"evidence\\\"][:5]]\\n\",\n \" })\\n\",\n \"\\n\",\n \"print(len(final))\\n\",\n \"with open(\\\"/private/home/xwhan/data/fever/results/dense_top4_mhop_dev.json\\\", \\\"w\\\") as g:\\n\",\n \" for l in final:\\n\",\n \" g.write(json.dumps(l) + \\\"\\\\n\\\")\"\n ]\n },\n {\n \"cell_type\": \"code\",\n \"execution_count\": 92,\n \"metadata\": {},\n \"outputs\": [\n {\n \"name\": \"stdout\",\n \"output_type\": \"stream\",\n \"text\": [\n \"1960\\n\",\n \"1960\\n\"\n ]\n }\n ],\n \"source\": [\n \"# build EL final prediction for evaluation\\n\",\n \"final_pred = [json.loads(l) for l in open(\\\"/private/home/xwhan/code/KernelGAT/kgat/output/el_bert_dev_mhop.json\\\").readlines()]\\n\",\n \"final_retrieval = [json.loads(l) for l in open(\\\"/private/home/xwhan/code/KernelGAT/data/bert_dev_multi_el.json\\\").readlines()]\\n\",\n \"all_dev_gold = [json.loads(l) for l in open(\\\"/private/home/xwhan/data/fever/shared_task_dev.jsonl\\\").readlines()]\\n\",\n \"id2gold = {_[\\\"id\\\"]:_ for _ in all_dev_gold}\\n\",\n \"\\n\",\n \"print(len(final_retrieval))\\n\",\n \"final = []\\n\",\n \"for pred, retrieval in zip(final_pred, final_retrieval):\\n\",\n \" assert pred[\\\"id\\\"] == retrieval[\\\"id\\\"]\\n\",\n \" final.append({\\n\",\n \" \\\"id\\\": pred[\\\"id\\\"],\\n\",\n \" \\\"label\\\": id2gold[pred[\\\"id\\\"]][\\\"label\\\"],\\n\",\n \" \\\"evidence\\\": id2gold[pred[\\\"id\\\"]][\\\"evidence\\\"],\\n\",\n \" \\\"predicted_label\\\": pred[\\\"predicted_label\\\"],\\n\",\n \" \\\"predicted_evidence\\\": [[normalize(e[0]), int(e[1])] for e in retrieval[\\\"evidence\\\"][:5]]\\n\",\n \" })\\n\",\n \"\\n\",\n \"print(len(final))\\n\",\n \"with open(\\\"/private/home/xwhan/data/fever/results/el_mhop_dev.json\\\", \\\"w\\\") as g:\\n\",\n \" for l in final:\\n\",\n \" g.write(json.dumps(l) + \\\"\\\\n\\\")\"\n ]\n },\n {\n \"cell_type\": \"code\",\n \"execution_count\": 93,\n \"metadata\": {},\n \"outputs\": [\n {\n \"name\": \"stdout\",\n \"output_type\": \"stream\",\n \"text\": [\n \"1960\\n\"\n ]\n }\n ],\n \"source\": [\n \"final_pred = [json.loads(l) for l in open(\\\"/private/home/xwhan/code/KernelGAT/kgat/output/esim_mhop_dev.json\\\").readlines()]\\n\",\n \"final_retrieval = [json.loads(l) for l in open(\\\"/private/home/xwhan/data/fever/retrieval/esim_mhop_dev.json\\\").readlines()]\\n\",\n \"all_dev_gold = [json.loads(l) for l in open(\\\"/private/home/xwhan/data/fever/shared_task_dev.jsonl\\\").readlines()]\\n\",\n \"id2gold = {_[\\\"id\\\"]:_ for _ in all_dev_gold}\\n\",\n \"\\n\",\n \"final = []\\n\",\n \"for pred, retrieval in zip(final_pred, final_retrieval):\\n\",\n \" assert pred[\\\"id\\\"] == retrieval[\\\"id\\\"]\\n\",\n \" final.append({\\n\",\n \" \\\"id\\\": pred[\\\"id\\\"],\\n\",\n \" \\\"label\\\": id2gold[pred[\\\"id\\\"]][\\\"label\\\"],\\n\",\n \" \\\"evidence\\\": id2gold[pred[\\\"id\\\"]][\\\"evidence\\\"],\\n\",\n \" \\\"predicted_label\\\": pred[\\\"predicted_label\\\"],\\n\",\n \" \\\"predicted_evidence\\\": [[e[0], int(e[1])] for e in retrieval[\\\"evidence\\\"][:5]]\\n\",\n \" })\\n\",\n \"\\n\",\n \"print(len(final))\\n\",\n \"with open(\\\"/private/home/xwhan/data/fever/results/esim_mhop_dev.json\\\", \\\"w\\\") as g:\\n\",\n \" for l in final:\\n\",\n \" g.write(json.dumps(l) + \\\"\\\\n\\\")\"\n ]\n }\n ],\n \"metadata\": {\n \"kernelspec\": {\n \"display_name\": \"Python 3\",\n \"language\": \"python\",\n \"name\": \"python3\"\n },\n \"language_info\": {\n \"codemirror_mode\": {\n \"name\": \"ipython\",\n \"version\": 3\n },\n \"file_extension\": \".py\",\n \"mimetype\": \"text/x-python\",\n \"name\": \"python\",\n \"nbconvert_exporter\": \"python\",\n \"pygments_lexer\": \"ipython3\",\n \"version\": \"3.6.3\"\n }\n },\n \"nbformat\": 4,\n \"nbformat_minor\": 2\n}\n" }, { "path": "mdr/retrieval/hotpot.ipynb", "content": "{\n \"cells\": [\n {\n \"cell_type\": \"code\",\n \"execution_count\": 2,\n \"metadata\": {},\n \"outputs\": [],\n \"source\": [\n \"import json\\n\",\n \"\\n\",\n \"test_qas = json.load(open(\\\"/private/home/xwhan/data/hotpot/hotpot_test_fullwiki_v1.json\\\"))\\n\",\n \"test_results = json.load(open(\\\"/private/home/xwhan/data/hotpot/results/hotpot_test_b200_k500.json\\\"))\"\n ]\n },\n {\n \"cell_type\": \"code\",\n \"execution_count\": 4,\n \"metadata\": {},\n \"outputs\": [\n {\n \"data\": {\n \"text/plain\": [\n \"True\"\n ]\n },\n \"execution_count\": 4,\n \"metadata\": {},\n \"output_type\": \"execute_result\"\n }\n ],\n \"source\": [\n \"len(test_qas) == len(test_results[\\\"answer\\\"])\"\n ]\n },\n {\n \"cell_type\": \"code\",\n \"execution_count\": 15,\n \"metadata\": {},\n \"outputs\": [\n {\n \"name\": \"stdout\",\n \"output_type\": \"stream\",\n \"text\": [\n \"Who has been in more bands, Deron Miller or Steve Marriott?\\n\",\n \"Deron John Miller\\n\"\n ]\n }\n ],\n \"source\": [\n \"import random\\n\",\n \"qid2question = {_[\\\"_id\\\"]:_[\\\"question\\\"] for _ in test_qas}\\n\",\n \"qids = list(test_results[\\\"answer\\\"].keys())\\n\",\n \"random.shuffle(qids)\\n\",\n \"\\n\",\n \"\\n\",\n \"\\n\",\n \"print(qid2question[qids[0]])\\n\",\n \"print(test_results[\\\"answer\\\"][qids[0]])\"\n ]\n },\n {\n \"cell_type\": \"code\",\n \"execution_count\": 63,\n \"metadata\": {},\n \"outputs\": [\n {\n \"name\": \"stdout\",\n \"output_type\": \"stream\",\n \"text\": [\n \"Collecting seaborn\\r\\n\",\n \" Using cached https://files.pythonhosted.org/packages/c7/e6/54aaaafd0b87f51dfba92ba73da94151aa3bc179e5fe88fc5dfb3038e860/seaborn-0.10.1-py3-none-any.whl\\r\\n\",\n \"Collecting matplotlib>=2.1.2 (from seaborn)\\r\\n\",\n \" Using cached https://files.pythonhosted.org/packages/96/a7/b6fa244fd8a8814ef9408c8a5a7e4ed0340e232a6f0ce2046b42e50672c0/matplotlib-3.3.1-cp36-cp36m-manylinux1_x86_64.whl\\r\\n\",\n \"Requirement already satisfied: scipy>=1.0.1 in /public/apps/anaconda3/5.0.1/lib/python3.6/site-packages (from seaborn)\\r\\n\",\n \"Requirement already satisfied: numpy>=1.13.3 in /public/apps/anaconda3/5.0.1/lib/python3.6/site-packages (from seaborn)\\r\\n\",\n \"Requirement already satisfied: pandas>=0.22.0 in /public/apps/anaconda3/5.0.1/lib/python3.6/site-packages (from seaborn)\\r\\n\",\n \"Requirement already satisfied: python-dateutil>=2.1 in /public/apps/anaconda3/5.0.1/lib/python3.6/site-packages (from matplotlib>=2.1.2->seaborn)\\r\\n\",\n \"Collecting pillow>=6.2.0 (from matplotlib>=2.1.2->seaborn)\\r\\n\",\n \" Using cached https://files.pythonhosted.org/packages/30/bf/92385b4262178ca22b34f82e0e09c2922eb351fe39f3cc7b8ba9ea555b41/Pillow-7.2.0-cp36-cp36m-manylinux1_x86_64.whl\\r\\n\",\n \"Collecting certifi>=2020.06.20 (from matplotlib>=2.1.2->seaborn)\\r\\n\",\n \" Using cached https://files.pythonhosted.org/packages/5e/c4/6c4fe722df5343c33226f0b4e0bb042e4dc13483228b4718baf286f86d87/certifi-2020.6.20-py2.py3-none-any.whl\\r\\n\",\n \"Requirement already satisfied: pyparsing!=2.0.4,!=2.1.2,!=2.1.6,>=2.0.3 in /public/apps/anaconda3/5.0.1/lib/python3.6/site-packages (from matplotlib>=2.1.2->seaborn)\\r\\n\",\n \"Requirement already satisfied: cycler>=0.10 in /public/apps/anaconda3/5.0.1/lib/python3.6/site-packages (from matplotlib>=2.1.2->seaborn)\\r\\n\",\n \"Collecting kiwisolver>=1.0.1 (from matplotlib>=2.1.2->seaborn)\\r\\n\",\n \" Using cached https://files.pythonhosted.org/packages/ae/23/147de658aabbf968324551ea22c0c13a00284c4ef49a77002e91f79657b7/kiwisolver-1.2.0-cp36-cp36m-manylinux1_x86_64.whl\\r\\n\",\n \"Requirement already satisfied: pytz>=2011k in /public/apps/anaconda3/5.0.1/lib/python3.6/site-packages (from pandas>=0.22.0->seaborn)\\r\\n\",\n \"Requirement already satisfied: six>=1.5 in /public/apps/anaconda3/5.0.1/lib/python3.6/site-packages (from python-dateutil>=2.1->matplotlib>=2.1.2->seaborn)\\r\\n\",\n \"Installing collected packages: pillow, certifi, kiwisolver, matplotlib, seaborn\\r\\n\",\n \"Successfully installed certifi-2020.6.20 kiwisolver-1.2.0 matplotlib-3.3.1 pillow-7.2.0 seaborn-0.10.1\\r\\n\",\n \"\\u001b[33mYou are using pip version 9.0.1, however version 20.2.2 is available.\\r\\n\",\n \"You should consider upgrading via the 'pip install --upgrade pip' command.\\u001b[0m\\r\\n\"\n ]\n }\n ],\n \"source\": [\n \"# Install a pip package in the current Jupyter kernel\\n\",\n \"import sys\\n\",\n \"!{sys.executable} -m pip install seaborn --user\"\n ]\n },\n {\n \"cell_type\": \"code\",\n \"execution_count\": 119,\n \"metadata\": {},\n \"outputs\": [\n {\n \"data\": {\n \"image/png\": \"iVBORw0KGgoAAAANSUhEUgAAAYsAAAESCAYAAAAMifkAAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz\\nAAALEgAACxIB0t1+/AAAADl0RVh0U29mdHdhcmUAbWF0cGxvdGxpYiB2ZXJzaW9uIDIuMS4xLCBo\\ndHRwOi8vbWF0cGxvdGxpYi5vcmcvAOZPmwAAIABJREFUeJzs3XlcVFX/wPHPzMAM44LgArK5pIlY\\nLii5ofaIuS+4Zy6ZO6W2KKZmaWpqrpVZz5M/pSxNKyXMLSrNJTOXwiVFMzUUEFQUkXVg5v7+QG5O\\n7Mhi8n2/Xrxkzj333HOver5zzz33HI2iKApCCCFEHrRlXQEhhBAPPgkWQggh8iXBQgghRL4kWAgh\\nhMiXBAshhBD5kmAhhBAiXxIshBBC5EuChRBCiHzZlMZBIiMjmThxovr5zp07JCYmcuTIES5dusSM\\nGTOIj4/HwcGBxYsXU6dOndKolhBCiALSlMUb3AsWLMBsNjN79myeffZZBgwYgL+/P1u3bmXLli18\\n+umnpV0lIYQQeSj1biiTycS2bdsYMGAAcXFxnDlzhl69egHQq1cvzpw5w82bN0u7WkIIIfJQKt1Q\\n99qzZw/Ozs489thj/P777zg7O6PT6QDQ6XQ4OTlx9epVqlatarVfQkICCQkJVmkmk4krV65Qp04d\\ntQwhhBB5M5vNXL9+nccffxw7O7sC7VPqwWLLli0MGDCg0PutW7eOVatWlUCNhBCifNqwYQM+Pj4F\\nyluqwSI2NpajR4+yZMkSAFxcXIiNjcVsNqPT6TCbzVy7dg0XF5ds+44cOZJ+/fpZpUVFRfHss8+y\\nYcMGatasWSrnIIQQ/3YxMTEMGzaMGjVqFHifUg0WX3/9NU8++SSOjo4AVKtWDS8vL7Zv346/vz/b\\nt2/Hy8srWxcUgL29Pfb29jmWW7NmTdzd3Uu07kII8bApTPd9qT7g/vrrr7N1Qb355pusX7+erl27\\nsn79eubOnVuaVRJCCFEApXpnERoami2tXr16fPXVV6VZDSGEEIUkb3ALIYTIlwQLIYQQ+ZJgIYQQ\\nIl8SLIQQQuRLgoUQQoh8SbAQQgiRLwkWQggh8iXBQgghRL4kWAghhMhXqc86K4QQouQdGTmG9Pj4\\nbOm2Dg64Lij8tEpyZyGEEA8ZRVFyDBRArun5kTsLIYQoBYqigMWCRqfDkp5O2vXrWNJMWEwmzGlp\\nWEwmKtTywM7JCdPNW1w/8BMWkwnL3W2WtDScOz9Fpfr1SPzzAn+t+0zdZk7L3N7glRep0vhx4n7+\\npdjrL8FCCFFuKYqS2RCbTFjSTGhtbbCtUgVFUbh96ve/G+u0zAa9Qi0PqjzWCIvJRMT6z+9prDP/\\nrN7OF6f/PIkpPp5TM19Xg4ElLQ1Lejp1Rj2Lm38fUmNjCZv4Urb61HthAjW7dsF08yZ/BX2SmajV\\notXr0Rn0OHg3o1L9egBY0tPR2tlhW6UKWr0ercGATeVKAFSsU7vYr5UECyHEAykjKQlzSqpVg6y1\\ntaVSvUcAiPvlMKZb8Vbfvg1OTtTs8hQAf374Eaa4OKvtVRo/Tt0xowA48twY0m9Zd8k4PeXHo5Mn\\nAnB6zjywWKy2u/TsQZXHGgEQE/o9OkNmI63V69Hq9ZhTUgHQ6vVUqvcIWr0BrUF/t7E3ULlBAwD0\\nVavx6Csvosva9+6fdjWdAahYtw6tPv8UrV6PxsYGjUZjVY9K9evR5O0FuV47o5troa93fiRYCCHy\\npVgsWNLTsaSZsLWvDEDa9euYbt7K/FZ9tzFWzBZqdGgHwI2Dh0i6ePHvbhKTCZ2dHfUCxgFwcU0Q\\nCb+fvru/CYspDYOTM81WZK6keWbuAu6cO2dVj0qPPkrTZW8DcHnjFyT/FfH3Rq0WR++marBIu3aN\\n9ISEzIbaaMS2ShVs7y68BlCzW1ewWP5urA16KtxdRE2j0fD4W3PR2tqqDbnOYEBXwZh5KL2eNl9s\\nyPV62VSogGfglDy2G3H6z5O5btfodNhUrJjr9rIgwUKIIsprtEnLdWtL/PhqF8o9XR3mtDQquLuh\\n1etJuXqVpIuX1C4UiymzUXbp1RObCkbiDh8l7tAvf/eJ3y3jsflvojMYuLzpS65u36GWn6VtyGY0\\nGg1XvtxM7Hc/WNVJa2enBou4X37hxk8/q10oWoMBQ/Xqal6bihUxONWw+mZtqFZN3e7Wrw/pd+6g\\n1RvU/W0qV1a3N5o9C41Gq35z/+c38MfefCPP61dryOA8t2fdQfxb2To45PrvsygkWAhRRPmNNrGY\\nTKRdv2HVp21JS6PiI3UxVKtG6rVrdxtrk9WDStdePahQy4Pbp09zZeOX/9jfhNesGVSq9wjXftjN\\nn6v+m+343u+/Q4Vatbh17DcurQnKtr3Gkx2wqWAk7do1Ek6fvttVktkg64xGFLMZgIq1a1G9fTv1\\nW3VWg47FAjodLj27U61N67uNtUENClkavPwiDaa8nK0LJUutZ57O8/pWa9M6z+33BhaRXV5fWCIj\\nIwtdngQLUW4pioI5JQVzSirmlBQsqZl/6qtXx+hSk4zkFGK//17dbk5JxZKaSvUO7ajq0yLf8pMu\\n/cXJV2dmS28w9RVqdGhHakwsfwWty0xUH2IaqO7bhgq1PDLraDZndqE4VPm7O8RoB0ClR+tTe8Qw\\ntQsl8xu4Af3dRrRGh3Y4NG18tz/dYPUNHMC1d09ce/fMtf7V2rTOs8GuWKcOFevUyXW7phDrO4sH\\nnwQL8a+gKApKejrm1MwHiLb29gDEnzhJRlKS2tCbU1IxerhTrVVLFEXh3NIVWFKzAkJmnhr/6UCt\\nZ57GkprK4WdGZDuW++CB1B72DBaTSW3Ms/q9dUY77B9/rEB1tnNxocGUl+92s+jVBt3oUhMA+0Ze\\neT7ErPLYYzRe9Fau5efXWNtWqYJtlSoFqqsQ+ZFg8ZAq6/70LOm3b5N+JzGzMb/boGttbXFo1hSA\\nq7tCSYuNzWzo7243urlRZ2RmI35y+mukREZhTklRu0eqtmqJ12vTAfhj+Tuk306wOmaNJztQrVVL\\nNBoNqdFX0ei06IxGDDWqo7Wzw+5uY601GKgzaiQ6ox06u8xAoLWzUxtzW/vKtPr8U3R2dkX6lmxr\\nX5kaT7bPdbvWxgatjfwXFP8O8i/1IVXYtzct6el/d8fcHU8OcOfcH6RERd/91p7ZoKPRUHvYM0Dm\\niJTbJ0/93VWTmore0YFm7y4H4NzSFdw+9bvVsSrWraNuv7b7R5IjIjIbbKMxc9y4w98jVqo0fpxK\\n9epZbTe6uqjbvd6YhdbG5m5Df7fB1//db97s3WW5XiONVotb3z55bn/QRqQIUVYkWJRDJ6bNoPHC\\n+Whtbfnrk0+J3rYDJSPj7wxaLW2Dv0Sj0RAT+j3Xdu+x2qZ3dFCDhZKRkZlWvZraoOurVlWzu/Xz\\nx+mpTmoXjs5oxKZSJXV7kyUL0Whzn3Wm9vCheZ5L5UfrF/Lsi09xjzYR4kFWasEiLS2NhQsXcujQ\\nIQwGA82aNWP+/Pn4+fmh1+sxGAwABAYG0r597rfuImem+HjuhJ8lIfwsCWfO5pnXpmLFzEbe1hb7\\nRl5odDq0dpkNue7unygKaDTUGvYMHoMHWH1rv7dvvfaIYXkey7FF8zy35xUoHnSl2Z0nRFkrtWCx\\ndOlSDAYDoaGhaDQabty4oW5buXIlDe6+2SjypygKitmM1saGhPCznF+5itToqwBobG2p3ODRPPe/\\nd/x51ZZPULXlE7nmNVSrmus2IUT5USrBIikpiZCQEPbt26d+K61+z8s5Im+W9HSSLl7KvGsIP8ud\\n8HBqDX2Gmt26oK/qSAV3d5w7P4V9I6/MKQZsbTnoP6Csqy2EeIiUSrC4cuUKDg4OrFq1isOHD1Ox\\nYkVeeuklfHx8gMyuJ0VRaNGiBVOmTMH+7rDIeyUkJJCQYD3qJSYmpjSqX+oykpPJuJOInbMT5pQU\\njowcgyUtDQC7mjVxaN4cu7sPee2cnfGaNSNbGdKfLoQoTqUSLDIyMrhy5QqNGjVi+vTpnDhxgoCA\\nAL7//ns2bNiAi4sLJpOJBQsWMG/ePJYtyz6CZd26daxatao0qlvq0uLiSDiTeceQcOYsSRERODb3\\nptEbr6EzGnEf2B+jmxv2Xg3RV3XMv0CkP10IUbxKJVi4urpiY2NDr169AGjatCmOjo5cunSJxo0b\\nA6DX6xk6dCjPP/98jmWMHDmSfv36WaXFxMQwbFjeD1gfNIrFQvKVSFKioqjetg2QObz0TvhZtHZ2\\nVPZsgMfTg6jS+HF1H4/BA8uqukIIAZRSsKhatSqtWrXi4MGDtGvXjkuXLhEXF4eTkxN37tyhcuXK\\nKIrCzp078fLyyrEMe3v7HLun/g2SIi5z6+gx9ZmDOSkJjY0Nji2aozMYqDNyBFpbWyrWrSNTJAgh\\nHkilNhpq7ty5vPbaayxevBgbGxuWLFmCyWRiwoQJmM1mLBYL9erVY86cOaVVpUIryFvR6Ql3Mh9C\\nnz2LWz9/bO3tufXrb0R8tgGjuzvVfdtg7+WFfaOG6stj9l4NS/U8hBCisEotWHh4ePDZZ59lSw8J\\nCSmtKty3vN6K/vOD/5Jw5iwpd2dz1NjY4NjcmyqNH8e5cyecn/JT5zMSQoh/G3mDu5jcOHgIe6+G\\nOHV8kspeDalUvx66uy8a2t4zB78QQvwbSbAoJq3Wf/KvfhtZCCHyIq1bAWUkJuW5XQKFEOJhJi1c\\nAd089mtZV0EIIcqMBIs8KBYLyZcvA1DjyfZW6//eS96KFkI87OSZRS7SExI4/94qbp/6neYfvIeh\\nRg1arf+krKslhBBlotwHi9zenUCjQaPTUXf0c+hl0kMhRDlX7oNFbu9OoCg0WbqISo88UroVEkKI\\nB5A8s8iDBAohhMgkwUIIIUS+JFgIIYTIlwQLIYQQ+Sr3wUJrZ5djurw7IYQQfyvXo6FSrsaAxUK1\\nNq1oOOPVsq6OEEI8sMpdsMjpvYq4Q4c5MnKMLEUqBBASFsXS0HNEx6fg6mBkWldP+nq7lXW1RBkr\\nd8EirzUphCjvQsKimBl8ipR0MwBR8SnMDD4FIAGjnCv3zyyEEH9bGnpODRRZUtLNLA09V0Y1Eg8K\\nCRZCCFV0fEqh0kX5IcFCCKFydTAWKl2UHxIshBCqaV09MdrqrNKMtjqmdfUsoxqJB0W5Cxa5vT9R\\nnO9V+Pn58fPPP1ulBQcH88wzz6ifd+zYwaBBg2jWrBlt2rRh0KBBbNiwAUVRAJgxYwaenp6cPHlS\\n3SciIgJPT/lPK0pOX283FvVvjJuDEQ3g5mBkUf/G8nBblL/RUC3XreXMWwtJu3Yd75XvlEkdgoKC\\nWLNmDbNnz6Zdu3ZUrFiR8PBw1q5dy6BBg9Dr9QA4ODjw7rvvEhQUVCb1FOVTX283CQ4im1ILFmlp\\naSxcuJBDhw5hMBho1qwZ8+fP59KlS8yYMYP4+HgcHBxYvHgxderUKdG6VGvTGiU9o0SPkZs7d+6w\\ncuVKFi9eTNeuXdX0Ro0asXz5cqu8ffv2Zfv27Rw5coSWLVuWdlWFEEJVasFi6dKlGAwGQkND0Wg0\\n3LhxA4A5c+YwdOhQ/P392bp1K7Nnz+bTTz8t0bo4d/K7r/3v56WlsLAwTCYTnTp1yjevnZ0dEyZM\\n4J133mHjxo33VWchhLgfpRIskpKSCAkJYd++fWg0GgCqV69OXFwcZ86c4eOPPwagV69ezJ8/n5s3\\nb1K1atUSqUtGcgoZdxIw1KiBRlv4RzYFfWlp4sSJ6HR/PyhMT0+nUaNG3Lp1C0dHR2xs/r70Q4YM\\n4c8//8RkMrF27VqeeOIJq21BQUHs27evxO+4hBAiN6XygPvKlSs4ODiwatUq+vfvz4gRIzh27BhX\\nr17F2dlZbVR1Oh1OTk5cvXo1WxkJCQlERkZa/cTExBS6LrdPnOTX8S+QeOFikc6loC8tffDBBxw7\\ndkz9mTNnDpD5HOLWrVtkZPzdDbZp0yaOHTuGg4MDFovFqhy9Xs8LL7zAe++9pz78FkKI0lYqwSIj\\nI4MrV67QqFEjgoODCQwMZPLkySQnJxe4jHXr1tGpUyern2HDhhW6LilRUQAY3VwLvS/c/0tL3t7e\\n6PV6du/eXeBj9u/fn8TERL7//vsC7yPKt/fff5/AwMCyroa469ixY1bPKP+NSqUbytXVFRsbG3r1\\n6gVA06ZNcXR0xM7OjtjYWMxmMzqdDrPZzLVr13BxcclWxsiRI+nXr59VWkxMTKEDRkpUNLaOjthU\\nqFC0c3EwEpVDYCjoS0v29vZMnDiRuXPnoigK7du3x2g0cu7cOVJScg44NjY2TJo0iQULFhSpzqLs\\n7dixg08++YTz589jNBpxd3enb9++DB06VO2aLUuenp4YjUY0Gg2VKlWiR48evPrqq1Zdqf9Ghw8f\\nZtq0aezfvz/XPDNmzGD79u3Y2tpia2vLY489xuuvv069evUKdAxPT0++++47ateunWseHx8fQkND\\nC13/B0mp3FlUrVqVVq1acfDgQQAuXbpEXFwcderUwcvLi+3btwOwfft2vLy8cnxeYW9vj7u7u9VP\\nzZo1C12XlKioIt9VQPG8tDRu3DhmzJjBmjVraNu2LW3btmX27NkEBgbi7e2d4z69evWiRo0aRa63\\nKDtBQUEsWLCAMWPG8NNPP/Hzzz8zd+5cfvvtN9LT03Pcx2w255hekrZu3UpYWBjr169n586dbNmy\\npVSPf2/XbF5pJWHMmDGEhYWxf/9+nJ2dmTVrVrGVXVrnUOLHVUrJ5cuXleHDhyu9evVS+vbtq+zd\\nu1dRFEX5888/lYEDBypdunRRBg4cqFy4cKHAZV65ckVp0KCBcuXKlQLlt1gsyi9Dn1X+/PB/RTqH\\nLF//Fqm0XbRbqTN9u9J20W7l698i76s88fBKSEhQmjZtqnz77bd55ps+fboye/ZsZezYsUrTpk2V\\ngwcPKj/++KPi7++veHt7Kx06dFBWrlyp5s/6t79p0ybF19dX8fX1VdauXatuX7lypfLiiy8q06ZN\\nU5o1a6b06NFDOXnyZK7Hb9CggfLXX3+pn1988UXlzTfftDqPmTNnKr6+vkq7du2UFStWKBkZGer2\\nL774QunWrZvSrFkzpXv37srvv/+eY7nTp09XVqxYoSiKovzyyy9K+/btlY8++khp27atEhgYmGOa\\noijKnj17lD59+igtWrRQnn76aSU8PFwts2PHjsqaNWuUXr16Kc2bN1deeuklJTU1VUlKSlIaN26s\\neHp6Ks2aNVOaNWumxMTE5Hjts+qkKIqyd+9epWnTplZ5vvrqK6Vbt26Kj4+PMnr0aCUyMvP//NCh\\nQ5UGDRooTZs2VZo1a6bs2LEjz/PKEhMTo0yaNElp1aqV0rFjR2XdunVqeuPGjZVbt26peU+fPq20\\nbNlSMZlMedYl63qvX79e6dy5s9KxY8dc/rYL33YqiqKUWrAoCYUOFmazcv3AT0rC2XMlXDMhMu3b\\nt0/x8vJS0tPT88w3ffp0pXnz5sqxY8cUs9mspKamKr/88oty9uxZxWw2K+Hh4UqbNm2U77//XlGU\\nv//tv/LKK0pSUpJy9uxZpVWrVsrBgwcVRckMFo8//riyd+9eJSMjQ1m2bJkyaNCgXI9/b6P+559/\\nKr6+vsrHH3+sbn/++eeVN954Q0lKSlJu3LihDBgwQNm4caOiKIqyc+dOpV27dsqJEycUi8Wi/PXX\\nX2oDll+w8PLyUpYsWaKkpaUpKSkpOab9/vvvSuvWrZXjx48rGRkZSnBwsNKxY0clLS1NUZTMYDFg\\nwAAlJiZGuXXrltKtWzfl888/V49xbyOd27XPqlNSUpISGBio9O7dW93+/fffK0899ZTy559/Kunp\\n6coHH3ygPP300zleu7zOK6seZrNZ6devn/L+++8raWlpyuXLlxU/Pz9l//79iqIoyogRI5QvvvhC\\nLe/tt99W3njjjQLX5bnnnlNu3bqlpKSk5HrORQkW5Wq6D41WS/V2vlT2bFDWVRHlRG5DpX18fGjS\\npAlHjx5V0zt16kSLFi3QarUYDAZatWqFp6cnWq2Whg0b0rNnT44cOWJV/sSJE6lQoQKenp70799f\\n7dIFaNGiBU8++SQ6nQ5/f3/Onj2bZ1379etHs2bN6NGjBy1btmTo0KEA3Lhxg/379/Paa69RoUIF\\nqlWrxnPPPceOHTsA2Lx5M2PHjqVJkyZoNBpq166Nm1vB3jvSarW8+OKL6PV67O4ucfzPtC+//JKn\\nn36apk2botPp6NevH7a2thw/flwtZ8SIETg7O+Pg4EDHjh0JDw8v0PGzBAUF4ePjQ/Pmzfn1119Z\\nsmSJum3Tpk2MHz+eevXqYWNjQ0BAAOHh4UTdHSxT0PPKcurUKW7evMmkSZPQ6/V4eHgwePBgdu7c\\nCUDv3r3Vv0dFUdi5cye9e/cucF3Gjx+Pg4NDtuPer3I13UdyZBQZCQlU9myA5l/+4E78O9w7VDor\\nYGzatAmADh06WA2V/ufAjhMnTrBs2TLOnz9Peno6JpOJbt26WeW5dx83Nzf++OMP9XP16tXV3+3s\\n7EhLS7Oqxz99/fXX1KpVi127drF8+XKSk5PR6/VER0eTkZFBu3bt1LwWi0U99tWrV6lVq1ahrksW\\nR0dHDAZDnmnR0dGEhISwfv16NS09PZ1r166pn+99nmc0Gq22FcTo0aN55ZVXiI6OZuzYsVy6dImG\\nDRuqx1+4cCGLFy9W8yuKQmxsbK5BMafzyhIVFcW1a9fw8fFR08xms/q5a9euzJ8/n9jYWCIiItBo\\nNOq2gtQlpwFCxaFcBYvY738gZue3tN60Pv/MQhSDe4dKF3bo5NSpUxk+fDhr1qzBYDCwYMECbt26\\nZZXn6tWr6qid6OhonJyc7qu+Go2GHj16sHv3bj744ANmzZpFzZo10ev1/PLLLzkGGhcXFy5fvpxj\\neUaj0WqU3/Xr13F2drY6Xk51+Gf5AQEBPP/880U6n8JwdXVl1qxZTJ8+nY4dO2JnZ6cev0+fPsVy\\nXBcXF9zd3fnuu+9y3G5vb4+vry+7du3i4sWL9OzZUy2vIHUpqdF15aobKiUqGjuXmnJXIUrNvUOl\\nv/32W5KSkrBYLISHh+c6VDpLUlISVapUwWAwcPLkSasupiwffvghKSkpnD9/nuDgYHr06FEs9R4/\\nfjxffvkl169fx8nJCV9fX95++20SExOxWCxcvnxZ7RIbOHAgQUFB/P777yiKQkREhNot0rBhQ7Zv\\n347ZbGb//v1W3W4FNWjQIDZt2sSJEydQFIXk5GT27t1LYmJivvtWq1aN+Ph47ty5U+Dj+fr64uTk\\nxBdffAFkdhuuXr2a8+fPA5nzu+3atUvNX716da5cuVLg8ps0aUKlSpVYvXo1qampmM1m/vjjD6sZ\\npnv37s3WrVsJDQ1Vu6AKUpeSVK7uLFKioqgoU2aIUjZu3DicnZ1Zs2YN06dPx2g04uHhkedQacic\\nN23x4sXMmzePli1b0r17dxISEqzytGzZks6dO6MoCqNHj7bqKrofnp6ePPHEE6xdu5YZM2awZMkS\\nli1bRo8ePUhKSsLDw4Nx48YB0L17d+Lj45k6dSrXrl3Dzc2NJUuW4ObmxqxZs5gxYwYbNmzgqaee\\n4qmnnip0XRo3bsz8+fOZN28eERER2NnZ0bx5c6tunNzUq1ePnj178tRTT2E2m9mxY4fVnU1uxo4d\\ny6JFi3jmmWfo3LkzSUlJTJkyhaioKCpXrkzbtm3p3r07AJMmTWLGjBmkpqYyb948qlWrlmfZOp2O\\n//73vyxevJhOnTphMpmoW7cuL7/8sprHz8+PWbNm4erqqnaHAfnWpSRpFOXfO4dEZGQknTp1Yvfu\\n3bi7u+eZ15KezqHBQ3Ef2J/aw57JM68QD7qsf/unT5/O9RmEELkpTNuZpdx0Q6XGxILFcl8v5Akh\\nRHlVbr6SGJxq0PjtBdiV0EgBIYR4mJWbYKEzGLD3aph/RiH+Bdzd3Tl37lz+GYUoJuWmGyrul8Pc\\nPFL4kRhCCCHK0Z1FVHAIGltbqrZ8Iv/MQgghrJSLOwtFUUiJisZYwCkIhBBCWCsXwSIjIYGMxEQq\\nuEuwEEKIoigXwSIlKhoo+up4QghR3pWTYHF3KVW5sxBCiCIpFw+4nTr54dCsKfp8XsMXQgiRs3IR\\nLDRaLQZZklQIIYqsXHRDRXy2gbjD8o6FEEIU1UMfLCzp6UQGh5B4d0pfIYQQhffQBwuZQFAIIe7f\\nQ/vM4sjIMaTHx6ufz7/7PufffR9bBwdarltbhjUTQoh/nyLfWaSnp/Pss88WZ12K1b2BoiDpQggh\\nclfkOwtFUQq1RKKfnx96vV5dxDwwMJD27dvj6elJgwYN0Goz49aSJUvw9PQsarWEEEKUgDyDRadO\\nnXLdVpQF9lauXEmDBg2ypW/atImKFSsWujwhhBClI89gcfv2baZPn57jsnsmk4mAgIASq5gQQogH\\nR57BolGjRhgMBtq0aZNtm8lkKvTdRWBgIIqi0KJFC6ZMmYK9vT0AI0aMwGw206FDByZPnoxer8+2\\nb0JCQrbF6mNiYgp1fCGEEEWTZ7CYOHEiRqMxx222trZ8+umnBT7Qhg0bcHFxwWQysWDBAubNm8ey\\nZcvYu3cvLi4uJCYmMm3aND744ANeeeWVbPuvW7eOVatWFfh4tg4OOT7MtnVwKHAZQgghMmmUojx8\\nuE/nzp3j+eefZ8+ePVbpe/bs4eOPP+azzz7Ltk9udxbDhg1j9+7dOXaVCSGEyC4yMpJOnToVqu3M\\nc+jsW2+9ZfX55MmTRapYcnIyd+7cATIfjO/cuRMvLy9u375NamoqABkZGYSGhuLl5ZVjGfb29ri7\\nu1v91KxZs0j1EUIIUTh5dkMFBwfz+uuvq5/Hjh3LkSNHCn2QuLg4Jk+ejNlsxmKxUK9ePebMmcPF\\nixeZPXs2Go2GjIwMvL29eemllwp/FkIIIUpUnsHinz1URe2x8vDwICQkJFu6k5MT27ZtK1KZQggh\\nSk+e3VAajSbPz0IIIcqHPO8sUlNTGTZsmPo5KSnJ6jNkjnISQgjxcMszWCxYsMDq88CBA0u0MuLh\\ncOzYMWbOnMmNGzdYsWIFHTt2LOsqCSHuU57Bol+/fqVVD3GfvL291d9TUlLQ6/XodDoA5s6dS58+\\nfUqtLu+++y7PPfdctrvQB8WIRI7AAAAgAElEQVQnn3xCcHAwW7ZswdbWFoC1a9eyY8cOvvrqKyIj\\nI+nSpQvnzp2z2i8wMJDatWszefLksqi2EGXqoZ2ivLwJCwtTf/fz8+Ott96ibdu2uebPyMjAxqZk\\n/vqjo6OpX79+kfYtyXpllf/ss8+ya9cuVq9ezcSJE4mIiODDDz9kw4YNaoAVQliTYPEvEhIWxdLQ\\nc0THp+DqYGRaV0/6ersVaN933nmHiIgItFotP/74I2+88QZ169Zl0aJFXLx4ETs7O7p168b06dOx\\ntbUlIyODxx57jLlz57J27Vri4+Px9/dXh1JfunSJ119/nbNnz2JjY0O7du1Yvnw5fn5+REdHM27c\\nOHQ6HceOHeP69evMmTOHsLAwHBwcGD9+vNqlmVO9IiIiuHz5MhqNhh9//BEPDw9WrVrF9u3b+fTT\\nT7Gzs2PhwoVqMExISGDRokUcOHAArVbLgAEDmDx5Mlqtlq+++oqQkBC8vLz45ptvGDFiBJMnT2bB\\nggUMHjyYLl26MHfuXEaMGEHDhg1L5i9OiIfAQ79S3sMiJCyKmcGniIpPQQGi4lOYGXyKkLCoApfx\\nww8/0KtXL3799Vd69OiBTqdj1qxZ/PLLL2zcuJEDBw7wxRdfWO2zb98+goOD+frrr/nmm2/4+eef\\ngcxG/sknn+To0aPs37+foUOHAplv4Ts5OfF///d/hIWFodPpePnll3F3d+fAgQO88847LF261Op9\\nnX/WC2D37t0MHDiQo0eP8uijj/Lcc89hY2PDTz/9xPjx43nzzTfV/adNm4bBYOD7779ny5Yt7Nu3\\njy1btqjbw8LCeOSRRzh06BDjxo0DoH79+owZM4bhw4dz8+ZNXnjhhUL9fQhR3hQoWOzatSvH9G+/\\n/bZYKyNytzT0HCnpZqu0lHQzS0PP5bJHds2bN8fPzw+tVoudnR1NmjShadOm2NjY4OHhweDBg7O9\\ndDlhwgQqV66Mu7s7LVu2JDw8HMicGywqKorr169jMBho0aJFjse8cuUKp06dYurUqRgMBh577DH6\\n9+/P1q1bc60XQMuWLWnbti02NjZ069aN27dvM3bsWGxsbOjZsycREREkJSURGxvLoUOHeO211zAa\\njdSoUYNnn32WnTt3quW7uLgwdOhQdDqdWj6Aj48P8fHxdOvWLcfJK318fKx+5N+7KM8K1A01a9Ys\\nunfvni199uzZdOvWrdgrJbKLjk8pVHpOXFxcrD5fuHCBxYsXc/r0aVJSUjCbzTRp0sQqT/Xq1dXf\\n7ezsSE5OBmD69Om89957DBgwAEdHR0aPHp3jgIhr167h6OhIhQoV1DQ3NzfOnz+fa73+eVyDwUDV\\nqlXVBbKyGvzk5GSioqIwmUxWz2csFgtubn93z+VUflpaGnPmzGHEiBGsW7eOAQMGWO0DmaO67hUY\\nGJitHCHKizyDxZUrV4DMN7ezfr93W07fxkTJcHUwEpVDYHB1yHlW4Jz886XKOXPm0LRpU9555x0q\\nVqzI2rVr2bt3b4HKcnJyUodWHz16lFGjRuHj44OHh0e2fLdu3SI5OVkNGNHR0Tg7O+dar8JwcXHB\\naDRy5MgRNZj8U07lr1q1ipo1a/L6669jY2PD7NmzWbtW1mYXIjd5dkN17tyZLl26kJKSQufOna1+\\nXn31VRlCWIqmdfXEaGs9Usdoq2Na16IvQZuUlETlypWpUKECFy5cyPa8Ii87d+4kNjYWgMqVK6PR\\naHIcSeTh4cHjjz/OihUrMJlMhIeHExwcTO/evYtc73u5uLjwxBNPsHjxYhITE7FYLEREROS55O/p\\n06f5/PPPmT9/PgAvvfQSEREROU5JI4TIlOedxdmzZwEYPnw469evL5UKiZxljXoq6mionEyfPp05\\nc+bw0Ucf0ahRI7p3785vv/1WoH1PnjzJwoULSUxMpEaNGsyePRtXV9cc877zzjvMmTMHX19fHBwc\\nmDJlCq1bty5yvf9p6dKlLF++nB49epCUlISHhwfjx4/PMW9GRgavvfYakyZNUu+CjEYj8+fP55VX\\nXqFDhw7FVi8hHiZlsp5FcSnKnOxCCFHeFaXtLNAD7itXrvDuu+8SHh6uPuDMUtA+biGEEP9eBQoW\\ngYGBeHh4MH369FyXWRVCCPHwKlCwOH/+PBs3bsx1tIkQQoiHW4Fa/yeeeIIzZ86UdF2EEEI8oAp0\\nZ+Hm5saYMWPo0qWL1ctSgCyDKoQQ5UCBgkVKSgp+fn5kZGQQExNT0nUSQgjxgClQsFi0aFFJ10MI\\nIcQDLNdgERkZqY6//edUH/f65/QOQgghHj65BovevXurC+p07twZjUbDP9/f02g06iykQgghHl65\\nBot7V17Lmvbjfvj5+aHX6zEYDEDmuxvt27fn+PHjzJ49m7S0NNzc3Fi6dCnVqlW77+MJIYQoPoVa\\nKS86OprY2Fhq1qyZ47TP+Vm5ciUNGjRQPyuKwrRp01i0aBE+Pj58+OGHLFu2TJ6RCCHEA6ZA71lc\\nu3aN4cOH06VLFyZPnkznzp0ZNmyYOutoUZ06dQqDwYCPjw8AQ4YMkQVmhBDiAVSgO4s333yThg0b\\nsnr1aipUqEBycjIrVqxgzpw5/O9//yvwwQIDA1EUhRYtWjBlyhSuXr1qNVNp1apVsVgsxMfH4+Dg\\nYLVvQkICCQkJVmkyjFcIIUpHgYLFr7/+ynvvvYetrS0AFSpU4NVXX6V9+/YFPtCGDRtwcXHBZDKx\\nYMEC5s2bR+fOnQu8/7p161i1alWB8wshhCg+BeqGqlKlChcuXLBKu3jxIvb29gU+UNYzDr1ez9Ch\\nQ/ntt99wcXEhOjpazXPz5k00Gk22uwqAkSNHsnv3bqufDRs2FPj4Qgghiq5AdxZjx47lueeeY+DA\\ngbi6uhIdHU1wcHCBp/pITk7GbDZTuXJlFEVh586deHl58fjjj5OamsqxY8fw8fFh06ZNOa71DWBv\\nb1+o4CSEEKL4FChYDB48mFq1arFt2zbOnTuHs7Mzy5cvp02bNgU6SFxcHJMnT8ZsNmOxWKhXrx5z\\n5sxBq9WyZMkS5syZYzV0VgghxIMlz5XyNm/enPfOd7uMGjVqVKShtPdLVsoTQojCK/aV8rZu3Zpv\\nAUlJSVy8eJFp06YxbNiwgtVUCCHEv0qeweKzzz4rUCHnz59n7NixEiyEEOIhVSxL3z366KP07t27\\nOIoSQgjxACq2dVIDAwOLqyhRRr755htGjx5d1tXI0bFjx+jatWuJH8fPz4+ff/65xI8jxL+NLKpd\\nAo4dO8aQIUNo0aIFLVu2ZMiQIZw8ebKsq2UlMjIST09PMjIy1LQ+ffoQFBRU6LJmzJjB448/jre3\\nNy1btmTUqFHZ3svJi6enJxEREXnm8fHxITQ0tNB1E0IUDwkWxSwxMZGAgACGDx/OkSNH2L9/P5Mm\\nTUKv15d11UrUmDFjCAsLY//+/Tg7OzNr1qxiK/vegCaEKBsSLIrZpUuXAOjVqxc6nQ47OzvatWtH\\nw4YN1TybN2+me/fuPPHEE4wZM4aoqCh1m6enJxs2bKBLly54e3vz7rvvcvnyZZ5++mmaN2/OSy+9\\nhMlkAuD27dtMmDCB1q1b88QTTzBhwgSr+bJGjBjBu+++y5AhQ/D29mb06NHcvHkTgOHDhwPwxBNP\\n4O3tTVhYGMHBwTzzzDPq/ufPn2fUqFG0bNmStm3bFmgeMDs7O7p3755tWvvczjlrUIS/vz/e3t7s\\n3LmTw4cP06FDB1avXo2vry8zZ85U07LExsYyefJkWrdujZ+fH59++qma3qRJE+Lj49W8Z86coVWr\\nVqSnp3P58mWeffZZWrVqRatWrZg6dWq2OceEENlJsCiCkLAofN/eQ90ZO/B9ew8hYX839nXr1kWn\\n0zF9+nT27dvH7du3rfb94Ycf+Oijj1i1ahWHDh2iRYsWTJ061SrPgQMHCA4O5ssvv2TNmjW88cYb\\nLFu2jH379nH+/Hl27NgBgMVioX///vz444/8+OOPGAwG5s2bZ1XW9u3bWbRoEYcOHSI9PV3tZlq/\\nfj0AR48eJSwsDG9vb6v9EhMTGTVqFO3bt+fAgQN89913BXoJMzk5me3bt1OrVq0CnXPWlC1bt24l\\nLCyMHj16AHDjxg1u377Njz/+yPz5862OYbFYeP755/H09GT//v2sW7eOdevWceDAAZydnWnWrBnf\\nffedmn/btm107doVW1tbFEVhwoQJHDhwgF27dhETE8P777+f73kJUd5JsCikkLAoZgafIio+BQWI\\nik9hZvApNWBUqlSJzz//HI1GwxtvvEGbNm0ICAjgxo0bAGzatInx48dTr149bGxsCAgIIDw83Oru\\nYty4cVSqVIlHH32UBg0a4Ovri4eHB5UrV6ZDhw6cOXMGAEdHR7p27YrRaKRSpUo8//zzHD161Kq+\\n/fv3p27dutjZ2dGtW7cCr2y4d+9eqlevzujRozEYDFSqVImmTZvmmj8oKAgfHx+aN2/Or7/+ypIl\\nS9RtBTnnf9Jqtbz44ovo9Xrs7Oystp06dYqbN2+q3XseHh4MHjyYnTt3ApmrPG7fvh1AnV4ma7Re\\n7dq18fX1Ra/XU7VqVUaNGpXtmgkhsivU4kcCloaeIyXdbJWWkm5maeg5+nq7AVCvXj3efvttAC5c\\nuMC0adNYuHAhK1asIDo6moULF7J48WJ1f0VRiI2Nxc0tc//q1aur2wwGQ7bPWYEnJSWFRYsWceDA\\nAfUOJikpCbPZjE6nA6BGjRrqvkajkeTk5AKd59WrV63uDvIzevRoXnnlFaKjoxk7diyXLl1Su94K\\ncs7/5OjoqK6q+E9RUVFcu3ZNXQcFwGw2q5+7du3K/PnziY2NJSIiAo1Go26Li4vjrbfe4tixYyQl\\nJaEoisw5JkQBSLAopOj4lEKl16tXj/79+/PFF18AmbPvBgQE0KdPn/uuS1BQEJcuXeLLL7+kRo0a\\nhIeH07dv32xrpedEo9Hkud3FxUXt7ioMV1dXZs2axfTp0+nYsSN2dnZFOue86ufi4oK7u7tVV9O9\\n7O3t8fX1ZdeuXVy8eJGePXuq5S1fvhyNRsM333yDo6MjP/zwQ7auOyFEdtINVUiuDsY80y9cuEBQ\\nUJD6oPnq1ats375d7cIZMmQIq1ev5vz58wDcuXOHXbt2FakuSUlJGAwG7O3tiY+PL9R6H1WrVkWr\\n1XLlypUct//nP//hxo0bfPLJJ5hMJhITEzlx4kSByvb19cXJyUkNkPmdc/Xq1XOtR06aNGlCpUqV\\nWL16NampqZjNZv744w+r4cm9e/dm69athIaGWr0wmpSURIUKFbC3tyc2NpY1a9YU+LhClGcSLApp\\nWldPjLY6qzSjrY5pXT2BzGcWJ06cYNCgQTRr1ozBgwfToEEDZsyYAUDnzp0ZO3YsU6ZMoXnz5vTq\\n1Yv9+/cXqS4jR44kLS2N1q1b8/TTTxdqMSqj0UhAQADPPPMMPj4+HD9+3Gp7pUqVCAoK4scff8TX\\n15euXbty+PDhApc/duxY1qxZg8lkyvecJ02axIwZM/Dx8VGfO+RFp9Px3//+l7Nnz9KpUydat27N\\n66+/TmJioprHz8+Pv/76i+rVq1uNRJs0aRJnzpzBx8eH8ePH06VLlwKfkxDlWZ6zzj7oymrW2ZCw\\nKJaGniM6PgVXByPTunqqzyuEEOJBV+yzzoqc9fV2k+AghChXpBtKCCFEviRYCCGEyJcECyGEEPmS\\nYCGEECJfEiyEEELkS4KFEEKIfEmwEEIIkS8JFkIIIfJV6sFi1apVeHp68scffwCZi/307t0bf39/\\n/P39OXfuXGlXSQghRD5K9Q3u06dPc/z4cVxdXa3SN23aRMWKFUuzKkIIIQqh1IKFyWRi3rx5LFu2\\njJEjRxZ6/4SEhGzLX967hKgQQoiSU2rB4r333qNPnz54eHhk2zZixAjMZjMdOnRg8uTJ6PX6bHnW\\nrVtXqCm4hRBCFJ9SCRZhYWGcOnWKwMDAbNv27t2Li4sLiYmJTJs2jQ8++IBXXnklW76RI0fSr18/\\nq7SYmBiGDRtWYvUWQgiRqVSCxdGjR7l48SKdOnUCMhv5MWPGsGjRItq1awdkrp8waNAgPv744xzL\\nsLe3l+UvhRCijJRKsBg/fjzjx49XP/v5+fG///0PZ2dnUlNTsbOzIyMjg9DQULy8vEqjSkIIIQqh\\nTNezuHjxIrNnz0aj0ZCRkYG3tzcvvfRSWVZJCCFEDsokWOzZs0f9fdu2bWVRBSGEEIUgb3ALIYTI\\nlwQLIYQQ+Sq3a3D7+flx48YNdDodOp2O+vXr4+/vz9NPP41Wm3MMnTFjBs7OzurQ3vPnzzNq1ChG\\njx7N6NGj8fPz46233uLkyZN89NFHAGRkZJCRkYGdnR0Arq6u7Nixo3ROUgghikm5DRYA//vf/2jb\\nti137tzhyJEjLFiwgJMnT7Jo0aJsec1ms9Xn8PBwRo8ezcSJExk+fLjVtoCAAAICAgAIDg7mq6++\\nYuPGjSV3IkIIUcIe2mAREhbF0tBzRMen4OpgZFpXT/p6u+WYt3LlynTq1IkaNWowePBgRo0aRVBQ\\nEAaDgejoaI4ePcqHH36o5j958iTjxo0jMDCQQYMGldYpCSFEmXkon1mEhEUxM/gUUfEpKEBUfAoz\\ng08REhaV535NmjShZs2aHDt2DIDt27cTEBDAb7/9RosWLQA4deoUY8eOZebMmRIohBDlxkMZLJaG\\nniMl3brbKCXdzNLQ/Kc/d3Jy4vbt2wB06tSJFi1aoNVqMRgMABw/fpxKlSrRoUOH4q+4EEI8oB7K\\nYBEdn1Ko9HvFxsZSpUoVAFxcXLJtHzZsGI0bN2b06NFqUBFCiIfdQxksXB2MhUrPcvLkSWJjY9Uu\\np5xotVqWLVuGi4sLY8aMITEx8b7qKoQQ/wYPZbCY1tUTo63OKs1oq2NaV88c8ycmJvLjjz8yZcoU\\n+vTpg6dnzvmy2Nra8t577+Ho6Mi4ceNITk4utroLIcSD6KEcDZU16im/0VABAQHodDq0Wi3169dn\\n1KhRDBkypEDH0Ov1rFq1igkTJhAQEMDq1auL/TyEEOJBoVEURSnrShRVZGQknTp1Yvfu3bi7u5d1\\ndYQQ4l+hKG3nQ9kNJYQQonhJsBBCCJEvCRZCCCHyJcFCCCFEviRYCCGEyJcECyGEEPmSYCGEECJf\\nEiyEEELkS4KFEEKIfEmwEEIIka9SDxarVq3C09OTP/74A8hcH6JPnz507dqV0aNHExcXV9pVEkII\\nkY9SDRanT5/m+PHjuLq6AqAoCtOmTWP27NmEhobi4+PDsmXLSrNKQgghCqDUgoXJZGLevHnMmTMH\\njUYDZC5RajAY8PHxAWDIkCF8++23Oe6fkJBAZGSk1U9MTExpVT+b4OBgnnnmmVy3jxgxgq+++qoU\\naySEECWn1KYof++99+jTpw8eHh5q2tWrV9W7DICqVatisViIj4/HwcHBav9169axatWqYq/Xjh07\\n+OSTTzh//jxGoxF3d3f69u3L0KFD1aB2v4KDg5k1axZ2dnZW6d9++y3Ozs7FcgwhhChJpRIswsLC\\nOHXqFIGBgUUuY+TIkfTr188qLSYmhmHDhhW5zKCgINasWcPs2bNp164dFStWJDw8nLVr1zJo0CD0\\nen2Ry/6nZs2asXHjxmIrTwghSlOpBIujR49y8eJFOnXqBGQ28mPGjGHEiBFER0er+W7evIlGo8l2\\nVwFgb2+Pvb19sdXpzp07rFy5ksWLF9O1a1c1vVGjRixfvlzNM3/+fPbv34/RaGTQoEEEBASg1Wbv\\nvTt48CDz58/n+vXr+Pv7U5hlQvz8/Bg2bBghISFER0fTvn17Fi9ejMFguP8TFUKIYlAqzyzGjx/P\\nTz/9xJ49e9izZw81a9Zk7dq1jB07ltTUVI4dOwbApk2b6N69e7EcMyQsCt+391B3xg58395DSFiU\\n1fawsDBMJpMawHIyf/587ty5ww8//MBnn33G1q1b2bJlS7Z8N2/eZPLkybz88sv88ssv1KpVi99+\\n+61Q9d21axdr1qxh9+7dnDt3juDg4ELtL4QQJalMl1XVarUsWbKEOXPmkJaWhpubG0uXLr3vckPC\\nopgZfIqUdDMAUfEpzAw+Bfy95OqtW7dwdHTExubvSzBkyBD+/PNPTCYTa9asYefOnYSEhFCpUiUq\\nVarEqFGj+Oabbxg0aJDV8fbv30/9+vXp1q0bkNllFhQUZJXnxIkT6oN8AAcHB3744Qf184gRI9Tn\\nFx07diQ8PPy+r4MQQhSXMgkWe/bsUX9v3rw527ZtK9byl4aeUwNFlpR0M0tDz6nBwsHBgVu3bpGR\\nkaEGjE2bNgHQoUMHbty4QXp6utUDeFdXV2JjY7Md79q1a9SsWVP9rNFocHFxscrTtGnTPJ9Z1KhR\\nQ/3daDRy7dq1gp6uEEKUuIfyDe7o+JR80729vdHr9ezevTvHvI6Ojtja2lo9U7l69WqOo5dq1Khh\\nNYxXURSuXr1a1OoLIcQD56EMFq4OxnzT7e3tmThxInPnzuXbb78lKSkJi8VCeHg4KSkpaLVaunXr\\nxjvvvENiYiJRUVF8/PHH9OnTJ1u5Tz75JOfPn+e7774jIyODTz/9lBs3bpTY+QkhRGkr02cWJWVa\\nV0+rZxYARlsd07p6WuUbN24czs7OrFmzhunTp2M0GvHw8CAwMBBvb28aNmzI/PnzeeqppzAYDAwa\\nNIgBAwZkO17VqlV57733WLBgATNnzsTf35/mzZtb5Tl+/Dje3t5WaevWraNJkybFeOYCID09ncjI\\nSFJTU8u6KkKUOTs7O9zd3bG1tb2vcjRKYcZ4PmAiIyPp1KkTu3fvxt3d3WpbSFgUS0PPER2fgquD\\nkWldPdXnFeLhdunSJSpXrky1atWK7cVKIf6NFEUhLi6OO3fuULduXTU9r7YzNw/lnQVkjnqS4FA+\\npaamUqdOHQkUotzTaDRUq1aN69ev33dZD+UzCyEkUAiRqbj+L0iwEEIIka+HthtKiLLk5eVFgwYN\\nMJvNuLu7s2TJkmKdrqagZs2axahRo6hfv/59lTNixAiuXbuGXq8nPT2dtm3b8vLLL5fJORX22iYk\\nJLBt27Y855EbMmSI+p5VSZs4cSKRkZEkJydz8+ZN9ZnBnDlzsg2MyU9gYCDdunXjqaeeKomqWpFg\\nIcq1IyPHkB4fny3d1sGBluvWFrlcOzs7tm7dCsD06dPZsGEDzz//fJHLy4vZbEan0+W4bcGCBcV2\\nnGXLltG4cWNMJhMrVqzghRdeYP369cVWfkEV9tomJCSwcePGHINF1rUryUChKAqKoqhzyn3wwQcA\\nHD58mKCgID766KMSO3Zxkm4oUa7lFCjySi+KZs2aWb35v2bNGgYMGEDv3r1ZuXKlmh4SEkLv3r3p\\n06cP06ZNA2DGjBlWa7xkDb8+fPgwI0aMYOrUqfTu3Zvk5GTGjx9Pnz596NWrFzt37gQy7whOnTrF\\n559/zpIlS9RygoODmT9/PgBbt25l4MCB+Pv7M3v2bMxm69kP/kmv1zNt2jSio6M5e/ZsnmV4e3vz\\nzjvv0KdPHwYPHqy+f7Rr1y569epFnz591EbcbDazePFi9doUpAEvyLVdvnw5ly9fxt/fn8WLF2e7\\ndvde19zKWLp0KRs2bFDzvP/+++qUPjnlj4yMpHv37rz55pv069evwC/pHjx4EH9/f3r37s3rr7+O\\nyWQCMmeVWLZsGQMHDmTQoEFcuXIl277Lly/ntddew2KxFOhYhSV3FuKhd2rW7Gxp1X3b4tKjW777\\npickcHax9eqNjRfMK/CxzWYzhw4dYuDAgQD89NNPREREsHnzZhRF4fnnn+fo0aM4ODjw3//+l40b\\nN1K1alXiCxCsTp06xbZt2/Dw8CA0NBQnJydWr14NZM6YfK9u3brx9NNP8+qrrwKwc+dOAgICuHDh\\nArt27WLjxo3Y2try5ptvsm3bNvr27ZvnsXU6HQ0bNuTixYvY2trmWkZycjJNmzbllVdeYcmSJXz5\\n5Ze88MILfPjhh6xduxZnZ2cSEhIA2Lx5M5UrV2bLli2YTCaGDBmCr6+v1Ro4Rbm2U6dO5fz58+rd\\nyOHDh62u3b1yK6Nnz54sXLhQDWxZE3/mlt/FxYVLly6xaNEi3nzzzXz/LgFSUlJ47bXX+Oyzz6hV\\nqxZTp07lyy+/ZPjw4UDmi8SbN29m8+bNLFq0iA8//FDdd9GiRZhMJhYsWFBigzskWAhRAlJTU/H3\\n9ycqKorHHnsMX19fIPOb48GDB9XGODk5mb/++ovU1FS6detG1apVAXKcpv+fGjdurDZ2DRo0YPHi\\nxSxdupSOHTtaTVoJmS+Oenh4cPz4cWrXrs2lS5do0aIFGzZs4Pfff1cb3NTUVKpVq1agc8x6RevQ\\noUO5lmFra0vHjh0BePzxxzl48CCQ+U1+xowZdO/enc6dO6vX5ty5c4SGhgKZAS8iIiJbg17Ya/vP\\nedr+ee3ulVsZgwYNIi4ujtjYWG7duoW9vT2urq589tlnuR7T1dWVZs2aFehaAly4cIHatWtTq1Yt\\nAPr27cvmzZvVYNGrVy8A+vTpoy6jALBy5Uq8vb2ZO3dugY9VFBIsxEOvMHcC/2Rrb1+k/bP61e/c\\nucOECRPYsGEDzz77LIqiMH78eIYMGWKV/9NPP82xHJ1Op3YrKIpCenq6uq1ChQrq73Xr1iU4OJh9\\n+/axfPlyfH19mTRpklVZ3bt3Z9euXTzyyCN07twZjUaDoij069ePqVOnFur8zGYzf/zxB4888ghx\\ncXG5lmFra6t+09VqtWr31Lx58zhx4gR79+6lb9++hISEoCgKr7/+Ou3bt8/z2IW9tpGRkdnKuPfa\\n3Su3MgC6du1KaGgoN27coGfPnnnmj4yMzPUYucnv/ejc7hiaNGnC77//zu3bt6lSpUqhjlkY8sxC\\niBJUuXJlXn/9dYKCgkhPT6ddu3Zs2bKFpKQkAGJjY4mLi6NNmzZ8++233Lp1C0DthnJzc+P06dMA\\n7N692ypY3Cs2Nhaj0aA2EG4AAAySSURBVIi/vz9jxozhzJkz2fJ06dKFH374ge3bt9OjRw8A2rRp\\nQ2hoKHFxcepxo6Kisu17r/T0dJYvX46LiwsNGzYsUhmXL1+madOmvPTSSzg6OhITE0O7du3YuHGj\\neo6XLl0iOTk51zIKem0rVqyopuUntzIAevbsyc6dOwkNDVUXTMsrf2HVr1+fiIgI9XnEN998Q8uW\\nLdXtWc+htm/fbjVq6j//+Q+jR49mwoQJBT7PopA7C1Gu2To45Doaqrg0atSIhg0bsmPHDvr27cuF\\nCxfUb6IVKlRg6dKlPProowQEBDBixAi0Wi2NGjXi7bffZvDgwbzwwgsMHDiQNm3a5Ppt9Y8//mDJ\\nkiVotVpsbGxy7CevUqUK9evX588//1TnJKtfvz4vv/wyo0ePxmKxYGtry+zZs3Fzyz77QWBgIHq9\\nHpPJRNu2bdU+88KUkWXJkiVERESgKAqtW7emYcOGeHp6EhUVRf/+/VEUBUdHR6t++aJe21q1atG8\\neXN69epF+/bt+c9//pNree3atcuxjGrVqvHoo4+SlJSEk5MTTk5OeebPaTXN/BiNRhYsWMCkSZOw\\nWCw0adLEau2clJQUBg4ciEajYcWKFVb79uzZk6SkJF544QVWr15dIqtsPrRzQ4nyKzw8HC8vr7Ku\\nhhDFpkOHDmzfvr3I77X88/9EUdpO6YYSQgiRL+mGEkKIB9z+/fvLugpyZyEeTv/i3lUhilVx/V+Q\\nYCEeOnZ2dsTFxUnAEOVe1noWdnZ2912WdEOJh467uzuRkZHFMoe/EP92WSvl3S8JFuKhY2tra7Uq\\nmBDi/pVasHjhhReIjIxEq9VSoUIF3njjDby8vPDz80Ov16vjggMDA/N9g1MIIUTpKrVgsXjxYipX\\nrgzADz/8wGuvvcbXX38NZM5t0qBBg9KqihBCiEIqtWCRFSgAEhMTCz0zYkJCgjo7ZZasKQViYmLu\\nv4JCCFFOZLWZ+U1Hf69SfWYxa9YsDh48iKIorFmzRk0PDAxEURRatGjBlClTcnxLcd26daxatSrH\\ncvNaAUsIIUTOrl+/Tu3atQuUt0ym+wgJCWHHjh383//9H//f3r2GNPm+cQD/6siBmefDppKWoEkG\\nDkUjUrEEM5QCCcw0oROBiUUgZjJ/mUQrQyMGovUiSTqQNdJMw8QwMlI0SuzklDwtNZ2Wp8bc838R\\nPv/U7ZmZPrNxfd7Nze27+77o7n7crlulUkEsFrO92CcmJpCfn7/gd/TtLDQaDXp6euDt7b3gpLCv\\nX7/iwIEDKCsrg0gkWtH386co29JQtqWhbEtjztlmZmYwNDSEgICARX+s1iSfhtq7dy+kUinUajXb\\na97KygqJiYkGj0e0tbXVu+PYuHEj52uJRKJV2zeKsi0NZVsayrY05pptsTuKWbx8KW9iYmLOsYJ1\\ndXWws7ODUChkT/RiGAZVVVXUAI4QQlYhXnYWU1NTSE9Px9TUFCwtLWFnZ4eioiIMDw8jLS0NMzMz\\n0Ol08PHxQU5ODh+RCCGE/AFeFgtnZ2fcu3dP730KhYKPCIQQQv6C4L/Fnib+DxIKhQgNDV2Rg0D+\\nFmVbGsq2NJRtaSjb//3Thx8RQgjhB3WdJYQQYhQtFoQQQowy266zXV1dyMzMxOjoKOzt7SGTyeDt\\n7c17DrVajYyMDHR3d8PKygpeXl7Izc2Fo6Mj/Pz84Ovryx7ufunSJfj5+fGaz1Ajxzdv3kAqleLn\\nz5/w8PBgD63nS29vL1JTU9nbP378wPj4OF6/fm2S5pMymQw1NTXo6+tDRUUF28uMq874qkF92bjq\\nDgBvtWdo3LjmkK/a05eNq+6M5V4uXHPHNTYrPm6MmUpOTmYUCgXDMAyjUCiY5ORkk+RQq9XMq1ev\\n2NsXL15kzpw5wzAMw/j6+jLj4+MmyTUrMjKS+fjx45yf6XQ6JioqimlqamIYhmHkcjmTmZlpinis\\nvLw85ty5cwzD6M+80pqampj+/v4Fr81VZ3zVoL5sXHXHMPzVnqFxMzSHfNaeoWy/+73uuHIvJ0Nz\\nxzU2fIybWV6GGh4eRnt7O2JjYwEAsbGxaG9vx8jICO9Z7O3tERoayt4ODAxEf38/7zn+xLt37yAU\\nChEcHAwASEhIQHV1tcnyaDQaVFRUID4+3mQZgoOD2W4Ds7jqjM8a1JdttdSdvmxc+Kw9Y9lMVXeG\\n5o5rbPgYN7O8DKVSqeDm5sb2ixIIBHB1dYVKpWK34aag0+lw+/Zt7Nixg/1ZcnIyZmZmEB4ejrS0\\nNFhZWfGea34jR5VKBXd3d/Z+R0dH6HQ69nIK3+rq6uDm5obNmzcbzKyvFcxK46ozhmFWTQ3qqzvA\\n9LWnbw5XU+3pqztDuVfK73PHNTZ8jJtZ7ixWq/Pnz8Pa2hpJSUkAgPr6ejx48ABlZWXo6OiAXC7n\\nPVNZWRkePXqE8vJyMAyD3Nxc3jMYU15ePud/d/9C5tVkft0Bpq+9f2EO59cdwH9ufXNnKma5WIjF\\nYgwMDLC92mdmZjA4OPhH2+HlJpPJ8OXLFxQWFrJ/VJzNY2Njg3379qGlpYX3XPMbOba0tEAsFs+5\\nZDEyMgILCwuT7CoGBgbQ1NSEuLg4zsymwFVnq6UG9dXdbHbAdLVnaA5XS+3pq7vZfAA/tTd/7rjG\\nho9xM8vFwsnJCf7+/qisrAQAVFZWwt/f32SXoAoKCtDW1ga5XM5u9cfGxjA9PQ0A0Gq1qKmp4b2J\\n4uTkpN5GjgEBAZienkZzczMA4M6dO4iJieE126yHDx8iIiICDg4OnJlNgavOVkMN6qs7wPS1xzWH\\nq6X25tedsdzLTd/ccY0NH+Nmtt/gViqVyMzMxPfv32FrawuZTGa0nflK+Pz5M2JjY+Ht7c32jff0\\n9MSRI0cglUphYWEBrVYLiUSCrKwsrF27lrdsPT09Cxo5Zmdnw9XVFS0tLcjJyZnzMTxnZ2fess2K\\njo7G2bNnER4ebjTzSsrLy8PTp0/x7ds3ODg4wN7eHo8fP+asM75qUF+2wsJCvXUnl8vR2trKW+3p\\ny1ZUVMQ5h3zVnqE5BRbWHcBf7Rn6N0Mul3OOzUqPm9kuFoQQQpaPWV6GIoQQsrxosSCEEGIULRaE\\nEEKMosWCEEKIUbRYEEIIMYoWC0IIIUbRYkEIIcQoWiwI4Vl0dDS6uroA/Poi1cGDByGRSBAUFITj\\nx49DqVSaOCEhC9FiQQiPuru7odPpsGHDBrS2tuLw4cPYuXMnGhoa8OzZM/j5+SExMRG9vb2mjkrI\\nHLRYEMKhuLgYYWFhkEgkiI6ORmNjI3Q6HYqLixEVFYXQ0FCkp6djdHSU/Z3m5mYkJCQgODgYERER\\nePDgAXtffX09IiIiAACXL1/Gnj17kJKSAhsbG9jb2+PUqVPYsmWLSToQE8KFFgtCDOjs7ERZWRnu\\n37+P1tZW3LhxAx4eHigtLUVtbS1u3bqFhoYG2NnZsa2q+/v7cfToUSQlJaGxsREKhWJOs7nnz58j\\nIiICU1NTaG1txa5duxa8bkxMDF68eMHb+yRkMczy8CNCloNAIIBGo4FSqYSjoyM8PT0BAHfv3oVU\\nKoVIJAIAnDhxApGRkdBqtaioqMC2bdvYE/IcHBzYzqVTU1Noa2tDSEgI1Go1dDodXFxcFryui4sL\\n1Go1T++SkMWhxYIQA7y8vJCVlYVr166ho6MD27dvR2ZmJvr7+5GamjrnfAhLS0sMDw9DpVJh/fr1\\nep+vsbEREokEQqEQtra2sLS0xNDQEHx8fOY8bmhoaE5rbEJWA1osCOEQFxeHuLg4jI+PQyqVIj8/\\nHyKRCBcuXEBQUNCCx4vFYrx9+1bvc81eggIAa2trBAYGorq6Glu3bp3zuCdPniAkJGT53wwhf4H+\\nZkGIAZ2dnWhsbIRGo4GVlRWEQiEEAgH279+PwsJC9PX1Afh1KlltbS2AX4vLy5cvUVVVBa1WC7Va\\njffv3wMAGhoa2MUCAE6fPg2FQoHS0lKMj49jbGwMBQUFaG5uxrFjx/h/w4RwoJ0FIQZoNBpcuXIF\\nSqUSa9asgUQiQW5uLlxcXMAwDA4dOoTBwUE4OTlh9+7diIqKgru7O0pKSiCTyZCdnY1169bh5MmT\\nEAgEsLa2hru7O/v8wcHBuH79Oq5evYqCggJMTk5CJBLh5s2b8PPzM+E7J2QhOvyIEB6UlJRArVYj\\nIyPD4GM+fPiAlJQU5OfnIywsjMd0hBhHl6EI4YGHhwfi4+M5H7Np0ybI5XJ8+vQJWq2Wp2SELA7t\\nLAghhBhFOwtCCCFG0WJBCCHEKFosCCGEGEWLBSGEEKNosSCEEGIULRaEEEKM+h82MoeBZjQ7SAAA\\nAABJRU5ErkJggg==\\n\",\n \"text/plain\": [\n \"<matplotlib.figure.Figure at 0x7f1cd8117748>\"\n ]\n },\n \"metadata\": {},\n \"output_type\": \"display_data\"\n }\n ],\n \"source\": [\n \"# figure showing the efficieicnty trade off using 16-core CPUs\\n\",\n \"import matplotlib.pyplot as plt\\n\",\n \"import seaborn as sns\\n\",\n \"sns.set(style='ticks')\\n\",\n \"# Asai et al.2020\\n\",\n \"plt.xlim(-5, 210)\\n\",\n \"plt.ylim(35, 70)\\n\",\n \"\\n\",\n \"# Asai et al. 500 + 8*3\\n\",\n \"plt.scatter(133, 61.4, c='#1f77b4')\\n\",\n \"plt.text(133 - 30, 61.4 - 3, \\\"Graph Recurrent Retriever\\\")\\n\",\n \"\\n\",\n \"# Gold \\n\",\n \"plt.scatter(0.5, 39.1, c='#1f77b4')\\n\",\n \"plt.text(3.5, 39.1 - 0.5, \\\"GoldEn\\\")\\n\",\n \"\\n\",\n \"# DiKIT\\n\",\n \"plt.scatter(0.5, 42.9, c='#1f77b4')\\n\",\n \"plt.text(3.5, 42.9 - 0.5, \\\"DrKIT\\\")\\n\",\n \"\\n\",\n \"# Semmantic Retrieval\\n\",\n \"plt.scatter(50*0.3, 47.6, c='#1f77b4')\\n\",\n \"plt.text(50*0.3 + 3, 47.6 - 0.5, \\\"Semantic Retrieval\\\")\\n\",\n \"\\n\",\n \"# HGN\\n\",\n \"plt.scatter(50*0.3, 60.0, c='#1f77b4')\\n\",\n \"plt.text(50*0.3 + 3, 60 - 0.5, \\\"HGN\\\")\\n\",\n \"\\n\",\n \"# TransformerXH 115 cross attention 52.9\\n\",\n \"plt.scatter(115*0.3, 52.9, c='#1f77b4')\\n\",\n \"plt.text(115*0.3 + 3, 52.9 - 0.5, \\\"TransformerXH\\\")\\n\",\n \"\\n\",\n \"\\n\",\n \"time_f1s = [(1.02, 55.5), (4.7, 61.5), (7.7, 62.7), (14.7, 64.0), (42, 65.6), (99.2, 66.3), (207, 67.3)]\\n\",\n \"xs = [_[0] for _ in time_f1s]\\n\",\n \"ys =[_[1] for _ in time_f1s]\\n\",\n \"plt.plot(xs,ys, linestyle='--', marker='s', c='r', label='Recursive Dense Retriever Topk')\\n\",\n \"\\n\",\n \"plt.xlabel('sec/Q', fontsize=12)\\n\",\n \"plt.ylabel('Joint F1', fontsize=12)\\n\",\n \"\\n\",\n \"lg = plt.legend(loc='lower right', fontsize=10)\\n\",\n \"frame = lg.get_frame()\\n\",\n \"# frame.set_edgecolor('black')\\n\",\n \"lg.draw_frame(True)\\n\",\n \"\\n\",\n \"\\n\",\n \"plt.savefig('efficiency.pdf')\\n\",\n \"plt.show()\"\n ]\n },\n {\n \"cell_type\": \"code\",\n \"execution_count\": 431,\n \"metadata\": {},\n \"outputs\": [],\n \"source\": [\n \"def autolabel(rects):\\n\",\n \" \\\"\\\"\\\"Attach a text label above each bar in *rects*, displaying its height.\\\"\\\"\\\"\\n\",\n \" for rect in rects:\\n\",\n \" height = rect.get_height()\\n\",\n \" ax.annotate('{}'.format(height),\\n\",\n \" xy=(rect.get_x() + rect.get_width() / 2, height),\\n\",\n \" xytext=(0, 3), # 3 points vertical offset\\n\",\n \" textcoords=\\\"offset points\\\",\\n\",\n \" ha='center', va='bottom')\"\n ]\n },\n {\n \"cell_type\": \"code\",\n \"execution_count\": 450,\n \"metadata\": {},\n \"outputs\": [\n {\n \"name\": \"stderr\",\n \"output_type\": \"stream\",\n \"text\": [\n \"/public/apps/anaconda3/5.0.1/lib/python3.6/site-packages/matplotlib/figure.py:418: UserWarning: matplotlib is currently using a non-GUI backend, so cannot show the figure\\n\",\n \" \\\"matplotlib is currently using a non-GUI backend, \\\"\\n\"\n ]\n },\n {\n \"data\": {\n \"image/png\": \"iVBORw0KGgoAAAANSUhEUgAAAb0AAAEtCAYAAACcSL4HAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\\nrPQkSZIksyErPUmSJMls/B8DEzApxeEo9gAAAABJRU5ErkJggg==\\n\",\n \"text/plain\": [\n \"<matplotlib.figure.Figure at 0x7f1c4e731278>\"\n ]\n },\n \"metadata\": {},\n \"output_type\": \"display_data\"\n }\n ],\n \"source\": [\n \"# lines comparing the performance of bridge and comparison questions, use beam search to get top100\\n\",\n \"import matplotlib.pyplot as plt\\n\",\n \"import seaborn as sns\\n\",\n \"%matplotlib inline\\n\",\n \"sns.color_palette(\\\"tab10\\\")\\n\",\n \"sns.set(style='ticks')\\n\",\n \"\\n\",\n \"k = [1,2,5,10,20,50,80,100]\\n\",\n \"comparison = [95.8, 98.3, 99.3, 99.5, 99.7, 99.9, 99.9,100]\\n\",\n \"bridge = [61.2, 66.8, 72.8, 75.7, 78.1, 80.4, 81.7, 82.0]\\n\",\n \"\\n\",\n \"f, (a0, a1) = plt.subplots(1, 2, gridspec_kw={'width_ratios': [2.5, 1]})\\n\",\n \"a0.plot(k, comparison, marker=\\\"o\\\", linestyle='--', label=\\\"Comparison\\\")\\n\",\n \"a0.plot(k, bridge, marker=\\\"s\\\", linestyle='--', label=\\\"Bridge\\\")\\n\",\n \"a0.set_xlabel('Top-k Chain')\\n\",\n \"a0.set_ylabel('Recall')\\n\",\n \"a0.legend()\\n\",\n \"\\n\",\n \"a1.yaxis.tick_right()\\n\",\n \"a1.yaxis.set_label_position(\\\"right\\\")\\n\",\n \"p = a1.bar(np.arange(2), [97.8, 79.0], 0.5, color=('#1f77b4', '#ff7f0e'))\\n\",\n \"autolabel(p)\\n\",\n \"a1.set_xlabel('Q type')\\n\",\n \"a1.set_ylabel('Reranked Top1 EM Accuracy')\\n\",\n \"plt.sca(a1)\\n\",\n \"plt.xticks(np.arange(2), [\\\"Comparison\\\", \\\"Bridge\\\"])\\n\",\n \"f.tight_layout(pad=0.05)\\n\",\n \"plt.savefig('retrieva_types.pdf')\\n\",\n \"f.show()\\n\"\n ]\n },\n {\n \"cell_type\": \"code\",\n \"execution_count\": 451,\n \"metadata\": {},\n \"outputs\": [\n {\n \"data\": {\n \"image/png\": \"iVBORw0KGgoAAAANSUhEUgAAAYsAAAESCAYAAAAMifkAAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\\nogAAAABJRU5ErkJggg==\\n\",\n \"text/plain\": [\n \"<matplotlib.figure.Figure at 0x7f1c4e6e3a58>\"\n ]\n },\n \"metadata\": {},\n \"output_type\": \"display_data\"\n }\n ],\n \"source\": [\n \"# close-book QA diagnosis experiments\\n\",\n \"import matplotlib.pyplot as plt\\n\",\n \"import seaborn as sns\\n\",\n \"import numpy as np\\n\",\n \"import pandas as pd\\n\",\n \"sns.set_style(\\\"white\\\")\\n\",\n \"sns.set_palette(\\\"colorblind\\\", 10)\\n\",\n \"# sns.set_color_codes(\\\"g\\\")\\n\",\n \"bart_ems = [26.5, 27.4, 14.5]\\n\",\n \"datasets = [\\\"NQ\\\", \\\"WebQ\\\", \\\"HotpotQA\\\"]\\n\",\n \"sota_ems = [51.4, 45.5, 60.0]\\n\",\n \"\\n\",\n \"width = 0.2\\n\",\n \"ind = np.arange(3)\\n\",\n \"fig, ax = plt.subplots()\\n\",\n \"p1 = ax.bar(ind - width/2, bart_ems, width, label=\\\"retrieval-free BART\\\", color='tab:blue')\\n\",\n \"p2 = ax.bar(ind + width/2, sota_ems, width, label=\\\"retrieved-based SoTA\\\", color='tab:red')\\n\",\n \"plt.xticks(ind, datasets)\\n\",\n \"# plt.yticks(np.arange(0, 60, 10))\\n\",\n \"plt.ylabel('Answer EM', fontsize=12)\\n\",\n \"plt.xlabel('Datasets', fontsize=12)\\n\",\n \"\\n\",\n \"def autolabel(rects):\\n\",\n \" \\\"\\\"\\\"Attach a text label above each bar in *rects*, displaying its height.\\\"\\\"\\\"\\n\",\n \" for rect in rects:\\n\",\n \" height = rect.get_height()\\n\",\n \" ax.annotate('{}'.format(height),\\n\",\n \" xy=(rect.get_x() + rect.get_width() / 2, height),\\n\",\n \" xytext=(0, 3), # 3 points vertical offset\\n\",\n \" textcoords=\\\"offset points\\\",\\n\",\n \" ha='center', va='bottom')\\n\",\n \" \\n\",\n \"autolabel(p1)\\n\",\n \"autolabel(p2)\\n\",\n \"\\n\",\n \"lg = plt.legend(loc='upper left', fontsize=10)\\n\",\n \"frame = lg.get_frame()\\n\",\n \"lg.draw_frame(True)\\n\",\n \"plt.ylim(0, 70)\\n\",\n \"plt.savefig('retrieval_free.pdf')\\n\",\n \"plt.show()\"\n ]\n },\n {\n \"cell_type\": \"code\",\n \"execution_count\": 177,\n \"metadata\": {},\n \"outputs\": [\n {\n \"name\": \"stdout\",\n \"output_type\": \"stream\",\n \"text\": [\n \"7405\\n\",\n \"0.3683997299122215\\n\",\n \"0.6267386900742742\\n\"\n ]\n }\n ],\n \"source\": [\n \"# evalute TFIDF hotpotQA retrieval results\\n\",\n \"import json\\n\",\n \"val_inputs = [json.loads(l) for l in open(\\\"/private/home/xwhan/data/hotpot/hotpot_qas_val.json\\\").readlines()]\\n\",\n \"tfidf_results = json.load(open(\\\"/private/home/xwhan/data/hotpot/tfidf/hotpot_dev_tfidf_results.json\\\"))\\n\",\n \"k = 20\\n\",\n \"\\n\",\n \"tfidf_covered = []\\n\",\n \"bm25_covered = []\\n\",\n \"for gold, res in zip(val_inputs, tfidf_results):\\n\",\n \" assert gold[\\\"question\\\"] == res[\\\"question\\\"]\\n\",\n \" \\n\",\n \" gold_sp = gold[\\\"sp\\\"]\\n\",\n \"\\n\",\n \" tfidf_topk = res[\\\"tfidf_topk\\\"][:k]\\n\",\n \" bm25_topk = res[\\\"bm25_topk\\\"][:k]\\n\",\n \" \\n\",\n \" tfidf_covered.append(np.sum([int(_ in tfidf_topk) for _ in gold_sp]) == len(gold_sp))\\n\",\n \" bm25_covered.append(np.sum([int(_ in bm25_topk) for _ in gold_sp]) == len(gold_sp))\\n\",\n \"\\n\",\n \"print(len(tfidf_covered))\\n\",\n \"print(np.mean(tfidf_covered))\\n\",\n \"print(np.mean(bm25_covered))\"\n ]\n },\n {\n \"cell_type\": \"code\",\n \"execution_count\": 188,\n \"metadata\": {},\n \"outputs\": [],\n \"source\": [\n \"# get transformerXH P EM\\n\",\n \"dev_transformer_xh = [json.loads(l) for l in open(\\\"/private/home/xwhan/code/Transformer-XH/data/hotpot_dev_graph.json\\\").readlines()]\\n\",\n \"qid2goldsp = {_[\\\"_id\\\"]:_[\\\"sp\\\"] for _ in val_inputs}\"\n ]\n },\n {\n \"cell_type\": \"code\",\n \"execution_count\": 192,\n \"metadata\": {},\n \"outputs\": [\n {\n \"name\": \"stdout\",\n \"output_type\": \"stream\",\n \"text\": [\n \"5904\\n\",\n \"0.8162262872628726\\n\"\n ]\n }\n ],\n \"source\": [\n \"len(dev_transformer_xh[2][\\\"node\\\"])\\n\",\n \"p_covered = []\\n\",\n \"for item in dev_transformer_xh:\\n\",\n \" node_names = set([_[\\\"name\\\"] for _ in item[\\\"node\\\"]])\\n\",\n \" qid = item[\\\"qid\\\"]\\n\",\n \" gold_sp = qid2goldsp[qid]\\n\",\n \" gold_sp = [_.lower().replace(\\\" \\\", \\\"_\\\") for _ in gold_sp]\\n\",\n \" p_covered.append(np.sum([int(_ in node_names) for _ in gold_sp]) == len(gold_sp))\\n\",\n \"print(len(p_covered))\\n\",\n \"print(np.mean(p_covered))\"\n ]\n },\n {\n \"cell_type\": \"code\",\n \"execution_count\": 294,\n \"metadata\": {},\n \"outputs\": [],\n \"source\": [\n \"id2doc = json.load(open(\\\"/private/home/xwhan/Mhop-Pretrain/retrieval/index/abstracts_id2doc.json\\\"))\\n\",\n \"title2text = {v[0]:v[1] for v in id2doc.values()}\"\n ]\n },\n {\n \"cell_type\": \"code\",\n \"execution_count\": 332,\n \"metadata\": {},\n \"outputs\": [],\n \"source\": [\n \"# pick an intro example to use \\n\",\n \"val_inputs = [json.loads(l) for l in open(\\\"/private/home/xwhan/data/hotpot/hotpot_qas_train.json\\\").readlines()]\\n\",\n \"bridge_val = [_ for _ in val_inputs if _[\\\"type\\\"] == \\\"bridge\\\" and len(_[\\\"question\\\"].split()) < 10]\"\n ]\n },\n {\n \"cell_type\": \"code\",\n \"execution_count\": 371,\n \"metadata\": {},\n \"outputs\": [\n {\n \"name\": \"stdout\",\n \"output_type\": \"stream\",\n \"text\": [\n \"3826\\n\",\n \"{'question': 'In what country is the Holtermann collection located?', '_id': '5adff69955429925eb1afba6', 'answer': ['Australia'], 'sp': ['Holtermann collection', 'Hill End, New South Wales'], 'type': 'bridge'}\\n\"\n ]\n }\n ],\n \"source\": [\n \"print(len(bridge_val))\\n\",\n \"import random\\n\",\n \"random.shuffle(bridge_val)\\n\",\n \"print(bridge_val[0])\"\n ]\n },\n {\n \"cell_type\": \"code\",\n \"execution_count\": 328,\n \"metadata\": {},\n \"outputs\": [\n {\n \"data\": {\n \"text/plain\": [\n \"'Girl 27 is a 2007 documentary film about the 1937 rape of dancer and sometime movie extra Patricia Douglas (1917-2003) at an M-G-M exhibitors\\\\' convention, the front-page news stories that followed, and the studio\\\\'s subsequent cover-up of the crime. Also covered in the film are a similar assault on singer Eloise Spann and her subsequent suicide, and the better-known scandal involving actress Loretta Young and her \\\"adopted\\\" daughter Judy Lewis, the product of her affair with Clark Gable during the production of \\\"The Call of the Wild\\\".'\"\n ]\n },\n \"execution_count\": 328,\n \"metadata\": {},\n \"output_type\": \"execute_result\"\n }\n ],\n \"source\": [\n \"title2text['Clark Gable']\"\n ]\n },\n {\n \"cell_type\": \"code\",\n \"execution_count\": null,\n \"metadata\": {},\n \"outputs\": [],\n \"source\": [\n \"{'question': \\\"What was the nickname of Judy Lewis's father?\\\", '_id': '5a80b9df5542992bc0c4a7ec', 'answer': ['\\\"The King of Hollywood\\\"'], 'sp': ['Girl 27', 'Clark Gable'], 'type': 'bridge'}\\n\",\n \"{'question': 'In what county is The Third Fitzwilliam Meetinghouse located?', '_id': '5a888d535542997e5c09a617', 'answer': ['Cheshire County, New Hampshire'], 'sp': ['Third Fitzwilliam Meetinghouse', 'Fitzwilliam, New Hampshire'], 'type': 'bridge'}\\n\",\n \"{'question': 'Where does the descendant of the Red Setter originate? ', '_id': '5abde4595542991f66106095', 'answer': ['Scotland'], 'sp': ['Irish Setter', 'Scotch Collie'], 'type': 'bridge'}\"\n ]\n },\n {\n \"cell_type\": \"code\",\n \"execution_count\": 453,\n \"metadata\": {},\n \"outputs\": [],\n \"source\": [\n \"corpus = json.load(open(\\\"index/hotpotQA_corpus_dict.json\\\"))\\n\",\n \"title2text = {v[\\\"title\\\"]:v[\\\"text\\\"] for v in corpus.values()}\"\n ]\n },\n {\n \"cell_type\": \"code\",\n \"execution_count\": 456,\n \"metadata\": {},\n \"outputs\": [],\n \"source\": [\n \"# bridge errors after reranking\\n\",\n \"val_inputs = [json.loads(l) for l in open(\\\"/private/home/xwhan/data/hotpot/hotpot_qas_val.json\\\").readlines()]\\n\",\n \"id2goldsp = {_[\\\"_id\\\"]:_[\\\"sp\\\"] for _ in val_inputs}\\n\",\n \"id2goldans = {_[\\\"_id\\\"]:_[\\\"answer\\\"] for _ in val_inputs}\\n\",\n \"id2type = {_[\\\"_id\\\"]:_[\\\"type\\\"] for _ in val_inputs}\\n\",\n \"id2item = {_[\\\"_id\\\"]:_ for _ in val_inputs}\\n\",\n \"results = json.load(open(\\\"/private/home/xwhan/data/hotpot/results/hotpot_val_b250_k250.json\\\"))\"\n ]\n },\n {\n \"cell_type\": \"code\",\n \"execution_count\": 469,\n \"metadata\": {},\n \"outputs\": [\n {\n \"name\": \"stdout\",\n \"output_type\": \"stream\",\n \"text\": [\n \"0.20986819871578236\\n\",\n \"7405\\n\"\n ]\n }\n ],\n \"source\": [\n \"bridge_errors = []\\n\",\n \"bridge_c = 0\\n\",\n \"for qid in results[\\\"titles\\\"].keys():\\n\",\n \" type_ = id2type[qid]\\n\",\n \" if type_ != \\\"bridge\\\":\\n\",\n \" continue\\n\",\n \" chain = results[\\\"titles\\\"][qid]\\n\",\n \" sp = id2goldsp[qid]\\n\",\n \" sp_covered = int(np.sum([int(_ in chain) for _ in sp]) == len(sp))\\n\",\n \" if not sp_covered:\\n\",\n \" bridge_errors.append({\\n\",\n \" \\\"id\\\":qid,\\n\",\n \" \\\"question\\\": id2item[qid][\\\"question\\\"],\\n\",\n \" \\\"sp\\\": sp,\\n\",\n \" \\\"error\\\": chain,\\n\",\n \" \\\"answer\\\": id2goldans[qid],\\n\",\n \" })\\n\",\n \" bridge_c += 1\\n\",\n \"print(len(bridge_errors)/ bridge_c)\\n\",\n \"print(len(results[\\\"titles\\\"].keys()))\"\n ]\n },\n {\n \"cell_type\": \"code\",\n \"execution_count\": 470,\n \"metadata\": {},\n \"outputs\": [],\n \"source\": [\n \"random.shuffle(bridge_errors)\"\n ]\n },\n {\n \"cell_type\": \"code\",\n \"execution_count\": 546,\n \"metadata\": {},\n \"outputs\": [],\n \"source\": [\n \"id2error = {_[\\\"id\\\"]: idx for idx, _ in enumerate(bridge_errors[:50])}\\n\",\n \"json.dump(id2error, open(\\\"/private/home/xwhan/data/hotpot/retrieval_errors_50sampled.json\\\", \\\"w\\\"))\"\n ]\n },\n {\n \"cell_type\": \"code\",\n \"execution_count\": 564,\n \"metadata\": {},\n \"outputs\": [\n {\n \"name\": \"stdout\",\n \"output_type\": \"stream\",\n \"text\": [\n \"{'id': '5abba91e554299642a094b10', 'question': 'What languages did the son of Sacagawea speak?', 'sp': ['Charbonneau, Oregon', 'Jean Baptiste Charbonneau'], 'error': ['Museum of Human Beings', 'Jean Baptiste Charbonneau'], 'answer': ['French and English']}\\n\",\n \"Gold:\\n\",\n \"Charbonneau (also known as the Charbonneau District) is a private planned community within the city limits of Wilsonville in Clackamas County, Oregon, United States. It is on the opposite side the Willamette River from the main area of the city. The development was named for Jean Baptiste Charbonneau, the son of Sacagawea.\\n\",\n \"Jean Baptiste Charbonneau (February 11, 1805 – May 16, 1866) was an American Indian explorer, guide, fur trapper trader, military scout during the Mexican-American War, \\\"alcalde\\\" (mayor) of Mission San Luis Rey de Francia and a gold prospector and hotel operator in Northern California. He spoke French and English, and learned German and Spanish during his six years in Europe from 1823 to 1829. He spoke Shoshone, his mother tongue, and other western American Indian languages, which he picked up during his years of trapping and guiding.\\n\",\n \"Predicted:\\n\",\n \"Museum of Human Beings, included in the National American Indian Heritage Month Booklist, November 2012 and 2013 is a novel written by Colin Sargent, which delves into the heart-rending life of Jean-Baptiste Charbonneau, the son of Sacagawea. Sacagawea was the Native American guide, who at 16 led the Lewis and Clark expedition.\\n\",\n \"Jean Baptiste Charbonneau (February 11, 1805 – May 16, 1866) was an American Indian explorer, guide, fur trapper trader, military scout during the Mexican-American War, \\\"alcalde\\\" (mayor) of Mission San Luis Rey de Francia and a gold prospector and hotel operator in Northern California. He spoke French and English, and learned German and Spanish during his six years in Europe from 1823 to 1829. He spoke Shoshone, his mother tongue, and other western American Indian languages, which he picked up during his years of trapping and guiding.\\n\"\n ]\n }\n ],\n \"source\": [\n \"idx = id2error['5abba91e554299642a094b10']\\n\",\n \"item = bridge_errors[idx]\\n\",\n \"print(item)\\n\",\n \"print(\\\"Gold:\\\")\\n\",\n \"for t in item[\\\"sp\\\"]:\\n\",\n \" print(title2text[t])\\n\",\n \"print(\\\"Predicted:\\\")\\n\",\n \"for t in item[\\\"error\\\"]:\\n\",\n \" print(title2text[t])\"\n ]\n },\n {\n \"cell_type\": \"code\",\n \"execution_count\": null,\n \"metadata\": {},\n \"outputs\": [],\n \"source\": []\n }\n ],\n \"metadata\": {\n \"kernelspec\": {\n \"display_name\": \"Python 3\",\n \"language\": \"python\",\n \"name\": \"python3\"\n },\n \"language_info\": {\n \"codemirror_mode\": {\n \"name\": \"ipython\",\n \"version\": 3\n },\n \"file_extension\": \".py\",\n \"mimetype\": \"text/x-python\",\n \"name\": \"python\",\n \"nbconvert_exporter\": \"python\",\n \"pygments_lexer\": \"ipython3\",\n \"version\": \"3.6.3\"\n }\n },\n \"nbformat\": 4,\n \"nbformat_minor\": 2\n}\n" }, { "path": "mdr/retrieval/interactive_retrieval.py", "content": "# Copyright (c) Facebook, Inc. and its affiliates.\n# All rights reserved.\n#\n# This source code is licensed under the license found in the \n# LICENSE file in the root directory of this source tree.\nfrom models.mhop_retriever import MhopRetriever\n\n\nimport faiss\nimport numpy as np\nimport torch\nfrom tqdm import tqdm\nfrom transformers import AutoConfig, AutoTokenizer\nimport json\nimport logging\nimport argparse\nfrom .utils.utils import (load_saved, move_to_cuda)\n\nparser = argparse.ArgumentParser()\nparser.add_argument('--topk', type=int, default=2, help=\"topk paths\")\nparser.add_argument('--num-workers', type=int, default=10)\nparser.add_argument('--max-q-len', type=int, default=70)\nparser.add_argument('--max-c-len', type=int, default=300)\nparser.add_argument('--max-q-sp-len', type=int, default=350)\nparser.add_argument('--model-name', type=str, default='bert-base-uncased')\nparser.add_argument('--gpu', action=\"store_true\")\nparser.add_argument('--shared-encoder', action=\"store_true\")\nparser.add_argument(\"--stop-drop\", default=0, type=float)\nargs = parser.parse_args()\n\nindex_path = \"index/abstracts_v0_fixed.npy\"\ncorpus_path = \"index/abstracts_id2doc.json\"\nmodel_path = \"logs/08-05-2020/baseline_v0_fixed-seed16-bsz150-fp16True-lr2e-05-decay0.0-warm0.1-valbsz3000-sharedTrue-multi1-schemenone/checkpoint_best.pt\"\n\nprint(f\"Loading corpus and index...\")\nid2doc = json.load(open(corpus_path))\nindex_vectors = np.load(index_path).astype('float32')\n\nindex = faiss.IndexFlatIP(768)\nindex.add(index_vectors)\nres = faiss.StandardGpuResources()\nindex = faiss.index_cpu_to_gpu(res, 1, index)\n\nprint(f\"Loading retrieval model...\")\nbert_config = AutoConfig.from_pretrained(\"bert-base-uncased\")\ntokenizer = AutoTokenizer.from_pretrained(\"bert-base-uncased\")\nmodel = MhopRetriever(bert_config, args)\nmodel = load_saved(model, args.model_path, exact=False)\n\ncuda = torch.device('cuda')\nmodel.to(cuda)\nfrom apex import amp\nmodel = amp.initialize(model, opt_level='O1')\nmodel.eval()\n\nwhile True:\n question = input(\"Type Question:\")\n question = \"the Danish musicians who died in 1931\"\n batch_q_encodes = tokenizer.batch_encode_plus([\"question\"], max_length=args.max_q_len, pad_to_max_length=True, return_tensors=\"pt\")\n batch_q_encodes = move_to_cuda(dict(batch_q_encodes))\n q_embeds = model.encode_q(batch_q_encodes[\"input_ids\"], batch_q_encodes[\"attention_mask\"], batch_q_encodes.get(\"token_type_ids\", None))\n q_embeds_numpy = q_embeds.cpu().contiguous().numpy() \n D, I = index.search(q_embeds_numpy, 1)\n\n print(I)\n\n" }, { "path": "mdr/retrieval/mhop_trainer.py", "content": "# Copyright (c) Facebook, Inc. and its affiliates.\n# All rights reserved.\n#\n# This source code is licensed under the license found in the \n# LICENSE file in the root directory of this source tree.\n\n\"\"\"\nsubmitit trainer for hyperparameter tuning\n\"\"\"\n\nimport os\nimport os.path as osp\nfrom typing import Optional, NamedTuple\nimport torch\nimport torch.distributed\nimport torch.nn as nn\nimport torch.optim as optim\nimport attr\nimport submitit\nfrom functools import partial\nimport numpy as np\nimport random\nfrom torch.utils.tensorboard import SummaryWriter\nfrom pathlib import Path\nimport json\nfrom transformers import (\n AdamW, AutoConfig, AutoTokenizer, get_linear_schedule_with_warmup)\n\nfrom torch.optim import Adam\nfrom .utils.utils import move_to_cuda, AverageMeter\nfrom .config import ClusterConfig\nfrom .data.mhop_dataset import MhopDataset, mhop_collate\nfrom .models.mhop_retriever import (MhopRetriever, RobertaRetriever)\nfrom .criterions import (mhop_loss, mhop_eval)\n\nfrom tqdm import tqdm\nimport apex\napex.amp.register_half_function(torch, 'einsum')\nfrom apex import amp\n\n@attr.s(auto_attribs=True)\nclass TrainerState:\n \"\"\"\n Contains the state of the Trainer.\n It can be saved to checkpoint the training and loaded to resume it.\n \"\"\"\n\n epoch: int\n model: nn.Module\n optimizer: optim.Optimizer\n lr_scheduler: torch.optim.lr_scheduler._LRScheduler\n global_step: int\n\n def save(self, filename: str) -> None:\n data = attr.asdict(self)\n # store only the state dict\n data[\"model\"] = self.model.state_dict()\n data[\"optimizer\"] = self.optimizer.state_dict()\n data[\"lr_scheduler\"] = self.lr_scheduler.state_dict()\n torch.save(data, filename)\n\n @classmethod\n def load(cls, filename: str, default: \"TrainerState\", gpu: int) -> \"TrainerState\":\n data = torch.load(filename, map_location=lambda storage, loc: storage.cuda(gpu))\n # We need this default to load the state dict\n model = default.model\n model.load_state_dict(data[\"model\"])\n data[\"model\"] = model\n\n optimizer = default.optimizer\n optimizer.load_state_dict(data[\"optimizer\"])\n data[\"optimizer\"] = optimizer\n\n lr_scheduler = default.lr_scheduler\n lr_scheduler.load_state_dict(data[\"lr_scheduler\"])\n data[\"lr_scheduler\"] = lr_scheduler\n\n return cls(**data)\n\nclass Trainer:\n def __init__(self, train_cfg: NamedTuple, cluster_cfg: ClusterConfig) -> None:\n self._train_cfg = train_cfg\n self._cluster_cfg = cluster_cfg\n\n def __call__(self) -> Optional[float]:\n \"\"\"\n Called by submitit for each task.\n :return: The master task return the final accuracy of the model.\n \"\"\"\n self._setup_process_group()\n self._init_state()\n final_acc = self._train()\n return final_acc\n\n def log(self, log_data: dict):\n job_env = submitit.JobEnvironment()\n # z = {**vars(self._train_cfg), **log_data}\n save_dir = Path(self._train_cfg.output_dir)\n os.makedirs(save_dir, exist_ok=True)\n with open(save_dir / 'log.txt', 'a') as f:\n f.write(json.dumps(log_data) + '\\n')\n\n def checkpoint(self, rm_init=True) -> submitit.helpers.DelayedSubmission:\n # will be called by submitit in case of preemption\n job_env = submitit.JobEnvironment()\n save_dir = osp.join(self._train_cfg.output_dir, str(job_env.job_id))\n os.makedirs(save_dir, exist_ok=True)\n self._state.save(osp.join(save_dir, \"checkpoint.pth\"))\n\n # Trick here: when the job will be requeue, we will use the same init file\n # but it must not exist when we initialize the process group\n # so we delete it, but only when this method is called by submitit for requeue\n if rm_init and osp.exists(self._cluster_cfg.dist_url[7:]):\n os.remove(self._cluster_cfg.dist_url[7:]) # remove file:// at the beginning\n # This allow to remove any non-pickable part of the Trainer instance.\n empty_trainer = Trainer(self._train_cfg, self._cluster_cfg)\n return submitit.helpers.DelayedSubmission(empty_trainer)\n\n def _setup_process_group(self) -> None:\n job_env = submitit.JobEnvironment()\n torch.cuda.set_device(job_env.local_rank)\n torch.distributed.init_process_group(\n backend=self._cluster_cfg.dist_backend,\n init_method=self._cluster_cfg.dist_url,\n world_size=job_env.num_tasks,\n rank=job_env.global_rank,\n )\n print(f\"Process group: {job_env.num_tasks} tasks, rank: {job_env.global_rank}\")\n\n def _init_state(self) -> None:\n \"\"\"\n Initialize the state and load it from an existing checkpoint if any\n \"\"\"\n job_env = submitit.JobEnvironment()\n\n if job_env.global_rank == 0:\n # config_path = Path(args.save_folder) / str(job_env.job_id) / 'config.json'\n os.makedirs(self._train_cfg.output_dir, exist_ok=True)\n config_path = Path(self._train_cfg.output_dir) / 'config.json'\n with open(config_path, \"w\") as g:\n g.write(json.dumps(self._train_cfg._asdict()))\n\n print(f\"Setting random seed {self._train_cfg.seed}\", flush=True)\n random.seed(self._train_cfg.seed)\n np.random.seed(self._train_cfg.seed)\n torch.manual_seed(self._train_cfg.seed)\n torch.cuda.manual_seed_all(self._train_cfg.seed)\n\n print(\"Create data loaders\", flush=True)\n tokenizer = AutoTokenizer.from_pretrained(self._train_cfg.model_name)\n collate_fc = partial(mhop_collate, pad_id=tokenizer.pad_token_id)\n train_set = MhopDataset(tokenizer, self._train_cfg.train_file, self._train_cfg.max_q_len, self._train_cfg.max_q_sp_len, self._train_cfg.max_c_len, train=True)\n\n self._train_loader = torch.utils.data.DataLoader(train_set, batch_size=self._train_cfg.train_batch_size, num_workers=self._train_cfg.num_workers, collate_fn=collate_fc, shuffle=True)\n test_set = MhopDataset(tokenizer, self._train_cfg.predict_file, self._train_cfg.max_q_len, self._train_cfg.max_q_sp_len, self._train_cfg.max_c_len)\n\n self._test_loader = torch.utils.data.DataLoader(\n test_set,\n batch_size=self._train_cfg.predict_batch_size,\n num_workers=self._train_cfg.num_workers, collate_fn=collate_fc, pin_memory=True\n )\n\n print(\"Create model\", flush=True)\n print(f\"Local rank {job_env.local_rank}\", flush=True)\n bert_config = AutoConfig.from_pretrained(self._train_cfg.model_name)\n if \"roberta\" in self._train_cfg.model_name:\n model = RobertaRetriever(bert_config, self._train_cfg)\n else:\n model = MhopRetriever(bert_config, self._train_cfg)\n model.cuda(job_env.local_rank)\n\n no_decay = ['bias', 'LayerNorm.weight']\n optimizer_parameters = [\n {'params': [p for n, p in model.named_parameters() if not any(\n nd in n for nd in no_decay)], 'weight_decay': self._train_cfg.weight_decay},\n {'params': [p for n, p in model.named_parameters() if any(\n nd in n for nd in no_decay)], 'weight_decay': 0.0}\n ]\n optimizer = Adam(optimizer_parameters, lr=self._train_cfg.learning_rate, eps=self._train_cfg.adam_epsilon)\n\n if self._train_cfg.fp16:\n model, optimizer = amp.initialize(\n model, optimizer, opt_level=self._train_cfg.fp16_opt_level)\n\n t_total = len(self._train_loader) // self._train_cfg.gradient_accumulation_steps * self._train_cfg.num_train_epochs\n warmup_steps = t_total * self._train_cfg.warmup_ratio\n lr_scheduler = get_linear_schedule_with_warmup(\n optimizer, num_warmup_steps=warmup_steps, num_training_steps=t_total\n )\n model = torch.nn.DataParallel(model)\n self._state = TrainerState(\n epoch=0, model=model, optimizer=optimizer, lr_scheduler=lr_scheduler, global_step=0\n )\n\n self.tb_logger = SummaryWriter(self._train_cfg.output_dir.replace(\"logs\", \"tflogs\"))\n\n checkpoint_fn = osp.join(self._train_cfg.output_dir, str(job_env.job_id), \"checkpoint.pth\")\n # checkpoint_fn = osp.join(self._train_cfg.output_dir, \"checkpoint.pth\")\n if os.path.isfile(checkpoint_fn):\n print(f\"Load existing checkpoint from {checkpoint_fn}\", flush=True)\n self._state = TrainerState.load(\n checkpoint_fn, default=self._state, gpu=job_env.local_rank)\n\n def _train(self) -> Optional[float]:\n job_env = submitit.JobEnvironment()\n\n batch_step = 0 # forward batch count\n best_mrr = 0\n train_loss_meter = AverageMeter()\n print(f\"Start training\", flush=True)\n # Start from the loaded epoch\n start_epoch = self._state.epoch\n global_step = self._state.global_step\n for epoch in range(start_epoch, self._train_cfg.num_train_epochs):\n print(f\"Start epoch {epoch}\", flush=True)\n self._state.model.train()\n self._state.epoch = epoch\n for batch in self._train_loader:\n batch_step += 1\n batch = move_to_cuda(batch)\n loss = mhop_loss(self._state.model, batch, self._train_cfg)\n\n if self._train_cfg.gradient_accumulation_steps > 1:\n loss = loss / self._train_cfg.gradient_accumulation_steps\n if self._train_cfg.fp16:\n with amp.scale_loss(loss, self._state.optimizer) as scaled_loss:\n scaled_loss.backward()\n else:\n loss.backward()\n \n train_loss_meter.update(loss.item())\n\n if (batch_step + 1) % self._train_cfg.gradient_accumulation_steps == 0:\n if self._train_cfg.fp16:\n torch.nn.utils.clip_grad_norm_(\n amp.master_params(self._state.optimizer), self._train_cfg.max_grad_norm)\n else:\n torch.nn.utils.clip_grad_norm_(\n self._state.model.parameters(), self._train_cfg.max_grad_norm)\n self._state.optimizer.step()\n self._state.lr_scheduler.step()\n self._state.model.zero_grad()\n\n global_step += 1\n self._state.global_step = global_step\n\n self.tb_logger.add_scalar('batch_train_loss',\n loss.item(), global_step)\n self.tb_logger.add_scalar('smoothed_train_loss',\n train_loss_meter.avg, global_step)\n\n # Checkpoint only on the master\n # if job_env.global_rank == 0:\n self.checkpoint(rm_init=False)\n mrrs = self._eval()\n mrr = mrrs[\"mrr_avg\"]\n self.tb_logger.add_scalar('dev_mrr', mrr*100, epoch)\n self._state.lr_scheduler.step(mrr)\n if best_mrr < mrr:\n print(\"Saving model with best MRR %.2f -> MRR %.2f on epoch=%d\" % (best_mrr*100, mrr*100, epoch))\n torch.save(self._state.model.state_dict(), os.path.join(self._train_cfg.output_dir, f\"checkpoint_best.pt\"))\n best_mrr = mrr\n self.log({\n \"best_mrr\": best_mrr,\n \"curr_mrr\": mrr,\n \"smoothed_loss\": train_loss_meter.avg,\n \"epoch\": epoch\n })\n return best_mrr\n\n def _eval(self) -> float:\n print(\"Start evaluation of the model\", flush=True)\n job_env = submitit.JobEnvironment()\n args = self._train_cfg\n eval_dataloader = self._test_loader\n self._state.model.eval()\n rrs_1, rrs_2 = [], [] # reciprocal rank\n for batch in tqdm(eval_dataloader):\n batch_to_feed = move_to_cuda(batch)\n with torch.no_grad():\n outputs = self._state.model(batch_to_feed)\n eval_results = mhop_eval(outputs, args)\n _rrs_1, _rrs_2 = eval_results[\"rrs_1\"], eval_results[\"rrs_2\"]\n rrs_1 += _rrs_1\n rrs_2 += _rrs_2\n mrr_1 = np.mean(rrs_1)\n mrr_2 = np.mean(rrs_2)\n print(f\"evaluated {len(rrs_1)} examples...\")\n print(f'MRR-1: {mrr_1}')\n print(f'MRR-2: {mrr_2}')\n self._state.model.train()\n return {\"mrr_1\": mrr_1, \"mrr_2\": mrr_2, \"mrr_avg\": (mrr_1 + mrr_2) / 2}\n" }, { "path": "mdr/retrieval/single_trainer.py", "content": "# Copyright (c) Facebook, Inc. and its affiliates.\n# All rights reserved.\n#\n# This source code is licensed under the license found in the \n# LICENSE file in the root directory of this source tree.\n\n\n\"\"\"\ntrainer defined for submitit hyperparameter tuning\n\"\"\"\n\nimport os\nimport os.path as osp\nfrom typing import Optional, NamedTuple\nimport torch\nimport torch.distributed\nimport torch.nn as nn\nimport torch.optim as optim\nimport attr\nimport submitit\nimport argparse\nfrom functools import partial\nfrom torch.nn import CrossEntropyLoss\nimport numpy as np\nimport random\nfrom torch.utils.tensorboard import SummaryWriter\nfrom pathlib import Path\n\nfrom .utils import move_to_cuda, convert_to_half, AverageMeter\nfrom .config import ClusterConfig\nfrom .data.sp_datasets import SPDataset, sp_collate\nfrom .models.retriever import BertForRetrieverSP\nfrom transformers import AdamW, BertConfig, BertTokenizer\nimport json\n\nimport apex\napex.amp.register_half_function(torch, 'einsum')\nfrom apex import amp\n\n@attr.s(auto_attribs=True)\nclass TrainerState:\n \"\"\"\n Contains the state of the Trainer.\n It can be saved to checkpoint the training and loaded to resume it.\n \"\"\"\n\n epoch: int\n model: nn.Module\n optimizer: optim.Optimizer\n lr_scheduler: torch.optim.lr_scheduler._LRScheduler\n global_step: int\n\n def save(self, filename: str) -> None:\n data = attr.asdict(self)\n # store only the state dict\n data[\"model\"] = self.model.state_dict()\n data[\"optimizer\"] = self.optimizer.state_dict()\n data[\"lr_scheduler\"] = self.lr_scheduler.state_dict()\n torch.save(data, filename)\n\n @classmethod\n def load(cls, filename: str, default: \"TrainerState\", gpu: int) -> \"TrainerState\":\n data = torch.load(filename, map_location=lambda storage, loc: storage.cuda(gpu))\n # We need this default to load the state dict\n model = default.model\n model.load_state_dict(data[\"model\"])\n data[\"model\"] = model\n\n optimizer = default.optimizer\n optimizer.load_state_dict(data[\"optimizer\"])\n data[\"optimizer\"] = optimizer\n\n lr_scheduler = default.lr_scheduler\n lr_scheduler.load_state_dict(data[\"lr_scheduler\"])\n data[\"lr_scheduler\"] = lr_scheduler\n\n return cls(**data)\n\nclass Trainer:\n def __init__(self, train_cfg: NamedTuple, cluster_cfg: ClusterConfig) -> None:\n self._train_cfg = train_cfg\n self._cluster_cfg = cluster_cfg\n\n def __call__(self) -> Optional[float]:\n \"\"\"\n Called by submitit for each task.\n :return: The master task return the final accuracy of the model.\n \"\"\"\n self._setup_process_group()\n self._init_state()\n final_acc = self._train()\n return final_acc\n\n def log(self, log_data: dict):\n job_env = submitit.JobEnvironment()\n # z = {**vars(self._train_cfg), **log_data}\n save_dir = Path(self._train_cfg.output_dir)\n os.makedirs(save_dir, exist_ok=True)\n with open(save_dir / 'log.txt', 'a') as f:\n f.write(json.dumps(log_data) + '\\n')\n\n def checkpoint(self, rm_init=True) -> submitit.helpers.DelayedSubmission:\n # will be called by submitit in case of preemption\n job_env = submitit.JobEnvironment()\n save_dir = osp.join(self._train_cfg.output_dir, str(job_env.job_id))\n os.makedirs(save_dir, exist_ok=True)\n self._state.save(osp.join(save_dir, \"checkpoint.pth\"))\n\n # Trick here: when the job will be requeue, we will use the same init file\n # but it must not exist when we initialize the process group\n # so we delete it, but only when this method is called by submitit for requeue\n if rm_init and osp.exists(self._cluster_cfg.dist_url[7:]):\n os.remove(self._cluster_cfg.dist_url[7:]) # remove file:// at the beginning\n # This allow to remove any non-pickable part of the Trainer instance.\n empty_trainer = Trainer(self._train_cfg, self._cluster_cfg)\n return submitit.helpers.DelayedSubmission(empty_trainer)\n\n def _setup_process_group(self) -> None:\n job_env = submitit.JobEnvironment()\n torch.cuda.set_device(job_env.local_rank)\n torch.distributed.init_process_group(\n backend=self._cluster_cfg.dist_backend,\n init_method=self._cluster_cfg.dist_url,\n world_size=job_env.num_tasks,\n rank=job_env.global_rank,\n )\n print(f\"Process group: {job_env.num_tasks} tasks, rank: {job_env.global_rank}\")\n\n def _init_state(self) -> None:\n \"\"\"\n Initialize the state and load it from an existing checkpoint if any\n \"\"\"\n job_env = submitit.JobEnvironment()\n\n if job_env.global_rank == 0:\n # config_path = Path(args.save_folder) / str(job_env.job_id) / 'config.json'\n os.makedirs(self._train_cfg.output_dir, exist_ok=True)\n config_path = Path(self._train_cfg.output_dir) / 'config.json'\n with open(config_path, \"w\") as g:\n g.write(json.dumps(self._train_cfg._asdict()))\n\n print(f\"Setting random seed {self._train_cfg.seed}\", flush=True)\n random.seed(self._train_cfg.seed)\n np.random.seed(self._train_cfg.seed)\n torch.manual_seed(self._train_cfg.seed)\n\n print(\"Create data loaders\", flush=True)\n tokenizer = BertTokenizer.from_pretrained(self._train_cfg.bert_model_name)\n collate_fc = sp_collate\n train_set = SPDataset(tokenizer, self._train_cfg.train_file, self._train_cfg.max_q_len, self._train_cfg.max_c_len, train=True)\n # train_sampler = torch.utils.data.distributed.DistributedSampler(\n # train_set, num_replicas=job_env.num_tasks, rank=job_env.global_rank\n # )\n # self._train_loader = torch.utils.data.DataLoader(\n # train_set,\n # batch_size=self._train_cfg.train_batch_size,\n # num_workers=4,\n # sampler=train_sampler, collate_fn=collate_fc\n # )\n self._train_loader = torch.utils.data.DataLoader(train_set, batch_size=self._train_cfg.train_batch_size, num_workers=4, collate_fn=collate_fc)\n test_set = SPDataset(tokenizer, self._train_cfg.predict_file, self._train_cfg.max_q_len, self._train_cfg.max_c_len)\n self._test_loader = torch.utils.data.DataLoader(\n test_set,\n batch_size=self._train_cfg.predict_batch_size,\n num_workers=4, collate_fn=collate_fc\n )\n print(f\"Per Node batch_size: {self._train_cfg.train_batch_size // job_env.num_tasks}\", flush=True)\n\n print(\"Create model\", flush=True)\n print(f\"Local rank {job_env.local_rank}\", flush=True)\n bert_config = BertConfig.from_pretrained(self._train_cfg.bert_model_name)\n model = BertForRetrieverSP(bert_config, self._train_cfg)\n model.cuda(job_env.local_rank)\n\n no_decay = ['bias', 'LayerNorm.weight']\n optimizer_parameters = [\n {'params': [p for n, p in model.named_parameters() if not any(\n nd in n for nd in no_decay)], 'weight_decay': self._train_cfg.weight_decay},\n {'params': [p for n, p in model.named_parameters() if any(\n nd in n for nd in no_decay)], 'weight_decay': 0.0}\n ]\n optimizer = AdamW(optimizer_parameters,\n lr=self._train_cfg.learning_rate)\n lr_scheduler = optim.lr_scheduler.ReduceLROnPlateau(optimizer, mode='max', factor=0.5)\n\n if self._train_cfg.fp16:\n model, optimizer = amp.initialize(\n model, optimizer, opt_level=self._train_cfg.fp16_opt_level)\n model = torch.nn.DataParallel(model) # \n self._state = TrainerState(\n epoch=0, model=model, optimizer=optimizer, lr_scheduler=lr_scheduler, global_step=0\n )\n\n self.tb_logger = SummaryWriter(os.path.join(self._train_cfg.output_dir, \"tblog\"))\n\n checkpoint_fn = osp.join(self._train_cfg.output_dir, str(job_env.job_id), \"checkpoint.pth\")\n # checkpoint_fn = osp.join(self._train_cfg.output_dir, \"checkpoint.pth\")\n if os.path.isfile(checkpoint_fn):\n print(f\"Load existing checkpoint from {checkpoint_fn}\", flush=True)\n self._state = TrainerState.load(\n checkpoint_fn, default=self._state, gpu=job_env.local_rank)\n\n def _train(self) -> Optional[float]:\n job_env = submitit.JobEnvironment()\n\n loss_fct = CrossEntropyLoss()\n batch_step = 0 # forward batch count\n best_mrr = 0\n train_loss_meter = AverageMeter()\n print(f\"Start training\", flush=True)\n # Start from the loaded epoch\n start_epoch = self._state.epoch\n global_step = self._state.global_step\n for epoch in range(start_epoch, self._train_cfg.num_train_epochs):\n print(f\"Start epoch {epoch}\", flush=True)\n self._state.model.train()\n self._state.epoch = epoch\n\n for batch in self._train_loader:\n batch_step += 1\n batch = move_to_cuda(batch)\n outputs = self._state.model(batch)\n q = outputs['q']\n c = outputs['c']\n neg_c = outputs['neg_c']\n product_in_batch = torch.mm(q, c.t())\n product_neg = (q * neg_c).sum(-1).unsqueeze(1)\n product = torch.cat([product_in_batch, product_neg], dim=-1)\n target = torch.arange(product.size(0)).to(product.device)\n loss = loss_fct(product, target)\n\n if self._train_cfg.gradient_accumulation_steps > 1:\n loss = loss / self._train_cfg.gradient_accumulation_steps\n if self._train_cfg.fp16:\n with amp.scale_loss(loss, self._state.optimizer) as scaled_loss:\n scaled_loss.backward()\n else:\n loss.backward()\n \n train_loss_meter.update(loss.item())\n self.tb_logger.add_scalar('batch_train_loss',\n loss.item(), global_step)\n self.tb_logger.add_scalar('smoothed_train_loss',\n train_loss_meter.avg, global_step)\n\n if (batch_step + 1) % self._train_cfg.gradient_accumulation_steps == 0:\n if self._train_cfg.fp16:\n torch.nn.utils.clip_grad_norm_(\n amp.master_params(self._state.optimizer), self._train_cfg.max_grad_norm)\n else:\n torch.nn.utils.clip_grad_norm_(\n self._state.model.parameters(), self._train_cfg.max_grad_norm)\n self._state.optimizer.step() # We have accumulated enought gradients\n self._state.model.zero_grad()\n global_step += 1\n self._state.global_step = global_step\n\n # Checkpoint only on the master\n # if job_env.global_rank == 0:\n self.checkpoint(rm_init=False)\n mrr = self._eval()\n self.tb_logger.add_scalar('dev_mrr', mrr*100, epoch)\n self._state.lr_scheduler.step(mrr)\n if best_mrr < mrr:\n print(\"Saving model with best MRR %.2f -> MRR %.2f on epoch=%d\" % (best_mrr*100, mrr*100, epoch))\n torch.save(self._state.model.state_dict(), os.path.join(self._train_cfg.output_dir, f\"checkpoint_best.pt\"))\n best_mrr = mrr\n self.log({\n \"best_mrr\": best_mrr,\n \"curr_mrr\": mrr,\n \"smoothed_loss\": train_loss_meter.avg,\n \"epoch\": epoch\n })\n return best_mrr\n\n def _eval(self) -> float:\n print(\"Start evaluation of the model\", flush=True)\n job_env = submitit.JobEnvironment()\n args = self._train_cfg\n eval_dataloader = self._test_loader\n num_correct = 0\n num_total = 0.0\n rrs = [] # reciprocal rank\n self._state.model.eval()\n for batch in self._test_loader:\n batch_to_feed = move_to_cuda(batch)\n with torch.no_grad():\n outputs = self._state.model(batch_to_feed)\n q = outputs['q']\n c = outputs['c']\n neg_c = outputs['neg_c']\n\n product_in_batch = torch.mm(q, c.t()) \n product_neg = (q * neg_c).sum(-1).unsqueeze(1)\n product = torch.cat([product_in_batch, product_neg], dim=-1)\n\n target = torch.arange(product.size(0)).to(product.device)\n ranked = product.argsort(dim=1, descending=True)\n\n # MRR\n idx2rank = ranked.argsort(dim=1)\n for idx, t in enumerate(target.tolist()):\n rrs.append(1 / (idx2rank[idx][t].item() +1))\n\n prediction = product.argmax(-1)\n pred_res = prediction == target\n\n num_total += pred_res.size(0)\n num_correct += pred_res.sum(0)\n\n acc = num_correct/num_total\n mrr = np.mean(rrs)\n print(f\"evaluated {num_total} examples...\", flush=True)\n print(f\"avg. Acc: {acc}\", flush=True)\n print(f'MRR: {mrr}', flush=True)\n self._state.model.train()\n return mrr\n" }, { "path": "mdr/retrieval/train_single.py", "content": "# Copyright (c) Facebook, Inc. and its affiliates.\n# All rights reserved.\n#\n# This source code is licensed under the license found in the \n# LICENSE file in the root directory of this source tree.\n\"\"\"\n# DPR baseline shared encoder\n\nCUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python train_single.py \\\n --do_train \\\n --prefix nq_dpr_shared \\\n --predict_batch_size 5000 \\\n --model_name bert-base-uncased \\\n --train_batch_size 256 \\\n --gradient_accumulation_steps 1 \\\n --accumulate_gradients 1 \\\n --learning_rate 2e-5 \\\n --fp16 \\\n --train_file /private/home/xwhan/data/nq-dpr/nq-with-neg-train.txt \\\n --predict_file /private/home/xwhan/data/nq-dpr/nq-with-neg-dev.txt \\\n --seed 16 \\\n --eval-period -1 \\\n --max_c_len 300 \\\n --max_q_len 50 \\\n --warmup-ratio 0.1 \\\n --shared-encoder \\\n --num_train_epochs 50\n\n# WebQ single train\nCUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python train_single.py \\\n --do_train \\\n --prefix wq_dpr_shared \\\n --predict_batch_size 5000 \\\n --model_name bert-base-uncased \\\n --train_batch_size 256 \\\n --gradient_accumulation_steps 1 \\\n --accumulate_gradients 1 \\\n --learning_rate 2e-5 \\\n --fp16 \\\n --train_file /private/home/xwhan/data/WebQ/wq-train-simplified.txt \\\n --predict_file /private/home/xwhan/data/WebQ/wq-dev-simplified.txt \\\n --seed 16 \\\n --eval-period -1 \\\n --max_c_len 300 \\\n --max_q_len 50 \\\n --warmup-ratio 0.1 \\\n --shared-encoder \\\n --num_train_epochs 50\n\n# FEVER single-hop retrieval\nCUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python train_single.py \\\n --do_train \\\n --prefix fever_single \\\n --predict_batch_size 5000 \\\n --model_name bert-base-uncased \\\n --train_batch_size 256 \\\n --gradient_accumulation_steps 1 \\\n --accumulate_gradients 1 \\\n --learning_rate 2e-5 \\\n --fp16 \\\n --train_file /private/home/xwhan/data/fever/retrieval/train_tfidf_neg.txt \\\n --predict_file /private/home/xwhan/data/fever/retrieval/dev_tfidf_neg.txt \\\n --seed 16 \\\n --eval-period -1 \\\n --max_c_len 400 \\\n --max_q_len 45 \\\n --shared-encoder \\\n --num_train_epochs 40\n\n# HotpotQA single-hop\nCUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python train_single.py \\\n --do_train \\\n --prefix hotpot_single \\\n --predict_batch_size 5000 \\\n --model_name roberta-base \\\n --train_batch_size 256 \\\n --gradient_accumulation_steps 1 \\\n --accumulate_gradients 1 \\\n --learning_rate 2e-5 \\\n --fp16 \\\n --train_file /private/home/xwhan/data/hotpot/hotpot_train_with_neg_v0.json \\\n --predict_file /private/home/xwhan/data/hotpot/hotpot_val_with_neg_v0.json \\\n --seed 16 \\\n --eval-period -1 \\\n --max_c_len 300 \\\n --max_q_len 70 \\\n --shared-encoder \\\n --warmup-ratio 0.1 \\\n --num_train_epochs 50\n \n\"\"\"\nimport logging\nimport os\nimport random\nfrom tqdm import tqdm\nimport numpy as np\nimport torch\nfrom datetime import date\nfrom torch.utils.data import DataLoader\nfrom models.retriever import BertRetrieverSingle, RobertaRetrieverSingle, MomentumRetriever\nfrom transformers import AdamW, AutoConfig, AutoTokenizer, get_linear_schedule_with_warmup\nfrom torch.utils.tensorboard import SummaryWriter\nfrom data.sp_datasets import SPDataset, sp_collate, NQMhopDataset, FeverSingleDataset\nfrom utils.utils import move_to_cuda, AverageMeter, load_saved\nfrom config import train_args\nfrom criterions import loss_single\nfrom torch.optim import Adam\nfrom functools import partial\nimport apex\n\ndef main():\n args = train_args()\n\n if args.fp16:\n apex.amp.register_half_function(torch, 'einsum')\n\n date_curr = date.today().strftime(\"%m-%d-%Y\")\n model_name = f\"{args.prefix}-seed{args.seed}-bsz{args.train_batch_size}-fp16{args.fp16}-lr{args.learning_rate}-decay{args.weight_decay}-warm{args.warmup_ratio}-{args.model_name}\"\n args.output_dir = os.path.join(args.output_dir, date_curr, model_name)\n tb_logger = SummaryWriter(os.path.join(args.output_dir.replace(\"logs\",\"tflogs\")))\n\n if os.path.exists(args.output_dir) and os.listdir(args.output_dir):\n print(\n f\"output directory {args.output_dir} already exists and is not empty.\")\n os.makedirs(args.output_dir, exist_ok=True)\n\n logging.basicConfig(format='%(asctime)s - %(levelname)s - %(name)s - %(message)s', datefmt='%m/%d/%Y %H:%M:%S',\n level=logging.INFO,\n handlers=[logging.FileHandler(os.path.join(args.output_dir, \"log.txt\")),\n logging.StreamHandler()])\n logger = logging.getLogger(__name__)\n logger.info(args)\n\n if args.local_rank == -1 or args.no_cuda:\n device = torch.device(\n \"cuda\" if torch.cuda.is_available() and not args.no_cuda else \"cpu\")\n n_gpu = torch.cuda.device_count()\n else:\n device = torch.device(\"cuda\", args.local_rank)\n n_gpu = 1\n torch.distributed.init_process_group(backend='nccl')\n logger.info(\"device %s n_gpu %d distributed training %r\",\n device, n_gpu, bool(args.local_rank != -1))\n\n args.train_batch_size = int(\n args.train_batch_size / args.accumulate_gradients)\n random.seed(args.seed)\n np.random.seed(args.seed)\n torch.manual_seed(args.seed)\n if n_gpu > 0:\n torch.cuda.manual_seed_all(args.seed)\n if not args.do_train and not args.do_predict:\n raise ValueError(\n \"At least one of `do_train` or `do_predict` must be True.\")\n\n bert_config = AutoConfig.from_pretrained(args.model_name)\n if args.momentum:\n model = MomentumRetriever(bert_config, args)\n elif \"roberta\" in args.model_name:\n model = RobertaRetrieverSingle(bert_config, args)\n else:\n model = BertRetrieverSingle(bert_config, args)\n \n tokenizer = AutoTokenizer.from_pretrained(args.model_name)\n collate_fc = partial(sp_collate, pad_id=tokenizer.pad_token_id)\n\n if args.do_train and args.max_c_len > bert_config.max_position_embeddings:\n raise ValueError(\n \"Cannot use sequence length %d because the BERT model \"\n \"was only trained up to sequence length %d\" %\n (args.max_c_len, bert_config.max_position_embeddings))\n\n if \"fever\" in args.predict_file:\n eval_dataset = FeverSingleDataset(tokenizer, args.predict_file, args.max_q_len, args.max_c_len)\n else:\n eval_dataset = SPDataset(tokenizer, args.predict_file, args.max_q_len, args.max_c_len)\n eval_dataloader = DataLoader(\n eval_dataset, batch_size=args.predict_batch_size, collate_fn=collate_fc, pin_memory=True, num_workers=args.num_workers)\n logger.info(f\"Num of dev batches: {len(eval_dataloader)}\")\n\n if args.init_checkpoint != \"\":\n model = load_saved(model, args.init_checkpoint)\n\n model.to(device)\n print(f\"number of trainable parameters: {sum(p.numel() for p in model.parameters() if p.requires_grad)}\")\n\n if args.do_train:\n no_decay = ['bias', 'LayerNorm.weight']\n optimizer_parameters = [\n {'params': [p for n, p in model.named_parameters() if not any(\n nd in n for nd in no_decay)], 'weight_decay': args.weight_decay},\n {'params': [p for n, p in model.named_parameters() if any(\n nd in n for nd in no_decay)], 'weight_decay': 0.0}\n ]\n optimizer = Adam(optimizer_parameters, lr=args.learning_rate, eps=args.adam_epsilon)\n\n if args.fp16:\n model, optimizer = apex.amp.initialize(model, optimizer, opt_level=args.fp16_opt_level)\n else:\n if args.fp16:\n model = apex.amp.initialize(model, opt_level=args.fp16_opt_level)\n\n if n_gpu > 1:\n model = torch.nn.DataParallel(model)\n\n if args.do_train:\n global_step = 0 # gradient update step\n batch_step = 0 # forward batch count\n best_mrr = 0\n train_loss_meter = AverageMeter()\n model.train()\n if \"fever\" in args.predict_file:\n train_dataset = FeverSingleDataset(tokenizer, args.train_file, args.max_q_len, args.max_c_len, train=True)\n else:\n train_dataset = SPDataset(tokenizer, args.train_file, args.max_q_len, args.max_c_len, train=True)\n train_dataloader = DataLoader(train_dataset, batch_size=args.train_batch_size, pin_memory=True, collate_fn=collate_fc, num_workers=args.num_workers, shuffle=True)\n\n t_total = len(train_dataloader) // args.gradient_accumulation_steps * args.num_train_epochs\n warmup_steps = t_total * args.warmup_ratio\n scheduler = get_linear_schedule_with_warmup(\n optimizer, num_warmup_steps=warmup_steps, num_training_steps=t_total\n )\n\n\n logger.info('Start training....')\n for epoch in range(int(args.num_train_epochs)):\n\n for batch in tqdm(train_dataloader):\n batch_step += 1\n batch = move_to_cuda(batch)\n loss = loss_single(model, batch, args.momentum)\n\n if args.gradient_accumulation_steps > 1:\n loss = loss / args.gradient_accumulation_steps\n\n if args.fp16:\n with apex.amp.scale_loss(loss, optimizer) as scaled_loss:\n scaled_loss.backward()\n else:\n loss.backward()\n train_loss_meter.update(loss.item())\n\n \n if (batch_step + 1) % args.gradient_accumulation_steps == 0:\n if args.fp16:\n torch.nn.utils.clip_grad_norm_(\n apex.amp.master_params(optimizer), args.max_grad_norm)\n else:\n torch.nn.utils.clip_grad_norm_(\n model.parameters(), args.max_grad_norm)\n optimizer.step()\n scheduler.step()\n model.zero_grad()\n global_step += 1\n\n tb_logger.add_scalar('batch_train_loss',\n loss.item(), global_step)\n tb_logger.add_scalar('smoothed_train_loss',\n train_loss_meter.avg, global_step)\n\n if args.eval_period != -1 and global_step % args.eval_period == 0:\n mrr = predict(args, model, eval_dataloader,\n device, logger)\n logger.info(\"Step %d Train loss %.2f MRR %.2f on epoch=%d\" % (global_step, train_loss_meter.avg, mrr*100, epoch))\n\n if best_mrr < mrr:\n logger.info(\"Saving model with best MRR %.2f -> MRR %.2f on epoch=%d\" %\n (best_mrr*100, mrr*100, epoch))\n torch.save(model.state_dict(), os.path.join(\n args.output_dir, f\"checkpoint_best.pt\"))\n model = model.to(device)\n best_mrr = mrr\n\n mrr = predict(args, model, eval_dataloader, device, logger)\n logger.info(\"Step %d Train loss %.2f MRR %.2f on epoch=%d\" % (\n global_step, train_loss_meter.avg, mrr*100, epoch))\n tb_logger.add_scalar('dev_mrr', mrr*100, epoch)\n if best_mrr < mrr:\n torch.save(model.state_dict(), os.path.join(\n args.output_dir, f\"checkpoint_last.pt\"))\n logger.info(\"Saving model with best MRR %.2f -> MRR %.2f on epoch=%d\" %\n (best_mrr*100, mrr*100, epoch))\n torch.save(model.state_dict(), os.path.join(\n args.output_dir, f\"checkpoint_best.pt\"))\n model = model.to(device)\n best_mrr = mrr\n\n logger.info(\"Training finished!\")\n\n elif args.do_predict:\n acc = predict(args, model, eval_dataloader, device, logger)\n logger.info(f\"test performance {acc}\")\n\n\ndef predict(args, model, eval_dataloader, device, logger):\n model.eval()\n num_correct = 0\n num_total = 0.0\n rrs = [] # reciprocal rank\n for batch in tqdm(eval_dataloader):\n batch_to_feed = move_to_cuda(batch)\n with torch.no_grad():\n outputs = model(batch_to_feed)\n\n q = outputs['q']\n c = outputs['c']\n neg_c = outputs['neg_c']\n\n product_in_batch = torch.mm(q, c.t()) \n product_neg = (q * neg_c).sum(-1).unsqueeze(1)\n product = torch.cat([product_in_batch, product_neg], dim=-1)\n\n target = torch.arange(product.size(0)).to(product.device)\n ranked = product.argsort(dim=1, descending=True)\n prediction = product.argmax(-1)\n\n # MRR\n idx2rank = ranked.argsort(dim=1)\n for idx, t in enumerate(target.tolist()):\n rrs.append(1 / (idx2rank[idx][t].item() +1))\n\n pred_res = prediction == target\n num_total += pred_res.size(0)\n num_correct += pred_res.sum(0)\n\n acc = num_correct/num_total\n mrr = np.mean(rrs)\n logger.info(f\"evaluated {num_total} examples...\")\n logger.info(f\"avg. Acc: {acc}\")\n logger.info(f'MRR: {mrr}')\n model.train()\n return mrr\n\n\nif __name__ == \"__main__\":\n main()" }, { "path": "mdr/retrieval/utils/basic_tokenizer.py", "content": "# Copyright (c) Facebook, Inc. and its affiliates.\n# All rights reserved.\n#\n# This source code is licensed under the license found in the \n# LICENSE file in the root directory of this source tree.\n#!/usr/bin/env python3\n# Copyright 2017-present, Facebook, Inc.\n# All rights reserved.\n#\n# This source code is licensed under the license found in the\n# LICENSE file in the root directory of this source tree.\n\"\"\"Base tokenizer/tokens classes and utilities.\"\"\"\n\nimport copy\n\n\n\nclass Tokens(object):\n \"\"\"A class to represent a list of tokenized text.\"\"\"\n TEXT = 0\n TEXT_WS = 1\n SPAN = 2\n POS = 3\n LEMMA = 4\n NER = 5\n\n def __init__(self, data, annotators, opts=None):\n self.data = data\n self.annotators = annotators\n self.opts = opts or {}\n\n def __len__(self):\n \"\"\"The number of tokens.\"\"\"\n return len(self.data)\n\n def slice(self, i=None, j=None):\n \"\"\"Return a view of the list of tokens from [i, j).\"\"\"\n new_tokens = copy.copy(self)\n new_tokens.data = self.data[i: j]\n return new_tokens\n\n def untokenize(self):\n \"\"\"Returns the original text (with whitespace reinserted).\"\"\"\n return ''.join([t[self.TEXT_WS] for t in self.data]).strip()\n\n def words(self, uncased=False):\n \"\"\"Returns a list of the text of each token\n\n Args:\n uncased: lower cases text\n \"\"\"\n if uncased:\n return [t[self.TEXT].lower() for t in self.data]\n else:\n return [t[self.TEXT] for t in self.data]\n\n def offsets(self):\n \"\"\"Returns a list of [start, end) character offsets of each token.\"\"\"\n return [t[self.SPAN] for t in self.data]\n\n def pos(self):\n \"\"\"Returns a list of part-of-speech tags of each token.\n Returns None if this annotation was not included.\n \"\"\"\n if 'pos' not in self.annotators:\n return None\n return [t[self.POS] for t in self.data]\n\n def lemmas(self):\n \"\"\"Returns a list of the lemmatized text of each token.\n Returns None if this annotation was not included.\n \"\"\"\n if 'lemma' not in self.annotators:\n return None\n return [t[self.LEMMA] for t in self.data]\n\n def entities(self):\n \"\"\"Returns a list of named-entity-recognition tags of each token.\n Returns None if this annotation was not included.\n \"\"\"\n if 'ner' not in self.annotators:\n return None\n return [t[self.NER] for t in self.data]\n\n def ngrams(self, n=1, uncased=False, filter_fn=None, as_strings=True):\n \"\"\"Returns a list of all ngrams from length 1 to n.\n\n Args:\n n: upper limit of ngram length\n uncased: lower cases text\n filter_fn: user function that takes in an ngram list and returns\n True or False to keep or not keep the ngram\n as_string: return the ngram as a string vs list\n \"\"\"\n def _skip(gram):\n if not filter_fn:\n return False\n return filter_fn(gram)\n\n words = self.words(uncased)\n ngrams = [(s, e + 1)\n for s in range(len(words))\n for e in range(s, min(s + n, len(words)))\n if not _skip(words[s:e + 1])]\n\n # Concatenate into strings\n if as_strings:\n ngrams = ['{}'.format(' '.join(words[s:e])) for (s, e) in ngrams]\n\n return ngrams\n\n def entity_groups(self):\n \"\"\"Group consecutive entity tokens with the same NER tag.\"\"\"\n entities = self.entities()\n if not entities:\n return None\n non_ent = self.opts.get('non_ent', 'O')\n groups = []\n idx = 0\n while idx < len(entities):\n ner_tag = entities[idx]\n # Check for entity tag\n if ner_tag != non_ent:\n # Chomp the sequence\n start = idx\n while (idx < len(entities) and entities[idx] == ner_tag):\n idx += 1\n groups.append((self.slice(start, idx).untokenize(), ner_tag))\n else:\n idx += 1\n return groups\n\n\nclass Tokenizer(object):\n \"\"\"Base tokenizer class.\n Tokenizers implement tokenize, which should return a Tokens class.\n \"\"\"\n\n def tokenize(self, text):\n raise NotImplementedError\n\n def shutdown(self):\n pass\n\n def __del__(self):\n self.shutdown()\n\n\nimport regex\nimport logging\n\nlogger = logging.getLogger(__name__)\n\n\nclass RegexpTokenizer(Tokenizer):\n DIGIT = r'\\p{Nd}+([:\\.\\,]\\p{Nd}+)*'\n TITLE = (r'(dr|esq|hon|jr|mr|mrs|ms|prof|rev|sr|st|rt|messrs|mmes|msgr)'\n r'\\.(?=\\p{Z})')\n ABBRV = r'([\\p{L}]\\.){2,}(?=\\p{Z}|$)'\n ALPHA_NUM = r'[\\p{L}\\p{N}\\p{M}]++'\n HYPHEN = r'{A}([-\\u058A\\u2010\\u2011]{A})+'.format(A=ALPHA_NUM)\n NEGATION = r\"((?!n't)[\\p{L}\\p{N}\\p{M}])++(?=n't)|n't\"\n CONTRACTION1 = r\"can(?=not\\b)\"\n CONTRACTION2 = r\"'([tsdm]|re|ll|ve)\\b\"\n START_DQUOTE = r'(?<=[\\p{Z}\\(\\[{<]|^)(``|[\"\\u0093\\u201C\\u00AB])(?!\\p{Z})'\n START_SQUOTE = r'(?<=[\\p{Z}\\(\\[{<]|^)[\\'\\u0091\\u2018\\u201B\\u2039](?!\\p{Z})'\n END_DQUOTE = r'(?<!\\p{Z})(\\'\\'|[\"\\u0094\\u201D\\u00BB])'\n END_SQUOTE = r'(?<!\\p{Z})[\\'\\u0092\\u2019\\u203A]'\n DASH = r'--|[\\u0096\\u0097\\u2013\\u2014\\u2015]'\n ELLIPSES = r'\\.\\.\\.|\\u2026'\n PUNCT = r'\\p{P}'\n NON_WS = r'[^\\p{Z}\\p{C}]'\n\n def __init__(self, **kwargs):\n \"\"\"\n Args:\n annotators: None or empty set (only tokenizes).\n substitutions: if true, normalizes some token types (e.g. quotes).\n \"\"\"\n self._regexp = regex.compile(\n '(?P<digit>%s)|(?P<title>%s)|(?P<abbr>%s)|(?P<neg>%s)|(?P<hyph>%s)|'\n '(?P<contr1>%s)|(?P<alphanum>%s)|(?P<contr2>%s)|(?P<sdquote>%s)|'\n '(?P<edquote>%s)|(?P<ssquote>%s)|(?P<esquote>%s)|(?P<dash>%s)|'\n '(?<ellipses>%s)|(?P<punct>%s)|(?P<nonws>%s)' %\n (self.DIGIT, self.TITLE, self.ABBRV, self.NEGATION, self.HYPHEN,\n self.CONTRACTION1, self.ALPHA_NUM, self.CONTRACTION2,\n self.START_DQUOTE, self.END_DQUOTE, self.START_SQUOTE,\n self.END_SQUOTE, self.DASH, self.ELLIPSES, self.PUNCT,\n self.NON_WS),\n flags=regex.IGNORECASE + regex.UNICODE + regex.MULTILINE\n )\n if len(kwargs.get('annotators', {})) > 0:\n logger.warning('%s only tokenizes! Skipping annotators: %s' %\n (type(self).__name__, kwargs.get('annotators')))\n self.annotators = set()\n self.substitutions = kwargs.get('substitutions', True)\n\n def tokenize(self, text):\n data = []\n matches = [m for m in self._regexp.finditer(text)]\n for i in range(len(matches)):\n # Get text\n token = matches[i].group()\n\n # Make normalizations for special token types\n if self.substitutions:\n groups = matches[i].groupdict()\n if groups['sdquote']:\n token = \"``\"\n elif groups['edquote']:\n token = \"''\"\n elif groups['ssquote']:\n token = \"`\"\n elif groups['esquote']:\n token = \"'\"\n elif groups['dash']:\n token = '--'\n elif groups['ellipses']:\n token = '...'\n\n # Get whitespace\n span = matches[i].span()\n start_ws = span[0]\n if i + 1 < len(matches):\n end_ws = matches[i + 1].span()[0]\n else:\n end_ws = span[1]\n\n # Format data\n data.append((\n token,\n text[start_ws: end_ws],\n span,\n ))\n return Tokens(data, self.annotators)\n\n\nclass SimpleTokenizer(Tokenizer):\n ALPHA_NUM = r'[\\p{L}\\p{N}\\p{M}]+'\n NON_WS = r'[^\\p{Z}\\p{C}]'\n\n def __init__(self, **kwargs):\n \"\"\"\n Args:\n annotators: None or empty set (only tokenizes).\n \"\"\"\n self._regexp = regex.compile(\n '(%s)|(%s)' % (self.ALPHA_NUM, self.NON_WS),\n flags=regex.IGNORECASE + regex.UNICODE + regex.MULTILINE\n )\n if len(kwargs.get('annotators', {})) > 0:\n logger.warning('%s only tokenizes! Skipping annotators: %s' %\n (type(self).__name__, kwargs.get('annotators')))\n self.annotators = set()\n\n def tokenize(self, text):\n data = []\n matches = [m for m in self._regexp.finditer(text)]\n for i in range(len(matches)):\n # Get text\n token = matches[i].group()\n\n # Get whitespace\n span = matches[i].span()\n start_ws = span[0]\n if i + 1 < len(matches):\n end_ws = matches[i + 1].span()[0]\n else:\n end_ws = span[1]\n\n # Format data\n data.append((\n token,\n text[start_ws: end_ws],\n span,\n ))\n return Tokens(data, self.annotators)\n\n\n\nSTOPWORDS = {\n 'i', 'me', 'my', 'myself', 'we', 'our', 'ours', 'ourselves', 'you', 'your',\n 'yours', 'yourself', 'yourselves', 'he', 'him', 'his', 'himself', 'she',\n 'her', 'hers', 'herself', 'it', 'its', 'itself', 'they', 'them', 'their',\n 'theirs', 'themselves', 'what', 'which', 'who', 'whom', 'this', 'that',\n 'these', 'those', 'am', 'is', 'are', 'was', 'were', 'be', 'been', 'being',\n 'have', 'has', 'had', 'having', 'do', 'does', 'did', 'doing', 'a', 'an',\n 'the', 'and', 'but', 'if', 'or', 'because', 'as', 'until', 'while', 'of',\n 'at', 'by', 'for', 'with', 'about', 'against', 'between', 'into', 'through',\n 'during', 'before', 'after', 'above', 'below', 'to', 'from', 'up', 'down',\n 'in', 'out', 'on', 'off', 'over', 'under', 'again', 'further', 'then',\n 'once', 'here', 'there', 'when', 'where', 'why', 'how', 'all', 'any',\n 'both', 'each', 'few', 'more', 'most', 'other', 'some', 'such', 'no', 'nor',\n 'not', 'only', 'own', 'same', 'so', 'than', 'too', 'very', 's', 't', 'can',\n 'will', 'just', 'don', 'should', 'now', 'd', 'll', 'm', 'o', 're', 've',\n 'y', 'ain', 'aren', 'couldn', 'didn', 'doesn', 'hadn', 'hasn', 'haven',\n 'isn', 'ma', 'mightn', 'mustn', 'needn', 'shan', 'shouldn', 'wasn', 'weren',\n 'won', 'wouldn', \"'ll\", \"'re\", \"'ve\", \"n't\", \"'s\", \"'d\", \"'m\", \"''\", \"``\"\n}\n\nimport unicodedata\n\ndef normalize(text):\n \"\"\"Resolve different type of unicode encodings.\"\"\"\n return unicodedata.normalize('NFD', text)\n\n\ndef filter_word(text):\n \"\"\"Take out english stopwords, punctuation, and compound endings.\"\"\"\n text = normalize(text)\n if regex.match(r'^\\p{P}+$', text):\n return True\n if text.lower() in STOPWORDS:\n return True\n return False\n\ndef filter_ngram(gram, mode='any'):\n \"\"\"Decide whether to keep or discard an n-gram.\n\n Args:\n gram: list of tokens (length N)\n mode: Option to throw out ngram if\n 'any': any single token passes filter_word\n 'all': all tokens pass filter_word\n 'ends': book-ended by filterable tokens\n \"\"\"\n filtered = [filter_word(w) for w in gram]\n if mode == 'any':\n return any(filtered)\n elif mode == 'all':\n return all(filtered)\n elif mode == 'ends':\n return filtered[0] or filtered[-1]\n else:\n raise ValueError('Invalid mode: %s' % mode)\n" }, { "path": "mdr/retrieval/utils/gen_index_id_map.py", "content": "# Copyright (c) Facebook, Inc. and its affiliates.\n# All rights reserved.\n#\n# This source code is licensed under the license found in the \n# LICENSE file in the root directory of this source tree.\nimport json\n\nmapping = {}\nwith open('../data/para_doc.db') as f_in:\n for idx, line in enumerate(f_in):\n sample = json.loads(line.strip())\n mapping[idx] = sample['id']\nwith open('index_data/idx_id.json', 'w') as f_out:\n json.dump(mapping, f_out)\n\n" }, { "path": "mdr/retrieval/utils/mhop_utils.py", "content": "# Copyright (c) Facebook, Inc. and its affiliates.\n# All rights reserved.\n#\n# This source code is licensed under the license found in the \n# LICENSE file in the root directory of this source tree.\nimport sys\nimport json\nfrom tqdm import tqdm\nimport collections\nimport re\nfrom collections import Counter\nimport string\nfrom basic_tokenizer import SimpleTokenizer, filter_ngram\nimport csv\n\ndef pick_bridge_v0(title2linked, title2doc, titles, q, ans):\n \"\"\"\n 1. mainly based on if the passage includes the answer (assuming that only the 2nd hop passage has the answer)\n 2. if 1 fails, then resort the linking structure, if A links to B, then B is the \n \"\"\"\n # check answer\n if (ans in titles[0] + \" \" + title2doc[titles[0]]) and ans not in titles[1] + \" \" + title2doc[titles[1]]:\n return titles[0]\n elif (ans in titles[1] + \" \" + title2doc[titles[1]]) and (ans not in titles[0] + \" \" + title2doc[titles[0]]):\n return titles[1]\n elif titles[0] in title2linked[titles[1]] and titles[1] not in title2linked[titles[0]]:\n return titles[0]\n else:\n return titles[1]\n\ndef load_annotated(path=\"/private/home/xwhan/data/hotpot/tfidf/abstracts.txt\"):\n content = [json.loads(l) for l in open(path).readlines()]\n title2doc = {item[\"title\"]:item[\"text\"] for item in content}\n title2linked = {item[\"title\"]:item[\"linked\"] for item in content}\n return title2doc, title2linked\n\ndef normalize_answer(s):\n \"\"\"Lower text and remove punctuation, articles and extra whitespace.\"\"\"\n def remove_articles(text):\n return re.sub(r'\\b(a|an|the)\\b', ' ', text)\n\n def white_space_fix(text):\n return ' '.join(text.split())\n\n def remove_punc(text):\n exclude = set(string.punctuation)\n return ''.join(ch for ch in text if ch not in exclude)\n\n def lower(text):\n return text.lower()\n\n return white_space_fix(remove_articles(remove_punc(lower(s))))\n\n\ndef hotpot_sp_data(raw_path):\n train = json.load(open(raw_path + '/hotpot_train_v1.1.json'))\n dev = json.load(open(raw_path + '/hotpot_dev_distractor_v1.json'))\n title2doc, title2linked = load_annotated()\n\n for split_name, split in {\"train\": train, \"val\": dev}.items():\n data_to_save = []\n for item in tqdm(split):\n title2passage = {_[0]: _[1] for _ in item[\"context\"]}\n sp_titles = list(set([_[0] for _ in item[\"supporting_facts\"]]))\n question = item[\"question\"]\n if item[\"type\"] == \"comparison\":\n pos_paras = []\n for title in sp_titles:\n pos_paras.append({\n \"title\": title,\n \"text\": \"\".join(title2passage[title])\n })\n data_to_save.append({\n \"question\": question,\n \"pos_paras\": pos_paras,\n \"neg_paras\": [],\n \"type\": item[\"type\"],\n \"answers\": item[\"answer\"]\n })\n \n else:\n bridge = pick_bridge(title2linked, title2doc, sp_titles, question, item[\"answer\"])\n if sp_titles[0] == bridge:\n sp_titles = sp_titles[::-1]\n start, bridge = sp_titles[0], sp_titles[1]\n pos_paras = []\n for title in sp_titles:\n pos_paras.append({\n \"title\": title,\n \"text\": \"\".join(title2passage[title])\n })\n data_to_save.append({\n \"question\": question,\n \"pos_paras\": pos_paras,\n \"neg_paras\": [],\n \"type\": item[\"type\"],\n \"answers\": item[\"answer\"],\n \"bridge\": bridge\n })\n\n with open(raw_path + f\"/hotpot_retrieval_{split_name}.json\", \"w\") as g:\n for line in data_to_save:\n g.write(json.dumps(line) + \"\\n\")\n\n\ndef add_qid(raw_path):\n title2doc, title2linked = load_annotated()\n raw_train = json.load(open(raw_path + '/hotpot_train_v1.1.json'))\n raw_val = json.load(open(raw_path + '/hotpot_dev_distractor_v1.json'))\n for s, raw_data in zip([\"train\", \"val\"], [raw_train, raw_val]):\n qas_data = []\n for item in raw_data:\n question = item[\"question\"]\n _id = item[\"_id\"]\n _type = item[\"type\"]\n answer = [item[\"answer\"]]\n sp = list(set([f[0] for f in item[\"supporting_facts\"]]))\n if _type == \"bridge\":\n # make sure the sp order follows reasoning process\n bridge_title = pick_bridge_v0(title2linked, title2doc, sp, question, answer[0])\n if sp[0] == bridge_title:\n sp = sp[::-1]\n qas_data.append({\n \"question\": question,\n \"_id\": _id,\n \"answer\": answer,\n \"sp\": sp,\n \"type\": _type\n })\n\n with open(raw_path + f\"/hotpot_qas_{s}.json\", \"w\") as g1:\n for _ in qas_data:\n g1.write(json.dumps(_) + \"\\n\")\n\ndef add_bridge_ann(raw_path):\n title2doc, title2linked = load_annotated()\n\n raw_train = json.load(open(raw_path + '/hotpot_train_v1.1.json'))\n raw_val = json.load(open(raw_path + '/hotpot_dev_distractor_v1.json'))\n for split, raw in zip([\"train\", \"val\"], [raw_train, raw_val]):\n \n data = json.load(open(raw_path + f\"/hotpot_{split}_with_neg.txt\"))\n for idx, item in enumerate(data):\n assert item[\"question\"] == raw[idx][\"question\"]\n item[\"_id\"] = raw[idx][\"_id\"]\n if item[\"type\"] == \"bridge\":\n ans = item[\"answers\"][0]\n sp_titles = [p[\"title\"] for p in item[\"pos_paras\"]]\n bridge_title = pick_bridge_v0(title2linked, title2doc, sp_titles, item[\"question\"], ans)\n item[\"bridge\"] = bridge_title\n with open(raw_path + f\"/hotpot_{split}_with_neg_v0.json\", \"w\") as g:\n for _ in data:\n g.write(json.dumps(_) + \"\\n\")\n # data = [json.loads(l) for l in open(raw_path + f\"/hotpot_qas_{split}.json\").readlines()]\n # for item in data:\n # if item[\"type\"] == \"bridge\":\n # ans = item[\"answer\"][0]\n # sp_titles = item[\"sp\"]\n # bridge_title = pick_bridge(title2linked, title2doc, sp_titles, item[\"question\"], ans)\n # item[\"bridge\"] = bridge_title\n # with open(raw_path + f\"/hotpot_qas_{split}_bridge_label.json\", \"w\") as g:\n # for _ in data:\n # g.write(json.dumps(_) + \"\\n\")\nimport numpy as np\n\ndef check_2hop(raw_path):\n data = [json.loads(l) for l in open(raw_path + \"/bridge_val.json\").readlines()]\n\n target_in_query = [int(item[\"pos_para\"][\"title\"].strip() in item[\"sp1\"] or item[\"pos_para\"][\"title\"].lower().strip() in item[\"sp1\"]) for item in data]\n print(np.mean(target_in_query))\n\n\ndef add_sp_labels(raw_path, input_file, save_path,\n title2sent_map=\"data/hotpot_index/title2sents.txt\"):\n \"\"\"\n Add sp sentence supervision for QA model training\n \n Inputs:\n raw_path: original HotpotQA data\n input_file: from MDR retrieval\n save_path: retrieved results with sentence level annotation\n \"\"\"\n # raw_train = json.load(open(raw_path + '/hotpot_train_v1.1.json'))\n # train = [json.loads(l) for l in open(raw_path + \"/dense_train_b100_k100.json\").readlines()]\n raw_data = json.load(open(raw_path))\n retrieved = [json.loads(l) for l in open(input_file).readlines()]\n\n # title2sents\n title_and_sents = [json.loads(l) for l in open(title2sent_map).readlines()]\n title2sents = {_['title']:_['sents'] for _ in title_and_sents}\n\n for instance, raw in zip(retrieved, raw_data):\n assert instance[\"question\"] == raw[\"question\"]\n\n if \"supporting_facts\" in raw:\n orig_sp = raw[\"supporting_facts\"]\n sptitle2sentids = collections.defaultdict(list)\n for _ in orig_sp:\n sptitle2sentids[_[0]].append(_[1])\n\n instance[\"sp\"] = []\n\n for title in sptitle2sentids.keys():\n instance[\"sp\"].append({\"title\": title, \"sents\": title2sents[title], \"sp_sent_ids\": sptitle2sentids[title]})\n \n instance[\"answer\"] = [raw[\"answer\"]]\n\n with open(save_path, \"w\") as out:\n for l in retrieved:\n out.write(json.dumps(l) + \"\\n\")\n\ndef explore_QDMR(path=\"/private/home/xwhan/data/Break-dataset/QDMR-high-level\"):\n \"\"\"\n question decomposition from the BREAK dataset\n try to get the single-hop questions without introduce different question styles (hopefully the questions with have similar style as the original hotpotQA data), such the model is not only learning to distinguish the styles of the questions\n \"\"\"\n\n hotpot_retrieval_train = [json.loads(l) for l in open(\"/private/home/xwhan/data/hotpot/hotpot_train_with_neg_v0.json\").readlines()]\n hotpot_retrieval_val = [json.loads(l) for l in open(\"/private/home/xwhan/data/hotpot/hotpot_val_with_neg_v0.json\").readlines()]\n qid2data = {}\n for item in hotpot_retrieval_train + hotpot_retrieval_val:\n qid2data[item[\"_id\"]] = item\n\n for split in [\"train\", \"dev\"]:\n break_data = []\n with open(f\"{path}/{split}.csv\") as csvfile:\n reader = csv.reader(csvfile, quotechar='\"', delimiter=',')\n for row in reader:\n if row[0] != \"question_id\":\n q_id, orig_q, decom = row[0], row[1], row[2]\n if q_id.startswith(\"HOTPOT\"):\n orig_split, orig_id = q_id.split(\"_\")[1], q_id.split(\"_\")[2]\n sp_paras = qid2data[orig_id][\"pos_paras\"]\n break_data.append({\n \"id\": orig_id,\n \"split\": orig_split,\n \"q\": orig_q,\n \"q_decom\": decom,\n \"sp\": sp_paras,\n \"type\": qid2data[orig_id][\"type\"]\n })\n \n with open(f\"/private/home/xwhan/data/QDMR/{split}.json\", \"w\") as out:\n for _ in break_data:\n out.write(json.dumps(_) + \"\\n\")\n\n\n\ndef add_sents_to_corpus_dict():\n id2doc = json.load(open(\"/private/home/xwhan/Mhop-Pretrain/retrieval/index/abstracts_id2doc.json\"))\n title_and_sents = [json.loads(l) for l in open(\"/private/home/xwhan/data/hotpot/tfidf/title_sents.txt\").readlines()]\n title2sents = {_['title']:_['sents'] for _ in title_and_sents}\n\n for k in id2doc.keys():\n title, text = id2doc[k][0], id2doc[k][1]\n sents = title2sents[title]\n id2doc[k] = {\n \"title\": title,\n \"text\": text,\n \"sents\": sents\n }\n \n json.dump(id2doc, open(\"/private/home/xwhan/Mhop-Pretrain/retrieval/index/hotpotQA_corpus_dict.json\", \"w\"))\n\nif __name__ == \"__main__\":\n original_hotpot_data, retrieved, output_path = sys.argv[1], sys.argv[2], sys.argv[3]\n add_sp_labels(original_hotpot_data, retrieved, output_path)\n" }, { "path": "mdr/retrieval/utils/tokenizer.py", "content": "# Copyright (c) Facebook, Inc. and its affiliates.\n# All rights reserved.\n#\n# This source code is licensed under the license found in the \n# LICENSE file in the root directory of this source tree.\n# coding=utf-8\n# Copyright 2018 The Google AI Language Team Authors.\n#\n# Licensed under the Apache License, Version 2.0 (the \"License\");\n# you may not use this file except in compliance with the License.\n# You may obtain a copy of the License at\n#\n# http://www.apache.org/licenses/LICENSE-2.0\n#\n# Unless required by applicable law or agreed to in writing, software\n# distributed under the License is distributed on an \"AS IS\" BASIS,\n# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n# See the License for the specific language governing permissions and\n# limitations under the License.\n\"\"\"Tokenization classes.\"\"\"\n\nfrom __future__ import absolute_import\nfrom __future__ import division\nfrom __future__ import print_function\n\nimport collections\nimport unicodedata\nimport six\nimport tensorflow as tf\n\n\ndef convert_tokens_to_ids(vocab, tokens):\n \"\"\"Converts a sequence of tokens into ids using the vocab.\"\"\"\n ids = []\n for token in tokens:\n ids.append(vocab[token])\n return ids\n\ndef whitespace_tokenize(text):\n \"\"\"Runs basic whitespace cleaning and splitting on a peice of text.\"\"\"\n text = text.strip()\n if not text:\n return []\n tokens = text.split()\n return tokens\n\n\ndef convert_to_unicode(text):\n \"\"\"Converts `text` to Unicode (if it's not already), assuming utf-8 input.\"\"\"\n if six.PY3:\n if isinstance(text, str):\n return text\n elif isinstance(text, bytes):\n return text.decode(\"utf-8\", \"ignore\")\n else:\n raise ValueError(\"Unsupported string type: %s\" % (type(text)))\n elif six.PY2:\n if isinstance(text, str):\n return text.decode(\"utf-8\", \"ignore\")\n elif isinstance(text, unicode):\n return text\n else:\n raise ValueError(\"Unsupported string type: %s\" % (type(text)))\n else:\n raise ValueError(\"Not running on Python2 or Python 3?\")\n\n\ndef _is_whitespace(char):\n \"\"\"Checks whether `chars` is a whitespace character.\"\"\"\n # \\t, \\n, and \\r are technically contorl characters but we treat them\n # as whitespace since they are generally considered as such.\n if char == \" \" or char == \"\\t\" or char == \"\\n\" or char == \"\\r\":\n return True\n cat = unicodedata.category(char)\n if cat == \"Zs\":\n return True\n return False\n\n\ndef _is_control(char):\n \"\"\"Checks whether `chars` is a control character.\"\"\"\n # These are technically control characters but we count them as whitespace\n # characters.\n if char == \"\\t\" or char == \"\\n\" or char == \"\\r\":\n return False\n cat = unicodedata.category(char)\n if cat.startswith(\"C\"):\n return True\n return False\n\nclass BasicTokenizer(object):\n \"\"\"Runs basic tokenization (punctuation splitting, lower casing, etc.).\"\"\"\n\n def __init__(self, do_lower_case=True):\n \"\"\"Constructs a BasicTokenizer.\n Args:\n do_lower_case: Whether to lower case the input.\n \"\"\"\n self.do_lower_case = do_lower_case\n\n def tokenize(self, text):\n \"\"\"Tokenizes a piece of text.\"\"\"\n text = convert_to_unicode(text)\n text = self._clean_text(text)\n orig_tokens = whitespace_tokenize(text)\n split_tokens = []\n for token in orig_tokens:\n if self.do_lower_case:\n token = token.lower()\n token = self._run_strip_accents(token)\n split_tokens.extend(self._run_split_on_punc(token))\n\n output_tokens = whitespace_tokenize(\" \".join(split_tokens))\n return output_tokens\n\n def _run_strip_accents(self, text):\n \"\"\"Strips accents from a piece of text.\"\"\"\n text = unicodedata.normalize(\"NFD\", text)\n output = []\n for char in text:\n cat = unicodedata.category(char)\n if cat == \"Mn\":\n continue\n output.append(char)\n return \"\".join(output)\n\n def _run_split_on_punc(self, text):\n \"\"\"Splits punctuation on a piece of text.\"\"\"\n chars = list(text)\n i = 0\n start_new_word = True\n output = []\n while i < len(chars):\n char = chars[i]\n if _is_punctuation(char):\n output.append([char])\n start_new_word = True\n else:\n if start_new_word:\n output.append([])\n start_new_word = False\n output[-1].append(char)\n i += 1\n\n return [\"\".join(x) for x in output]\n\n def _clean_text(self, text):\n \"\"\"Performs invalid character removal and whitespace cleanup on text.\"\"\"\n output = []\n for char in text:\n cp = ord(char)\n if cp == 0 or cp == 0xfffd or _is_control(char):\n continue\n if _is_whitespace(char):\n output.append(\" \")\n else:\n output.append(char)\n return \"\".join(output)\n\n\ndef _is_punctuation(char):\n \"\"\"Checks whether `chars` is a punctuation character.\"\"\"\n cp = ord(char)\n # We treat all non-letter/number ASCII as punctuation.\n # Characters such as \"^\", \"$\", and \"`\" are not in the Unicode\n # Punctuation class but we treat them as punctuation anyways, for\n # consistency.\n if ((cp >= 33 and cp <= 47) or (cp >= 58 and cp <= 64) or\n (cp >= 91 and cp <= 96) or (cp >= 123 and cp <= 126)):\n return True\n cat = unicodedata.category(char)\n if cat.startswith(\"P\"):\n return True\n return False\n\n\ndef process(s, tokenizer):\n try:\n return tokenizer.tokenize(s)\n except:\n print('failed on', s)\n raise\n\nif __name__ == \"__main__\":\n _is_whitespace(\"a\")\n" }, { "path": "mdr/retrieval/utils/utils.py", "content": "# Copyright (c) Facebook, Inc. and its affiliates.\n# All rights reserved.\n#\n# This source code is licensed under the license found in the \n# LICENSE file in the root directory of this source tree.\nimport torch\nimport sqlite3\nimport unicodedata\n\ndef load_saved(model, path, exact=True):\n try:\n state_dict = torch.load(path)\n except:\n state_dict = torch.load(path, map_location=torch.device('cpu'))\n \n def filter(x): return x[7:] if x.startswith('module.') else x\n if exact:\n state_dict = {filter(k): v for (k, v) in state_dict.items()}\n else:\n state_dict = {filter(k): v for (k, v) in state_dict.items() if filter(k) in model.state_dict()}\n model.load_state_dict(state_dict)\n return model\n\ndef move_to_cuda(sample):\n if len(sample) == 0:\n return {}\n\n def _move_to_cuda(maybe_tensor):\n if torch.is_tensor(maybe_tensor):\n return maybe_tensor.cuda()\n elif isinstance(maybe_tensor, dict):\n return {\n key: _move_to_cuda(value)\n for key, value in maybe_tensor.items()\n }\n elif isinstance(maybe_tensor, list):\n return [_move_to_cuda(x) for x in maybe_tensor]\n else:\n return maybe_tensor\n\n return _move_to_cuda(sample)\n\ndef convert_to_half(sample):\n if len(sample) == 0:\n return {}\n\n def _convert_to_half(maybe_floatTensor):\n if torch.is_tensor(maybe_floatTensor) and maybe_floatTensor.type() == \"torch.FloatTensor\":\n return maybe_floatTensor.half()\n elif isinstance(maybe_floatTensor, dict):\n return {\n key: _convert_to_half(value)\n for key, value in maybe_floatTensor.items()\n }\n elif isinstance(maybe_floatTensor, list):\n return [_convert_to_half(x) for x in maybe_floatTensor]\n else:\n return maybe_floatTensor\n\n return _convert_to_half(sample)\n\n\nclass AverageMeter(object):\n \"\"\"Computes and stores the average and current value\"\"\"\n\n def __init__(self):\n self.reset()\n\n def reset(self):\n self.val = 0\n self.avg = 0\n self.sum = 0\n self.count = 0\n\n def update(self, val, n=1):\n self.val = val\n self.sum += val * n\n self.count += n\n self.avg = self.sum / self.count\n\n\ndef normalize(text):\n \"\"\"Resolve different type of unicode encodings.\"\"\"\n return unicodedata.normalize('NFD', text)\n\n\nclass DocDB(object):\n \"\"\"Sqlite backed document storage.\n\n Implements get_doc_text(doc_id).\n \"\"\"\n\n def __init__(self, db_path=None):\n self.path = db_path\n self.connection = sqlite3.connect(self.path, check_same_thread=False)\n\n def __enter__(self):\n return self\n\n def __exit__(self, *args):\n self.close()\n\n def close(self):\n \"\"\"Close the connection to the database.\"\"\"\n self.connection.close()\n\n def get_doc_ids(self):\n \"\"\"Fetch all ids of docs stored in the db.\"\"\"\n cursor = self.connection.cursor()\n cursor.execute(\"SELECT id FROM documents\")\n results = [r[0] for r in cursor.fetchall()]\n cursor.close()\n return results\n\n def get_doc_text(self, doc_id):\n \"\"\"Fetch the raw text of the doc for 'doc_id'.\"\"\"\n cursor = self.connection.cursor()\n cursor.execute(\n \"SELECT text FROM documents WHERE id = ?\",\n (normalize(doc_id),)\n )\n result = cursor.fetchone()\n cursor.close()\n return result if result is None else result[0]\n\ndef para_has_answer(answer, para, tokenizer):\n assert isinstance(answer, list)\n text = normalize(para)\n tokens = tokenizer.tokenize(text)\n text = tokens.words(uncased=True)\n assert len(text) == len(tokens)\n for single_answer in answer:\n single_answer = normalize(single_answer)\n single_answer = tokenizer.tokenize(single_answer)\n single_answer = single_answer.words(uncased=True)\n for i in range(0, len(text) - len(single_answer) + 1):\n if single_answer == text[i: i + len(single_answer)]:\n return True\n return False\n\n\ndef complex_ans_recall():\n \"\"\"\n calculate retrieval metrics for complexwebQ\n \"\"\"\n import json\n import numpy as np\n from basic_tokenizer import SimpleTokenizer\n tok = SimpleTokenizer()\n\n predictions = json.load(open(\"/private/home/xwhan/code/learning_to_retrieve_reasoning_paths/results/complexwebq_retrieval_res.json\"))\n raw_dev = [json.loads(l) for l in open(\"/private/home/xwhan/data/ComplexWebQ/complexwebq_dev_qas.txt\").readlines()]\n id2qas = {_[\"id\"]:_ for _ in raw_dev}\n\n assert len(predictions) == len(raw_dev)\n answer_recalls = []\n for item in predictions:\n qid = item[\"q_id\"]\n title2passage = item[\"context\"]\n gold_answers = id2qas[qid][\"answer\"]\n\n chain_coverage = []\n for chain in item[\"topk_titles\"]:\n chain_text = \" \".join([title2passage[_] for _ in chain])\n chain_coverage.append(para_has_answer(gold_answers, chain_text, tok))\n answer_recalls.append(np.sum(chain_coverage) > 0)\n print(len(answer_recalls))\n print(np.mean(answer_recalls))\n\nif __name__ == \"__main__\":\n complex_ans_recall()" }, { "path": "requirements.txt", "content": "transformers==2.11.0\ntensorboard>=1.15.0\nnumpy\ntqdm\nujson\nstreamlit" }, { "path": "scripts/add_sp_label.sh", "content": "#!/bin/bash\n# Copyright (c) Facebook, Inc. and its affiliates.\n# All rights reserved.\n#\n# This source code is licensed under the license found in the \n# LICENSE file in the root directory of this source tree.\n#!/bin/bash \n\n# to add SP sentence labels to the retrieved passages for dev data ./scripts/add_sp_label.sh data/hotpot/hotpot_dev_distractor_v1.json data/hotpot/dev_retrieved_top100.json data/hotpot/dev_retrieved_top100_with_sp.json\n\nORIGINAL_DATA=$1\nRETRIEVED_DATA=$2\nSAVED_PATH=$3\n\npython mdr/retrieval/utils/mhop_utils.py ${ORIGINAL_DATA} ${RETRIEVED_DATA}\n${SASAVED_PATH}\n" }, { "path": "scripts/demo.py", "content": "# Copyright (c) Facebook, Inc. and its affiliates.\n# All rights reserved.\n#\n# This source code is licensed under the license found in the\n# LICENSE file in the root directory of this source tree.\n\nimport streamlit as st\nimport torch\nimport os\nimport numpy as np\nfrom apex import amp\nimport faiss\nimport json\nimport argparse\nfrom functools import partial\nfrom transformers import AutoConfig, AutoTokenizer\nfrom torch.utils.data import DataLoader\n\nfrom mdr.retrieval.models.mhop_retriever import RobertaRetriever\nfrom mdr.retrieval.utils.basic_tokenizer import SimpleTokenizer\nfrom mdr.retrieval.utils.utils import load_saved, move_to_cuda\n\nfrom mdr.qa.qa_model import QAModel\nfrom mdr.qa.qa_dataset import qa_collate, QAEvalDataset\nfrom train_qa import eval_final\n\n@st.cache(allow_output_mutation=True)\ndef init_retrieval(args):\n print(\"Initializing retrieval module...\")\n bert_config = AutoConfig.from_pretrained(args.model_name)\n tokenizer = AutoTokenizer.from_pretrained(args.model_name)\n retriever = RobertaRetriever(bert_config, args)\n retriever = load_saved(retriever, args.model_path, exact=False)\n cuda = torch.device('cuda')\n retriever.to(cuda)\n retriever = amp.initialize(retriever, opt_level='O1')\n retriever.eval()\n\n print(\"Loading index...\")\n index = faiss.IndexFlatIP(768)\n xb = np.load(args.indexpath).astype('float32')\n index.add(xb)\n if args.index_gpu != -1:\n res = faiss.StandardGpuResources()\n index = faiss.index_cpu_to_gpu(res, args.index_gpu, index)\n\n print(\"Loading documents...\")\n id2doc = json.load(open(args.corpus_dict))\n\n print(\"Index ready...\")\n return retriever, index, id2doc, tokenizer\n\n@st.cache(allow_output_mutation=True)\ndef init_reader(args):\n qa_config = AutoConfig.from_pretrained(\n 'google/electra-large-discriminator')\n qa_tokenizer = AutoTokenizer.from_pretrained(\n 'google/electra-large-discriminator')\n retriever_name = args.model_name\n args.model_name = \"google/electra-large-discriminator\"\n reader = QAModel(qa_config, args)\n reader = load_saved(reader, args.reader_path, False) \n cuda = torch.device('cuda')\n reader.to(cuda)\n reader = amp.initialize(reader, opt_level='O1')\n reader.eval()\n args.model_name = retriever_name\n return reader, qa_tokenizer\n\n\nst.markdown(\n \"# Multi-hop Open-domain QA with [MDR](https://github.com/facebookresearch/multihop_dense_retrieval)\")\n\nparser = argparse.ArgumentParser()\nparser.add_argument('--indexpath', type=str,\n default='data/hotpot_index/wiki_index.npy')\nparser.add_argument('--corpus_dict', type=str,\n default='data/hotpot_index/wiki_id2doc.json')\nparser.add_argument('--model_path', type=str, default='models/q_encoder.pt')\nparser.add_argument('--topk', type=int, default=20, help=\"topk paths\")\nparser.add_argument('--max-q-len', type=int, default=70)\nparser.add_argument('--max-c-len', type=int, default=300)\nparser.add_argument('--max-q-sp-len', type=int, default=350)\nparser.add_argument('--model-name', type=str, default='roberta-base')\nparser.add_argument('--reader_path', type=str, default=\"models/qa_electra.pt\")\n\nparser.add_argument(\"--sp-pred\", action=\"store_true\",\n help=\"whether to predict sentence sp\")\nparser.add_argument(\"--sp-weight\", default=0, type=float,\n help=\"weight of the sp loss\")\nparser.add_argument(\"--max-ans-len\", default=30, type=int)\nparser.add_argument(\"--save-prediction\", default=\"\", type=str)\nparser.add_argument(\"--index-gpu\", default=-1, type=int)\n\nargs = parser.parse_args()\n\nreader, qa_tokenizer = init_reader(args)\n\nretriever, index, id2doc, retriever_tokenizer = init_retrieval(args)\n\nst.markdown(\"*Trick: Due to the case sensitive tokenization we used during training, try to use capitalized entity names in your question, e.g., type United States instead of united states.*\")\n\nquery = st.text_input('Enter your question')\nif query:\n\n query = query[:-1] if query.endswith(\"?\") else query\n with torch.no_grad():\n\n print(\"Retrieving\")\n q_encodes = retriever_tokenizer.batch_encode_plus(\n [query], max_length=args.max_q_len, pad_to_max_length=True, return_tensors=\"pt\")\n q_encodes = move_to_cuda(dict(q_encodes))\n q_embeds = retriever.encode_q(\n q_encodes[\"input_ids\"], q_encodes[\"attention_mask\"], q_encodes.get(\"token_type_ids\", None)).cpu().numpy()\n scores_1, docid_1 = index.search(q_embeds, args.topk)\n query_pairs = [] # for 2nd hop\n for _, doc_id in enumerate(docid_1[0]):\n doc = id2doc[str(doc_id)][\"text\"]\n if doc.strip() == \"\":\n # roberta tokenizer does not accept empty string as segment B\n doc = id2doc[str(doc_id)][\"title\"]\n scores_1[b_idx][_] = float(\"-inf\")\n query_pairs.append((query, doc))\n \n q_sp_encodes = retriever_tokenizer.batch_encode_plus(\n query_pairs, max_length=args.max_q_sp_len, pad_to_max_length=True, return_tensors=\"pt\")\n q_sp_encodes = move_to_cuda(dict(q_sp_encodes))\n q_sp_embeds = retriever.encode_q(\n q_sp_encodes[\"input_ids\"], q_sp_encodes[\"attention_mask\"],q_sp_encodes.get(\"token_type_ids\", None)).cpu().numpy()\n scores_2, docid_2 = index.search(q_sp_embeds, args.topk)\n\n scores_2 = scores_2.reshape(1, args.topk, args.topk)\n docid_2 = docid_2.reshape(1, args.topk, args.topk)\n path_scores = np.expand_dims(scores_1, axis=2) + scores_2\n search_scores = path_scores[0]\n ranked_pairs = np.vstack(np.unravel_index(np.argsort(search_scores.ravel())[::-1], (args.topk, args.topk))).transpose()\n chains = []\n topk_docs = {}\n for _ in range(args.topk):\n path_ids = ranked_pairs[_]\n doc1_id = str(docid_1[0, path_ids[0]])\n doc2_id = str(docid_2[0, path_ids[0], path_ids[1]])\n chains.append([id2doc[doc1_id], id2doc[doc2_id]])\n topk_docs[id2doc[doc1_id]['title']] = id2doc[doc1_id]['text']\n topk_docs[id2doc[doc2_id]['title']] = id2doc[doc2_id]['text']\n\n\n reader_input = [{\n \"_id\": 0,\n \"question\": query,\n \"candidate_chains\": chains\n }]\n\n print(f\"Reading {len(chains)} chains...\")\n collate_fc = partial(qa_collate, pad_id=qa_tokenizer.pad_token_id)\n qa_eval_dataset = QAEvalDataset(\n qa_tokenizer, reader_input, max_seq_len=512, max_q_len=64)\n qa_eval_dataloader = DataLoader(\n qa_eval_dataset, batch_size=args.topk, collate_fn=collate_fc, pin_memory=True, num_workers=0)\n qa_results = eval_final(args, reader, qa_eval_dataloader, gpu=True)\n\n answer_pred = qa_results['answer'][0]\n sp_pred = qa_results['sp'][0]\n titles_pred = qa_results['titles'][0]\n\n\n st.markdown(f'**Answer**: {answer_pred}')\n st.markdown(f'**Supporting passages**:')\n st.markdown(f'> **{titles_pred[0]}**: {topk_docs[titles_pred[0]].replace(answer_pred, \"**\" + answer_pred + \"**\")}')\n st.markdown(\n f'> **{titles_pred[1]}**: {topk_docs[titles_pred[1]].replace(answer_pred, \"**\" + answer_pred + \"**\")}')\n\n # st.write(qa_results)\n" }, { "path": "scripts/download_hotpot.sh", "content": "#!/bin/bash\n# Copyright (c) Facebook, Inc. and its affiliates.\n# All rights reserved.\n#\n# This source code is licensed under the license found in the \n# LICENSE file in the root directory of this source tree.\n#!/bin/bash \n# Make data and model folder. \nmkdir data\nmkdir models\n\n# Download data \ncd data\nmkdir hotpot\ncd hotpot\nwget https://dl.fbaipublicfiles.com/mdpr/data/hotpot/hotpot_train_with_neg_v0.json\nwget https://dl.fbaipublicfiles.com/mdpr/data/hotpot/hotpot_dev_with_neg_v0.json\nwget https://dl.fbaipublicfiles.com/mdpr/data/hotpot/hotpot_qas_val.json\nwget https://dl.fbaipublicfiles.com/mdpr/data/hotpot/train_retrieval_b100_k100_sp.json\nwget https://dl.fbaipublicfiles.com/mdpr/data/hotpot/dev_retrieval_b50_k50_sp.json\nwget https://dl.fbaipublicfiles.com/mdpr/data/hotpot/dev_retrieval_top100_sp.json\n\ncd ..\nmkdir hotpot_index\ncd hotpot_index\nwget https://dl.fbaipublicfiles.com/mdpr/data/hotpot_index/wiki_id2doc.json\nwget https://dl.fbaipublicfiles.com/mdpr/data/hotpot_index/wiki_index.npy\n\necho \"Finished downloading data!\"\n\n# Download models\ncd ../../models\nwget https://dl.fbaipublicfiles.com/mdpr/models/doc_encoder.pt\nwget https://dl.fbaipublicfiles.com/mdpr/models/q_encoder.pt\nwget https://dl.fbaipublicfiles.com/mdpr/models/qa_electra.pt\n" }, { "path": "scripts/encode_corpus.py", "content": "# Copyright (c) Facebook, Inc. and its affiliates.\n# All rights reserved.\n#\n# This source code is licensed under the license found in the \n# LICENSE file in the root directory of this source tree.\n\n\"\"\"\nDescription: encode text corpus into a store of dense vectors. \n\nUsage (adjust the batch size according to your GPU memory):\n\nCUDA_VISIBLE_DEVICES=0,1,2,3 python scripts/encode_corpus.py \\\n --do_predict \\\n --predict_batch_size 1000 \\\n --model_name roberta-base \\\n --predict_file ${CORPUS_PATH} \\\n --init_checkpoint ${MODEL_CHECKPOINT} \\\n --embed_save_path ${SAVE_PATH} \\\n --fp16 \\\n --max_c_len 300 \\\n --num_workers 20 \n\n\"\"\"\n\nimport collections\nimport logging\nimport json\nimport os\nimport random\nfrom tqdm import tqdm\nimport numpy as np\nimport torch\nfrom transformers import AutoConfig, AutoTokenizer\nfrom torch.utils.data import DataLoader\n\nfrom mdr.retrieval.data.encode_datasets import EmDataset, em_collate\nfrom mdr.retrieval.models.retriever import CtxEncoder, RobertaCtxEncoder\nfrom mdr.retrieval.config import encode_args\nfrom mdr.retrieval.utils.utils import move_to_cuda, load_saved\n\ndef main():\n args = encode_args()\n if args.fp16:\n import apex\n apex.amp.register_half_function(torch, 'einsum')\n\n if args.local_rank == -1 or args.no_cuda:\n device = torch.device(\n \"cuda\" if torch.cuda.is_available() and not args.no_cuda else \"cpu\")\n n_gpu = torch.cuda.device_count()\n else:\n device = torch.device(\"cuda\", args.local_rank)\n n_gpu = 1\n torch.distributed.init_process_group(backend='nccl')\n\n if not args.predict_file:\n raise ValueError(\n \"If `do_predict` is True, then `predict_file` must be specified.\")\n\n bert_config = AutoConfig.from_pretrained(args.model_name)\n\n if \"roberta\" in args.model_name:\n model = RobertaCtxEncoder(bert_config, args)\n else:\n model = CtxEncoder(bert_config, args)\n tokenizer = AutoTokenizer.from_pretrained(args.model_name)\n\n eval_dataset = EmDataset(\n tokenizer, args.predict_file, args.max_q_len, args.max_c_len, args.is_query_embed, args.embed_save_path)\n eval_dataloader = DataLoader(\n eval_dataset, batch_size=args.predict_batch_size, collate_fn=em_collate, pin_memory=True, num_workers=args.num_workers)\n\n assert args.init_checkpoint != \"\"\n model = load_saved(model, args.init_checkpoint, exact=False)\n model.to(device)\n\n if args.fp16:\n try:\n from apex import amp\n except ImportError:\n raise ImportError(\n \"Please install apex from https://www.github.com/nvidia/apex to use fp16 training.\")\n model = amp.initialize(model, opt_level=args.fp16_opt_level)\n\n if args.local_rank != -1:\n model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[args.local_rank],\n output_device=args.local_rank)\n elif n_gpu > 1:\n model = torch.nn.DataParallel(model)\n\n embeds = predict(model, eval_dataloader)\n print(embeds.size())\n np.save(args.embed_save_path, embeds.cpu().numpy())\n\ndef predict(model, eval_dataloader):\n if type(model) == list:\n model = [m.eval() for m in model]\n else:\n model.eval()\n\n embed_array = []\n for batch in tqdm(eval_dataloader):\n batch_to_feed = move_to_cuda(batch)\n with torch.no_grad():\n results = model(batch_to_feed)\n embed = results['embed'].cpu()\n embed_array.append(embed)\n\n ## linear combination tuning on dev data\n embed_array = torch.cat(embed_array)\n\n model.train()\n return embed_array\n\n\nif __name__ == \"__main__\":\n main()\n" }, { "path": "scripts/end2end.py", "content": "#!/usr/bin/env python\n# Copyright (c) Facebook, Inc. and its affiliates.\n# All rights reserved.\n#\n# This source code is licensed under the license found in the\n# LICENSE file in the root directory of this source tree.\n\n\"\"\"\nEfficient end2end QA with HNSW index\n\ntaskset --cpu-list 0-15 python end2end.py ../data/hotpot/hotpot_qas_val.json\n\"\"\"\nimport argparse\nimport json\nimport logging\nfrom functools import partial\nimport time\n\nimport argparse\nimport collections\nimport json\nimport logging\nfrom torch.utils.data import DataLoader\n\nimport faiss\nimport numpy as np\nimport torch\nfrom tqdm import tqdm\nfrom transformers import AutoConfig, AutoTokenizer\n\nfrom retrieval.models.mhop_retriever import RobertaRetriever\nfrom retrieval.utils.utils import load_saved\n\nfrom qa.qa_model import QAModel\nfrom mdr.qa.qa_dataset import qa_collate, QAEvalDataset\nfrom .train_qa import eval_final\nfrom qa.hotpot_evaluate_v1 import f1_score, exact_match_score\nfrom qa.utils import set_global_logging_level\n\nlogger = logging.getLogger()\nlogger.setLevel(logging.INFO)\nif (logger.hasHandlers()):\n logger.handlers.clear()\nconsole = logging.StreamHandler()\nlogger.addHandler(console)\n\nset_global_logging_level(logging.ERROR, [\"transformers\", \"nlp\", \"torch\", \"tensorflow\", \"tensorboard\", \"wandb\"])\n\ndef convert_hnsw_query(query_vectors):\n aux_dim = np.zeros(len(query_vectors), dtype='float32')\n query_nhsw_vectors = np.hstack((query_vectors, aux_dim.reshape(-1, 1)))\n return query_nhsw_vectors\n\nif __name__ == '__main__':\n\n parser = argparse.ArgumentParser()\n parser.add_argument('raw_data', type=str, default=None)\n parser.add_argument('--indexpath', type=str, default=\"retrieval/index/wiki_index_hnsw_roberta\")\n parser.add_argument('--corpus_dict', type=str, default='retrieval/index/hotpotQA_corpus_dict.json')\n parser.add_argument('--retriever_path', type=str, default=\"retrieval/logs/08-16-2020/roberta_momentum_freeze_k-seed16-bsz150-fp16True-lr1e-05-decay0.0-warm0-valbsz3000-m0.999-k76800/checkpoint_q_best.pt\")\n parser.add_argument('--reader_path', type=str, default=\"qa/logs/08-10-2020/electra_val_top30-epoch7-lr5e-05-seed42-rdrop0-qadrop0-decay0-qpergpu2-aggstep8-clip2-evalper250-evalbsize1024-negnum5-warmup0.1-adamTrue-spweight0.025/checkpoint_best.pt\")\n parser.add_argument('--topk', type=int, default=1, help=\"topk paths\")\n parser.add_argument('--num-workers', type=int, default=10)\n parser.add_argument('--max-q-len', type=int, default=70)\n parser.add_argument('--max-q-sp-len', type=int, default=350)\n parser.add_argument('--batch-size', type=int, default=1)\n parser.add_argument(\"--max-ans-len\", default=35, type=int)\n parser.add_argument(\"--save-prediction\", default=\"\", type=str)\n parser.add_argument(\"--model-name\", type=str, default=\"\")\n parser.add_argument(\"--sp-pred\", action=\"store_true\", help=\"whether to predict sentence sp\")\n parser.add_argument(\"--sp-weight\", default=0, type=float, help=\"weight of the sp loss\")\n # parser.add_argument('--hnsw', action=\"store_true\")\n args = parser.parse_args()\n\n logger.info(\"Loading trained models...\")\n retrieval_config = AutoConfig.from_pretrained('roberta-base')\n retrieval_tokenizer = AutoTokenizer.from_pretrained('roberta-base')\n args.model_name = \"roberta-base\"\n retriever = RobertaRetriever(retrieval_config, args)\n retriever = load_saved(retriever, args.retriever_path)\n retriever.eval()\n \n qa_config = AutoConfig.from_pretrained('google/electra-large-discriminator')\n qa_tokenizer = AutoTokenizer.from_pretrained('google/electra-large-discriminator')\n args.model_name = \"google/electra-large-discriminator\"\n reader = QAModel(qa_config, args)\n reader = load_saved(reader, args.reader_path, False)\n reader.eval()\n\n logger.info(\"Loading index...\")\n index = faiss.read_index(args.indexpath)\n\n logger.info(f\"Loading corpus...\")\n id2doc = json.load(open(args.corpus_dict))\n logger.info(f\"Corpus size {len(id2doc)}\")\n\n logger.info(\"Loading queries...\")\n qas_items = [json.loads(_) for _ in open(args.raw_data).readlines()[:5]]\n questions = [_[\"question\"][:-1] if _[\"question\"].endswith(\"?\") else _[\"question\"] for _ in qas_items]\n id2gold_ans = {_[\"_id\"]: _[\"answer\"][0] for _ in qas_items}\n\n\n start = time.time()\n logger.info(\"Retrieving...\")\n retrieval_results = []\n\n encode_times = []\n search_times = []\n\n\n with torch.no_grad():\n for b_start in tqdm(range(0, len(questions), args.batch_size)):\n # 1-hop retrieval\n batch_q = questions[b_start:b_start + args.batch_size]\n batch_qas = qas_items[b_start:b_start + args.batch_size]\n batch_q_encodes = retrieval_tokenizer.batch_encode_plus(batch_q, max_length=args.max_q_len, pad_to_max_length=True, return_tensors=\"pt\")\n q_embeds = retriever.encode_q(batch_q_encodes[\"input_ids\"], batch_q_encodes[\"attention_mask\"], batch_q_encodes.get(\"token_type_ids\", None))\n q_embeds_numpy = q_embeds.numpy()\n q_embeds_numpy = convert_hnsw_query(q_embeds_numpy)\n scores_1, docid_1 = index.search(q_embeds_numpy, args.topk)\n\n # construct 2hop queries\n bsize = len(batch_q)\n query_pairs = []\n for b_idx in range(bsize):\n for _, doc_id in enumerate(docid_1[b_idx]):\n doc = id2doc[str(doc_id)][\"text\"]\n if doc.strip() == \"\":\n # roberta tokenizer does not accept empty string as segment B\n doc = id2doc[str(doc_id)][\"title\"]\n scores_1[b_idx][_] = float(\"-inf\")\n query_pairs.append((batch_q[b_idx], doc))\n \n # 2-hop retrieval\n s1 = time.time()\n batch_q_sp_encodes = retrieval_tokenizer.batch_encode_plus(query_pairs, max_length=args.max_q_sp_len, pad_to_max_length=True, return_tensors=\"pt\")\n q_sp_embeds = retriever.encode_q(batch_q_sp_encodes[\"input_ids\"], batch_q_sp_encodes[\"attention_mask\"], batch_q_sp_encodes.get(\"token_type_ids\", None))\n encode_times.append(time.time() - s1)\n\n s2 = time.time()\n q_sp_embeds = q_sp_embeds.numpy()\n q_sp_embeds = convert_hnsw_query(q_sp_embeds)\n scores_2, docid_2 = index.search(q_sp_embeds, args.topk)\n search_times.append(time.time() - s2)\n\n # aggregate chain scores\n scores_2 = scores_2.reshape(bsize, args.topk, args.topk)\n docid_2 = docid_2.reshape(bsize, args.topk, args.topk)\n path_scores = - (np.expand_dims(scores_1, axis=2) + scores_2)\n\n for idx in range(bsize):\n search_scores = path_scores[idx]\n ranked_pairs = np.vstack(np.unravel_index(np.argsort(search_scores.ravel())[::-1], (args.topk, args.topk))).transpose()\n\n chains = []\n for _ in range(args.topk):\n path_ids = ranked_pairs[_]\n doc1_id = str(docid_1[idx, path_ids[0]])\n doc2_id = str(docid_2[idx, path_ids[0], path_ids[1]])\n chains.append([id2doc[doc1_id], id2doc[doc2_id]])\n\n retrieval_results.append({\n \"_id\": batch_qas[idx][\"_id\"],\n \"question\": batch_qas[idx][\"question\"],\n \"candidate_chains\": chains\n })\n\n\n logger.info(\"Reading...\")\n collate_fc = partial(qa_collate, pad_id=qa_tokenizer.pad_token_id)\n qa_eval_dataset = QAEvalDataset(qa_tokenizer, retrieval_results, max_seq_len=512, max_q_len=64)\n qa_eval_dataloader = DataLoader(qa_eval_dataset, batch_size=args.topk, collate_fn=collate_fc, pin_memory=True, num_workers=0)\n qa_results = eval_final(args, reader, qa_eval_dataloader, gpu=False)\n print(f\"Finishing evaluation in {time.time() - start}s\")\n\n\n ems = [exact_match_score(qa_results[\"answer\"][k], id2gold_ans[k]) for k in qa_results[\"answer\"].keys()]\n f1s = [f1_score(qa_results[\"answer\"][k], id2gold_ans[k]) for k in qa_results[\"answer\"].keys()]\n\n logger.info(f\"Answer EM {np.mean(ems)}, F1 {np.mean(f1s)}\")\n" }, { "path": "scripts/end2end.sh", "content": "#!/bin/bash\n# Copyright (c) Facebook, Inc. and its affiliates.\n# All rights reserved.\n#\n# This source code is licensed under the license found in the \n# LICENSE file in the root directory of this source tree.\n\n#SBATCH --cpus-per-task=16\n#SBATCH --nodes=1\n#SBATCH --ntasks-per-node=1\n#SBATCH --time=12:00:00\n#SBATCH --job-name=hotpot_eval\n#SBATCH --output=/private/home/xwhan/Mhop-Pretrain/eval_logs\n#SBATCH --partition=dev\n#SBATCH --error=/checkpoint/%u/hotpot-jobs/sample-%j.err\n#SBATCH --mem=500GB\n#SBATCH --signal=USR1@140\n#SBATCH --open-mode=append\n\n--wrap=\"srun python end2end.py \\\n ../data/hotpot/hotpot_qas_val.json\"\n\n\n\nsbatch --job-name=hotpot_eval \\\n--error=/checkpoint/hotpot-jobs/hotpot-%j.err \\\n--output=/checkpoint/hotpot-jobs/hotpot-%j.out \\\n--partition=dev --nodes=1 --ntasks-per-node=1 \\\n--cpus-per-task=16 \\\n--gpus-per-node=1 --open-mode=append \\\n--time=12:00:00 \\\n--wrap=\"srun python end2end.py ../data/hotpot/hotpot_qas_val.json\"" }, { "path": "scripts/eval/eval_mhop_fever.py", "content": "# Copyright (c) Facebook, Inc. and its affiliates.\n# All rights reserved.\n#\n# This source code is licensed under the license found in the \n# LICENSE file in the root directory of this source tree.\n\"\"\"\npython eval_mhop_fever.py /private/home/xwhan/data/fever/retrieval/dev_multi_evidence.txt index/fever.npy index/fever_corpus_id2doc.json logs/08-27-2020/fever-seed16-bsz96-fp16True-lr2e-05-decay0.0-warm0.1-valbsz3000-sharedTrue/checkpoint_best.pt --batch-size 100 --beam-size-1 2 --beam-size-2 10 --topk 20 --shared-encoder --model-name roberta-base --gpu --save-path dense_fever_b2_10_k20.json\n\n# unified retrieval\npython eval_mhop_fever.py /private/home/xwhan/data/fever/retrieval/dev.txt index/fever_unified.npy index/fever_corpus_id2doc.json logs/08-30-2020/fever_unified_roberta-seed16-bsz96-fp16True-lr2e-05-decay0.0-adamTrue/checkpoint_best.pt --batch-size 100 --beam-size-1 1 --beam-size-2 20 --topk 20 --shared-encoder --model-name roberta-base --gpu --save-path dense_all_b1_k10_unified.json\n\n\npython eval_mhop_fever.py /private/home/xwhan/data/fever/retrieval/dev_multi_evidence.txt index/fever_unified.npy index/fever_corpus_id2doc.json logs/08-30-2020/fever_unified_roberta-seed16-bsz96-fp16True-lr2e-05-decay0.0-adamTrue/checkpoint_best.pt --batch-size 100 --beam-size-1 1 --beam-size-2 20 --topk 20 --shared-encoder --model-name roberta-base --gpu --save-path dense_multi_b1_k10_unified.json\n\n# fix parenthesis\npython eval_mhop_fever.py /private/home/xwhan/data/fever/retrieval/multi_dev.txt index/fever_.npy index/fever_corpus_id2doc.json logs/08-27-2020/fever_-seed16-bsz96-fp16True-lr2e-05-decay0.0-warm0.1-valbsz3000-sharedTrue/checkpoint_best.pt --batch-size 100 --beam-size-1 1 --beam-size-2 20 --topk 20 --shared-encoder --model-name roberta-base --gpu --save-path dense_fever_b1_k20_fix_brc.json\n\n\"\"\"\nimport argparse\nimport json\nimport logging\nfrom os import path\n\nimport faiss\nimport numpy as np\nimport torch\nfrom tqdm import tqdm\nfrom transformers import AutoConfig, AutoTokenizer\n\nfrom models.mhop_retriever import RobertaRetriever\nfrom models.unified_retriever import UnifiedRetriever\nfrom utils.basic_tokenizer import SimpleTokenizer\nfrom utils.utils import (load_saved, move_to_cuda, para_has_answer)\n\nlogger = logging.getLogger()\nlogger.setLevel(logging.INFO)\nif (logger.hasHandlers()):\n logger.handlers.clear()\nconsole = logging.StreamHandler()\nlogger.addHandler(console)\n\n\nif __name__ == '__main__':\n parser = argparse.ArgumentParser()\n parser.add_argument('raw_data', type=str, default=None)\n parser.add_argument('indexpath', type=str, default=None)\n parser.add_argument('corpus_dict', type=str, default=None)\n parser.add_argument('model_path', type=str, default=None)\n parser.add_argument('--topk', type=int, default=2, help=\"topk paths\")\n parser.add_argument('--num-workers', type=int, default=10)\n parser.add_argument('--max-q-len', type=int, default=45)\n parser.add_argument('--max-c-len', type=int, default=350)\n parser.add_argument('--max-q-sp-len', type=int, default=400)\n parser.add_argument('--batch-size', type=int, default=100)\n parser.add_argument('--beam-size-1', type=int, default=5)\n parser.add_argument('--beam-size-2', type=int, default=5)\n parser.add_argument('--model-name', type=str, default='bert-base-uncased')\n parser.add_argument('--gpu', action=\"store_true\")\n parser.add_argument('--shared-encoder', action=\"store_true\")\n parser.add_argument(\"--save-path\", type=str, default=\"\")\n parser.add_argument(\"--stop-drop\", default=0, type=float)\n args = parser.parse_args()\n \n logger.info(\"Loading data...\")\n ds_items = [json.loads(_) for _ in open(args.raw_data).readlines()]\n\n logger.info(\"Building index...\")\n d = 768\n xb = np.load(args.indexpath).astype('float32')\n print(xb.shape)\n\n index = faiss.IndexFlatIP(d)\n index.add(xb)\n if args.gpu:\n res = faiss.StandardGpuResources()\n index = faiss.index_cpu_to_gpu(res, 1, index)\n\n logger.info(f\"Loading corpus...\")\n id2doc = json.load(open(args.corpus_dict))\n title2doc = {item[0]:item[1] for item in id2doc.values()}\n logger.info(f\"Corpus size {len(id2doc)}\")\n \n logger.info(\"Loading trained model...\")\n bert_config = AutoConfig.from_pretrained(args.model_name)\n tokenizer = AutoTokenizer.from_pretrained(args.model_name)\n model = RobertaRetriever(bert_config, args)\n # model = UnifiedRetriever(bert_config, args)\n model = load_saved(model, args.model_path, exact=False)\n simple_tokenizer = SimpleTokenizer()\n\n cuda = torch.device('cuda')\n model.to(cuda)\n from apex import amp\n model = amp.initialize(model, opt_level='O1')\n model.eval()\n\n logger.info(\"Encoding claims and searching\")\n questions = [_[\"claim\"] for _ in ds_items]\n metrics = []\n retrieval_outputs = []\n for b_start in tqdm(range(0, len(questions), args.batch_size)):\n with torch.no_grad():\n batch_q = questions[b_start:b_start + args.batch_size]\n batch_ann = ds_items[b_start:b_start + args.batch_size]\n bsize = len(batch_q)\n batch_q_encodes = tokenizer.batch_encode_plus(batch_q, max_length=args.max_q_len, pad_to_max_length=True, return_tensors=\"pt\")\n batch_q_encodes = move_to_cuda(dict(batch_q_encodes))\n q_embeds = model.encode_q(batch_q_encodes[\"input_ids\"], batch_q_encodes[\"attention_mask\"], batch_q_encodes.get(\"token_type_ids\", None))\n q_embeds_numpy = q_embeds.cpu().contiguous().numpy()\n\n D, I = index.search(q_embeds_numpy, args.beam_size_1)\n\n # 2hop search\n query_pairs = []\n for b_idx in range(bsize):\n for _, doc_id in enumerate(I[b_idx]):\n doc = id2doc[str(doc_id)][1]\n if \"roberta\" in args.model_name and doc.strip() == \"\":\n # doc = \"fadeaxsaa\" * 100\n doc = id2doc[str(doc_id)][0]\n D[b_idx][_] = float(\"-inf\")\n query_pairs.append((batch_q[b_idx], doc))\n\n batch_q_sp_encodes = tokenizer.batch_encode_plus(query_pairs, max_length=args.max_q_sp_len, pad_to_max_length=True, return_tensors=\"pt\")\n\n batch_q_sp_encodes = move_to_cuda(dict(batch_q_sp_encodes))\n q_sp_embeds = model.encode_q(batch_q_sp_encodes[\"input_ids\"], batch_q_sp_encodes[\"attention_mask\"], batch_q_sp_encodes.get(\"token_type_ids\", None))\n q_sp_embeds = q_sp_embeds.contiguous().cpu().numpy()\n # search_start = time.time()\n D_, I_ = index.search(q_sp_embeds, args.beam_size_2)\n # logger.info(f\"MIPS searching: {time.time() - search_start}\")\n D_ = D_.reshape(bsize, args.beam_size_1, args.beam_size_2)\n I_ = I_.reshape(bsize, args.beam_size_1, args.beam_size_2)\n\n # aggregate path scores\n path_scores = np.expand_dims(D, axis=2) + D_\n\n # path_scores = D_\n # eval\n for idx in range(bsize):\n search_scores = path_scores[idx]\n ranked_pairs = np.vstack(np.unravel_index(np.argsort(search_scores.ravel())[::-1], (args.beam_size_1, args.beam_size_2))).transpose()\n retrieved_titles = []\n hop1_titles = []\n paths, path_titles = [], []\n paths_both_are_intro = []\n for _ in range(args.topk):\n path_ids = ranked_pairs[_]\n hop_1_id = I[idx, path_ids[0]]\n hop_2_id = I_[idx, path_ids[0], path_ids[1]]\n retrieved_titles.append(id2doc[str(hop_1_id)][0])\n retrieved_titles.append(id2doc[str(hop_2_id)][0])\n\n paths.append([str(hop_1_id), str(hop_2_id)])\n path_titles.append([id2doc[str(hop_1_id)][0], id2doc[str(hop_2_id)][0]])\n paths_both_are_intro.append(id2doc[str(hop_1_id)][2] and id2doc[str(hop_2_id)][2])\n hop1_titles.append(id2doc[str(hop_1_id)][0])\n\n # saving when there's no annotations\n if args.save_path != \"\":\n candidaite_chains = []\n for path in path_titles:\n candidaite_chains.append([(path[0], title2doc[path[0]]), (path[1], title2doc[path[1]])])\n retrieval_outputs.append({\n \"id\": batch_ann[idx][\"id\"],\n \"claim\": batch_ann[idx][\"claim\"],\n \"candidate_chains\": candidaite_chains,\n\n })\n\n if args.save_path != \"\":\n with open(f\"/private/home/xwhan/data/fever/retrieval/{args.save_path}\", \"w\") as out:\n for l in retrieval_outputs:\n out.write(json.dumps(l) + \"\\n\")\n\n" }, { "path": "scripts/eval/eval_mhop_retrieval.py", "content": "# Copyright (c) Facebook, Inc. and its affiliates.\n# All rights reserved.\n#\n# This source code is licensed under the license found in the \n# LICENSE file in the root directory of this source tree.\n\"\"\"\nEvaluating trained retrieval model.\n\nUsage:\npython eval_mhop_retrieval.py ${EVAL_DATA} ${CORPUS_VECTOR_PATH} ${CORPUS_DICT} ${MODEL_CHECKPOINT} \\\n --batch-size 50 \\\n --beam-size-1 20 \\\n --beam-size-2 5 \\\n --topk 20 \\\n --shared-encoder \\\n --gpu \\\n --save-path ${PATH_TO_SAVE_RETRIEVAL}\n\n\"\"\"\nimport argparse\nimport collections\nimport json\nimport logging\nfrom os import path\nimport time\n\nimport faiss\nimport numpy as np\nimport torch\nfrom tqdm import tqdm\nfrom transformers import AutoConfig, AutoTokenizer\n\nfrom mdr.retrieval.models.mhop_retriever import RobertaRetriever\nfrom mdr.retrieval.utils.basic_tokenizer import SimpleTokenizer\nfrom mdr.retrieval.utils.utils import (load_saved, move_to_cuda, para_has_answer)\n\nlogger = logging.getLogger()\nlogger.setLevel(logging.INFO)\nif (logger.hasHandlers()):\n logger.handlers.clear()\nconsole = logging.StreamHandler()\nlogger.addHandler(console)\n\ndef convert_hnsw_query(query_vectors):\n aux_dim = np.zeros(len(query_vectors), dtype='float32')\n query_nhsw_vectors = np.hstack((query_vectors, aux_dim.reshape(-1, 1)))\n return query_nhsw_vectors\n\nif __name__ == '__main__':\n parser = argparse.ArgumentParser()\n parser.add_argument('raw_data', type=str, default=None)\n parser.add_argument('indexpath', type=str, default=None)\n parser.add_argument('corpus_dict', type=str, default=None)\n parser.add_argument('model_path', type=str, default=None)\n parser.add_argument('--topk', type=int, default=2, help=\"topk paths\")\n parser.add_argument('--num-workers', type=int, default=10)\n parser.add_argument('--max-q-len', type=int, default=70)\n parser.add_argument('--max-c-len', type=int, default=300)\n parser.add_argument('--max-q-sp-len', type=int, default=350)\n parser.add_argument('--batch-size', type=int, default=100)\n parser.add_argument('--beam-size', type=int, default=5)\n parser.add_argument('--model-name', type=str, default='roberta-base')\n parser.add_argument('--gpu', action=\"store_true\")\n parser.add_argument('--save-index', action=\"store_true\")\n parser.add_argument('--only-eval-ans', action=\"store_true\")\n parser.add_argument('--shared-encoder', action=\"store_true\")\n parser.add_argument(\"--save-path\", type=str, default=\"\")\n parser.add_argument(\"--stop-drop\", default=0, type=float)\n parser.add_argument('--hnsw', action=\"store_true\")\n args = parser.parse_args()\n \n logger.info(\"Loading data...\")\n ds_items = [json.loads(_) for _ in open(args.raw_data).readlines()]\n\n # filter\n if args.only_eval_ans:\n ds_items = [_ for _ in ds_items if _[\"answer\"][0] not in [\"yes\", \"no\"]]\n\n logger.info(\"Loading trained model...\")\n bert_config = AutoConfig.from_pretrained(args.model_name)\n tokenizer = AutoTokenizer.from_pretrained(args.model_name)\n model = RobertaRetriever(bert_config, args)\n model = load_saved(model, args.model_path, exact=False)\n simple_tokenizer = SimpleTokenizer()\n\n cuda = torch.device('cuda')\n model.to(cuda)\n from apex import amp\n model = amp.initialize(model, opt_level='O1')\n model.eval()\n\n logger.info(\"Building index...\")\n d = 768\n xb = np.load(args.indexpath).astype('float32')\n\n if args.hnsw:\n if path.exists(\"data/hotpot_index/wiki_index_hnsw.index\"):\n index = faiss.read_index(\"index/wiki_index_hnsw.index\")\n else:\n index = faiss.IndexHNSWFlat(d + 1, 512)\n index.hnsw.efSearch = 128\n index.hnsw.efConstruction = 200\n phi = 0\n for i, vector in enumerate(xb):\n norms = (vector ** 2).sum()\n phi = max(phi, norms)\n logger.info('HNSWF DotProduct -> L2 space phi={}'.format(phi))\n\n data = xb\n buffer_size = 50000\n n = len(data)\n print(n)\n for i in tqdm(range(0, n, buffer_size)):\n vectors = [np.reshape(t, (1, -1)) for t in data[i:i + buffer_size]]\n norms = [(doc_vector ** 2).sum() for doc_vector in vectors]\n aux_dims = [np.sqrt(phi - norm) for norm in norms]\n hnsw_vectors = [np.hstack((doc_vector, aux_dims[idx].reshape(-1, 1))) for idx, doc_vector in enumerate(vectors)]\n hnsw_vectors = np.concatenate(hnsw_vectors, axis=0)\n index.add(hnsw_vectors)\n else:\n index = faiss.IndexFlatIP(d)\n index.add(xb)\n if args.gpu:\n res = faiss.StandardGpuResources()\n index = faiss.index_cpu_to_gpu(res, 6, index)\n\n if args.save_index:\n faiss.write_index(index, \"data/hotpot_index/wiki_index_hnsw_roberta\")\n \n logger.info(f\"Loading corpus...\")\n id2doc = json.load(open(args.corpus_dict))\n if isinstance(id2doc[\"0\"], list):\n id2doc = {k: {\"title\":v[0], \"text\": v[1]} for k, v in id2doc.items()}\n # title2text = {v[0]:v[1] for v in id2doc.values()}\n logger.info(f\"Corpus size {len(id2doc)}\")\n \n\n logger.info(\"Encoding questions and searching\")\n questions = [_[\"question\"][:-1] if _[\"question\"].endswith(\"?\") else _[\"question\"] for _ in ds_items]\n metrics = []\n retrieval_outputs = []\n for b_start in tqdm(range(0, len(questions), args.batch_size)):\n with torch.no_grad():\n batch_q = questions[b_start:b_start + args.batch_size]\n batch_ann = ds_items[b_start:b_start + args.batch_size]\n bsize = len(batch_q)\n\n batch_q_encodes = tokenizer.batch_encode_plus(batch_q, max_length=args.max_q_len, pad_to_max_length=True, return_tensors=\"pt\")\n batch_q_encodes = move_to_cuda(dict(batch_q_encodes))\n q_embeds = model.encode_q(batch_q_encodes[\"input_ids\"], batch_q_encodes[\"attention_mask\"], batch_q_encodes.get(\"token_type_ids\", None))\n\n q_embeds_numpy = q_embeds.cpu().contiguous().numpy()\n if args.hnsw:\n q_embeds_numpy = convert_hnsw_query(q_embeds_numpy)\n D, I = index.search(q_embeds_numpy, args.beam_size)\n\n # 2hop search\n query_pairs = []\n for b_idx in range(bsize):\n for _, doc_id in enumerate(I[b_idx]):\n doc = id2doc[str(doc_id)][\"text\"]\n if \"roberta\" in args.model_name and doc.strip() == \"\":\n # doc = \"fadeaxsaa\" * 100\n doc = id2doc[str(doc_id)][\"title\"]\n D[b_idx][_] = float(\"-inf\")\n query_pairs.append((batch_q[b_idx], doc))\n\n batch_q_sp_encodes = tokenizer.batch_encode_plus(query_pairs, max_length=args.max_q_sp_len, pad_to_max_length=True, return_tensors=\"pt\")\n batch_q_sp_encodes = move_to_cuda(dict(batch_q_sp_encodes))\n s1 = time.time()\n q_sp_embeds = model.encode_q(batch_q_sp_encodes[\"input_ids\"], batch_q_sp_encodes[\"attention_mask\"], batch_q_sp_encodes.get(\"token_type_ids\", None))\n # print(\"Encoding time:\", time.time() - s1)\n\n \n q_sp_embeds = q_sp_embeds.contiguous().cpu().numpy()\n s2 = time.time()\n if args.hnsw:\n q_sp_embeds = convert_hnsw_query(q_sp_embeds)\n D_, I_ = index.search(q_sp_embeds, args.beam_size)\n\n D_ = D_.reshape(bsize, args.beam_size, args.beam_size)\n I_ = I_.reshape(bsize, args.beam_size, args.beam_size)\n\n # aggregate path scores\n path_scores = np.expand_dims(D, axis=2) + D_\n\n if args.hnsw:\n path_scores = - path_scores\n\n for idx in range(bsize):\n search_scores = path_scores[idx]\n ranked_pairs = np.vstack(np.unravel_index(np.argsort(search_scores.ravel())[::-1],\n (args.beam_size, args.beam_size))).transpose()\n retrieved_titles = []\n hop1_titles = []\n paths, path_titles = [], []\n for _ in range(args.topk):\n path_ids = ranked_pairs[_]\n hop_1_id = I[idx, path_ids[0]]\n hop_2_id = I_[idx, path_ids[0], path_ids[1]]\n retrieved_titles.append(id2doc[str(hop_1_id)][\"title\"])\n retrieved_titles.append(id2doc[str(hop_2_id)][\"title\"])\n\n paths.append([str(hop_1_id), str(hop_2_id)])\n path_titles.append([id2doc[str(hop_1_id)][\"title\"], id2doc[str(hop_2_id)][\"title\"]])\n hop1_titles.append(id2doc[str(hop_1_id)][\"title\"])\n \n if args.only_eval_ans:\n gold_answers = batch_ann[idx][\"answer\"]\n concat_p = \"yes no \"\n for p in paths:\n concat_p += \" \".join([id2doc[doc_id][\"title\"] + \" \" + id2doc[doc_id][\"text\"] for doc_id in p])\n metrics.append({\n \"question\": batch_ann[idx][\"question\"],\n \"ans_recall\": int(para_has_answer(gold_answers, concat_p, simple_tokenizer)),\n \"type\": batch_ann[idx].get(\"type\", \"single\")\n })\n \n else:\n sp = batch_ann[idx][\"sp\"]\n assert len(set(sp)) == 2\n type_ = batch_ann[idx][\"type\"]\n question = batch_ann[idx][\"question\"]\n p_recall, p_em = 0, 0\n sp_covered = [sp_title in retrieved_titles for sp_title in sp]\n if np.sum(sp_covered) > 0:\n p_recall = 1\n if np.sum(sp_covered) == len(sp_covered):\n p_em = 1\n path_covered = [int(set(p) == set(sp)) for p in path_titles]\n path_covered = np.sum(path_covered) > 0\n recall_1 = 0\n covered_1 = [sp_title in hop1_titles for sp_title in sp]\n if np.sum(covered_1) > 0: recall_1 = 1\n metrics.append({\n \"question\": question,\n \"p_recall\": p_recall,\n \"p_em\": p_em,\n \"type\": type_,\n 'recall_1': recall_1,\n 'path_covered': int(path_covered)\n })\n\n\n # saving when there's no annotations\n candidaite_chains = []\n for path in paths:\n candidaite_chains.append([id2doc[path[0]], id2doc[path[1]]])\n \n retrieval_outputs.append({\n \"_id\": batch_ann[idx][\"_id\"],\n \"question\": batch_ann[idx][\"question\"],\n \"candidate_chains\": candidaite_chains,\n # \"sp\": sp_chain,\n # \"answer\": gold_answers,\n # \"type\": type_,\n # \"coverd_k\": covered_k\n })\n\n if args.save_path != \"\":\n with open(args.save_path, \"w\") as out:\n for l in retrieval_outputs:\n out.write(json.dumps(l) + \"\\n\")\n\n logger.info(f\"Evaluating {len(metrics)} samples...\")\n type2items = collections.defaultdict(list)\n for item in metrics:\n type2items[item[\"type\"]].append(item)\n if args.only_eval_ans:\n logger.info(f'Ans Recall: {np.mean([m[\"ans_recall\"] for m in metrics])}')\n for t in type2items.keys():\n logger.info(f\"{t} Questions num: {len(type2items[t])}\")\n logger.info(f'Ans Recall: {np.mean([m[\"ans_recall\"] for m in type2items[t]])}')\n else:\n logger.info(f'\\tAvg PR: {np.mean([m[\"p_recall\"] for m in metrics])}')\n logger.info(f'\\tAvg P-EM: {np.mean([m[\"p_em\"] for m in metrics])}')\n logger.info(f'\\tAvg 1-Recall: {np.mean([m[\"recall_1\"] for m in metrics])}')\n logger.info(f'\\tPath Recall: {np.mean([m[\"path_covered\"] for m in metrics])}')\n for t in type2items.keys():\n logger.info(f\"{t} Questions num: {len(type2items[t])}\")\n logger.info(f'\\tAvg PR: {np.mean([m[\"p_recall\"] for m in type2items[t]])}')\n logger.info(f'\\tAvg P-EM: {np.mean([m[\"p_em\"] for m in type2items[t]])}')\n logger.info(f'\\tAvg 1-Recall: {np.mean([m[\"recall_1\"] for m in type2items[t]])}')\n logger.info(f'\\tPath Recall: {np.mean([m[\"path_covered\"] for m in type2items[t]])}')\n" }, { "path": "scripts/eval/eval_reranked.py", "content": "# Copyright (c) Facebook, Inc. and its affiliates.\n# All rights reserved.\n#\n# This source code is licensed under the license found in the \n# LICENSE file in the root directory of this source tree.\nimport json\nimport numpy as np\nfrom utils.utils import para_has_answer, normalize\nfrom utils.basic_tokenizer import SimpleTokenizer\nfrom tqdm import tqdm\n\n\ncorpus = json.load(open(\"../data/hotpot_index/wiki_id2doc.json\"))\ntitle2text = {v[\"title\"]:v[\"text\"] for v in corpus.values()}\n\nval_inputs = [json.loads(l) for l in open(\"../data/hotpot/hotpot_qas_val.json\").readlines()]\nid2goldsp = {_[\"_id\"]:_[\"sp\"] for _ in val_inputs}\nid2goldans = {_[\"_id\"]:_[\"answer\"] for _ in val_inputs}\nid2type = {_[\"_id\"]:_[\"type\"] for _ in val_inputs}\n\n\n\nsimple_tokenizer = SimpleTokenizer()\n\n# out best results\nresults = json.load(open(\"../data/hotpot/results/hotpot_val_top100.json\"))\n\n# # asai results\n# results = json.load(open(\"/private/home/xwhan/code/learning_to_retrieve_reasoning_paths/results/hotpot_dev_reader_titles.json\"))\n# for k in results.keys():\n# v = results[k]\n# v = [normalize(_[:-2]) for _ in v]\n# # import pdb; pdb.set_trace()\n# results[k] = v\n# results = {\"titles\":results}\n\nsp_ems = []\nans_recalls = []\nbridge_ems = []\ncompare_ems = []\nfor qid in tqdm(results[\"titles\"].keys()):\n chain = results[\"titles\"][qid]\n sp = id2goldsp[qid]\n answer = id2goldans[qid]\n type_ = id2type[qid]\n\n # if answer[0].strip() in [\"yes\", \"no\"]:\n # continue\n\n sp_covered = int(np.sum([int(_ in chain) for _ in sp]) == len(sp))\n concat_p = \"yes no \" + \" \".join([t + \" \" + title2text.get(t, \"\") for t in chain])\n ans_covered = para_has_answer(answer, concat_p, simple_tokenizer)\n ans_recalls.append(ans_covered)\n sp_ems.append(sp_covered)\n\n if type_ == \"bridge\":\n bridge_ems.append(sp_covered)\n else:\n compare_ems.append(sp_covered)\n \n\nprint(len(sp_ems))\nprint(np.mean(sp_ems))\nprint(f\"Answer Recall: {np.mean(ans_recalls)}, count: {len(ans_recalls)}\")\n\nprint(\"Bridge P EM:\", np.mean(bridge_ems))\nprint(\"Comparison P EM:\", np.mean(compare_ems))" }, { "path": "scripts/eval/eval_retrieval.py", "content": "# Copyright (c) Facebook, Inc. and its affiliates.\n# All rights reserved.\n#\n# This source code is licensed under the license found in the \n# LICENSE file in the root directory of this source tree.\n\"\"\"\nSingle-hop retrieval evaluation\n\n## Use the unified model (trained with both hotpotQA and NQ)\n\n\npython eval_retrieval.py /private/home/xwhan/data/nq-dpr/nq-val-simplified.txt index/psg100_unified.npy index/psgs_w100_id2doc.json logs/07-24-2020/unified_continue-seed16-bsz150-fp16True-lr1e-05-decay0.0/checkpoint_best.pt --batch-size 1000 --shared-encoder --model-name bert-base-uncased --unified --save-pred nq-val-filtered-top50.txt --topk 50\n\n\n# DPR shared-encoder baseline bsz256\npython eval_retrieval.py /private/home/xwhan/data/nq-dpr/nq-test-qas.txt index/psg100_dpr_shared_baseline.npy index/psgs_w100_id2doc.json logs/08-23-2020/nq_dpr_shared-seed16-bsz256-fp16True-lr2e-05-decay0.0-warm0.1-bert-base-uncased/checkpoint_best.pt --batch-size 1000 --model-name bert-base-uncased --shared-encoder --max-q-len 50 --save-pred nq-test-dpr-shared-b256-res.txt \n\n# shared encoder on merged corpus\npython eval_retrieval.py /private/home/xwhan/data/nq-dpr/nq-test-qas.txt index/merged_all_single_only.npy index/merged_all_id2doc.json logs/08-23-2020/nq_dpr_shared-seed16-bsz256-fp16True-lr2e-05-decay0.0-warm0.1-bert-base-uncased/checkpoint_best.pt --batch-size 1000 --model-name bert-base-uncased --shared-encoder --max-q-len 50\n\n# to get negatives from DPR shared baseline\npython eval_retrieval.py /private/home/xwhan/data/nq-dpr/nq-val-simplified.txt index/psg100_dpr_shared_baseline.npy index/psgs_w100_id2doc.json logs/08-25-2020/wq_mhop_1_shared_dpr_neg_from_scratch-seed16-bsz150-fp16True-lr2e-05-decay0.0-warm0.1-bert-base-uncased/checkpoint_best.pt --batch-size 1000 --model-name bert-base-uncased --shared-encoder --save-pred nq-val-shared-dpr-top100.txt --topk 100 \n\npython eval_retrieval.py /private/home/xwhan/data/WebQ/WebQuestions-test.txt index/psg100_mhop_wq_1_from_baseline.npy index/psgs_w100_id2doc.json logs/08-26-2020/wq_mhop_1_shared_dpr_neg_from_scratch-seed16-bsz150-fp16True-lr2e-05-decay0.0-warm0.1-bert-base-uncased/checkpoint_best.pt --batch-size 1000 --model-name bert-base-uncased --shared-encoder --save-pred wq-test-res-type1.txt\n\n\npython eval_retrieval.py /private/home/xwhan/data/nq-dpr/nq-test-qas.txt index/merged_all.npy index/merged_all_id2doc.json logs/07-24-2020/unified_continue-seed16-bsz150-fp16True-lr1e-05-decay0.0/checkpoint_best.pt --batch-size 1000 --shared-encoder --model-name bert-base-uncased --unified\n\n\n\n\"\"\"\n\nimport numpy as np\nimport json\nimport faiss\nimport argparse\nimport logging\nimport torch\nfrom tqdm import tqdm\n\nfrom multiprocessing import Pool as ProcessPool\nfrom multiprocessing.util import Finalize\nfrom functools import partial\nfrom collections import defaultdict\n\nfrom utils.utils import load_saved, move_to_cuda, para_has_answer\nfrom utils.basic_tokenizer import SimpleTokenizer\n\nfrom transformers import AutoConfig, AutoTokenizer\nfrom models.retriever import BertRetrieverSingle, RobertaRetrieverSingle\nfrom models.unified_retriever import UnifiedRetriever, BertNQRetriever\n\nlogger = logging.getLogger()\nlogger.setLevel(logging.INFO)\nif (logger.hasHandlers()):\n logger.handlers.clear()\nconsole = logging.StreamHandler()\nlogger.addHandler(console)\n\nPROCESS_TOK = None\ndef init():\n global PROCESS_TOK\n PROCESS_TOK = SimpleTokenizer()\n Finalize(PROCESS_TOK, PROCESS_TOK.shutdown, exitpriority=100)\n\ndef get_score(answer_doc, topk=20):\n \"\"\"Search through all the top docs to see if they have the answer.\"\"\"\n question, answer, docs = answer_doc\n top5doc_covered = 0\n global PROCESS_TOK\n topkpara_covered = []\n for p in docs:\n topkpara_covered.append(int(para_has_answer(answer, p[\"title\"] + \" \" + p[\"text\"], PROCESS_TOK)))\n\n return {\n \"5\": int(np.sum(topkpara_covered[:5]) > 0),\n \"10\": int(np.sum(topkpara_covered[:10]) > 0),\n \"20\": int(np.sum(topkpara_covered[:20]) > 0),\n \"50\": int(np.sum(topkpara_covered[:50]) > 0),\n \"100\": int(np.sum(topkpara_covered[:100]) > 0),\n \"covered\": topkpara_covered\n }\n\n\ndef add_marker_q(tokenizer, q):\n q_toks = tokenizer.tokenize(q)\n return ['[unused0]'] + q_toks\n\nif __name__ == '__main__':\n parser = argparse.ArgumentParser()\n parser.add_argument('raw_data', type=str, default=None)\n parser.add_argument('indexpath', type=str, default=None)\n parser.add_argument('corpus_dict', type=str, default=None)\n parser.add_argument('model_path', type=str, default=None)\n parser.add_argument('--batch-size', type=int, default=100)\n parser.add_argument('--topk', type=int, default=100)\n parser.add_argument('--max-q-len', type=int, default=100)\n parser.add_argument('--num-workers', type=int, default=10)\n parser.add_argument('--shared-encoder', action=\"store_true\")\n parser.add_argument('--model-name', type=str, default='bert-base-uncased')\n parser.add_argument(\"--stop-drop\", default=0, type=float)\n parser.add_argument(\"--gpu\", action=\"store_true\")\n parser.add_argument(\"--save-pred\", default=\"\", type=str)\n parser.add_argument(\"--unified\", action=\"store_true\", help=\"test with unified trained model\")\n args = parser.parse_args()\n\n logger.info(f\"Loading questions\")\n qas = [json.loads(line) for line in open(args.raw_data).readlines()]\n questions = [_[\"question\"][:-1] if _[\"question\"].endswith(\"?\") else _[\"question\"] for _ in qas]\n answers = [item[\"answer\"] for item in qas]\n\n logger.info(f\"Loading index\")\n d = 768\n xb = np.load(args.indexpath).astype('float32')\n index = faiss.IndexFlatIP(d)\n index.add(xb)\n\n if args.gpu: \n res = faiss.StandardGpuResources()\n index = faiss.index_cpu_to_gpu(res, 1, index)\n # logger.info(f\"Building GPU index\")\n # co = faiss.GpuMultipleClonerOptions()\n # co.useFloat16 = True\n # co.shards = True\n # index = faiss.index_cpu_to_gpus_list(index, co, [1,2,3,4,5,6,7])\n # index.add(xb)\n\n logger.info(\"Loading trained model...\")\n bert_config = AutoConfig.from_pretrained(args.model_name)\n tokenizer = AutoTokenizer.from_pretrained(args.model_name)\n if args.unified:\n model = UnifiedRetriever(bert_config, args)\n elif \"roberta\" in args.model_name:\n model = RobertaRetrieverSingle(bert_config, args)\n else:\n model = BertRetrieverSingle(bert_config, args)\n \n model = load_saved(model, args.model_path, exact=False)\n cuda = torch.device('cuda')\n model.to(cuda)\n from apex import amp\n model = amp.initialize(model, opt_level='O1')\n model.eval()\n\n logger.info(f\"Loading corpus\")\n id2doc = json.load(open(args.corpus_dict))\n logger.info(f\"Corpus size {len(id2doc)}\")\n\n retrieved_results = []\n retrieved_docids = []\n for b_start in tqdm(range(0, len(questions), args.batch_size)):\n with torch.no_grad():\n batch_q = questions[b_start:b_start + args.batch_size]\n batch_ans = answers[b_start:b_start + args.batch_size]\n\n # test retrieval model with marker\n # batch_q_toks = [add_marker_q(tokenizer, q) for q in batch_q]\n # batch_q_encodes = tokenizer.batch_encode_plus(batch_q_toks, max_length=args.max_q_len, pad_to_max_length=True, return_tensors=\"pt\", is_pretokenized=True)\n\n batch_q_encodes = tokenizer.batch_encode_plus(batch_q, max_length=args.max_q_len, pad_to_max_length=True, return_tensors=\"pt\", is_pretokenized=True) \n\n batch_q_encodes = move_to_cuda(dict(batch_q_encodes))\n q_embeds = model.encode_q(batch_q_encodes[\"input_ids\"], batch_q_encodes[\"attention_mask\"], batch_q_encodes.get(\"token_type_ids\", None))\n q_embeds_numpy = q_embeds.cpu().contiguous().numpy()\n D, I = index.search(q_embeds_numpy, args.topk)\n for b_idx in range(len(batch_q)):\n topk_docs = [{\"title\": id2doc[str(doc_id)][0],\"text\": id2doc[str(doc_id)][1]} for doc_id in I[b_idx]]\n retrieved_results.append(topk_docs)\n retrieved_docids.append([str(doc_id) for doc_id in I[b_idx]])\n\n answers_docs = list(zip(questions, answers, retrieved_results))\n processes = ProcessPool(\n processes=args.num_workers,\n initializer=init\n )\n get_score_partial = partial(\n get_score, topk=args.topk)\n results = processes.map(get_score_partial, answers_docs)\n\n if args.save_pred != \"\":\n to_save = []\n for inputs, metrics, topk_ids in zip(answers_docs, results, retrieved_docids):\n q, ans, topk_doc = inputs\n topk_covered = metrics[\"covered\"]\n assert len(topk_doc) == len(topk_covered)\n assert len(topk_doc) == len(topk_ids)\n to_save.append({\n \"question\": q,\n \"ans\": ans,\n \"topk\": list(zip(topk_doc, topk_covered)),\n \"topkdocs\": topk_doc,\n \"metrics\": metrics,\n \"topk_ids\": topk_ids\n })\n print(f\"Saving {len(to_save)} instances...\")\n with open(\"/private/home/xwhan/data/nq-dpr/results/\" + args.save_pred, \"w\") as g:\n for l in to_save:\n g.write(json.dumps(l) + \"\\n\")\n\n aggregate = defaultdict(list)\n for r in results:\n for k, v in r.items():\n aggregate[k].append(v)\n\n for k in aggregate:\n results = aggregate[k]\n print('Top {} Recall for {} QA pairs: {} ...'.format(\n k, len(results), np.mean(results)))\n" }, { "path": "scripts/eval/eval_single_fever.py", "content": "# Copyright (c) Facebook, Inc. and its affiliates.\n# All rights reserved.\n#\n# This source code is licensed under the license found in the \n# LICENSE file in the root directory of this source tree.\n\"\"\"\npython eval_single_fever.py /private/home/xwhan/data/fever/retrieval/dev_single_evidence.txt index/fever_single.npy index/fever_corpus_id2doc.json logs/08-30-2020/fever_single-seed16-bsz256-fp16True-lr2e-05-decay0.0-warm0-bert-base-uncased/checkpoint_best.pt --batch-size 1000 --shared-encoder --model-name bert-base-uncased --gpu --save-path dense_dev_single_k10.json --topk 10\n\n\npython eval_single_fever.py /private/home/xwhan/data/fever/retrieval/dev_single_evidence.txt index/fever_unified.npy index/fever_corpus_id2doc.json logs/08-30-2020/fever_unified_roberta-seed16-bsz96-fp16True-lr2e-05-decay0.0-adamTrue/checkpoint_best.pt --batch-size 1000 --shared-encoder --model-name roberta-base --gpu --save-path dense_unified_dev_single_k10.json --topk 10\n\n\"\"\"\nimport argparse\nimport json\nimport logging\n\nimport faiss\nimport numpy as np\nimport torch\nfrom tqdm import tqdm\nfrom transformers import AutoConfig, AutoTokenizer\n\nfrom models.retriever import BertRetrieverSingle\nfrom models.unified_retriever import UnifiedRetriever\nfrom utils.basic_tokenizer import SimpleTokenizer\nfrom utils.utils import (load_saved, move_to_cuda)\n\nlogger = logging.getLogger()\nlogger.setLevel(logging.INFO)\nif (logger.hasHandlers()):\n logger.handlers.clear()\nconsole = logging.StreamHandler()\nlogger.addHandler(console)\n\n\nif __name__ == '__main__':\n parser = argparse.ArgumentParser()\n parser.add_argument('raw_data', type=str, default=None)\n parser.add_argument('indexpath', type=str, default=None)\n parser.add_argument('corpus_dict', type=str, default=None)\n parser.add_argument('model_path', type=str, default=None)\n parser.add_argument('--topk', type=int, default=2, help=\"topk paths\")\n parser.add_argument('--num-workers', type=int, default=10)\n parser.add_argument('--max-q-len', type=int, default=45)\n parser.add_argument('--batch-size', type=int, default=100)\n parser.add_argument('--model-name', type=str, default='bert-base-uncased')\n parser.add_argument('--gpu', action=\"store_true\")\n parser.add_argument('--shared-encoder', action=\"store_true\")\n parser.add_argument(\"--save-path\", type=str, default=\"\")\n parser.add_argument(\"--stop-drop\", default=0, type=float)\n args = parser.parse_args()\n \n logger.info(\"Loading data...\")\n ds_items = [json.loads(_) for _ in open(args.raw_data).readlines()]\n\n logger.info(\"Building index...\")\n d = 768\n xb = np.load(args.indexpath).astype('float32')\n print(xb.shape)\n\n index = faiss.IndexFlatIP(d)\n index.add(xb)\n if args.gpu:\n res = faiss.StandardGpuResources()\n index = faiss.index_cpu_to_gpu(res, 1, index)\n\n logger.info(f\"Loading corpus...\")\n id2doc = json.load(open(args.corpus_dict))\n title2doc = {item[0]:item[1] for item in id2doc.values()}\n logger.info(f\"Corpus size {len(id2doc)}\")\n \n logger.info(\"Loading trained model...\")\n bert_config = AutoConfig.from_pretrained(args.model_name)\n tokenizer = AutoTokenizer.from_pretrained(args.model_name)\n # model = BertRetrieverSingle(bert_config, args)\n model = UnifiedRetriever(bert_config, args)\n\n model = load_saved(model, args.model_path, exact=False)\n simple_tokenizer = SimpleTokenizer()\n\n cuda = torch.device('cuda')\n model.to(cuda)\n from apex import amp\n model = amp.initialize(model, opt_level='O1')\n model.eval()\n\n logger.info(\"Encoding claims and searching\")\n questions = [_[\"claim\"] for _ in ds_items]\n metrics = []\n retrieval_outputs = []\n for b_start in tqdm(range(0, len(questions), args.batch_size)):\n with torch.no_grad():\n batch_q = questions[b_start:b_start + args.batch_size]\n batch_ann = ds_items[b_start:b_start + args.batch_size]\n bsize = len(batch_q)\n batch_q_encodes = tokenizer.batch_encode_plus(batch_q, max_length=args.max_q_len, pad_to_max_length=True, return_tensors=\"pt\")\n batch_q_encodes = move_to_cuda(dict(batch_q_encodes))\n q_embeds = model.encode_q(batch_q_encodes[\"input_ids\"], batch_q_encodes[\"attention_mask\"], batch_q_encodes.get(\"token_type_ids\", None))\n q_embeds_numpy = q_embeds.cpu().contiguous().numpy()\n\n D, I = index.search(q_embeds_numpy, args.topk)\n\n for b_idx in range(bsize):\n topk_docs = []\n for _, doc_id in enumerate(I[b_idx]):\n doc = id2doc[str(doc_id)]\n topk_docs.append({\"title\": doc[0], \"text\": doc[1]})\n\n # saving when there's no annotations\n if args.save_path != \"\":\n candidaite_chains = []\n retrieval_outputs.append({\n \"id\": batch_ann[b_idx][\"id\"],\n \"claim\": batch_ann[b_idx][\"claim\"],\n \"topk\": topk_docs,\n })\n\n if args.save_path != \"\":\n with open(f\"/private/home/xwhan/data/fever/retrieval/{args.save_path}\", \"w\") as out:\n for l in retrieval_outputs:\n out.write(json.dumps(l) + \"\\n\")\n\n" }, { "path": "scripts/train_mhop.py", "content": "# Copyright (c) Facebook, Inc. and its affiliates.\n# All rights reserved.\n#\n# This source code is licensed under the license found in the \n# LICENSE file in the root directory of this source tree.\n\n\n\"\"\"\nDescription: train a multi-hop dense retrieval from pretrained BERT/RoBERTa encoder\n\nUsage:\n\nCUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python scripts/train_mhop.py \\\n --do_train \\\n --prefix ${RUN_ID} \\\n --predict_batch_size 3000 \\\n --model_name roberta-base \\\n --train_batch_size 150 \\\n --learning_rate 2e-5 \\\n --fp16 \\\n --train_file ${TRAIN_DATA_PATH} \\\n --predict_file ${DEV_DATA_PATH} \\\n --seed 16 \\\n --eval-period -1 \\\n --max_c_len 300 \\\n --max_q_len 70 \\\n --max_q_sp_len 350 \\\n --shared-encoder \\\n --warmup-ratio 0.1\n\n\"\"\"\nimport logging\nimport os\nimport random\nfrom datetime import date\nfrom functools import partial\n\nimport numpy as np\nimport torch\nfrom torch.optim import Adam\nfrom torch.utils.data import DataLoader\nfrom torch.utils.tensorboard import SummaryWriter\nfrom tqdm import tqdm\nfrom transformers import (AdamW, AutoConfig, AutoTokenizer,\n get_linear_schedule_with_warmup)\n\nfrom mdr.retrieval.config import train_args\nfrom mdr.retrieval.criterions import (mhop_eval, mhop_loss)\nfrom mdr.retrieval.data.mhop_dataset import MhopDataset, mhop_collate\nfrom mdr.retrieval.models.mhop_retriever import RobertaRetriever\nfrom mdr.retrieval.utils.utils import AverageMeter, move_to_cuda, load_saved\n\n\ndef main():\n args = train_args()\n if args.fp16:\n import apex\n apex.amp.register_half_function(torch, 'einsum')\n date_curr = date.today().strftime(\"%m-%d-%Y\")\n model_name = f\"{args.prefix}-seed{args.seed}-bsz{args.train_batch_size}-fp16{args.fp16}-lr{args.learning_rate}-decay{args.weight_decay}-warm{args.warmup_ratio}-valbsz{args.predict_batch_size}-shared{args.shared_encoder}-multi{args.multi_vector}-scheme{args.scheme}\"\n args.output_dir = os.path.join(args.output_dir, date_curr, model_name)\n tb_logger = SummaryWriter(os.path.join(args.output_dir.replace(\"logs\",\"tflogs\")))\n\n if os.path.exists(args.output_dir) and os.listdir(args.output_dir):\n print(\n f\"output directory {args.output_dir} already exists and is not empty.\")\n if not os.path.exists(args.output_dir):\n os.makedirs(args.output_dir, exist_ok=True)\n\n logging.basicConfig(format='%(asctime)s - %(levelname)s - %(name)s - %(message)s', datefmt='%m/%d/%Y %H:%M:%S',\n level=logging.INFO,\n handlers=[logging.FileHandler(os.path.join(args.output_dir, \"log.txt\")),\n logging.StreamHandler()])\n logger = logging.getLogger(__name__)\n logger.info(args)\n\n if args.local_rank == -1 or args.no_cuda:\n device = torch.device(\n \"cuda\" if torch.cuda.is_available() and not args.no_cuda else \"cpu\")\n n_gpu = torch.cuda.device_count()\n else:\n device = torch.device(\"cuda\", args.local_rank)\n n_gpu = 1\n torch.distributed.init_process_group(backend='nccl')\n logger.info(\"device %s n_gpu %d distributed training %r\",\n device, n_gpu, bool(args.local_rank != -1))\n\n if args.accumulate_gradients < 1:\n raise ValueError(\"Invalid accumulate_gradients parameter: {}, should be >= 1\".format(\n args.accumulate_gradients))\n\n args.train_batch_size = int(\n args.train_batch_size / args.accumulate_gradients)\n random.seed(args.seed)\n np.random.seed(args.seed)\n torch.manual_seed(args.seed)\n if n_gpu > 0:\n torch.cuda.manual_seed_all(args.seed)\n\n bert_config = AutoConfig.from_pretrained(args.model_name)\n\n model = RobertaRetriever(bert_config, args)\n\n tokenizer = AutoTokenizer.from_pretrained(args.model_name)\n collate_fc = partial(mhop_collate, pad_id=tokenizer.pad_token_id)\n if args.do_train and args.max_c_len > bert_config.max_position_embeddings:\n raise ValueError(\n \"Cannot use sequence length %d because the BERT model \"\n \"was only trained up to sequence length %d\" %\n (args.max_c_len, bert_config.max_position_embeddings))\n\n eval_dataset = MhopDataset(\n tokenizer, args.predict_file, args.max_q_len, args.max_q_sp_len, args.max_c_len)\n eval_dataloader = DataLoader(\n eval_dataset, batch_size=args.predict_batch_size, collate_fn=collate_fc, pin_memory=True, num_workers=args.num_workers)\n logger.info(f\"Num of dev batches: {len(eval_dataloader)}\")\n\n if args.init_checkpoint != \"\":\n model = load_saved(model, args.init_checkpoint)\n\n model.to(device)\n print(f\"number of trainable parameters: {sum(p.numel() for p in model.parameters() if p.requires_grad)}\")\n\n if args.do_train:\n no_decay = ['bias', 'LayerNorm.weight']\n optimizer_parameters = [\n {'params': [p for n, p in model.named_parameters() if not any(\n nd in n for nd in no_decay)], 'weight_decay': args.weight_decay},\n {'params': [p for n, p in model.named_parameters() if any(\n nd in n for nd in no_decay)], 'weight_decay': 0.0}\n ]\n optimizer = Adam(optimizer_parameters,\n lr=args.learning_rate, eps=args.adam_epsilon)\n\n if args.fp16:\n from apex import amp\n model, optimizer = amp.initialize(\n model, optimizer, opt_level=args.fp16_opt_level)\n else:\n if args.fp16:\n from apex import amp\n model = amp.initialize(model, opt_level=args.fp16_opt_level)\n\n if args.local_rank != -1:\n model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[args.local_rank],\n output_device=args.local_rank)\n elif n_gpu > 1:\n model = torch.nn.DataParallel(model)\n\n if args.do_train:\n global_step = 0 # gradient update step\n batch_step = 0 # forward batch count\n best_mrr = 0\n train_loss_meter = AverageMeter()\n model.train()\n train_dataset = MhopDataset(tokenizer, args.train_file, args.max_q_len, args.max_q_sp_len, args.max_c_len, train=True)\n train_dataloader = DataLoader(train_dataset, batch_size=args.train_batch_size, pin_memory=True, collate_fn=collate_fc, num_workers=args.num_workers, shuffle=True)\n\n t_total = len(train_dataloader) // args.gradient_accumulation_steps * args.num_train_epochs\n warmup_steps = t_total * args.warmup_ratio\n scheduler = get_linear_schedule_with_warmup(\n optimizer, num_warmup_steps=warmup_steps, num_training_steps=t_total\n )\n\n logger.info('Start training....')\n for epoch in range(int(args.num_train_epochs)):\n for batch in tqdm(train_dataloader):\n batch_step += 1\n batch = move_to_cuda(batch)\n loss = mhop_loss(model, batch, args)\n if args.gradient_accumulation_steps > 1:\n loss = loss / args.gradient_accumulation_steps\n\n if args.fp16:\n with amp.scale_loss(loss, optimizer) as scaled_loss:\n scaled_loss.backward()\n else:\n loss.backward()\n train_loss_meter.update(loss.item())\n \n if (batch_step + 1) % args.gradient_accumulation_steps == 0:\n if args.fp16:\n torch.nn.utils.clip_grad_norm_(\n amp.master_params(optimizer), args.max_grad_norm)\n else:\n torch.nn.utils.clip_grad_norm_(\n model.parameters(), args.max_grad_norm)\n optimizer.step()\n scheduler.step()\n model.zero_grad()\n global_step += 1\n\n tb_logger.add_scalar('batch_train_loss',\n loss.item(), global_step)\n tb_logger.add_scalar('smoothed_train_loss',\n train_loss_meter.avg, global_step)\n\n if args.eval_period != -1 and global_step % args.eval_period == 0:\n mrrs = predict(args, model, eval_dataloader,\n device, logger)\n mrr = mrrs[\"mrr_avg\"]\n logger.info(\"Step %d Train loss %.2f MRR %.2f on epoch=%d\" % (global_step, train_loss_meter.avg, mrr*100, epoch))\n\n if best_mrr < mrr:\n logger.info(\"Saving model with best MRR %.2f -> MRR %.2f on epoch=%d\" %\n (best_mrr*100, mrr*100, epoch))\n torch.save(model.state_dict(), os.path.join(\n args.output_dir, f\"checkpoint_best.pt\"))\n model = model.to(device)\n best_mrr = mrr\n\n mrrs = predict(args, model, eval_dataloader, device, logger)\n mrr = mrrs[\"mrr_avg\"]\n logger.info(\"Step %d Train loss %.2f MRR-AVG %.2f on epoch=%d\" % (\n global_step, train_loss_meter.avg, mrr*100, epoch))\n for k, v in mrrs.items():\n tb_logger.add_scalar(k, v*100, epoch)\n torch.save(model.state_dict(), os.path.join(\n args.output_dir, f\"checkpoint_last.pt\"))\n\n if best_mrr < mrr:\n logger.info(\"Saving model with best MRR %.2f -> MRR %.2f on epoch=%d\" % (best_mrr*100, mrr*100, epoch))\n torch.save(model.state_dict(), os.path.join(\n args.output_dir, f\"checkpoint_best.pt\"))\n best_mrr = mrr\n\n logger.info(\"Training finished!\")\n\n elif args.do_predict:\n acc = predict(args, model, eval_dataloader, device, logger)\n logger.info(f\"test performance {acc}\")\n\ndef predict(args, model, eval_dataloader, device, logger):\n model.eval()\n rrs_1, rrs_2 = [], [] # reciprocal rank\n for batch in tqdm(eval_dataloader):\n batch_to_feed = move_to_cuda(batch)\n with torch.no_grad():\n outputs = model(batch_to_feed)\n eval_results = mhop_eval(outputs, args)\n _rrs_1, _rrs_2 = eval_results[\"rrs_1\"], eval_results[\"rrs_2\"]\n rrs_1 += _rrs_1\n rrs_2 += _rrs_2\n mrr_1 = np.mean(rrs_1)\n mrr_2 = np.mean(rrs_2)\n logger.info(f\"evaluated {len(rrs_1)} examples...\")\n logger.info(f'MRR-1: {mrr_1}')\n logger.info(f'MRR-2: {mrr_2}')\n model.train()\n return {\"mrr_1\": mrr_1, \"mrr_2\": mrr_2, \"mrr_avg\": (mrr_1 + mrr_2) / 2}\n\n\nif __name__ == \"__main__\":\n main()\n" }, { "path": "scripts/train_momentum.py", "content": "# Copyright (c) Facebook, Inc. and its affiliates.\n# All rights reserved.\n#\n# This source code is licensed under the license found in the \n# LICENSE file in the root directory of this source tree.\nimport logging\nimport os\nimport random\nfrom datetime import date\nfrom functools import partial\n\nimport numpy as np\nimport torch\nfrom torch.optim import Adam\nfrom torch.utils.data import DataLoader\nfrom torch.utils.tensorboard import SummaryWriter\nfrom tqdm import tqdm\nfrom transformers import (AdamW, AutoConfig, AutoTokenizer,\n get_linear_schedule_with_warmup)\n\nfrom mdr.retrieval.config import train_args\nfrom mdr.retrieval.criterions import (mhop_eval, mhop_loss)\nfrom mdr.retrieval.data.mhop_dataset import MhopDataset, mhop_collate\nfrom mdr.retrieval.models.mhop_retriever import RobertaMomentumRetriever\nfrom mdr.retrieval.utils.utils import AverageMeter, move_to_cuda\nfrom mdr.retrieval.data.fever_dataset import FeverDataset\n\ndef main():\n args = train_args()\n if args.fp16:\n import apex\n apex.amp.register_half_function(torch, 'einsum')\n date_curr = date.today().strftime(\"%m-%d-%Y\")\n model_name = f\"{args.prefix}-seed{args.seed}-bsz{args.train_batch_size}-fp16{args.fp16}-lr{args.learning_rate}-decay{args.weight_decay}-warm{args.warmup_ratio}-valbsz{args.predict_batch_size}-m{args.m}-k{args.k}-t{args.temperature}\"\n args.output_dir = os.path.join(args.output_dir, date_curr, model_name)\n tb_logger = SummaryWriter(os.path.join(args.output_dir.replace(\"logs\",\"tflogs\")))\n\n if os.path.exists(args.output_dir) and os.listdir(args.output_dir):\n print(\n f\"output directory {args.output_dir} already exists and is not empty.\")\n if not os.path.exists(args.output_dir):\n os.makedirs(args.output_dir, exist_ok=True)\n\n logging.basicConfig(format='%(asctime)s - %(levelname)s - %(name)s - %(message)s', datefmt='%m/%d/%Y %H:%M:%S',\n level=logging.INFO,\n handlers=[logging.FileHandler(os.path.join(args.output_dir, \"log.txt\")),\n logging.StreamHandler()])\n logger = logging.getLogger(__name__)\n logger.info(args)\n\n if args.local_rank == -1 or args.no_cuda:\n device = torch.device(\n \"cuda\" if torch.cuda.is_available() and not args.no_cuda else \"cpu\")\n n_gpu = torch.cuda.device_count()\n else:\n device = torch.device(\"cuda\", args.local_rank)\n n_gpu = 1\n torch.distributed.init_process_group(backend='nccl')\n logger.info(\"device %s n_gpu %d distributed training %r\",\n device, n_gpu, bool(args.local_rank != -1))\n\n if args.accumulate_gradients < 1:\n raise ValueError(\"Invalid accumulate_gradients parameter: {}, should be >= 1\".format(\n args.accumulate_gradients))\n\n args.train_batch_size = int(\n args.train_batch_size / args.accumulate_gradients)\n random.seed(args.seed)\n np.random.seed(args.seed)\n torch.manual_seed(args.seed)\n if n_gpu > 0:\n torch.cuda.manual_seed_all(args.seed)\n\n bert_config = AutoConfig.from_pretrained(args.model_name)\n model = RobertaMomentumRetriever(bert_config, args)\n\n tokenizer = AutoTokenizer.from_pretrained(args.model_name)\n collate_fc = partial(mhop_collate, pad_id=tokenizer.pad_token_id)\n if args.do_train and args.max_c_len > bert_config.max_position_embeddings:\n raise ValueError(\n \"Cannot use sequence length %d because the BERT model \"\n \"was only trained up to sequence length %d\" %\n (args.max_c_len, bert_config.max_position_embeddings))\n\n\n if \"fever\" in args.predict_file:\n eval_dataset = FeverDataset(\n tokenizer, args.predict_file, args.max_q_len, args.max_q_sp_len, args.max_c_len)\n else:\n eval_dataset = MhopDataset(\n tokenizer, args.predict_file, args.max_q_len, args.max_q_sp_len, args.max_c_len)\n eval_dataloader = DataLoader(\n eval_dataset, batch_size=args.predict_batch_size, collate_fn=collate_fc, pin_memory=True, num_workers=args.num_workers)\n logger.info(f\"Num of dev batches: {len(eval_dataloader)}\")\n\n model.to(device)\n print(f\"number of trainable parameters: {sum(p.numel() for p in model.parameters() if p.requires_grad)}\")\n\n if args.do_train:\n no_decay = ['bias', 'LayerNorm.weight']\n optimizer_parameters = [\n {'params': [p for n, p in model.named_parameters() if not any(\n nd in n for nd in no_decay)], 'weight_decay': args.weight_decay},\n {'params': [p for n, p in model.named_parameters() if any(\n nd in n for nd in no_decay)], 'weight_decay': 0.0}\n ]\n optimizer = Adam(optimizer_parameters,\n lr=args.learning_rate, eps=args.adam_epsilon)\n\n if args.fp16:\n from apex import amp\n model, optimizer = amp.initialize(\n model, optimizer, opt_level=args.fp16_opt_level)\n else:\n if args.fp16:\n from apex import amp\n model = amp.initialize(model, opt_level=args.fp16_opt_level)\n\n if args.local_rank != -1:\n model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[args.local_rank],\n output_device=args.local_rank)\n elif n_gpu > 1:\n model = torch.nn.DataParallel(model)\n\n if args.do_train:\n global_step = 0 # gradient update step\n batch_step = 0 # forward batch count\n best_mrr = 0\n train_loss_meter = AverageMeter()\n model.train()\n\n if \"fever\" in args.train_file:\n train_dataset = FeverDataset(tokenizer, args.train_file, args.max_q_len, args.max_q_sp_len, args.max_c_len, train=True)\n else:\n train_dataset = MhopDataset(tokenizer, args.train_file, args.max_q_len, args.max_q_sp_len, args.max_c_len, train=True)\n train_dataloader = DataLoader(train_dataset, batch_size=args.train_batch_size, pin_memory=True, collate_fn=collate_fc, num_workers=args.num_workers, shuffle=True)\n\n t_total = len(train_dataloader) // args.gradient_accumulation_steps * args.num_train_epochs\n warmup_steps = t_total * args.warmup_ratio\n scheduler = get_linear_schedule_with_warmup(\n optimizer, num_warmup_steps=warmup_steps, num_training_steps=t_total\n )\n\n logger.info('Start training....')\n for epoch in range(int(args.num_train_epochs)):\n for batch in tqdm(train_dataloader):\n batch_step += 1\n batch = move_to_cuda(batch)\n loss = mhop_loss(model, batch, args)\n if args.gradient_accumulation_steps > 1:\n loss = loss / args.gradient_accumulation_steps\n\n if args.fp16:\n with amp.scale_loss(loss, optimizer) as scaled_loss:\n scaled_loss.backward()\n else:\n loss.backward()\n train_loss_meter.update(loss.item())\n \n if (batch_step + 1) % args.gradient_accumulation_steps == 0:\n if args.fp16:\n torch.nn.utils.clip_grad_norm_(\n amp.master_params(optimizer), args.max_grad_norm)\n else:\n torch.nn.utils.clip_grad_norm_(\n model.parameters(), args.max_grad_norm)\n optimizer.step()\n scheduler.step()\n model.zero_grad()\n global_step += 1\n\n tb_logger.add_scalar('batch_train_loss',\n loss.item(), global_step)\n tb_logger.add_scalar('smoothed_train_loss',\n train_loss_meter.avg, global_step)\n\n if args.eval_period != -1 and global_step % args.eval_period == 0:\n mrrs = predict(args, model, eval_dataloader,\n device, logger)\n mrr = mrrs[\"mrr_avg\"]\n logger.info(\"Step %d Train loss %.2f MRR %.2f on epoch=%d\" % (global_step, train_loss_meter.avg, mrr*100, epoch))\n\n if best_mrr < mrr:\n logger.info(\"Saving model with best MRR %.2f -> MRR %.2f on epoch=%d\" %\n (best_mrr*100, mrr*100, epoch))\n torch.save(model.module.encoder_q.state_dict(), os.path.join(\n args.output_dir, f\"checkpoint_q_best.pt\"))\n torch.save(model.module.encoder_q.state_dict(), os.path.join(\n args.output_dir, f\"checkpoint_k_best.pt\"))\n model = model.to(device)\n best_mrr = mrr\n\n mrrs = predict(args, model, eval_dataloader, device, logger)\n mrr = mrrs[\"mrr_avg\"]\n logger.info(\"Step %d Train loss %.2f MRR-AVG %.2f on epoch=%d\" % (\n global_step, train_loss_meter.avg, mrr*100, epoch))\n for k, v in mrrs.items():\n tb_logger.add_scalar(k, v*100, epoch)\n\n if best_mrr < mrr:\n logger.info(\"Saving model with best MRR %.2f -> MRR %.2f on epoch=%d\" % (best_mrr*100, mrr*100, epoch))\n torch.save(model.module.encoder_q.state_dict(), os.path.join(\n args.output_dir, f\"checkpoint_q_best.pt\"))\n torch.save(model.module.encoder_q.state_dict(), os.path.join(\n args.output_dir, f\"checkpoint_k_best.pt\"))\n best_mrr = mrr\n\n logger.info(\"Training finished!\")\n\n elif args.do_predict:\n acc = predict(args, model, eval_dataloader, device, logger)\n logger.info(f\"test performance {acc}\")\n\ndef predict(args, model, eval_dataloader, device, logger):\n model.eval()\n rrs_1, rrs_2 = [], [] # reciprocal rank\n for batch in tqdm(eval_dataloader):\n batch_to_feed = move_to_cuda(batch)\n with torch.no_grad():\n outputs = model(batch_to_feed)\n eval_results = mhop_eval(outputs, args)\n _rrs_1, _rrs_2 = eval_results[\"rrs_1\"], eval_results[\"rrs_2\"]\n rrs_1 += _rrs_1\n rrs_2 += _rrs_2\n mrr_1 = np.mean(rrs_1)\n mrr_2 = np.mean(rrs_2)\n logger.info(f\"evaluated {len(rrs_1)} examples...\")\n logger.info(f'MRR-1: {mrr_1}')\n logger.info(f'MRR-2: {mrr_2}')\n model.train()\n return {\"mrr_1\": mrr_1, \"mrr_2\": mrr_2, \"mrr_avg\": (mrr_1 + mrr_2) / 2}\n\n\nif __name__ == \"__main__\":\n main()\n" }, { "path": "scripts/train_qa.py", "content": "# Copyright (c) Facebook, Inc. and its affiliates.\n# All rights reserved.\n#\n# This source code is licensed under the license found in the \n# LICENSE file in the root directory of this source tree.\nimport collections\nimport json\nimport logging\nimport os\nimport random\nfrom datetime import date\nfrom functools import partial\n\nimport numpy as np\nfrom numpy.core.defchararray import encode\nimport torch\nfrom torch import sparse_coo_tensor\nimport torch.nn.functional as F\nfrom torch.utils.data import DataLoader\nfrom torch.utils.tensorboard import SummaryWriter\nfrom torch.optim import Adam\nfrom tqdm import tqdm\nfrom transformers import (AdamW, AutoConfig, AutoTokenizer,\n get_linear_schedule_with_warmup)\n\nfrom mdr.qa.config import train_args\nfrom mdr.qa.qa_dataset import QADataset, qa_collate, MhopSampler\nfrom mdr.qa.qa_model import QAModel\nfrom mdr.qa.utils import AverageMeter, move_to_cuda, get_final_text\n\nfrom mdr.qa.hotpot_evaluate_v1 import f1_score, exact_match_score, update_sp\n\ndef load_saved(model, path):\n state_dict = torch.load(path)\n def filter(x): return x[7:] if x.startswith('module.') else x\n state_dict = {filter(k): v for (k, v) in state_dict.items()}\n model.load_state_dict(state_dict)\n return model\n\ndef main():\n args = train_args()\n if args.fp16:\n import apex\n apex.amp.register_half_function(torch, 'einsum')\n date_curr = date.today().strftime(\"%m-%d-%Y\")\n model_name = f\"{args.prefix}-seed{args.seed}-bsz{args.train_batch_size}-fp16{args.fp16}-lr{args.learning_rate}-decay{args.weight_decay}-neg{args.neg_num}-sn{args.shared_norm}-adam{args.use_adam}-warm{args.warmup_ratio}-sp{args.sp_weight}\"\n args.output_dir = os.path.join(args.output_dir, date_curr, model_name)\n tb_logger = SummaryWriter(os.path.join(args.output_dir.replace(\"logs\",\"tflogs\")))\n\n if os.path.exists(args.output_dir) and os.listdir(args.output_dir):\n print(\n f\"output directory {args.output_dir} already exists and is not empty.\")\n if not os.path.exists(args.output_dir):\n os.makedirs(args.output_dir, exist_ok=True)\n\n logging.basicConfig(format='%(asctime)s - %(levelname)s - %(name)s - %(message)s', datefmt='%m/%d/%Y %H:%M:%S', level=logging.INFO,\n handlers=[logging.FileHandler(os.path.join(args.output_dir, \"log.txt\")),\n logging.StreamHandler()])\n logger = logging.getLogger(__name__)\n logger.setLevel(logging.INFO)\n logger.info(args)\n\n if args.local_rank == -1 or args.no_cuda:\n device = torch.device(\n \"cuda\" if torch.cuda.is_available() and not args.no_cuda else \"cpu\")\n n_gpu = torch.cuda.device_count()\n else:\n device = torch.device(\"cuda\", args.local_rank)\n n_gpu = 1\n # Initializes the distributed backend which will take care of sychronizing nodes/GPUs\n torch.distributed.init_process_group(backend='nccl')\n logger.info(\"device %s n_gpu %d distributed training %r\",\n device, n_gpu, bool(args.local_rank != -1))\n\n if args.shared_norm:\n # chains of each question are on the same gpu\n assert (args.train_batch_size // n_gpu) == args.neg_num + 1\n random.seed(args.seed)\n np.random.seed(args.seed)\n torch.manual_seed(args.seed)\n if n_gpu > 0:\n torch.cuda.manual_seed_all(args.seed)\n\n # define model\n if args.model_name == \"spanbert\":\n bert_config = AutoConfig.from_pretrained(\"/private/home/span-bert\")\n tokenizer = AutoTokenizer.from_pretrained('bert-large-cased')\n else:\n bert_config = AutoConfig.from_pretrained(args.model_name)\n tokenizer = AutoTokenizer.from_pretrained(args.model_name)\n model = QAModel(bert_config, args)\n\n collate_fc = partial(qa_collate, pad_id=tokenizer.pad_token_id)\n eval_dataset = QADataset(tokenizer, args.predict_file, args.max_seq_len, args.max_q_len)\n eval_dataloader = DataLoader(eval_dataset, batch_size=args.predict_batch_size, collate_fn=collate_fc, pin_memory=True, num_workers=args.num_workers)\n logger.info(f\"Num of dev batches: {len(eval_dataloader)}\")\n\n if args.init_checkpoint != \"\":\n logger.info(f\"Loading model from {args.init_checkpoint}\")\n model = load_saved(model, args.init_checkpoint)\n\n model.to(device)\n print(f\"number of trainable parameters: {sum(p.numel() for p in model.parameters() if p.requires_grad)}\")\n\n if args.do_train:\n no_decay = ['bias', 'LayerNorm.weight']\n\n optimizer_parameters = [\n {'params': [p for n, p in model.named_parameters() if not any(\n nd in n for nd in no_decay)], 'weight_decay': args.weight_decay},\n {'params': [p for n, p in model.named_parameters() if any(\n nd in n for nd in no_decay)], 'weight_decay': 0.0}\n ]\n\n if args.use_adam:\n optimizer = Adam(optimizer_parameters,\n lr=args.learning_rate, eps=args.adam_epsilon)\n else:\n optimizer = AdamW(optimizer_parameters,\n lr=args.learning_rate, eps=args.adam_epsilon)\n\n if args.fp16:\n from apex import amp\n model, optimizer = amp.initialize(\n model, optimizer, opt_level=args.fp16_opt_level)\n else:\n if args.fp16:\n from apex import amp\n model = amp.initialize(model, opt_level=args.fp16_opt_level)\n\n\n if args.local_rank != -1:\n model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[args.local_rank],\n output_device=args.local_rank)\n elif n_gpu > 1:\n model = torch.nn.DataParallel(model)\n\n if args.do_train:\n global_step = 0 # gradient update step\n batch_step = 0 # forward batch count\n best_em = 0\n train_loss_meter = AverageMeter()\n model.train()\n train_dataset = QADataset(tokenizer, args.train_file, args.max_seq_len, args.max_q_len, train=True)\n train_sampler = MhopSampler(train_dataset, num_neg=args.neg_num)\n train_dataloader = DataLoader(train_dataset, batch_size=args.train_batch_size, pin_memory=True, collate_fn=collate_fc, num_workers=args.num_workers, sampler=train_sampler)\n\n t_total = len(train_dataloader) // args.gradient_accumulation_steps * args.num_train_epochs\n warmup_steps = t_total * args.warmup_ratio\n scheduler = get_linear_schedule_with_warmup(\n optimizer, num_warmup_steps=warmup_steps, num_training_steps=t_total\n )\n\n logger.info('Start training....')\n for epoch in range(int(args.num_train_epochs)):\n for batch in tqdm(train_dataloader):\n batch_step += 1\n batch_inputs = move_to_cuda(batch[\"net_inputs\"])\n loss = model(batch_inputs)\n if n_gpu > 1:\n loss = loss.mean()\n if args.gradient_accumulation_steps > 1:\n loss = loss / args.gradient_accumulation_steps\n if args.fp16:\n with amp.scale_loss(loss, optimizer) as scaled_loss:\n scaled_loss.backward()\n else:\n loss.backward()\n train_loss_meter.update(loss.item())\n if (batch_step + 1) % args.gradient_accumulation_steps == 0:\n if args.fp16:\n torch.nn.utils.clip_grad_norm_(\n amp.master_params(optimizer), args.max_grad_norm)\n else:\n torch.nn.utils.clip_grad_norm_(\n model.parameters(), args.max_grad_norm)\n optimizer.step()\n scheduler.step()\n model.zero_grad()\n global_step += 1\n\n # logger.info(f\"current batch loss: {loss.item()}\")\n tb_logger.add_scalar('batch_train_loss',\n loss.item(), global_step)\n tb_logger.add_scalar('smoothed_train_loss',\n train_loss_meter.avg, global_step)\n\n if args.eval_period != -1 and global_step % args.eval_period == 0:\n metrics = predict(args, model, eval_dataloader, logger)\n em = metrics[\"em\"]\n logger.info(\"Step %d Train loss %.2f em %.2f on epoch=%d\" % (global_step, train_loss_meter.avg, em*100, epoch))\n if best_em < em:\n logger.info(\"Saving model with best em %.2f -> em %.2f on step=%d\" %\n (best_em*100, em*100, global_step))\n torch.save(model.state_dict(), os.path.join(\n args.output_dir, f\"checkpoint_best.pt\"))\n model = model.to(device)\n best_em = em\n\n metrics = predict(args, model, eval_dataloader, logger)\n em = metrics[\"em\"]\n logger.info(\"Step %d Train loss %.2f em %.2f\" % (\n global_step, train_loss_meter.avg, em*100))\n tb_logger.add_scalar('dev_em', em*100, global_step)\n if best_em < em:\n logger.info(\"Saving model with best em %.2f -> em %.2f on epoch=%d\" % (best_em*100, em*100, epoch))\n torch.save(model.state_dict(), os.path.join(\n args.output_dir, f\"checkpoint_best.pt\"))\n best_em = em\n\n logger.info(\"Training finished!\")\n\n elif args.do_predict:\n metrics = predict(args, model, eval_dataloader, logger, fixed_thresh=0.8)\n logger.info(f\"test performance {metrics}\")\n\n elif args.do_test:\n eval_final(args, model, eval_dataloader, weight=0.8)\n\ndef predict(args, model, eval_dataloader, logger, fixed_thresh=None):\n model.eval()\n id2result = collections.defaultdict(list)\n id2answer = collections.defaultdict(list)\n id2gold = {}\n id2goldsp = {}\n for batch in tqdm(eval_dataloader):\n batch_to_feed = move_to_cuda(batch[\"net_inputs\"])\n batch_qids = batch[\"qids\"]\n batch_labels = batch[\"net_inputs\"][\"label\"].view(-1).tolist()\n with torch.no_grad():\n outputs = model(batch_to_feed)\n scores = outputs[\"rank_score\"]\n scores = scores.view(-1).tolist()\n if args.sp_pred:\n sp_scores = outputs[\"sp_score\"]\n sp_scores = sp_scores.float().masked_fill(batch_to_feed[\"sent_offsets\"].eq(0), float(\"-inf\")).type_as(sp_scores)\n batch_sp_scores = sp_scores.sigmoid()\n\n # ans_type_predicted = torch.argmax(outputs[\"ans_type_logits\"], dim=1).view(-1).tolist()\n outs = [outputs[\"start_logits\"], outputs[\"end_logits\"]]\n for qid, label, score in zip(batch_qids, batch_labels, scores):\n id2result[qid].append((label, score))\n\n # answer prediction\n span_scores = outs[0][:, :, None] + outs[1][:, None]\n max_seq_len = span_scores.size(1)\n span_mask = np.tril(np.triu(np.ones((max_seq_len, max_seq_len)), 0), args.max_ans_len)\n span_mask = span_scores.data.new(max_seq_len, max_seq_len).copy_(torch.from_numpy(span_mask))\n span_scores_masked = span_scores.float().masked_fill((1 - span_mask[None].expand_as(span_scores)).bool(), -1e10).type_as(span_scores)\n start_position = span_scores_masked.max(dim=2)[0].max(dim=1)[1]\n end_position = span_scores_masked.max(dim=2)[1].gather(\n 1, start_position.unsqueeze(1)).squeeze(1)\n answer_scores = span_scores_masked.max(dim=2)[0].max(dim=1)[0].tolist()\n para_offset = batch['para_offsets']\n start_position_ = list(\n np.array(start_position.tolist()) - np.array(para_offset))\n end_position_ = list(\n np.array(end_position.tolist()) - np.array(para_offset)) \n\n for idx, qid in enumerate(batch_qids):\n id2gold[qid] = batch[\"gold_answer\"][idx]\n id2goldsp[qid] = batch[\"sp_gold\"][idx]\n\n rank_score = scores[idx]\n start = start_position_[idx]\n end = end_position_[idx]\n span_score = answer_scores[idx]\n \n tok_to_orig_index = batch['tok_to_orig_index'][idx]\n doc_tokens = batch['doc_tokens'][idx]\n wp_tokens = batch['wp_tokens'][idx]\n orig_doc_start = tok_to_orig_index[start]\n orig_doc_end = tok_to_orig_index[end]\n orig_tokens = doc_tokens[orig_doc_start:(orig_doc_end + 1)]\n tok_tokens = wp_tokens[start:end+1]\n tok_text = \" \".join(tok_tokens)\n tok_text = tok_text.replace(\" ##\", \"\")\n tok_text = tok_text.replace(\"##\", \"\")\n tok_text = tok_text.strip()\n tok_text = \" \".join(tok_text.split())\n orig_text = \" \".join(orig_tokens)\n pred_str = get_final_text(tok_text, orig_text, do_lower_case=True, verbose_logging=False)\n\n # get the sp sentences\n pred_sp = []\n if args.sp_pred:\n sp_score = batch_sp_scores[idx].tolist()\n passages = batch[\"passages\"][idx]\n for passage, sent_offset in zip(passages, [0, len(passages[0][\"sents\"])]):\n for idx, _ in enumerate(passage[\"sents\"]):\n try:\n if sp_score[idx + sent_offset] >= 0.5:\n pred_sp.append([passage[\"title\"], idx])\n except:\n # logger.info(f\"sentence exceeds max lengths\")\n continue\n id2answer[qid].append({\n \"pred_str\": pred_str.strip(),\n \"rank_score\": rank_score,\n \"span_score\": span_score,\n \"pred_sp\": pred_sp\n })\n acc = []\n for qid, res in id2result.items():\n res.sort(key=lambda x: x[1], reverse=True)\n acc.append(res[0][0] == 1)\n logger.info(f\"evaluated {len(id2result)} questions...\")\n logger.info(f'chain ranking em: {np.mean(acc)}')\n\n best_em, best_f1, best_joint_em, best_joint_f1, best_sp_em, best_sp_f1 = 0, 0, 0, 0, 0, 0\n best_res = None\n if fixed_thresh:\n lambdas = [fixed_thresh]\n else:\n # selecting threshhold on the dev data\n lambdas = [0, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1]\n\n for lambda_ in lambdas:\n ems, f1s, sp_ems, sp_f1s, joint_ems, joint_f1s = [], [], [], [], [], []\n results = collections.defaultdict(dict)\n for qid, res in id2result.items():\n ans_res = id2answer[qid]\n ans_res.sort(key=lambda x: lambda_ * x[\"rank_score\"] + (1 - lambda_) * x[\"span_score\"], reverse=True)\n top_pred = ans_res[0][\"pred_str\"]\n top_pred_sp = ans_res[0][\"pred_sp\"]\n\n results[\"answer\"][qid] = top_pred\n results[\"sp\"][qid] = top_pred_sp\n\n ems.append(exact_match_score(top_pred, id2gold[qid][0]))\n f1, prec, recall = f1_score(top_pred, id2gold[qid][0])\n f1s.append(f1)\n\n if args.sp_pred:\n metrics = {'sp_em': 0, 'sp_f1': 0, 'sp_prec': 0, 'sp_recall': 0}\n update_sp(metrics, top_pred_sp, id2goldsp[qid])\n sp_ems.append(metrics['sp_em'])\n sp_f1s.append(metrics['sp_f1'])\n # joint metrics\n joint_prec = prec * metrics[\"sp_prec\"]\n joint_recall = recall * metrics[\"sp_recall\"]\n if joint_prec + joint_recall > 0:\n joint_f1 = 2 * joint_prec * joint_recall / (joint_prec + joint_recall)\n else:\n joint_f1 = 0.\n joint_em = ems[-1] * sp_ems[-1]\n joint_ems.append(joint_em)\n joint_f1s.append(joint_f1)\n\n if args.sp_pred:\n if best_joint_f1 < np.mean(joint_f1s):\n best_joint_f1 = np.mean(joint_f1s)\n best_joint_em = np.mean(joint_ems)\n best_sp_f1 = np.mean(sp_f1s)\n best_sp_em = np.mean(sp_ems)\n best_f1 = np.mean(f1s)\n best_em = np.mean(ems)\n best_res = results\n else:\n if best_f1 < np.mean(f1s):\n best_f1 = np.mean(f1s)\n best_em = np.mean(ems)\n\n logger.info(f\".......Using combination factor {lambda_}......\")\n logger.info(f'answer em: {np.mean(ems)}, count: {len(ems)}')\n logger.info(f'answer f1: {np.mean(f1s)}, count: {len(f1s)}')\n logger.info(f'sp em: {np.mean(sp_ems)}, count: {len(sp_ems)}')\n logger.info(f'sp f1: {np.mean(sp_f1s)}, count: {len(sp_f1s)}')\n logger.info(f'joint em: {np.mean(joint_ems)}, count: {len(joint_ems)}')\n logger.info(f'joint f1: {np.mean(joint_f1s)}, count: {len(joint_f1s)}')\n logger.info(f\"Best joint F1 from combination {best_f1}\")\n if args.save_prediction != \"\":\n json.dump(best_res, open(f\"{args.save_prediction}\", \"w\"))\n\n model.train()\n return {\"em\": best_em, \"f1\": best_f1, \"joint_em\": best_joint_em, \"joint_f1\": best_joint_f1, \"sp_em\": best_sp_em, \"sp_f1\": best_sp_f1}\n\nimport time\n\ndef eval_final(args, model, eval_dataloader, weight=0.8, gpu=True):\n \"\"\"\n for final submission\n \"\"\"\n model.eval()\n id2answer = collections.defaultdict(list)\n encode_times = []\n for batch in tqdm(eval_dataloader):\n batch_to_feed = move_to_cuda(batch[\"net_inputs\"]) if gpu else batch[\"net_inputs\"]\n batch_qids = batch[\"qids\"]\n with torch.no_grad():\n start = time.time()\n outputs = model(batch_to_feed)\n encode_times.append(time.time() - start)\n\n scores = outputs[\"rank_score\"]\n scores = scores.view(-1).tolist()\n\n if args.sp_pred:\n sp_scores = outputs[\"sp_score\"]\n sp_scores = sp_scores.float().masked_fill(batch_to_feed[\"sent_offsets\"].eq(0), float(\"-inf\")).type_as(sp_scores)\n batch_sp_scores = sp_scores.sigmoid()\n\n # ans_type_predicted = torch.argmax(outputs[\"ans_type_logits\"], dim=1).view(-1).tolist()\n outs = [outputs[\"start_logits\"], outputs[\"end_logits\"]]\n\n\n # answer prediction\n span_scores = outs[0][:, :, None] + outs[1][:, None]\n max_seq_len = span_scores.size(1)\n span_mask = np.tril(np.triu(np.ones((max_seq_len, max_seq_len)), 0), args.max_ans_len)\n span_mask = span_scores.data.new(max_seq_len, max_seq_len).copy_(torch.from_numpy(span_mask))\n span_scores_masked = span_scores.float().masked_fill((1 - span_mask[None].expand_as(span_scores)).bool(), -1e10).type_as(span_scores)\n start_position = span_scores_masked.max(dim=2)[0].max(dim=1)[1]\n end_position = span_scores_masked.max(dim=2)[1].gather(\n 1, start_position.unsqueeze(1)).squeeze(1)\n answer_scores = span_scores_masked.max(dim=2)[0].max(dim=1)[0].tolist()\n para_offset = batch['para_offsets']\n start_position_ = list(\n np.array(start_position.tolist()) - np.array(para_offset))\n end_position_ = list(\n np.array(end_position.tolist()) - np.array(para_offset)) \n\n for idx, qid in enumerate(batch_qids):\n rank_score = scores[idx]\n start = start_position_[idx]\n end = end_position_[idx]\n span_score = answer_scores[idx]\n tok_to_orig_index = batch['tok_to_orig_index'][idx]\n doc_tokens = batch['doc_tokens'][idx]\n wp_tokens = batch['wp_tokens'][idx]\n orig_doc_start = tok_to_orig_index[start]\n orig_doc_end = tok_to_orig_index[end]\n orig_tokens = doc_tokens[orig_doc_start:(orig_doc_end + 1)]\n tok_tokens = wp_tokens[start:end+1]\n tok_text = \" \".join(tok_tokens)\n tok_text = tok_text.replace(\" ##\", \"\")\n tok_text = tok_text.replace(\"##\", \"\")\n tok_text = tok_text.strip()\n tok_text = \" \".join(tok_text.split())\n orig_text = \" \".join(orig_tokens)\n pred_str = get_final_text(tok_text, orig_text, do_lower_case=True, verbose_logging=False)\n\n chain_titles = [_[\"title\"] for _ in batch[\"passages\"][idx]]\n\n # get the sp sentences\n pred_sp = []\n if args.sp_pred:\n sp_score = batch_sp_scores[idx].tolist()\n passages = batch[\"passages\"][idx]\n for passage, sent_offset in zip(passages, [0, len(passages[0][\"sents\"])]):\n for idx, _ in enumerate(passage[\"sents\"]):\n try:\n if sp_score[idx + sent_offset] > 0.5:\n pred_sp.append([passage[\"title\"], idx])\n except:\n # logger.info(f\"sentence exceeds max lengths\")\n continue\n id2answer[qid].append({\n \"pred_str\": pred_str.strip(),\n \"rank_score\": rank_score,\n \"span_score\": span_score,\n \"pred_sp\": pred_sp,\n \"chain_titles\": chain_titles\n })\n lambda_ = weight\n results = collections.defaultdict(dict)\n for qid in id2answer.keys():\n ans_res = id2answer[qid]\n ans_res.sort(key=lambda x: lambda_ * x[\"rank_score\"] + (1 - lambda_) * x[\"span_score\"], reverse=True)\n top_pred = ans_res[0][\"pred_str\"]\n top_pred_sp = ans_res[0][\"pred_sp\"]\n\n results[\"answer\"][qid] = top_pred\n results[\"sp\"][qid] = top_pred_sp\n results[\"titles\"][qid] = ans_res[0][\"chain_titles\"]\n\n\n if args.save_prediction != \"\":\n json.dump(results, open(f\"{args.save_prediction}\", \"w\"))\n\n return results\n\nif __name__ == \"__main__\":\n main()\n" }, { "path": "setup.py", "content": "#!/usr/bin/env python\n# Copyright (c) Facebook, Inc. and its affiliates.\n# All rights reserved.\n#\n# This source code is licensed under the license found in the\n# LICENSE file in the root directory of this source tree.\n\nfrom setuptools import setup, find_packages\nimport sys\nimport subprocess\n\nwith open('README.md') as f:\n readme = f.read()\n\n# with open('LICENSE') as f:\n# license = f.read()\n\nwith open('requirements.txt') as f:\n reqs = f.read()\n\nsetup(\n name='mdr',\n version='0.0.1',\n description='Multi-hop dense retrieval for complex open-domain question answering',\n long_description='text/markdown',\n # license=license,\n python_requires='>=3.6',\n packages=find_packages(exclude=('data')),\n install_requires=reqs.strip().split('\\n'),\n)\n" }, { "path": "setup.sh", "content": "#!/bin/bash\n# Copyright (c) Facebook, Inc. and its affiliates.\n# All rights reserved.\n#\n# This source code is licensed under the license found in the \n# LICENSE file in the root directory of this source tree.\n\npip install -r requirements.txt\nconda install faiss-gpu cudatoolkit=10.2 -c pytorch\nconda install pytorch cudatoolkit=10.2 -c pytorch\n\ngit clone https://github.com/NVIDIA/apex\ncd apex\npip install -v --disable-pip-version-check --no-cache-dir ./\n\npython setup.py develop" }, { "path": "submitit/submit_retrieval.sh", "content": "#!/bin/bash\n# Copyright (c) Facebook, Inc. and its affiliates.\n# All rights reserved.\n#\n# This source code is licensed under the license found in the \n# LICENSE file in the root directory of this source tree.\n#!/bin/bash\n\nMKL_THREADING_LAYER=GNU python submitit_train.py --prefix mhop_retrieval_roberta \\\n --train_file /private/home/xwhan/data/hotpot/hotpot_train_with_neg_v0.json \\\n --predict_file /private/home/xwhan/data/hotpot/hotpot_val_with_neg_v0.json \\\n --model_name roberta-base \\\n --max_c_len 300 \\\n --max_q_len 70 \\\n --max_q_sp_len 350 \\\n --fp16\n" }, { "path": "submitit/submitit_qa.sh", "content": "#!/bin/bash\n# Copyright (c) Facebook, Inc. and its affiliates.\n# All rights reserved.\n#\n# This source code is licensed under the license found in the \n# LICENSE file in the root directory of this source tree.\n#!/bin/bash\n\nPREFIX=wwm_val_top50\nMODEL_BACKEND=bert-large-uncased-whole-word-masking\n\nMKL_THREADING_LAYER=GNU python submitit_train_qa.py --prefix ${PREFIX} \\\n --train_file /private/home/xwhan/data/hotpot/dense_train_b100_k100_sents.json \\\n --predict_file /private/home/xwhan/data/hotpot/dense_val_b50_k50_roberta_sents.json \\\n --model_name ${MODEL_BACKEND} \\\n --fp16 \\\n --sp-pred \\\n " }, { "path": "submitit/submitit_train.py", "content": "# Copyright (c) Facebook, Inc. and its affiliates.\n# All rights reserved.\n#\n# This source code is licensed under the license found in the \n# LICENSE file in the root directory of this source tree.\nimport os\nimport numpy as np\nimport uuid\nimport itertools\nfrom typing import Dict\nimport submitit\nfrom collections import Iterable, namedtuple\nfrom pathlib import Path\nfrom datetime import date\n\n# from trainer import Trainer\nfrom mhop_trainer import Trainer\nfrom config import ClusterConfig, train_args\n\ndef get_shared_folder() -> Path:\n return Path(\"/checkpoint/xwhan/mhop-dense-retrieval\")\n\ndef get_init_file() -> Path:\n # Init file must not exist, but it's parent dir must exist.\n os.makedirs(str(get_shared_folder()), exist_ok=True)\n init_file = get_shared_folder() / f\"{uuid.uuid4().hex}_init\"\n if init_file.exists():\n os.remove(str(init_file))\n return init_file\n\ndef grid_parameters(grid: Dict):\n \"\"\"\n Yield all combinations of parameters in the grid (as a dict)\n \"\"\"\n grid_copy = dict(grid)\n # Turn single value in an Iterable\n for k in grid_copy:\n if not isinstance(grid_copy[k], Iterable):\n grid_copy[k] = [grid_copy[k]]\n for p in itertools.product(*grid_copy.values()):\n yield dict(zip(grid.keys(), p))\n\ndef grid_search(args):\n cluster_cfg = ClusterConfig(dist_backend=\"nccl\", dist_url=\"\")\n\n date_curr = date.today().strftime(\"%m-%d-%Y\")\n log_dir = os.path.join(args.output_dir, date_curr)\n \n TrainerConfig = namedtuple(\"TrainerConfig\", sorted(vars(args)))\n train_cfg = TrainerConfig(**vars(args))\n\n # Create the executor\n print(\"Create the submitit Executor (can take time on FB cluster)\")\n # Note that the folder will depend on the job_id, to easily track experiments\n executor = submitit.AutoExecutor(folder=get_shared_folder() / \"%j\")\n num_gpus_per_node = 8\n executor.update_parameters(\n mem_gb=500,\n gpus_per_node=num_gpus_per_node,\n tasks_per_node=1,\n cpus_per_task=80,\n nodes=1,\n timeout_min=60*48,\n slurm_partition=\"learnfair\",\n slurm_signal_delay_s=120,\n slurm_constraint='volta32gb'\n )\n\n # Launch one job per grid position\n grid_meta = {\n \"num_train_epochs\": (50, lambda val: f'epoch{val}'), \n \"learning_rate\": ([2e-5, 1e-5, 3e-5], lambda val: f'lr{val}'), \n \"seed\": (16, lambda val: f'seed{val}'),\n \"predict_batch_size\": (3000, lambda val: f'evalbsize{val}'),\n \"train_batch_size\": (150, lambda val: f'trainbsize{val}'),\n \"temperature\": ([1, 0.5, 0.07], lambda val: f'tem{val}'),\n \"warmup_ratio\": ([0.1, 0.15], lambda val: f'warmup{val}'), \n }\n grid = {k:v[0] for k, v in grid_meta.items()}\n save_key = {k:v[1] for k, v in grid_meta.items()}\n\n hyper_parameters = list(grid_parameters(grid))\n jobs = []\n for i, grid_data in enumerate(hyper_parameters):\n cluster_cfg = cluster_cfg._replace(dist_url=get_init_file().as_uri())\n train_cfg = train_cfg._replace(**grid_data)\n\n run_name = f\"{train_cfg.prefix}\"\n for k, v in grid_data.items():\n run_name += \"-\" + save_key[k](v)\n train_cfg = train_cfg._replace(output_dir=os.path.join(log_dir, run_name))\n\n # Chronos needs a different job name each time\n executor.update_parameters(name=f\"sweep_{i:02d}_{uuid.uuid4().hex}\")\n trainer = Trainer(train_cfg, cluster_cfg)\n job = executor.submit(trainer)\n jobs.append(job)\n print(f\"Run {i:02d} submitted with train cfg: {train_cfg}, cluster cfg: {cluster_cfg}\")\n print(f\"Submitted jobs ids: {','.join([str(job.job_id) for job in jobs])}\")\n\n # Wait for the master's results of each job\n results = [job.task(0).result() for job in jobs]\n print(f\"Jobs results: {results}\")\n best_job = np.argmax(results)\n print(f\"Best configuration: {hyper_parameters[best_job]} (val acc = {results[best_job]:.1%})\")\n\n\nif __name__ == \"__main__\":\n args = train_args()\n grid_search(args)" }, { "path": "submitit/submitit_train_qa.py", "content": "# Copyright (c) Facebook, Inc. and its affiliates.\n# All rights reserved.\n#\n# This source code is licensed under the license found in the \n# LICENSE file in the root directory of this source tree.\nimport os\nimport numpy as np\nimport uuid\nimport itertools\nfrom typing import Dict\nimport submitit\nfrom collections import Iterable, namedtuple\nfrom pathlib import Path\nfrom datetime import date\n\nfrom qa_trainer import Trainer\nfrom config import ClusterConfig, train_args\n\ndef get_shared_folder() -> Path:\n return Path(\"/checkpoint/xwhan/mhop-qa\")\n\ndef get_init_file() -> Path:\n # Init file must not exist, but it's parent dir must exist.\n os.makedirs(str(get_shared_folder()), exist_ok=True)\n init_file = get_shared_folder() / f\"{uuid.uuid4().hex}_init\"\n if init_file.exists():\n os.remove(str(init_file))\n return init_file\n\ndef grid_parameters(grid: Dict):\n \"\"\"\n Yield all combinations of parameters in the grid (as a dict)\n \"\"\"\n grid_copy = dict(grid)\n # Turn single value in an Iterable\n for k in grid_copy:\n if not isinstance(grid_copy[k], Iterable):\n grid_copy[k] = [grid_copy[k]]\n for p in itertools.product(*grid_copy.values()):\n yield dict(zip(grid.keys(), p))\n\ndef grid_search(args):\n cluster_cfg = ClusterConfig(dist_backend=\"nccl\", dist_url=\"\")\n\n date_curr = date.today().strftime(\"%m-%d-%Y\")\n log_dir = os.path.join(args.output_dir, date_curr)\n \n TrainerConfig = namedtuple(\"TrainerConfig\", sorted(vars(args)))\n train_cfg = TrainerConfig(**vars(args))\n\n # Create the executor\n print(\"Create the submitit Executor (can take time on FB cluster)\")\n # Note that the folder will depend on the job_id, to easily track experiments\n executor = submitit.AutoExecutor(folder=get_shared_folder() / \"%j\")\n num_gpus_per_node = 8\n executor.update_parameters(\n mem_gb=400,\n gpus_per_node=num_gpus_per_node,\n tasks_per_node=1, # one task per GPU\n cpus_per_task=10,\n nodes=1,\n timeout_min=60*72,\n slurm_partition=\"learnfair\",\n slurm_signal_delay_s=120,\n slurm_constraint='volta32gb'\n )\n\n # Launch one job per grid position\n grid_meta = {\n \"num_train_epochs\": (7, lambda val: f'epoch{val}'), \n \"learning_rate\": ([2e-5, 5e-5, 3e-5], lambda val: f'lr{val}'), \n \"seed\": ([42,5], lambda val: f'seed{val}'),\n \"rank_drop\": (0, lambda val: f'rdrop{val}'),\n \"qa_drop\": (0, lambda val: f'qadrop{val}'),\n # \"max_seq_len\": (512, lambda val: f'c_len{val}'),\n # \"max_q_len\": (100, lambda val: f'q_len{val}'),\n \"weight_decay\": (0, lambda val: f'decay{val}'),\n \"num_q_per_gpu\": (2, lambda val: f'qpergpu{val}'), # how many questions per gpu\n \"gradient_accumulation_steps\": (8, lambda val: f'aggstep{val}'),\n \"max_grad_norm\": (2, lambda val: f'clip{val}'),\n \"eval_period\": (250, lambda val: f'evalper{val}'),\n \"predict_batch_size\": (1024, lambda val: f'evalbsize{val}'),\n \"neg_num\": (5, lambda val: f'negnum{val}'),\n \"warmup_ratio\": ([0.1, 0.2], lambda val: f'warmup{val}'),\n \"use_adam\": (True, lambda val: f'adam{val}'),\n \"sp_weight\": ([0.05, 0.025], lambda val: f'spweight{val}'),\n \"shared_norm\": (False, lambda val: f'sn{val}'),\n }\n grid = {k:v[0] for k, v in grid_meta.items()}\n save_key = {k:v[1] for k, v in grid_meta.items()}\n \n hyper_parameters = list(grid_parameters(grid))\n jobs = []\n for i, grid_data in enumerate(hyper_parameters):\n cluster_cfg = cluster_cfg._replace(dist_url=get_init_file().as_uri())\n train_cfg = train_cfg._replace(**grid_data)\n\n run_name = f\"{train_cfg.prefix}\"\n for k, v in grid_data.items():\n run_name += \"-\" + save_key[k](v)\n train_cfg = train_cfg._replace(output_dir=os.path.join(log_dir, run_name))\n\n # Chronos needs a different job name each time\n executor.update_parameters(name=f\"sweep_{i:02d}_{uuid.uuid4().hex}\")\n trainer = Trainer(train_cfg, cluster_cfg)\n job = executor.submit(trainer)\n jobs.append(job)\n print(f\"Run {i:02d} submitted with train cfg: {train_cfg}, cluster cfg: {cluster_cfg}\")\n print(f\"Submitted jobs ids: {','.join([str(job.job_id) for job in jobs])}\")\n\n # Wait for the master's results of each job\n results = [job.task(0).result() for job in jobs]\n print(f\"Jobs results: {results}\")\n best_job = np.argmax(results)\n print(f\"Best configuration: {hyper_parameters[best_job]} (val acc = {results[best_job]:.1%})\")\n\n\nif __name__ == \"__main__\":\n args = train_args()\n grid_search(args)" } ]