Repository: facebookresearch/multihop_dense_retrieval Branch: main Commit: 62eb2427e36a Files: 53 Total size: 446.1 KB Directory structure: gitextract_529i5290/ ├── .gitignore ├── CODE_OF_CONDUCT.md ├── CONTRIBUTING.md ├── LICENSE ├── README.md ├── mdr/ │ ├── __init__.py │ ├── qa/ │ │ ├── __init__.py │ │ ├── basic_tokenizer.py │ │ ├── config.py │ │ ├── data_utils.py │ │ ├── hotpot_evaluate_v1.py │ │ ├── qa_dataset.py │ │ ├── qa_model.py │ │ ├── qa_trainer.py │ │ ├── train.md │ │ ├── train_ranker.py │ │ └── utils.py │ └── retrieval/ │ ├── __init__.py │ ├── config.py │ ├── criterions.py │ ├── decomposed_analysis.py │ ├── fever.ipynb │ ├── hotpot.ipynb │ ├── interactive_retrieval.py │ ├── mhop_trainer.py │ ├── single_trainer.py │ ├── train_single.py │ └── utils/ │ ├── basic_tokenizer.py │ ├── gen_index_id_map.py │ ├── mhop_utils.py │ ├── tokenizer.py │ └── utils.py ├── requirements.txt ├── scripts/ │ ├── add_sp_label.sh │ ├── demo.py │ ├── download_hotpot.sh │ ├── encode_corpus.py │ ├── end2end.py │ ├── end2end.sh │ ├── eval/ │ │ ├── eval_mhop_fever.py │ │ ├── eval_mhop_retrieval.py │ │ ├── eval_reranked.py │ │ ├── eval_retrieval.py │ │ └── eval_single_fever.py │ ├── train_mhop.py │ ├── train_momentum.py │ └── train_qa.py ├── setup.py ├── setup.sh └── submitit/ ├── submit_retrieval.sh ├── submitit_qa.sh ├── submitit_train.py └── submitit_train_qa.py ================================================ FILE CONTENTS ================================================ ================================================ FILE: .gitignore ================================================ data/ mdr.egg*/ apex/ models/ logs/ .DS_Store *.pyc *.swp ================================================ FILE: CODE_OF_CONDUCT.md ================================================ # Code of Conduct ## Our Pledge In the interest of fostering an open and welcoming environment, we as contributors and maintainers pledge to make participation in our project and our community a harassment-free experience for everyone, regardless of age, body size, disability, ethnicity, sex characteristics, gender identity and expression, level of experience, education, socio-economic status, nationality, personal appearance, race, religion, or sexual identity and orientation. ## Our Standards Examples of behavior that contributes to creating a positive environment include: * Using welcoming and inclusive language * Being respectful of differing viewpoints and experiences * Gracefully accepting constructive criticism * Focusing on what is best for the community * Showing empathy towards other community members Examples of unacceptable behavior by participants include: * The use of sexualized language or imagery and unwelcome sexual attention or advances * Trolling, insulting/derogatory comments, and personal or political attacks * Public or private harassment * Publishing others' private information, such as a physical or electronic address, without explicit permission * Other conduct which could reasonably be considered inappropriate in a professional setting ## Our Responsibilities Project maintainers are responsible for clarifying the standards of acceptable behavior and are expected to take appropriate and fair corrective action in response to any instances of unacceptable behavior. Project maintainers have the right and responsibility to remove, edit, or reject comments, commits, code, wiki edits, issues, and other contributions that are not aligned to this Code of Conduct, or to ban temporarily or permanently any contributor for other behaviors that they deem inappropriate, threatening, offensive, or harmful. ## Scope This Code of Conduct applies within all project spaces, and it also applies when an individual is representing the project or its community in public spaces. Examples of representing a project or community include using an official project e-mail address, posting via an official social media account, or acting as an appointed representative at an online or offline event. Representation of a project may be further defined and clarified by project maintainers. ## Enforcement Instances of abusive, harassing, or otherwise unacceptable behavior may be reported by contacting the project team at . All complaints will be reviewed and investigated and will result in a response that is deemed necessary and appropriate to the circumstances. The project team is obligated to maintain confidentiality with regard to the reporter of an incident. Further details of specific enforcement policies may be posted separately. Project maintainers who do not follow or enforce the Code of Conduct in good faith may face temporary or permanent repercussions as determined by other members of the project's leadership. ## Attribution This Code of Conduct is adapted from the [Contributor Covenant][homepage], version 1.4, available at https://www.contributor-covenant.org/version/1/4/code-of-conduct.html [homepage]: https://www.contributor-covenant.org For answers to common questions about this code of conduct, see https://www.contributor-covenant.org/faq ================================================ FILE: CONTRIBUTING.md ================================================ # Contributing to multihop_dense_retrieval We want to make contributing to this project as easy and transparent as possible. ## Pull Requests We actively welcome your pull requests. 1. Fork the repo and create your branch from `master`. 2. If you've added code that should be tested, add tests. 3. If you've changed APIs, update the documentation. 4. Ensure the test suite passes. 5. Make sure your code lints. 6. If you haven't already, complete the Contributor License Agreement ("CLA"). ## Contributor License Agreement ("CLA") In order to accept your pull request, we need you to submit a CLA. You only need to do this once to work on any of Facebook's open source projects. Complete your CLA here: ## Issues We use GitHub issues to track public bugs. Please ensure your description is clear and has sufficient instructions to be able to reproduce the issue. Facebook has a [bounty program](https://www.facebook.com/whitehat/) for the safe disclosure of security bugs. In those cases, please go through the process outlined on that page and do not file a public issue. ## License By contributing to multihop_dense_retrieval, you agree that your contributions will be licensed under the LICENSE file in the root directory of this source tree. ================================================ FILE: LICENSE ================================================ Attribution-NonCommercial 4.0 International ======================================================================= Creative Commons Corporation ("Creative Commons") is not a law firm and does not provide legal services or legal advice. Distribution of Creative Commons public licenses does not create a lawyer-client or other relationship. Creative Commons makes its licenses and related information available on an "as-is" basis. Creative Commons gives no warranties regarding its licenses, any material licensed under their terms and conditions, or any related information. Creative Commons disclaims all liability for damages resulting from their use to the fullest extent possible. Using Creative Commons Public Licenses Creative Commons public licenses provide a standard set of terms and conditions that creators and other rights holders may use to share original works of authorship and other material subject to copyright and certain other rights specified in the public license below. The following considerations are for informational purposes only, are not exhaustive, and do not form part of our licenses. Considerations for licensors: Our public licenses are intended for use by those authorized to give the public permission to use material in ways otherwise restricted by copyright and certain other rights. Our licenses are irrevocable. Licensors should read and understand the terms and conditions of the license they choose before applying it. Licensors should also secure all rights necessary before applying our licenses so that the public can reuse the material as expected. Licensors should clearly mark any material not subject to the license. This includes other CC- licensed material, or material used under an exception or limitation to copyright. More considerations for licensors: wiki.creativecommons.org/Considerations_for_licensors Considerations for the public: By using one of our public licenses, a licensor grants the public permission to use the licensed material under specified terms and conditions. If the licensor's permission is not necessary for any reason--for example, because of any applicable exception or limitation to copyright--then that use is not regulated by the license. Our licenses grant only permissions under copyright and certain other rights that a licensor has authority to grant. Use of the licensed material may still be restricted for other reasons, including because others have copyright or other rights in the material. A licensor may make special requests, such as asking that all changes be marked or described. Although not required by our licenses, you are encouraged to respect those requests where reasonable. More_considerations for the public: wiki.creativecommons.org/Considerations_for_licensees ======================================================================= Creative Commons Attribution-NonCommercial 4.0 International Public License By exercising the Licensed Rights (defined below), You accept and agree to be bound by the terms and conditions of this Creative Commons Attribution-NonCommercial 4.0 International Public License ("Public License"). To the extent this Public License may be interpreted as a contract, You are granted the Licensed Rights in consideration of Your acceptance of these terms and conditions, and the Licensor grants You such rights in consideration of benefits the Licensor receives from making the Licensed Material available under these terms and conditions. Section 1 -- Definitions. a. Adapted Material means material subject to Copyright and Similar Rights that is derived from or based upon the Licensed Material and in which the Licensed Material is translated, altered, arranged, transformed, or otherwise modified in a manner requiring permission under the Copyright and Similar Rights held by the Licensor. For purposes of this Public License, where the Licensed Material is a musical work, performance, or sound recording, Adapted Material is always produced where the Licensed Material is synched in timed relation with a moving image. b. Adapter's License means the license You apply to Your Copyright and Similar Rights in Your contributions to Adapted Material in accordance with the terms and conditions of this Public License. c. Copyright and Similar Rights means copyright and/or similar rights closely related to copyright including, without limitation, performance, broadcast, sound recording, and Sui Generis Database Rights, without regard to how the rights are labeled or categorized. For purposes of this Public License, the rights specified in Section 2(b)(1)-(2) are not Copyright and Similar Rights. d. Effective Technological Measures means those measures that, in the absence of proper authority, may not be circumvented under laws fulfilling obligations under Article 11 of the WIPO Copyright Treaty adopted on December 20, 1996, and/or similar international agreements. e. Exceptions and Limitations means fair use, fair dealing, and/or any other exception or limitation to Copyright and Similar Rights that applies to Your use of the Licensed Material. f. Licensed Material means the artistic or literary work, database, or other material to which the Licensor applied this Public License. g. Licensed Rights means the rights granted to You subject to the terms and conditions of this Public License, which are limited to all Copyright and Similar Rights that apply to Your use of the Licensed Material and that the Licensor has authority to license. h. Licensor means the individual(s) or entity(ies) granting rights under this Public License. i. NonCommercial means not primarily intended for or directed towards commercial advantage or monetary compensation. For purposes of this Public License, the exchange of the Licensed Material for other material subject to Copyright and Similar Rights by digital file-sharing or similar means is NonCommercial provided there is no payment of monetary compensation in connection with the exchange. j. Share means to provide material to the public by any means or process that requires permission under the Licensed Rights, such as reproduction, public display, public performance, distribution, dissemination, communication, or importation, and to make material available to the public including in ways that members of the public may access the material from a place and at a time individually chosen by them. k. Sui Generis Database Rights means rights other than copyright resulting from Directive 96/9/EC of the European Parliament and of the Council of 11 March 1996 on the legal protection of databases, as amended and/or succeeded, as well as other essentially equivalent rights anywhere in the world. l. You means the individual or entity exercising the Licensed Rights under this Public License. Your has a corresponding meaning. Section 2 -- Scope. a. License grant. 1. Subject to the terms and conditions of this Public License, the Licensor hereby grants You a worldwide, royalty-free, non-sublicensable, non-exclusive, irrevocable license to exercise the Licensed Rights in the Licensed Material to: a. reproduce and Share the Licensed Material, in whole or in part, for NonCommercial purposes only; and b. produce, reproduce, and Share Adapted Material for NonCommercial purposes only. 2. Exceptions and Limitations. For the avoidance of doubt, where Exceptions and Limitations apply to Your use, this Public License does not apply, and You do not need to comply with its terms and conditions. 3. Term. The term of this Public License is specified in Section 6(a). 4. Media and formats; technical modifications allowed. The Licensor authorizes You to exercise the Licensed Rights in all media and formats whether now known or hereafter created, and to make technical modifications necessary to do so. The Licensor waives and/or agrees not to assert any right or authority to forbid You from making technical modifications necessary to exercise the Licensed Rights, including technical modifications necessary to circumvent Effective Technological Measures. For purposes of this Public License, simply making modifications authorized by this Section 2(a) (4) never produces Adapted Material. 5. Downstream recipients. a. Offer from the Licensor -- Licensed Material. Every recipient of the Licensed Material automatically receives an offer from the Licensor to exercise the Licensed Rights under the terms and conditions of this Public License. b. No downstream restrictions. You may not offer or impose any additional or different terms or conditions on, or apply any Effective Technological Measures to, the Licensed Material if doing so restricts exercise of the Licensed Rights by any recipient of the Licensed Material. 6. No endorsement. Nothing in this Public License constitutes or may be construed as permission to assert or imply that You are, or that Your use of the Licensed Material is, connected with, or sponsored, endorsed, or granted official status by, the Licensor or others designated to receive attribution as provided in Section 3(a)(1)(A)(i). b. Other rights. 1. Moral rights, such as the right of integrity, are not licensed under this Public License, nor are publicity, privacy, and/or other similar personality rights; however, to the extent possible, the Licensor waives and/or agrees not to assert any such rights held by the Licensor to the limited extent necessary to allow You to exercise the Licensed Rights, but not otherwise. 2. Patent and trademark rights are not licensed under this Public License. 3. To the extent possible, the Licensor waives any right to collect royalties from You for the exercise of the Licensed Rights, whether directly or through a collecting society under any voluntary or waivable statutory or compulsory licensing scheme. In all other cases the Licensor expressly reserves any right to collect such royalties, including when the Licensed Material is used other than for NonCommercial purposes. Section 3 -- License Conditions. Your exercise of the Licensed Rights is expressly made subject to the following conditions. a. Attribution. 1. If You Share the Licensed Material (including in modified form), You must: a. retain the following if it is supplied by the Licensor with the Licensed Material: i. identification of the creator(s) of the Licensed Material and any others designated to receive attribution, in any reasonable manner requested by the Licensor (including by pseudonym if designated); ii. a copyright notice; iii. a notice that refers to this Public License; iv. a notice that refers to the disclaimer of warranties; v. a URI or hyperlink to the Licensed Material to the extent reasonably practicable; b. indicate if You modified the Licensed Material and retain an indication of any previous modifications; and c. indicate the Licensed Material is licensed under this Public License, and include the text of, or the URI or hyperlink to, this Public License. 2. You may satisfy the conditions in Section 3(a)(1) in any reasonable manner based on the medium, means, and context in which You Share the Licensed Material. For example, it may be reasonable to satisfy the conditions by providing a URI or hyperlink to a resource that includes the required information. 3. If requested by the Licensor, You must remove any of the information required by Section 3(a)(1)(A) to the extent reasonably practicable. 4. If You Share Adapted Material You produce, the Adapter's License You apply must not prevent recipients of the Adapted Material from complying with this Public License. Section 4 -- Sui Generis Database Rights. Where the Licensed Rights include Sui Generis Database Rights that apply to Your use of the Licensed Material: a. for the avoidance of doubt, Section 2(a)(1) grants You the right to extract, reuse, reproduce, and Share all or a substantial portion of the contents of the database for NonCommercial purposes only; b. if You include all or a substantial portion of the database contents in a database in which You have Sui Generis Database Rights, then the database in which You have Sui Generis Database Rights (but not its individual contents) is Adapted Material; and c. You must comply with the conditions in Section 3(a) if You Share all or a substantial portion of the contents of the database. For the avoidance of doubt, this Section 4 supplements and does not replace Your obligations under this Public License where the Licensed Rights include other Copyright and Similar Rights. Section 5 -- Disclaimer of Warranties and Limitation of Liability. a. UNLESS OTHERWISE SEPARATELY UNDERTAKEN BY THE LICENSOR, TO THE EXTENT POSSIBLE, THE LICENSOR OFFERS THE LICENSED MATERIAL AS-IS AND AS-AVAILABLE, AND MAKES NO REPRESENTATIONS OR WARRANTIES OF ANY KIND CONCERNING THE LICENSED MATERIAL, WHETHER EXPRESS, IMPLIED, STATUTORY, OR OTHER. THIS INCLUDES, WITHOUT LIMITATION, WARRANTIES OF TITLE, MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE, NON-INFRINGEMENT, ABSENCE OF LATENT OR OTHER DEFECTS, ACCURACY, OR THE PRESENCE OR ABSENCE OF ERRORS, WHETHER OR NOT KNOWN OR DISCOVERABLE. WHERE DISCLAIMERS OF WARRANTIES ARE NOT ALLOWED IN FULL OR IN PART, THIS DISCLAIMER MAY NOT APPLY TO YOU. b. TO THE EXTENT POSSIBLE, IN NO EVENT WILL THE LICENSOR BE LIABLE TO YOU ON ANY LEGAL THEORY (INCLUDING, WITHOUT LIMITATION, NEGLIGENCE) OR OTHERWISE FOR ANY DIRECT, SPECIAL, INDIRECT, INCIDENTAL, CONSEQUENTIAL, PUNITIVE, EXEMPLARY, OR OTHER LOSSES, COSTS, EXPENSES, OR DAMAGES ARISING OUT OF THIS PUBLIC LICENSE OR USE OF THE LICENSED MATERIAL, EVEN IF THE LICENSOR HAS BEEN ADVISED OF THE POSSIBILITY OF SUCH LOSSES, COSTS, EXPENSES, OR DAMAGES. WHERE A LIMITATION OF LIABILITY IS NOT ALLOWED IN FULL OR IN PART, THIS LIMITATION MAY NOT APPLY TO YOU. c. The disclaimer of warranties and limitation of liability provided above shall be interpreted in a manner that, to the extent possible, most closely approximates an absolute disclaimer and waiver of all liability. Section 6 -- Term and Termination. a. This Public License applies for the term of the Copyright and Similar Rights licensed here. However, if You fail to comply with this Public License, then Your rights under this Public License terminate automatically. b. Where Your right to use the Licensed Material has terminated under Section 6(a), it reinstates: 1. automatically as of the date the violation is cured, provided it is cured within 30 days of Your discovery of the violation; or 2. upon express reinstatement by the Licensor. For the avoidance of doubt, this Section 6(b) does not affect any right the Licensor may have to seek remedies for Your violations of this Public License. c. For the avoidance of doubt, the Licensor may also offer the Licensed Material under separate terms or conditions or stop distributing the Licensed Material at any time; however, doing so will not terminate this Public License. d. Sections 1, 5, 6, 7, and 8 survive termination of this Public License. Section 7 -- Other Terms and Conditions. a. The Licensor shall not be bound by any additional or different terms or conditions communicated by You unless expressly agreed. b. Any arrangements, understandings, or agreements regarding the Licensed Material not stated herein are separate from and independent of the terms and conditions of this Public License. Section 8 -- Interpretation. a. For the avoidance of doubt, this Public License does not, and shall not be interpreted to, reduce, limit, restrict, or impose conditions on any use of the Licensed Material that could lawfully be made without permission under this Public License. b. To the extent possible, if any provision of this Public License is deemed unenforceable, it shall be automatically reformed to the minimum extent necessary to make it enforceable. If the provision cannot be reformed, it shall be severed from this Public License without affecting the enforceability of the remaining terms and conditions. c. No term or condition of this Public License will be waived and no failure to comply consented to unless expressly agreed to by the Licensor. d. Nothing in this Public License constitutes or may be interpreted as a limitation upon, or waiver of, any privileges and immunities that apply to the Licensor or You, including from the legal processes of any jurisdiction or authority. ======================================================================= Creative Commons is not a party to its public licenses. Notwithstanding, Creative Commons may elect to apply one of its public licenses to material it publishes and in those instances will be considered the “Licensor.” The text of the Creative Commons public licenses is dedicated to the public domain under the CC0 Public Domain Dedication. Except for the limited purpose of indicating that material is shared under a Creative Commons public license or as otherwise permitted by the Creative Commons policies published at creativecommons.org/policies, Creative Commons does not authorize the use of the trademark "Creative Commons" or any other trademark or logo of Creative Commons without its prior written consent including, without limitation, in connection with any unauthorized modifications to any of its public licenses or any other arrangements, understandings, or agreements concerning use of licensed material. For the avoidance of doubt, this paragraph does not form part of the public licenses. Creative Commons may be contacted at creativecommons.org. ================================================ FILE: README.md ================================================ # [

Multi-Hop Dense Text Retrieval (`MDR`)

](#p-aligncentermulti-hop-dense-text-retrieval-mdrp) **\*\*\*\*\* Update 3/4/2021: Adding simple demo code based on [streamlit](https://streamlit.io/) \*\*\*\*\*** `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)). More 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)

- [Use the trained models](#use-the-trained-models) - [Evaluating retrieval](#evaluating-retrieval) - [Evaluating QA](#evaluating-qa) - [Demo](#end-to-end-demo) - [Train models from scratch](#train-models-from-scratch) - [Retriever training](#retriever-training) - [Encoding the corpus for retrieval](#encoding-the-corpus-for-retrieval) - [ELECTRA QA model training](#electra-qa-model-training) ## Use the trained models 1. Set up the environment ```bash conda create --name MDR python=3.6 conda activate MDR git clone git@github.com:facebookresearch/multihop_dense_retrieval.git cd multihop_dense_retrieval bash setup.sh ``` 2. Download the necessary data files and pretrained retrieval models Simplified data files with **quesitons** and ground-truth **supporting passages**: ``` # save pretrained models to models/ and all processed hotpotQA into data/ # models will take about 2GB, and data will take 20GB since the pre-trained wikipedia index are included. bash ./scripts/download_hotpot.sh ``` ### Evaluating retrieval Evalauting direct retrieval performance (The printed statistics might not adhere to the metric names defined in the paper. * **PR**: whether one of the supporting passages is included in all retrieved passages; * **P-EM**: whether **both** supporting passages are included in all retrieval passages; * **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. Here's an example evaluating the top1 ranked passage chains: ``` python scripts/eval/eval_mhop_retrieval.py \ data/hotpot/hotpot_qas_val.json \ data/hotpot_index/wiki_index.npy \ data/hotpot_index/wiki_id2doc.json \ models/q_encoder.pt \ --batch-size 100 \ --beam-size 1 \ --topk 1 \ --shared-encoder \ --model-name roberta-base \ --gpu \ --save-path ${SAVE_RETRIEVAL_FOR_QA} ``` Sevaral important options includes * `--beam-size-n`: beam size at each hop; * `--topk`: topk passage chains from beam search * `--gpu`: move the dense index to GPU, resulting in much faster search Expected results (Top1): ``` Evaluating 7405 samples... Avg PR: 0.8428089128966915 Avg P-EM: 0.6592842673869007 Avg 1-Recall: 0.7906819716407832 Path Recall: 0.6592842673869007 comparison Questions num: 1487 Avg PR: 0.9932750504371217 Avg P-EM: 0.9482178883658372 Avg 1-Recall: 0.9643577673167452 Path Recall: 0.9482178883658372 bridge Questions num: 5918 Avg PR: 0.805001689760054 Avg P-EM: 0.5866846907739101 Avg 1-Recall: 0.7470429199053734 Path Recall: 0.5866846907739101 ``` **Note:** For more efficient retrieval on CPU, check out the `--hnsw` option in `scripts/eval/eval_mhop_retrieval.py`. ### Evaluating QA The 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. As 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`: ``` python scripts/train_qa.py \ --do_predict \ --predict_batch_size 200 \ --model_name google/electra-large-discriminator \ --fp16 \ --predict_file data/hotpot/dev_retrieval_top100_sp.json \ --max_seq_len 512 \ --max_q_len 64 \ --init_checkpoint models/qa_electra.pt \ --sp-pred \ --max_ans_len 30 \ --save-prediction hotpot_val_top100.json ``` Expected results: ``` 01/21/2021 17:01:49 - INFO - __main__ - evaluated 7405 questions... 01/21/2021 17:01:49 - INFO - __main__ - chain ranking em: 0.8113436866981769 01/21/2021 17:01:50 - INFO - __main__ - .......Using combination factor 0.8...... 01/21/2021 17:01:50 - INFO - __main__ - answer em: 0.6233625928426739, count: 7405 01/21/2021 17:01:50 - INFO - __main__ - answer f1: 0.7504594111976622, count: 7405 01/21/2021 17:01:50 - INFO - __main__ - sp em: 0.5654287643484133, count: 7405 01/21/2021 17:01:50 - INFO - __main__ - sp f1: 0.7942837708469039, count: 7405 01/21/2021 17:01:50 - INFO - __main__ - joint em: 0.42052667116812964, count: 7405 01/21/2021 17:01:50 - INFO - __main__ - joint f1: 0.6631669237532106, count: 7405 01/21/2021 17:01:50 - INFO - __main__ - Best joint F1 from combination 0.7504594111976622 01/21/2021 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} ``` ## End to end Demo A simple demo code using our pretrained models. ``` streamlit run scripts/demo.py ```

## Train models from scratch Our experiments are mostly run on 8 GPUs, however, we observed similar performance when using a smaller performance. ### Retriever training ``` CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python scripts/train_mhop.py \ --do_train \ --prefix ${RUN_ID} \ --predict_batch_size 3000 \ --model_name roberta-base \ --train_batch_size 150 \ --learning_rate 2e-5 \ --fp16 \ --train_file ${TRAIN_DATA_PATH} \ --predict_file ${DEV_DATA_PATH} \ --seed 16 \ --eval-period -1 \ --max_c_len 300 \ --max_q_len 70 \ --max_q_sp_len 350 \ --shared-encoder \ --warmup-ratio 0.1 ``` Processed train/validation data for retrieval training: * `${TRAIN_DATA_PATH}`: data/hotpot/hotpot_train_with_neg_v0.json * `${DEV_DATA_PATH}`: data/hotpot/hotpot_dev_with_neg_v0.json ### Finetune the question encoder with frozen memory bank This step happens after the previous training stage and reuses the checkpoint point. ``` CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python train_momentum.py \ --do_train \ --prefix {RUN_ID} \ --predict_batch_size 3000 \ --model_name roberta-base \ --train_batch_size 150 \ --learning_rate 1e-5 \ --fp16 \ --train_file {TRAIN_DATA_PATH} \ --predict_file {DEV_DATA_PATH} \ --seed 16 \ --eval-period -1 \ --max_c_len 300 \ --max_q_len 70 \ --max_q_sp_len 350 \ --momentum \ --k 76800 \ --m 0.999 \ --temperature 1 \ --init-retriever {CHECKPOINT_PT} ``` ## Encoding the corpus for retrieval ``` CUDA_VISIBLE_DEVICES=0,1,2,3 python scripts/encode_corpus.py \ --do_predict \ --predict_batch_size 1000 \ --model_name roberta-base \ --predict_file ${CORPUS_PATH} \ --init_checkpoint ${MODEL_CHECKPOINT} \ --embed_save_path ${SAVE_PATH} \ --fp16 \ --max_c_len 300 \ --num_workers 20 ``` * `${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). * `${SAVE_PATH}`: path to save the numpy vectors and ID2DOC lookup table. ### ELECTRA QA model training The 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: ``` CUDA_VISIBLE_DEVICES=0 python train_qa.py \ --do_train \ --prefix electra_large_debug_sn \ --predict_batch_size 1024 \ --model_name google/electra-large-discriminator \ --train_batch_size 12 \ --learning_rate 5e-5 \ --train_file ${QA_TRAIN_DATA} \ --predict_file ${QA_DEV_DATA} \ --seed 42 \ --eval-period 250 \ --max_seq_len 512 \ --max_q_len 64 \ --gradient_accumulation_steps 8 \ --neg-num 5 \ --fp16 \ --use-adam \ --warmup-ratio 0.1 \ --sp-weight 0.05 \ --sp-pred ``` Processed (ran [scripts/add_sp_label.sh](scripts/add_sp_label.sh)) train/validata data for QA training. * `${QA_TRAIN_DATA}`: data/hotpot/train_retrieval_b100_k100_sp.json * `${QA_DEV_DATA}`: data/hotpot/dev_retrieval_b50_k50_sp.json ## Cite ``` @article{xiong2020answering, title={Answering Complex Open-Domain Questions with Multi-Hop Dense Retrieval}, 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}, journal={International Conference on Learning Representations}, year={2021} } ``` ## License CC-BY-NC 4.0 ================================================ FILE: mdr/__init__.py ================================================ # Copyright (c) Facebook, Inc. and its affiliates. # All rights reserved. # # This source code is licensed under the license found in the # LICENSE file in the root directory of this source tree. from . import qa from . import retrieval ================================================ FILE: mdr/qa/__init__.py ================================================ # Copyright (c) Facebook, Inc. and its affiliates. # All rights reserved. # # This source code is licensed under the license found in the # LICENSE file in the root directory of this source tree. ================================================ FILE: mdr/qa/basic_tokenizer.py ================================================ # Copyright (c) Facebook, Inc. and its affiliates. # All rights reserved. # # This source code is licensed under the license found in the # LICENSE file in the root directory of this source tree. #!/usr/bin/env python3 # Copyright 2017-present, Facebook, Inc. # All rights reserved. # # This source code is licensed under the license found in the # LICENSE file in the root directory of this source tree. """Base tokenizer/tokens classes and utilities.""" import copy class Tokens(object): """A class to represent a list of tokenized text.""" TEXT = 0 TEXT_WS = 1 SPAN = 2 POS = 3 LEMMA = 4 NER = 5 def __init__(self, data, annotators, opts=None): self.data = data self.annotators = annotators self.opts = opts or {} def __len__(self): """The number of tokens.""" return len(self.data) def slice(self, i=None, j=None): """Return a view of the list of tokens from [i, j).""" new_tokens = copy.copy(self) new_tokens.data = self.data[i: j] return new_tokens def untokenize(self): """Returns the original text (with whitespace reinserted).""" return ''.join([t[self.TEXT_WS] for t in self.data]).strip() def words(self, uncased=False): """Returns a list of the text of each token Args: uncased: lower cases text """ if uncased: return [t[self.TEXT].lower() for t in self.data] else: return [t[self.TEXT] for t in self.data] def offsets(self): """Returns a list of [start, end) character offsets of each token.""" return [t[self.SPAN] for t in self.data] def pos(self): """Returns a list of part-of-speech tags of each token. Returns None if this annotation was not included. """ if 'pos' not in self.annotators: return None return [t[self.POS] for t in self.data] def lemmas(self): """Returns a list of the lemmatized text of each token. Returns None if this annotation was not included. """ if 'lemma' not in self.annotators: return None return [t[self.LEMMA] for t in self.data] def entities(self): """Returns a list of named-entity-recognition tags of each token. Returns None if this annotation was not included. """ if 'ner' not in self.annotators: return None return [t[self.NER] for t in self.data] def ngrams(self, n=1, uncased=False, filter_fn=None, as_strings=True): """Returns a list of all ngrams from length 1 to n. Args: n: upper limit of ngram length uncased: lower cases text filter_fn: user function that takes in an ngram list and returns True or False to keep or not keep the ngram as_string: return the ngram as a string vs list """ def _skip(gram): if not filter_fn: return False return filter_fn(gram) words = self.words(uncased) ngrams = [(s, e + 1) for s in range(len(words)) for e in range(s, min(s + n, len(words))) if not _skip(words[s:e + 1])] # Concatenate into strings if as_strings: ngrams = ['{}'.format(' '.join(words[s:e])) for (s, e) in ngrams] return ngrams def entity_groups(self): """Group consecutive entity tokens with the same NER tag.""" entities = self.entities() if not entities: return None non_ent = self.opts.get('non_ent', 'O') groups = [] idx = 0 while idx < len(entities): ner_tag = entities[idx] # Check for entity tag if ner_tag != non_ent: # Chomp the sequence start = idx while (idx < len(entities) and entities[idx] == ner_tag): idx += 1 groups.append((self.slice(start, idx).untokenize(), ner_tag)) else: idx += 1 return groups class Tokenizer(object): """Base tokenizer class. Tokenizers implement tokenize, which should return a Tokens class. """ def tokenize(self, text): raise NotImplementedError def shutdown(self): pass def __del__(self): self.shutdown() import regex import logging logger = logging.getLogger(__name__) class RegexpTokenizer(Tokenizer): DIGIT = r'\p{Nd}+([:\.\,]\p{Nd}+)*' TITLE = (r'(dr|esq|hon|jr|mr|mrs|ms|prof|rev|sr|st|rt|messrs|mmes|msgr)' r'\.(?=\p{Z})') ABBRV = r'([\p{L}]\.){2,}(?=\p{Z}|$)' ALPHA_NUM = r'[\p{L}\p{N}\p{M}]++' HYPHEN = r'{A}([-\u058A\u2010\u2011]{A})+'.format(A=ALPHA_NUM) NEGATION = r"((?!n't)[\p{L}\p{N}\p{M}])++(?=n't)|n't" CONTRACTION1 = r"can(?=not\b)" CONTRACTION2 = r"'([tsdm]|re|ll|ve)\b" START_DQUOTE = r'(?<=[\p{Z}\(\[{<]|^)(``|["\u0093\u201C\u00AB])(?!\p{Z})' START_SQUOTE = r'(?<=[\p{Z}\(\[{<]|^)[\'\u0091\u2018\u201B\u2039](?!\p{Z})' END_DQUOTE = r'(?%s)|(?P%s)|(?P<abbr>%s)|(?P<neg>%s)|(?P<hyph>%s)|' '(?P<contr1>%s)|(?P<alphanum>%s)|(?P<contr2>%s)|(?P<sdquote>%s)|' '(?P<edquote>%s)|(?P<ssquote>%s)|(?P<esquote>%s)|(?P<dash>%s)|' '(?<ellipses>%s)|(?P<punct>%s)|(?P<nonws>%s)' % (self.DIGIT, self.TITLE, self.ABBRV, self.NEGATION, self.HYPHEN, self.CONTRACTION1, self.ALPHA_NUM, self.CONTRACTION2, self.START_DQUOTE, self.END_DQUOTE, self.START_SQUOTE, self.END_SQUOTE, self.DASH, self.ELLIPSES, self.PUNCT, self.NON_WS), flags=regex.IGNORECASE + regex.UNICODE + regex.MULTILINE ) if len(kwargs.get('annotators', {})) > 0: logger.warning('%s only tokenizes! Skipping annotators: %s' % (type(self).__name__, kwargs.get('annotators'))) self.annotators = set() self.substitutions = kwargs.get('substitutions', True) def tokenize(self, text): data = [] matches = [m for m in self._regexp.finditer(text)] for i in range(len(matches)): # Get text token = matches[i].group() # Make normalizations for special token types if self.substitutions: groups = matches[i].groupdict() if groups['sdquote']: token = "``" elif groups['edquote']: token = "''" elif groups['ssquote']: token = "`" elif groups['esquote']: token = "'" elif groups['dash']: token = '--' elif groups['ellipses']: token = '...' # Get whitespace span = matches[i].span() start_ws = span[0] if i + 1 < len(matches): end_ws = matches[i + 1].span()[0] else: end_ws = span[1] # Format data data.append(( token, text[start_ws: end_ws], span, )) return Tokens(data, self.annotators) class SimpleTokenizer(Tokenizer): ALPHA_NUM = r'[\p{L}\p{N}\p{M}]+' NON_WS = r'[^\p{Z}\p{C}]' def __init__(self, **kwargs): """ Args: annotators: None or empty set (only tokenizes). """ self._regexp = regex.compile( '(%s)|(%s)' % (self.ALPHA_NUM, self.NON_WS), flags=regex.IGNORECASE + regex.UNICODE + regex.MULTILINE ) if len(kwargs.get('annotators', {})) > 0: logger.warning('%s only tokenizes! Skipping annotators: %s' % (type(self).__name__, kwargs.get('annotators'))) self.annotators = set() def tokenize(self, text): data = [] matches = [m for m in self._regexp.finditer(text)] for i in range(len(matches)): # Get text token = matches[i].group() # Get whitespace span = matches[i].span() start_ws = span[0] if i + 1 < len(matches): end_ws = matches[i + 1].span()[0] else: end_ws = span[1] # Format data data.append(( token, text[start_ws: end_ws], span, )) return Tokens(data, self.annotators) ================================================ FILE: mdr/qa/config.py ================================================ # Copyright (c) Facebook, Inc. and its affiliates. # All rights reserved. # # This source code is licensed under the license found in the # LICENSE file in the root directory of this source tree. import argparse from ast import parse from typing import NamedTuple from torch.nn import parallel class ClusterConfig(NamedTuple): dist_backend: str dist_url: str def common_args(): parser = argparse.ArgumentParser() # task parser.add_argument("--train_file", type=str, default="../data/nq-with-neg-train.txt") parser.add_argument("--predict_file", type=str, default="../data/nq-with-neg-dev.txt") parser.add_argument("--num_workers", default=10, type=int) parser.add_argument("--do_train", default=False, action='store_true', help="Whether to run training.") parser.add_argument("--do_predict", default=False, action='store_true', help="Whether to run eval on the dev set.") parser.add_argument("--do_test", default=False, action="store_true", help="for final test submission") # model parser.add_argument("--model_name", default="bert-base-uncased", type=str) parser.add_argument("--init_checkpoint", type=str, help="Initial checkpoint (usually from a pre-trained BERT model).", default="") parser.add_argument("--max_seq_len", default=512, type=int, help="The maximum total input sequence length after WordPiece tokenization. Sequences " "longer than this will be truncated, and sequences shorter than this will be padded.") parser.add_argument("--max_q_len", default=64, type=int) parser.add_argument("--max_ans_len", default=35, type=int) parser.add_argument('--fp16', action='store_true') parser.add_argument('--fp16_opt_level', type=str, default='O1', help="For fp16: Apex AMP optimization level selected in ['O0', 'O1', 'O2', and 'O3']." "See details at https://nvidia.github.io/apex/amp.html") parser.add_argument("--no_cuda", default=False, action='store_true', help="Whether not to use CUDA when available") parser.add_argument("--local_rank", type=int, default=-1, help="local_rank for distributed training on gpus") parser.add_argument("--predict_batch_size", default=256, type=int, help="Total batch size for predictions.") parser.add_argument("--save-prediction", default="", type=str) parser.add_argument("--sp-pred", action="store_true", help="whether to predict sentence sp") return parser def train_args(): parser = common_args() # optimization parser.add_argument('--prefix', type=str, default="eval") parser.add_argument("--weight_decay", default=0.0, type=float, help="Weight decay if we apply some.") parser.add_argument("--output_dir", default="./logs", type=str, help="The output directory where the model checkpoints will be written.") parser.add_argument("--train_batch_size", default=128, type=int, help="Total batch size for training.") parser.add_argument("--num_q_per_gpu", default=1) parser.add_argument("--learning_rate", default=1e-5, type=float, help="The initial learning rate for Adam.") parser.add_argument("--num_train_epochs", default=5, type=float, help="Total number of training epochs to perform.") parser.add_argument('--seed', type=int, default=3, help="random seed for initialization") parser.add_argument('--gradient_accumulation_steps', type=int, default=1, help="Number of updates steps to accumualte before performing a backward/update pass.") parser.add_argument('--eval-period', type=int, default=2500) parser.add_argument("--max_grad_norm", default=2.0, type=float, help="Max gradient norm.") parser.add_argument("--adam_epsilon", default=1e-8, type=float, help="Epsilon for Adam optimizer.") parser.add_argument("--neg-num", type=int, default=9, help="how many neg/distant passage chains to use") parser.add_argument("--shared-norm", action="store_true") parser.add_argument("--qa-drop", default=0, type=float) parser.add_argument("--rank-drop", default=0, type=float) parser.add_argument("--sp-drop", default=0, type=float) parser.add_argument("--final-metric", default="joint_f1") parser.add_argument("--use-adam", action="store_true", help="use adam or adamW") parser.add_argument("--warmup-ratio", default=0, type=float, help="Linear warmup over warmup_steps.") parser.add_argument("--sp-weight", default=0, type=float, help="weight of the sp loss") return parser.parse_args() ================================================ FILE: mdr/qa/data_utils.py ================================================ # Copyright (c) Facebook, Inc. and its affiliates. # All rights reserved. # # This source code is licensed under the license found in the # LICENSE file in the root directory of this source tree. import json from tqdm import tqdm import numpy as np def explore(path): train = json.load(open(path)) neg_counts = [] for item in train: tfidf_neg = item["tfidf_neg"] linked_neg = item["linked_neg"] neg_counts.append(len(tfidf_neg + linked_neg)) import pdb; pdb.set_trace() return def load_corpus(corpus_path="/private/home/xwhan/data/hotpot/tfidf/abstracts.txt"): content = [json.loads(l) for l in open(corpus_path).readlines()] title2doc = {item["title"]:item["text"] for item in content} if __name__ == "__main__": explore("/private/home/xwhan/data/hotpot/hotpot_rerank_train_2_neg_types.json") ================================================ FILE: mdr/qa/hotpot_evaluate_v1.py ================================================ # Copyright (c) Facebook, Inc. and its affiliates. # All rights reserved. # # This source code is licensed under the license found in the # LICENSE file in the root directory of this source tree. import sys import ujson as json import re import string from collections import Counter import pickle def normalize_answer(s): def remove_articles(text): return re.sub(r'\b(a|an|the)\b', ' ', text) def white_space_fix(text): return ' '.join(text.split()) def remove_punc(text): exclude = set(string.punctuation) return ''.join(ch for ch in text if ch not in exclude) def lower(text): return text.lower() return white_space_fix(remove_articles(remove_punc(lower(s)))) def f1_score(prediction, ground_truth): normalized_prediction = normalize_answer(prediction) normalized_ground_truth = normalize_answer(ground_truth) ZERO_METRIC = (0, 0, 0) if normalized_prediction in ['yes', 'no', 'noanswer'] and normalized_prediction != normalized_ground_truth: return ZERO_METRIC if normalized_ground_truth in ['yes', 'no', 'noanswer'] and normalized_prediction != normalized_ground_truth: return ZERO_METRIC prediction_tokens = normalized_prediction.split() ground_truth_tokens = normalized_ground_truth.split() common = Counter(prediction_tokens) & Counter(ground_truth_tokens) num_same = sum(common.values()) if num_same == 0: return ZERO_METRIC precision = 1.0 * num_same / len(prediction_tokens) recall = 1.0 * num_same / len(ground_truth_tokens) f1 = (2 * precision * recall) / (precision + recall) return f1, precision, recall def exact_match_score(prediction, ground_truth): return (normalize_answer(prediction) == normalize_answer(ground_truth)) def update_answer(metrics, prediction, gold): em = exact_match_score(prediction, gold) f1, prec, recall = f1_score(prediction, gold) metrics['em'] += float(em) metrics['f1'] += f1 metrics['prec'] += prec metrics['recall'] += recall return em, prec, recall def update_sp(metrics, prediction, gold): cur_sp_pred = set(map(tuple, prediction)) gold_sp_pred = set(map(tuple, gold)) tp, fp, fn = 0, 0, 0 for e in cur_sp_pred: if e in gold_sp_pred: tp += 1 else: fp += 1 for e in gold_sp_pred: if e not in cur_sp_pred: fn += 1 prec = 1.0 * tp / (tp + fp) if tp + fp > 0 else 0.0 recall = 1.0 * tp / (tp + fn) if tp + fn > 0 else 0.0 f1 = 2 * prec * recall / (prec + recall) if prec + recall > 0 else 0.0 em = 1.0 if fp + fn == 0 else 0.0 metrics['sp_em'] += em metrics['sp_f1'] += f1 metrics['sp_prec'] += prec metrics['sp_recall'] += recall return em, prec, recall def eval(prediction_file, gold_file): with open(prediction_file) as f: prediction = json.load(f) with open(gold_file) as f: gold = json.load(f) metrics = {'em': 0, 'f1': 0, 'prec': 0, 'recall': 0, 'sp_em': 0, 'sp_f1': 0, 'sp_prec': 0, 'sp_recall': 0, 'joint_em': 0, 'joint_f1': 0, 'joint_prec': 0, 'joint_recall': 0} for dp in gold: cur_id = dp['_id'] can_eval_joint = True if cur_id not in prediction['answer']: print('missing answer {}'.format(cur_id)) can_eval_joint = False else: em, prec, recall = update_answer( metrics, prediction['answer'][cur_id], dp['answer']) if cur_id not in prediction['sp']: print('missing sp fact {}'.format(cur_id)) can_eval_joint = False else: sp_em, sp_prec, sp_recall = update_sp( metrics, prediction['sp'][cur_id], dp['supporting_facts']) if can_eval_joint: joint_prec = prec * sp_prec joint_recall = recall * sp_recall if joint_prec + joint_recall > 0: joint_f1 = 2 * joint_prec * joint_recall / (joint_prec + joint_recall) else: joint_f1 = 0. joint_em = em * sp_em metrics['joint_em'] += joint_em metrics['joint_f1'] += joint_f1 metrics['joint_prec'] += joint_prec metrics['joint_recall'] += joint_recall N = len(gold) for k in metrics.keys(): metrics[k] /= N print(metrics) if __name__ == '__main__': eval(sys.argv[1], sys.argv[2]) ================================================ FILE: mdr/qa/qa_dataset.py ================================================ # Copyright (c) Facebook, Inc. and its affiliates. # All rights reserved. # # This source code is licensed under the license found in the # LICENSE file in the root directory of this source tree. import collections import json import random import torch from torch.utils.data import Dataset, Sampler from tqdm import tqdm from .basic_tokenizer import SimpleTokenizer from .utils import (find_ans_span_with_char_offsets, match_answer_span, para_has_answer, _is_whitespace) def collate_tokens(values, pad_idx, eos_idx=None, left_pad=False, move_eos_to_beginning=False): """Convert a list of 1d tensors into a padded 2d tensor.""" if len(values[0].size()) > 1: values = [v.view(-1) for v in values] size = max(v.size(0) for v in values) res = values[0].new(len(values), size).fill_(pad_idx) def copy_tensor(src, dst): assert dst.numel() == src.numel() if move_eos_to_beginning: assert src[-1] == eos_idx dst[0] = eos_idx dst[1:] = src[:-1] else: dst.copy_(src) for i, v in enumerate(values): copy_tensor(v, res[i][size - len(v):] if left_pad else res[i][:len(v)]) return res def prepare(item, tokenizer, special_toks=["[SEP]", "[unused1]", "[unused2]"]): """ tokenize the passages chains, add sentence start markers for SP sentence identification """ def _process_p(para): """ handle each para """ title, sents = para["title"].strip(), para["sents"] # return "[unused1] " + title + " [unused1] " + text # mark title # return title + " " + text pre_sents = [] for idx, sent in enumerate(sents): pre_sents.append("[unused1] " + sent.strip()) return title + " " + " ".join(pre_sents) # return " ".join(pre_sents) # mark passage boundary contexts = [] for para in item["passages"]: contexts.append(_process_p(para)) context = " [SEP] ".join(contexts) doc_tokens = [] char_to_word_offset = [] prev_is_whitespace = True context = "yes no [SEP] " + context for c in context: if _is_whitespace(c): prev_is_whitespace = True else: if prev_is_whitespace: doc_tokens.append(c) else: doc_tokens[-1] += c prev_is_whitespace = False char_to_word_offset.append(len(doc_tokens) - 1) sent_starts = [] orig_to_tok_index = [] tok_to_orig_index = [] all_doc_tokens = [] for (i, token) in enumerate(doc_tokens): orig_to_tok_index.append(len(all_doc_tokens)) if token in special_toks: if token == "[unused1]": sent_starts.append(len(all_doc_tokens)) sub_tokens = [token] else: sub_tokens = tokenizer.tokenize(token) for sub_token in sub_tokens: tok_to_orig_index.append(i) all_doc_tokens.append(sub_token) item["context_processed"] = { "doc_tokens": doc_tokens, "char_to_word_offset": char_to_word_offset, "orig_to_tok_index": orig_to_tok_index, "tok_to_orig_index": tok_to_orig_index, "all_doc_tokens": all_doc_tokens, "context": context, "sent_starts": sent_starts } return item class QAEvalDataset(Dataset): def __init__(self, tokenizer, retrievel_results, max_seq_len, max_q_len, ): retriever_outputs = retrievel_results self.tokenizer = tokenizer self.max_seq_len = max_seq_len self.max_q_len = max_q_len self.data = [] for item in retriever_outputs: if item["question"].endswith("?"): item["question"] = item["question"][:-1] # for validation, add target predictions sp_titles = None gold_answer = item.get("answer", []) sp_gold = [] for chain in item["candidate_chains"]: chain_titles = [_["title"] for _ in chain] if sp_titles: label = int(set(chain_titles) == sp_titles) else: label = -1 self.data.append({ "question": item["question"], "passages": chain, "label": label, "qid": item["_id"], "gold_answer": gold_answer, "sp_gold": sp_gold }) print(f"Total instances size {len(self.data)}") def __len__(self): return len(self.data) def __getitem__(self, index): item = prepare(self.data[index], self.tokenizer) context_ann = item["context_processed"] q_toks = self.tokenizer.tokenize(item["question"])[:self.max_q_len] para_offset = len(q_toks) + 2 # cls and seq item["wp_tokens"] = context_ann["all_doc_tokens"] assert item["wp_tokens"][0] == "yes" and item["wp_tokens"][1] == "no" item["para_offset"] = para_offset max_toks_for_doc = self.max_seq_len - para_offset - 1 if len(item["wp_tokens"]) > max_toks_for_doc: item["wp_tokens"] = item["wp_tokens"][:max_toks_for_doc] 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) item["paragraph_mask"] = torch.zeros(item["encodings"]["input_ids"].size()).view(-1) item["paragraph_mask"][para_offset:-1] = 1 item["doc_tokens"] = context_ann["doc_tokens"] item["tok_to_orig_index"] = context_ann["tok_to_orig_index"] # filter sentence offsets exceeding max sequence length sent_labels, sent_offsets = [], [] for idx, s in enumerate(item["context_processed"]["sent_starts"]): if s >= len(item["wp_tokens"]): break if "sp_sent_labels" in item: sent_labels.append(item["sp_sent_labels"][idx]) sent_offsets.append(s + para_offset) assert item["encodings"]["input_ids"].view(-1)[s+para_offset] == self.tokenizer.convert_tokens_to_ids("[unused1]") # supporting fact label item["sent_offsets"] = sent_offsets item["sent_offsets"] = torch.LongTensor(item["sent_offsets"]) item["label"] = torch.LongTensor([item["label"]]) return item class QADataset(Dataset): def __init__(self, tokenizer, data_path, max_seq_len, max_q_len, train=False, no_sent_label=False ): retriever_outputs = [json.loads(l) for l in tqdm(open(data_path).readlines())] self.tokenizer = tokenizer self.max_seq_len = max_seq_len self.max_q_len = max_q_len self.train = train self.no_sent_label = no_sent_label self.simple_tok = SimpleTokenizer() self.data = [] if train: self.qid2gold = collections.defaultdict(list) # idx self.qid2neg = collections.defaultdict(list) for item in retriever_outputs: if item["question"].endswith("?"): item["question"] = item["question"][:-1] sp_sent_labels = [] sp_gold = [] if not self.no_sent_label: for sp in item["sp"]: for _ in sp["sp_sent_ids"]: sp_gold.append([sp["title"], _]) for idx in range(len(sp["sents"])): sp_sent_labels.append(int(idx in sp["sp_sent_ids"])) question_type = item["type"] self.data.append({ "question": item["question"], "passages": item["sp"], "label": 1, "qid": item["_id"], "gold_answer": item["answer"], "sp_sent_labels": sp_sent_labels, "ans_covered": 1, # includes partial chains. "sp_gold": sp_gold }) self.qid2gold[item["_id"]].append(len(self.data) - 1) sp_titles = set([_["title"] for _ in item["sp"]]) if question_type == "bridge": ans_titles = set([p["title"] for p in item["sp"] if para_has_answer(item["answer"], "".join(p["sents"]), self.simple_tok)]) else: ans_titles = set() # top ranked negative chains ds_count = 0 # track how many distant supervised chain to use ds_limit = 5 for chain in item["candidate_chains"]: chain_titles = [_["title"] for _ in chain] if set(chain_titles) == sp_titles: continue if question_type == "bridge": answer_covered = int(len(set(chain_titles) & ans_titles) > 0) ds_count += answer_covered else: answer_covered = 0 self.data.append({ "question": item["question"], "passages": chain, "label": 0, "qid": item["_id"], "gold_answer": item["answer"], "ans_covered": answer_covered, "sp_gold": sp_gold }) self.qid2neg[item["_id"]].append(len(self.data) - 1) else: for item in retriever_outputs: if item["question"].endswith("?"): item["question"] = item["question"][:-1] # for validation, add target predictions sp_titles = set([_["title"] for _ in item["sp"]]) if "sp" in item else None gold_answer = item.get("answer", []) sp_gold = [] if "sp" in item: for sp in item["sp"]: for _ in sp["sp_sent_ids"]: sp_gold.append([sp["title"], _]) chain_seen = set() for chain in item["candidate_chains"]: chain_titles = [_["title"] for _ in chain] # title_set = frozenset(chain_titles) # if len(title_set) == 0 or title_set in chain_seen: # continue # chain_seen.add(title_set) if sp_titles: label = int(set(chain_titles) == sp_titles) else: label = -1 self.data.append({ "question": item["question"], "passages": chain, "label": label, "qid": item["_id"], "gold_answer": gold_answer, "sp_gold": sp_gold }) print(f"Data size {len(self.data)}") def __len__(self): return len(self.data) def __getitem__(self, index): item = prepare(self.data[index], self.tokenizer) context_ann = item["context_processed"] q_toks = self.tokenizer.tokenize(item["question"])[:self.max_q_len] para_offset = len(q_toks) + 2 # cls and seq item["wp_tokens"] = context_ann["all_doc_tokens"] assert item["wp_tokens"][0] == "yes" and item["wp_tokens"][1] == "no" item["para_offset"] = para_offset max_toks_for_doc = self.max_seq_len - para_offset - 1 if len(item["wp_tokens"]) > max_toks_for_doc: item["wp_tokens"] = item["wp_tokens"][:max_toks_for_doc] 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) item["paragraph_mask"] = torch.zeros(item["encodings"]["input_ids"].size()).view(-1) item["paragraph_mask"][para_offset:-1] = 1 if self.train: # if item["label"] == 1: if item["ans_covered"]: if item["gold_answer"][0] == "yes": # ans_type = 0 starts, ends= [para_offset], [para_offset] elif item["gold_answer"][0] == "no": # ans_type = 1 starts, ends= [para_offset + 1], [para_offset + 1] else: # ans_type = 2 matched_spans = match_answer_span(context_ann["context"], item["gold_answer"], self.simple_tok) ans_starts, ans_ends= [], [] for span in matched_spans: char_starts = [i for i in range(len(context_ann["context"])) if context_ann["context"].startswith(span, i)] if len(char_starts) > 0: char_ends = [start + len(span) - 1 for start in char_starts] answer = {"text": span, "char_spans": list(zip(char_starts, char_ends))} ans_spans = find_ans_span_with_char_offsets( answer, context_ann["char_to_word_offset"], context_ann["doc_tokens"], context_ann["all_doc_tokens"], context_ann["orig_to_tok_index"], self.tokenizer) for s, e in ans_spans: ans_starts.append(s) ans_ends.append(e) starts, ends = [], [] for s, e in zip(ans_starts, ans_ends): if s >= len(item["wp_tokens"]): continue else: s = min(s, len(item["wp_tokens"]) - 1) + para_offset e = min(e, len(item["wp_tokens"]) - 1) + para_offset starts.append(s) ends.append(e) if len(starts) == 0: starts, ends = [-1], [-1] else: starts, ends= [-1], [-1] # ans_type = -1 item["starts"] = torch.LongTensor(starts) item["ends"] = torch.LongTensor(ends) # item["ans_type"] = torch.LongTensor([ans_type]) if item["label"]: assert len(item["sp_sent_labels"]) == len(item["context_processed"]["sent_starts"]) else: # # for answer extraction item["doc_tokens"] = context_ann["doc_tokens"] item["tok_to_orig_index"] = context_ann["tok_to_orig_index"] # filter sentence offsets exceeding max sequence length sent_labels, sent_offsets = [], [] for idx, s in enumerate(item["context_processed"]["sent_starts"]): if s >= len(item["wp_tokens"]): break if "sp_sent_labels" in item: sent_labels.append(item["sp_sent_labels"][idx]) sent_offsets.append(s + para_offset) assert item["encodings"]["input_ids"].view(-1)[s+para_offset] == self.tokenizer.convert_tokens_to_ids("[unused1]") # supporting fact label item["sent_offsets"] = sent_offsets item["sent_offsets"] = torch.LongTensor(item["sent_offsets"]) if self.train: item["sent_labels"] = sent_labels if len(sent_labels) != 0 else [0] * len(sent_offsets) item["sent_labels"] = torch.LongTensor(item["sent_labels"]) item["ans_covered"] = torch.LongTensor([item["ans_covered"]]) item["label"] = torch.LongTensor([item["label"]]) return item class MhopSampler(Sampler): """ Shuffle QA pairs not context, make sure data within the batch are from the same QA pair """ def __init__(self, data_source, num_neg=9, n_gpu=8): # for each QA pair, sample negative paragraphs self.qid2gold = data_source.qid2gold self.qid2neg = data_source.qid2neg self.neg_num = num_neg self.n_gpu = n_gpu self.all_qids = list(self.qid2gold.keys()) assert len(self.qid2gold) == len(self.qid2neg) self.q_num_per_epoch = len(self.qid2gold) - len(self.qid2gold) % self.n_gpu self._num_samples = self.q_num_per_epoch * (self.neg_num + 1) def __len__(self): return self._num_samples def __iter__(self): sample_indice = [] random.shuffle(self.all_qids) # when use shared-normalization, passages for each question should be on the same GPU qids_to_use = self.all_qids[:self.q_num_per_epoch] for qid in qids_to_use: neg_samples = self.qid2neg[qid] random.shuffle(neg_samples) sample_indice += self.qid2gold[qid] sample_indice += neg_samples[:self.neg_num] return iter(sample_indice) def qa_collate(samples, pad_id=0): if len(samples) == 0: return {} batch = { 'input_ids': collate_tokens([s["encodings"]['input_ids'] for s in samples], pad_id), 'attention_mask': collate_tokens([s["encodings"]['attention_mask'] for s in samples], 0), 'paragraph_mask': collate_tokens([s['paragraph_mask'] for s in samples], 0), 'label': collate_tokens([s["label"] for s in samples], -1), "sent_offsets": collate_tokens([s["sent_offsets"] for s in samples], 0), } # training labels if "starts" in samples[0]: batch["starts"] = collate_tokens([s['starts'] for s in samples], -1) batch["ends"] = collate_tokens([s['ends'] for s in samples], -1) # batch["ans_types"] = collate_tokens([s['ans_type'] for s in samples], -1) batch["sent_labels"] = collate_tokens([s['sent_labels'] for s in samples], 0) batch["ans_covered"] = collate_tokens([s['ans_covered'] for s in samples], 0) # roberta does not use token_type_ids if "token_type_ids" in samples[0]["encodings"]: batch["token_type_ids"] = collate_tokens([s["encodings"]['token_type_ids']for s in samples], 0) batched = { "qids": [s["qid"] for s in samples], "passages": [s["passages"] for s in samples], "gold_answer": [s["gold_answer"] for s in samples], "sp_gold": [s["sp_gold"] for s in samples], "para_offsets": [s["para_offset"] for s in samples], "net_inputs": batch, } # for answer extraction if "doc_tokens" in samples[0]: batched["doc_tokens"] = [s["doc_tokens"] for s in samples] batched["tok_to_orig_index"] = [s["tok_to_orig_index"] for s in samples] batched["wp_tokens"] = [s["wp_tokens"] for s in samples] return batched ================================================ FILE: mdr/qa/qa_model.py ================================================ # Copyright (c) Facebook, Inc. and its affiliates. # All rights reserved. # # This source code is licensed under the license found in the # LICENSE file in the root directory of this source tree. from transformers import AutoModel, BertModel import torch.nn as nn from torch.nn import CrossEntropyLoss import torch import torch.nn.functional as F class BertPooler(nn.Module): def __init__(self, config): super().__init__() self.dense = nn.Linear(config.hidden_size, config.hidden_size) self.activation = nn.Tanh() def forward(self, hidden_states): # We "pool" the model by simply taking the hidden state corresponding # to the first token. first_token_tensor = hidden_states[:, 0] pooled_output = self.dense(first_token_tensor) pooled_output = self.activation(pooled_output) return pooled_output class QAModel(nn.Module): def __init__(self, config, args ): super().__init__() self.model_name = args.model_name self.sp_weight = args.sp_weight self.sp_pred = args.sp_pred self.encoder = AutoModel.from_pretrained(args.model_name) if "electra" in args.model_name: self.pooler = BertPooler(config) self.qa_outputs = nn.Linear(config.hidden_size, 2) self.rank = nn.Linear(config.hidden_size, 1) # noan if self.sp_pred: self.sp = nn.Linear(config.hidden_size, 1) self.loss_fct = CrossEntropyLoss(ignore_index=-1, reduction="none") def forward(self, batch): outputs = self.encoder(batch['input_ids'], batch['attention_mask'], batch.get('token_type_ids', None)) if "electra" in self.model_name: sequence_output = outputs[0] pooled_output = self.pooler(sequence_output) else: sequence_output, pooled_output = outputs[0], outputs[1] logits = self.qa_outputs(sequence_output) outs = [o.squeeze(-1) for o in logits.split(1, dim=-1)] outs = [o.float().masked_fill(batch["paragraph_mask"].ne(1), float("-inf")).type_as(o) for o in outs] start_logits, end_logits = outs[0], outs[1] rank_score = self.rank(pooled_output) if self.sp_pred: gather_index = batch["sent_offsets"].unsqueeze(2).expand(-1, -1, sequence_output.size()[-1]) sent_marker_rep = torch.gather(sequence_output, 1, gather_index) sp_score = self.sp(sent_marker_rep).squeeze(2) else: sp_score = None if self.training: rank_target = batch["label"] if self.sp_pred: sp_loss = F.binary_cross_entropy_with_logits(sp_score, batch["sent_labels"].float(), reduction="none") sp_loss = (sp_loss * batch["sent_offsets"]) * batch["label"] sp_loss = sp_loss.sum() start_positions, end_positions = batch["starts"], batch["ends"] rank_loss = F.binary_cross_entropy_with_logits(rank_score, rank_target.float(), reduction="sum") start_losses = [self.loss_fct(start_logits, starts) for starts in torch.unbind(start_positions, dim=1)] end_losses = [self.loss_fct(end_logits, ends) for ends in torch.unbind(end_positions, dim=1)] 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) log_prob = - loss_tensor log_prob = log_prob.float().masked_fill(log_prob == 0, float('-inf')).type_as(log_prob) marginal_probs = torch.sum(torch.exp(log_prob), dim=1) m_prob = [marginal_probs[idx] for idx in marginal_probs.nonzero()] if len(m_prob) == 0: span_loss = self.loss_fct(start_logits, start_logits.new_zeros( start_logits.size(0)).long()-1).sum() else: span_loss = - torch.log(torch.cat(m_prob)).sum() if self.sp_pred: loss = rank_loss + span_loss + sp_loss * self.sp_weight else: loss = rank_loss + span_loss return loss.unsqueeze(0) return { 'start_logits': start_logits, 'end_logits': end_logits, 'rank_score': rank_score, "sp_score": sp_score } ================================================ FILE: mdr/qa/qa_trainer.py ================================================ # Copyright (c) Facebook, Inc. and its affiliates. # All rights reserved. # # This source code is licensed under the license found in the # LICENSE file in the root directory of this source tree. import argparse import json import os import os.path as osp import random from functools import partial from pathlib import Path from typing import NamedTuple, Optional import collections from torch.optim import lr_scheduler from tqdm import tqdm import apex import attr import numpy as np import submitit import torch import torch.distributed import torch.nn as nn import torch.optim as optim from apex import amp from torch.utils.tensorboard import SummaryWriter from transformers import (AdamW, AutoConfig, AutoTokenizer, get_linear_schedule_with_warmup) from config import ClusterConfig from hotpot_evaluate_v1 import exact_match_score, f1_score, update_sp from qa_model import QAModel from reranking_datasets import RankingDataset, rank_collate, MhopSampler from utils import AverageMeter, move_to_cuda, get_final_text apex.amp.register_half_function(torch, 'einsum') @attr.s(auto_attribs=True) class TrainerState: """ Contains the state of the Trainer. It can be saved to checkpoint the training and loaded to resume it. """ epoch: int model: nn.Module optimizer: optim.Optimizer lr_scheduler: torch.optim.lr_scheduler._LRScheduler global_step: int def save(self, filename: str) -> None: data = attr.asdict(self) # store only the state dict data["model"] = self.model.state_dict() data["optimizer"] = self.optimizer.state_dict() data["lr_scheduler"] = self.lr_scheduler.state_dict() torch.save(data, filename) @classmethod def load(cls, filename: str, default: "TrainerState", gpu: int) -> "TrainerState": data = torch.load(filename, map_location=lambda storage, loc: storage.cuda(gpu)) # We need this default to load the state dict model = default.model model.load_state_dict(data["model"]) data["model"] = model optimizer = default.optimizer optimizer.load_state_dict(data["optimizer"]) data["optimizer"] = optimizer lr_scheduler = default.lr_scheduler lr_scheduler.load_state_dict(data["lr_scheduler"]) data["lr_scheduler"] = lr_scheduler return cls(**data) class Trainer: def __init__(self, train_cfg: NamedTuple, cluster_cfg: ClusterConfig) -> None: self._train_cfg = train_cfg self._cluster_cfg = cluster_cfg def __call__(self) -> Optional[float]: """ Called by submitit for each task. :return: The master task return the final accuracy of the model. """ self._setup_process_group() self._init_state() final_acc = self._train() return final_acc def log(self, log_data: dict): job_env = submitit.JobEnvironment() # z = {**vars(self._train_cfg), **log_data} save_dir = Path(self._train_cfg.output_dir) os.makedirs(save_dir, exist_ok=True) with open(save_dir / 'log.txt', 'a') as f: f.write(json.dumps(log_data) + '\n') def checkpoint(self, rm_init=True) -> submitit.helpers.DelayedSubmission: # will be called by submitit in case of preemption job_env = submitit.JobEnvironment() save_dir = osp.join(self._train_cfg.output_dir, str(job_env.job_id)) os.makedirs(save_dir, exist_ok=True) self._state.save(osp.join(save_dir, "checkpoint.pth")) # Trick here: when the job will be requeue, we will use the same init file # but it must not exist when we initialize the process group # so we delete it, but only when this method is called by submitit for requeue if rm_init and osp.exists(self._cluster_cfg.dist_url[7:]): os.remove(self._cluster_cfg.dist_url[7:]) # remove file:// at the beginning # This allow to remove any non-pickable part of the Trainer instance. empty_trainer = Trainer(self._train_cfg, self._cluster_cfg) return submitit.helpers.DelayedSubmission(empty_trainer) def _setup_process_group(self) -> None: job_env = submitit.JobEnvironment() torch.cuda.set_device(job_env.local_rank) torch.distributed.init_process_group( backend=self._cluster_cfg.dist_backend, init_method=self._cluster_cfg.dist_url, world_size=job_env.num_tasks, rank=job_env.global_rank, ) print(f"Process group: {job_env.num_tasks} tasks, rank: {job_env.global_rank}") def _init_state(self) -> None: """ Initialize the state and load it from an existing checkpoint if any """ job_env = submitit.JobEnvironment() if job_env.global_rank == 0: # config_path = Path(args.save_folder) / str(job_env.job_id) / 'config.json' os.makedirs(self._train_cfg.output_dir, exist_ok=True) config_path = Path(self._train_cfg.output_dir) / 'config.json' with open(config_path, "w") as g: g.write(json.dumps(self._train_cfg._asdict())) print(f"Setting random seed {self._train_cfg.seed}", flush=True) random.seed(self._train_cfg.seed) np.random.seed(self._train_cfg.seed) torch.manual_seed(self._train_cfg.seed) print("Create data loaders", flush=True) tokenizer = AutoTokenizer.from_pretrained(self._train_cfg.model_name) collate_fc = partial(rank_collate, pad_id=tokenizer.pad_token_id) train_set = RankingDataset(tokenizer, self._train_cfg.train_file, self._train_cfg.max_seq_len, self._train_cfg.max_q_len, train=True) train_sampler = MhopSampler(train_set, num_neg=self._train_cfg.neg_num) batch_size_per_gpu = (1 + self._train_cfg.neg_num) * self._train_cfg.num_q_per_gpu n_gpu = torch.cuda.device_count() print(f"Number of GPUs: {n_gpu}", flush=True) print(f"Batch size per node: {batch_size_per_gpu * n_gpu}", flush=True) 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) test_set = RankingDataset(tokenizer, self._train_cfg.predict_file, self._train_cfg.max_seq_len, self._train_cfg.max_q_len) self._test_loader = torch.utils.data.DataLoader( test_set, batch_size=self._train_cfg.predict_batch_size, num_workers=self._train_cfg.num_workers, collate_fn=collate_fc ) print("Create model", flush=True) print(f"Local rank {job_env.local_rank}", flush=True) bert_config = AutoConfig.from_pretrained(self._train_cfg.model_name) model = QAModel(bert_config, self._train_cfg) model.cuda(job_env.local_rank) no_decay = ['bias', 'LayerNorm.weight'] optimizer_parameters = [ {'params': [p for n, p in model.named_parameters() if not any( nd in n for nd in no_decay)], 'weight_decay': self._train_cfg.weight_decay}, {'params': [p for n, p in model.named_parameters() if any( nd in n for nd in no_decay)], 'weight_decay': 0.0} ] if self._train_cfg.use_adam: optimizer = optim.Adam(optimizer_parameters, lr=self._train_cfg.learning_rate) else: optimizer = AdamW(optimizer_parameters, lr=self._train_cfg.learning_rate) # lr_scheduler = optim.lr_scheduler.ReduceLROnPlateau(optimizer, mode='max', factor=0.5, patience=2) if self._train_cfg.fp16: model, optimizer = amp.initialize( model, optimizer, opt_level=self._train_cfg.fp16_opt_level) t_total = len(self._train_loader) // self._train_cfg.gradient_accumulation_steps * self._train_cfg.num_train_epochs warmup_steps = t_total * self._train_cfg.warmup_ratio lr_scheduler = get_linear_schedule_with_warmup( optimizer, num_warmup_steps=warmup_steps, num_training_steps=t_total ) model = torch.nn.DataParallel(model) self._state = TrainerState( epoch=0, model=model, optimizer=optimizer, lr_scheduler=lr_scheduler, global_step=0 ) self.tb_logger = SummaryWriter(self._train_cfg.output_dir.replace("logs", "tflogs")) checkpoint_fn = osp.join(self._train_cfg.output_dir, str(job_env.job_id), "checkpoint.pth") # checkpoint_fn = osp.join(self._train_cfg.output_dir, "checkpoint.pth") if os.path.isfile(checkpoint_fn): print(f"Load existing checkpoint from {checkpoint_fn}", flush=True) self._state = TrainerState.load( checkpoint_fn, default=self._state, gpu=job_env.local_rank) def _train(self) -> Optional[float]: job_env = submitit.JobEnvironment() batch_step = 0 # forward batch count best_metric = 0 train_loss_meter = AverageMeter() print(f"Start training", flush=True) # Start from the loaded epoch start_epoch = self._state.epoch global_step = self._state.global_step for epoch in range(start_epoch, self._train_cfg.num_train_epochs): print(f"Start epoch {epoch}", flush=True) self._state.model.train() self._state.epoch = epoch for batch in self._train_loader: batch_step += 1 batch_inputs = move_to_cuda(batch["net_inputs"]) loss = self._state.model(batch_inputs) if torch.cuda.device_count() > 1: loss = loss.mean() if self._train_cfg.gradient_accumulation_steps > 1: loss = loss / self._train_cfg.gradient_accumulation_steps if self._train_cfg.fp16: with amp.scale_loss(loss, self._state.optimizer) as scaled_loss: scaled_loss.backward() else: loss.backward() train_loss_meter.update(loss.item()) if (batch_step + 1) % self._train_cfg.gradient_accumulation_steps == 0: if self._train_cfg.fp16: torch.nn.utils.clip_grad_norm_( amp.master_params(self._state.optimizer), self._train_cfg.max_grad_norm) else: torch.nn.utils.clip_grad_norm_( self._state.model.parameters(), self._train_cfg.max_grad_norm) self._state.optimizer.step() self._state.lr_scheduler.step() self._state.model.zero_grad() global_step += 1 self._state.global_step = global_step self.tb_logger.add_scalar('batch_train_loss', loss.item(), global_step) self.tb_logger.add_scalar('smoothed_train_loss', train_loss_meter.avg, global_step) if job_env.global_rank == 0: if self._train_cfg.eval_period != -1 and global_step % self._train_cfg.eval_period == 0: metrics = self._eval() for k, v in metrics.items(): self.tb_logger.add_scalar(k, v*100, global_step) score = metrics[self._train_cfg.final_metric] if best_metric < score: print("Saving model with best %s %.2f -> em %.2f" % (self._train_cfg.final_metric, best_metric*100, score*100), flush=True) torch.save(self._state.model.state_dict(), os.path.join(self._train_cfg.output_dir, f"checkpoint_best.pt")) best_metric = score # Checkpoint only on the master if job_env.global_rank == 0: self.checkpoint(rm_init=False) metrics = self._eval() for k, v in metrics.items(): self.tb_logger.add_scalar(k, v*100, global_step) score = metrics[self._train_cfg.final_metric] if best_metric < score: print("Saving model with best %s %.2f -> em %.2f" % (self._train_cfg.final_metric, best_metric*100, score*100), flush=True) torch.save(self._state.model.state_dict(), os.path.join(self._train_cfg.output_dir, f"checkpoint_best.pt")) best_metric = score self.log({ "best_score": best_metric, "curr_score": score, "smoothed_loss": train_loss_meter.avg, "epoch": epoch }) return best_metric def _eval(self) -> dict: print("Start evaluation of the model", flush=True) job_env = submitit.JobEnvironment() args = self._train_cfg eval_dataloader = self._test_loader model = self._state.model model.eval() id2result = collections.defaultdict(list) id2answer = collections.defaultdict(list) id2gold = {} id2goldsp = {} for batch in tqdm(eval_dataloader): batch_to_feed = move_to_cuda(batch["net_inputs"]) batch_qids = batch["qids"] batch_labels = batch["net_inputs"]["label"].view(-1).tolist() with torch.no_grad(): outputs = model(batch_to_feed) scores = outputs["rank_score"] scores = scores.view(-1).tolist() sp_scores = outputs["sp_score"] sp_scores = sp_scores.float().masked_fill(batch_to_feed["sent_offsets"].eq(0), float("-inf")).type_as(sp_scores) batch_sp_scores = sp_scores.sigmoid() # ans_type_predicted = torch.argmax(outputs["ans_type_logits"], dim=1).view(-1).tolist() outs = [outputs["start_logits"], outputs["end_logits"]] for qid, label, score in zip(batch_qids, batch_labels, scores): id2result[qid].append((label, score)) # answer prediction span_scores = outs[0][:, :, None] + outs[1][:, None] max_seq_len = span_scores.size(1) span_mask = np.tril(np.triu(np.ones((max_seq_len, max_seq_len)), 0), args.max_ans_len) span_mask = span_scores.data.new(max_seq_len, max_seq_len).copy_(torch.from_numpy(span_mask)) span_scores_masked = span_scores.float().masked_fill((1 - span_mask[None].expand_as(span_scores)).bool(), -1e10).type_as(span_scores) start_position = span_scores_masked.max(dim=2)[0].max(dim=1)[1] end_position = span_scores_masked.max(dim=2)[1].gather( 1, start_position.unsqueeze(1)).squeeze(1) answer_scores = span_scores_masked.max(dim=2)[0].max(dim=1)[0].tolist() para_offset = batch['para_offsets'] start_position_ = list( np.array(start_position.tolist()) - np.array(para_offset)) end_position_ = list( np.array(end_position.tolist()) - np.array(para_offset)) for idx, qid in enumerate(batch_qids): id2gold[qid] = batch["gold_answer"][idx] id2goldsp[qid] = batch["sp_gold"][idx] rank_score = scores[idx] sp_score = batch_sp_scores[idx].tolist() start = start_position_[idx] end = end_position_[idx] span_score = answer_scores[idx] tok_to_orig_index = batch['tok_to_orig_index'][idx] doc_tokens = batch['doc_tokens'][idx] wp_tokens = batch['wp_tokens'][idx] orig_doc_start = tok_to_orig_index[start] orig_doc_end = tok_to_orig_index[end] orig_tokens = doc_tokens[orig_doc_start:(orig_doc_end + 1)] tok_tokens = wp_tokens[start:end+1] tok_text = " ".join(tok_tokens) tok_text = tok_text.replace(" ##", "") tok_text = tok_text.replace("##", "") tok_text = tok_text.strip() tok_text = " ".join(tok_text.split()) orig_text = " ".join(orig_tokens) pred_str = get_final_text(tok_text, orig_text, do_lower_case=True, verbose_logging=False) pred_sp = [] passages = batch["passages"][idx] for passage, sent_offset in zip(passages, [0, len(passages[0]["sents"])]): for idx, _ in enumerate(passage["sents"]): try: if sp_score[idx + sent_offset] > 0.5: pred_sp.append([passage["title"], idx]) except: continue id2answer[qid].append((pred_str.strip(), rank_score, span_score, pred_sp)) acc = [] for qid, res in id2result.items(): res.sort(key=lambda x: x[1], reverse=True) acc.append(res[0][0] == 1) print(f"evaluated {len(id2result)} questions...", flush=True) print(f'chain ranking em: {np.mean(acc)}', flush=True) best_em, best_f1, best_joint_em, best_joint_f1 = 0, 0, 0, 0 lambdas = [0, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1] for lambda_ in lambdas: ems, f1s = [], [] sp_ems, sp_f1s = [], [] joint_ems, joint_f1s = [], [] for qid, res in id2result.items(): ans_res = id2answer[qid] ans_res.sort(key=lambda x: lambda_ * x[1] + (1 - lambda_) * x[2], reverse=True) top_pred = ans_res[0][0] ems.append(exact_match_score(top_pred, id2gold[qid][0])) f1, prec, recall = f1_score(top_pred, id2gold[qid][0]) f1s.append(f1) top_pred_sp = ans_res[0][3] metrics = {'sp_em': 0, 'sp_f1': 0, 'sp_prec': 0, 'sp_recall': 0} update_sp(metrics, top_pred_sp, id2goldsp[qid]) sp_ems.append(metrics['sp_em']) sp_f1s.append(metrics['sp_f1']) # joint metrics joint_prec = prec * metrics["sp_prec"] joint_recall = recall * metrics["sp_recall"] if joint_prec + joint_recall > 0: joint_f1 = 2 * joint_prec * joint_recall / (joint_prec + joint_recall) else: joint_f1 = 0 joint_em = ems[-1] * sp_ems[-1] joint_ems.append(joint_em) joint_f1s.append(joint_f1) if best_joint_f1 < np.mean(joint_f1s): best_joint_f1 = np.mean(joint_f1s) best_joint_em = np.mean(joint_ems) best_f1 = np.mean(f1s) best_em = np.mean(ems) print(f".......Using combination factor {lambda_}......", flush=True) print(f'answer em: {np.mean(ems)}, count: {len(ems)}', flush=True) print(f'answer f1: {np.mean(f1s)}, count: {len(f1s)}', flush=True) print(f'sp em: {np.mean(sp_ems)}, count: {len(sp_ems)}', flush=True) print(f'sp f1: {np.mean(sp_f1s)}, count: {len(sp_f1s)}', flush=True) print(f'joint em: {np.mean(joint_ems)}, count: {len(joint_ems)}', flush=True) print(f'joint f1: {np.mean(joint_f1s)}, count: {len(joint_f1s)}', flush=True) print(f"Best joint EM/F1 from combination {best_em}/{best_f1}", flush=True) model.train() return {"em": best_em, "f1": best_f1, "joint_em": best_joint_em, "joint_f1": best_joint_f1} ================================================ FILE: mdr/qa/train.md ================================================ CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python train_qa.py \ --do_train \ --prefix qa_wwm_bert_title_mark_eval_debug \ --predict_batch_size 512 \ --model_name bert-large-uncased-whole-word-masking \ --train_batch_size 80 \ --learning_rate 3e-5 \ --fp16 \ --train_file /private/home/xwhan/data/hotpot/dense_train_b10_top20_outputs.json \ --predict_file /private/home/xwhan/data/hotpot/dense_val_outputs.json \ --seed 3 \ --eval-period 10 \ --max_seq_len 512 \ --max_q_len 100 \ --gradient_accumulation_steps 8 \ --neg-num 4 # spanbert debug, fp16 does not work CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python train_qa.py \ --do_train \ --prefix ranked_spanbert_debug \ --predict_batch_size 1024 \ --model_name spanbert \ --train_batch_size 48 \ --learning_rate 3e-5 \ --train_file /private/home/xwhan/data/hotpot/dense_train_b10_top20_outputs_sents.json \ --predict_file /private/home/xwhan/data/hotpot/dense_val_outputs_sents.json \ --seed 3 \ --eval-period 500 \ --max_seq_len 512 \ --max_q_len 64 \ --gradient_accumulation_steps 8 \ --neg-num 5 \ --use-adam # test electra CUDA_VISIBLE_DEVICES=0 python train_qa.py \ --do_train \ --prefix electra_large_debug_sn \ --predict_batch_size 1024 \ --model_name google/electra-large-discriminator \ --train_batch_size 12 \ --learning_rate 5e-5 \ --train_file /private/home/xwhan/data/hotpot/dense_train_b100_k100_sents.json \ --predict_file /private/home/xwhan/data/hotpot/dense_val_b30_k30_roberta_sents.json \ --seed 42 \ --eval-period 250 \ --max_seq_len 512 \ --max_q_len 64 \ --gradient_accumulation_steps 8 \ --neg-num 11 \ --fp16 \ --use-adam \ --warmup-ratio 0.1 \ --sp-weight 0.05 \ --sp-pred \ --shared-norm # QA evaluation CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python train_qa.py \ --do_predict \ --predict_batch_size 2000 \ --model_name google/electra-large-discriminator \ --fp16 \ --predict_file /private/home/xwhan/data/hotpot/dense_val_b100_k100_roberta_best_sents.json \ --max_seq_len 512 \ --max_q_len 64 \ --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 \ --sp-pred \ --max_ans_len 30 \ --save-prediction hotpot_val_top100.json # QA evaluation with wwm CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python train_qa.py \ --do_predict \ --predict_batch_size 1024 \ --model_name bert-large-uncased-whole-word-masking \ --fp16 \ --predict_file /private/home/xwhan/data/hotpot/dense_hotpot_val_b250_k250_roberta_best_sents.json \ --max_seq_len 512 \ --max_q_len 64 \ --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 \ --sp-pred \ --max_ans_len 30 \ --save-prediction hotpot_val_wwm_top250.json CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python train_qa.py \ --do_predict \ --predict_batch_size 1024 \ --model_name google/electra-large-discriminator \ --fp16 \ --predict_file /private/home/xwhan/data/hotpot/dense_val_b50_k50_roberta_best_sents.json \ --max_seq_len 512 \ --max_q_len 64 \ --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 \ --sp-pred \ --max_ans_len 30 \ --save-prediction hotpot_val_b5_k5.json \ srun --gres=gpu:8 --partition learnfair --time=48:00:00 --mem 500G --constraint volta32gb --cpus-per-task 80 --pty /bin/bash -l ================================================ FILE: mdr/qa/train_ranker.py ================================================ # Copyright (c) Facebook, Inc. and its affiliates. # All rights reserved. # # This source code is licensed under the license found in the # LICENSE file in the root directory of this source tree. import collections import json import logging import os import random from datetime import date from functools import partial import copy import numpy as np import torch import torch.nn.functional as F from torch.utils.data import DataLoader from torch.utils.tensorboard import SummaryWriter from torch.optim import Adam from tqdm import tqdm from transformers import (AdamW, AutoConfig, AutoTokenizer, get_linear_schedule_with_warmup) from config import train_args from reranking_datasets import RankingDataset, rank_collate from reranking_model import RankModel from utils import AverageMeter, convert_to_half, move_to_cuda def load_saved(model, path): state_dict = torch.load(path) def filter(x): return x[7:] if x.startswith('module.') else x state_dict = {filter(k): v for (k, v) in state_dict.items()} model.load_state_dict(state_dict) return model def main(): args = train_args() if args.fp16: import apex apex.amp.register_half_function(torch, 'einsum') date_curr = date.today().strftime("%m-%d-%Y") model_name = f"{args.prefix}-seed{args.seed}-bsz{args.train_batch_size}-fp16{args.fp16}-lr{args.learning_rate}-decay{args.weight_decay}" args.output_dir = os.path.join(args.output_dir, date_curr, model_name) tb_logger = SummaryWriter(os.path.join(args.output_dir.replace("logs","tflogs"))) if os.path.exists(args.output_dir) and os.listdir(args.output_dir): print( f"output directory {args.output_dir} already exists and is not empty.") if not os.path.exists(args.output_dir): os.makedirs(args.output_dir, exist_ok=True) logging.basicConfig(format='%(asctime)s - %(levelname)s - %(name)s - %(message)s', datefmt='%m/%d/%Y %H:%M:%S', level=logging.INFO, handlers=[logging.FileHandler(os.path.join(args.output_dir, "log.txt")), logging.StreamHandler()]) logger = logging.getLogger(__name__) logger.info(args) if args.local_rank == -1 or args.no_cuda: device = torch.device( "cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu") n_gpu = torch.cuda.device_count() else: device = torch.device("cuda", args.local_rank) n_gpu = 1 # Initializes the distributed backend which will take care of sychronizing nodes/GPUs torch.distributed.init_process_group(backend='nccl') logger.info("device %s n_gpu %d distributed training %r", device, n_gpu, bool(args.local_rank != -1)) if args.accumulate_gradients < 1: raise ValueError("Invalid accumulate_gradients parameter: {}, should be >= 1".format( args.accumulate_gradients)) args.train_batch_size = int( args.train_batch_size / args.accumulate_gradients) random.seed(args.seed) np.random.seed(args.seed) torch.manual_seed(args.seed) if n_gpu > 0: torch.cuda.manual_seed_all(args.seed) bert_config = AutoConfig.from_pretrained(args.model_name) model = RankModel(bert_config, args) tokenizer = AutoTokenizer.from_pretrained(args.model_name) collate_fc = partial(rank_collate, pad_id=tokenizer.pad_token_id) if args.do_train and args.max_seq_len > bert_config.max_position_embeddings: raise ValueError( "Cannot use sequence length %d because the BERT model " "was only trained up to sequence length %d" % (args.max_seq_len, bert_config.max_position_embeddings)) eval_dataset = RankingDataset( tokenizer, args.predict_file, args.max_seq_len, args.max_q_len) eval_dataloader = DataLoader( eval_dataset, batch_size=args.predict_batch_size, collate_fn=collate_fc, pin_memory=True, num_workers=args.num_workers) logger.info(f"Num of dev batches: {len(eval_dataloader)}") if args.init_checkpoint != "": model = load_saved(model, args.init_checkpoint) model.to(device) print(f"number of trainable parameters: {sum(p.numel() for p in model.parameters() if p.requires_grad)}") if args.do_train: no_decay = ['bias', 'LayerNorm.weight'] optimizer_parameters = [ {'params': [p for n, p in model.named_parameters() if not any( nd in n for nd in no_decay)], 'weight_decay': args.weight_decay}, {'params': [p for n, p in model.named_parameters() if any( nd in n for nd in no_decay)], 'weight_decay': 0.0} ] optimizer = AdamW(optimizer_parameters, lr=args.learning_rate, eps=args.adam_epsilon) if args.fp16: from apex import amp model, optimizer = amp.initialize( model, optimizer, opt_level=args.fp16_opt_level) else: if args.fp16: from apex import amp model = amp.initialize(model, opt_level=args.fp16_opt_level) if args.local_rank != -1: model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[args.local_rank], output_device=args.local_rank) elif n_gpu > 1: model = torch.nn.DataParallel(model) if args.do_train: global_step = 0 # gradient update step batch_step = 0 # forward batch count best_acc = 0 train_loss_meter = AverageMeter() model.train() train_dataset = RankingDataset(tokenizer, args.train_file, args.max_seq_len, args.max_q_len, train=True) 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) logger.info('Start training....') for epoch in range(int(args.num_train_epochs)): for batch in tqdm(train_dataloader): batch_step += 1 batch_inputs = move_to_cuda(batch["net_inputs"]) loss = model(batch_inputs) if n_gpu > 1: loss = loss.mean() if args.gradient_accumulation_steps > 1: loss = loss / args.gradient_accumulation_steps if args.fp16: with amp.scale_loss(loss, optimizer) as scaled_loss: scaled_loss.backward() else: loss.backward() train_loss_meter.update(loss.item()) tb_logger.add_scalar('batch_train_loss', loss.item(), global_step) tb_logger.add_scalar('smoothed_train_loss', train_loss_meter.avg, global_step) if (batch_step + 1) % args.gradient_accumulation_steps == 0: if args.fp16: torch.nn.utils.clip_grad_norm_( amp.master_params(optimizer), args.max_grad_norm) else: torch.nn.utils.clip_grad_norm_( model.parameters(), args.max_grad_norm) optimizer.step() # We have accumulated enought gradients model.zero_grad() global_step += 1 if args.eval_period != -1 and global_step % args.eval_period == 0: acc = predict(args, model, eval_dataloader, device, logger) logger.info("Step %d Train loss %.2f acc %.2f on epoch=%d" % (global_step, train_loss_meter.avg, acc*100, epoch)) # save most recent model torch.save(model.state_dict(), os.path.join( args.output_dir, f"checkpoint_last.pt")) if best_acc < acc: logger.info("Saving model with best acc %.2f -> acc %.2f on epoch=%d" % (best_acc*100, acc*100, epoch)) torch.save(model.state_dict(), os.path.join( args.output_dir, f"checkpoint_best.pt")) model = model.to(device) best_acc = acc acc = predict(args, model, eval_dataloader, device, logger) logger.info("Step %d Train loss %.2f acc %.2f on epoch=%d" % ( global_step, train_loss_meter.avg, acc*100, epoch)) tb_logger.add_scalar('dev_acc', acc*100, epoch) torch.save(model.state_dict(), os.path.join(args.output_dir, f"checkpoint_last.pt")) if best_acc < acc: logger.info("Saving model with best acc %.2f -> acc %.2f on epoch=%d" % (best_acc*100, acc*100, epoch)) torch.save(model.state_dict(), os.path.join( args.output_dir, f"checkpoint_best.pt")) best_acc = acc logger.info("Training finished!") elif args.do_predict: acc = predict(args, model, eval_dataloader, device, logger) logger.info(f"test performance {acc}") def predict(args, model, eval_dataloader, device, logger): model.eval() id2result = collections.defaultdict(list) for batch in tqdm(eval_dataloader): batch_to_feed = move_to_cuda(batch["net_inputs"]) batch_qids = batch["qids"] batch_labels = batch["net_inputs"]["label"].view(-1).tolist() with torch.no_grad(): scores = model(batch_to_feed) scores = scores.view(-1).tolist() for qid, label, score in zip(batch_qids, batch_labels, scores): id2result[qid].append((label, score)) acc = [] top_pred = {} for qid, res in id2result.items(): res.sort(key=lambda x: x[1], reverse=True) acc.append(res[0][0] == 1) logger.info(f"evaluated {len(id2result)} questions...") logger.info(f'acc: {np.mean(acc)}') model.train() return np.mean(acc) if __name__ == "__main__": main() ================================================ FILE: mdr/qa/utils.py ================================================ # Copyright (c) Facebook, Inc. and its affiliates. # All rights reserved. # # This source code is licensed under the license found in the # LICENSE file in the root directory of this source tree. import torch import sqlite3 import unicodedata import collections import logging import re def set_global_logging_level(level=logging.ERROR, prefices=[""]): """ Override logging levels of different modules based on their name as a prefix. It needs to be invoked after the modules have been loaded so that their loggers have been initialized. Args: - level: desired level. e.g. logging.INFO. Optional. Default is logging.ERROR - prefices: list of one or more str prefices to match (e.g. ["transformers", "torch"]). Optional. Default is `[""]` to match all active loggers. The match is a case-sensitive `module_name.startswith(prefix)` """ prefix_re = re.compile(fr'^(?:{ "|".join(prefices) })') for name in logging.root.manager.loggerDict: if re.match(prefix_re, name): logging.getLogger(name).setLevel(level) def load_saved(model, path, exact=True): try: state_dict = torch.load(path) except: state_dict = torch.load(path, map_location=torch.device('cpu')) def filter(x): return x[7:] if x.startswith('module.') else x if exact: state_dict = {filter(k): v for (k, v) in state_dict.items()} else: state_dict = {filter(k): v for ( k, v) in state_dict.items() if filter(k) in model.state_dict()} model.load_state_dict(state_dict) return model def move_to_cuda(sample): if len(sample) == 0: return {} def _move_to_cuda(maybe_tensor): if torch.is_tensor(maybe_tensor): return maybe_tensor.cuda() elif isinstance(maybe_tensor, dict): return { key: _move_to_cuda(value) for key, value in maybe_tensor.items() } elif isinstance(maybe_tensor, list): return [_move_to_cuda(x) for x in maybe_tensor] else: return maybe_tensor return _move_to_cuda(sample) def convert_to_half(sample): if len(sample) == 0: return {} def _convert_to_half(maybe_floatTensor): if torch.is_tensor(maybe_floatTensor) and maybe_floatTensor.type() == "torch.FloatTensor": return maybe_floatTensor.half() elif isinstance(maybe_floatTensor, dict): return { key: _convert_to_half(value) for key, value in maybe_floatTensor.items() } elif isinstance(maybe_floatTensor, list): return [_convert_to_half(x) for x in maybe_floatTensor] else: return maybe_floatTensor return _convert_to_half(sample) class AverageMeter(object): """Computes and stores the average and current value""" def __init__(self): self.reset() def reset(self): self.val = 0 self.avg = 0 self.sum = 0 self.count = 0 def update(self, val, n=1): self.val = val self.sum += val * n self.count += n self.avg = self.sum / self.count def normalize(text): """Resolve different type of unicode encodings.""" return unicodedata.normalize('NFD', text) def para_has_answer(answer, para, tokenizer): text = normalize(para) tokens = tokenizer.tokenize(text) text = tokens.words(uncased=True) assert len(text) == len(tokens) for single_answer in answer: single_answer = normalize(single_answer) single_answer = tokenizer.tokenize(single_answer) single_answer = single_answer.words(uncased=True) for i in range(0, len(text) - len(single_answer) + 1): if single_answer == text[i: i + len(single_answer)]: return True return False def match_answer_span(p, answer, tokenizer, match="string"): # p has been normalized if match == 'string': tokens = tokenizer.tokenize(p) text = tokens.words(uncased=True) matched = set() for single_answer in answer: single_answer = normalize(single_answer) single_answer = tokenizer.tokenize(single_answer) single_answer = single_answer.words(uncased=True) for i in range(0, len(text) - len(single_answer) + 1): if single_answer == text[i: i + len(single_answer)]: matched.add(tokens.slice( i, i + len(single_answer)).untokenize()) return list(matched) elif match == 'regex': # Answer is a regex single_answer = normalize(answer[0]) return regex_match(p, single_answer) def _is_whitespace(char): """Checks whether `chars` is a whitespace character.""" # \t, \n, and \r are technically contorl characters but we treat them # as whitespace since they are generally considered as such. if char == " " or char == "\t" or char == "\n" or char == "\r": return True cat = unicodedata.category(char) if cat == "Zs": return True return False def _improve_answer_span(doc_tokens, input_start, input_end, tokenizer, orig_answer_text): tok_answer_text = " ".join(tokenizer.tokenize(orig_answer_text)) for new_start in range(input_start, input_end + 1): for new_end in range(input_end, new_start - 1, -1): text_span = " ".join(doc_tokens[new_start:(new_end + 1)]) if text_span == tok_answer_text: return (new_start, new_end) return (input_start, input_end) def whitespace_tokenize(text): """Runs basic whitespace cleaning and splitting on a peice of text.""" text = text.strip() if not text: return [] tokens = text.split() return tokens def find_ans_span_with_char_offsets(detected_ans, char_to_word_offset, doc_tokens, all_doc_tokens, orig_to_tok_index, tokenizer): # could return mutiple spans for an answer string ans_text = detected_ans["text"] char_spans = detected_ans["char_spans"] ans_subtok_spans = [] for char_start, char_end in char_spans: tok_start = char_to_word_offset[char_start] # char_end points to the last char of the answer, not one after tok_end = char_to_word_offset[char_end] sub_tok_start = orig_to_tok_index[tok_start] if tok_end < len(doc_tokens) - 1: sub_tok_end = orig_to_tok_index[tok_end + 1] - 1 else: sub_tok_end = len(all_doc_tokens) - 1 actual_text = " ".join(doc_tokens[tok_start:(tok_end + 1)]) cleaned_answer_text = " ".join(whitespace_tokenize(ans_text)) if actual_text.find(cleaned_answer_text) == -1: print("Could not find answer: '{}' vs. '{}'".format( actual_text, cleaned_answer_text)) (sub_tok_start, sub_tok_end) = _improve_answer_span( all_doc_tokens, sub_tok_start, sub_tok_end, tokenizer, ans_text) ans_subtok_spans.append((sub_tok_start, sub_tok_end)) return ans_subtok_spans import six def convert_to_unicode(text): """Converts `text` to Unicode (if it's not already), assuming utf-8 input.""" if six.PY3: if isinstance(text, str): return text elif isinstance(text, bytes): return text.decode("utf-8", "ignore") else: raise ValueError("Unsupported string type: %s" % (type(text))) elif six.PY2: if isinstance(text, str): return text.decode("utf-8", "ignore") elif isinstance(text, unicode): return text else: raise ValueError("Unsupported string type: %s" % (type(text))) else: raise ValueError("Not running on Python2 or Python 3?") def _is_control(char): """Checks whether `chars` is a control character.""" # These are technically control characters but we count them as whitespace # characters. if char == "\t" or char == "\n" or char == "\r": return False cat = unicodedata.category(char) if cat.startswith("C"): return True return False def _is_punctuation(char): """Checks whether `chars` is a punctuation character.""" cp = ord(char) # We treat all non-letter/number ASCII as punctuation. # Characters such as "^", "$", and "`" are not in the Unicode # Punctuation class but we treat them as punctuation anyways, for # consistency. if ((cp >= 33 and cp <= 47) or (cp >= 58 and cp <= 64) or (cp >= 91 and cp <= 96) or (cp >= 123 and cp <= 126)): return True cat = unicodedata.category(char) if cat.startswith("P"): return True return False class BasicTokenizer(object): """Runs basic tokenization (punctuation splitting, lower casing, etc.).""" def __init__(self, do_lower_case=True): """Constructs a BasicTokenizer. Args: do_lower_case: Whether to lower case the input. """ self.do_lower_case = do_lower_case def tokenize(self, text): """Tokenizes a piece of text.""" text = convert_to_unicode(text) text = self._clean_text(text) orig_tokens = whitespace_tokenize(text) split_tokens = [] for token in orig_tokens: if self.do_lower_case: token = token.lower() token = self._run_strip_accents(token) split_tokens.extend(self._run_split_on_punc(token)) output_tokens = whitespace_tokenize(" ".join(split_tokens)) return output_tokens def _run_strip_accents(self, text): """Strips accents from a piece of text.""" text = unicodedata.normalize("NFD", text) output = [] for char in text: cat = unicodedata.category(char) if cat == "Mn": continue output.append(char) return "".join(output) def _run_split_on_punc(self, text): """Splits punctuation on a piece of text.""" chars = list(text) i = 0 start_new_word = True output = [] while i < len(chars): char = chars[i] if _is_punctuation(char): output.append([char]) start_new_word = True else: if start_new_word: output.append([]) start_new_word = False output[-1].append(char) i += 1 return ["".join(x) for x in output] def _clean_text(self, text): """Performs invalid character removal and whitespace cleanup on text.""" output = [] for char in text: cp = ord(char) if cp == 0 or cp == 0xfffd or _is_control(char): continue if _is_whitespace(char): output.append(" ") else: output.append(char) return "".join(output) def get_final_text(pred_text, orig_text, do_lower_case=False, verbose_logging=True): """Project the tokenized prediction back to the original text.""" def _strip_spaces(text): ns_chars = [] ns_to_s_map = collections.OrderedDict() for (i, c) in enumerate(text): if c == " ": continue ns_to_s_map[len(ns_chars)] = i ns_chars.append(c) ns_text = "".join(ns_chars) return (ns_text, ns_to_s_map) # We first tokenize `orig_text`, strip whitespace from the result # and `pred_text`, and check if they are the same length. If they are # NOT the same length, the heuristic has failed. If they are the same # length, we assume the characters are one-to-one aligned. tokenizer = BasicTokenizer(do_lower_case=do_lower_case) tok_text = " ".join(tokenizer.tokenize(orig_text)) start_position = tok_text.find(pred_text) if start_position == -1: if verbose_logging: print( "Unable to find text: '%s' in '%s'" % (pred_text, orig_text)) return orig_text end_position = start_position + len(pred_text) - 1 (orig_ns_text, orig_ns_to_s_map) = _strip_spaces(orig_text) (tok_ns_text, tok_ns_to_s_map) = _strip_spaces(tok_text) if len(orig_ns_text) != len(tok_ns_text): if verbose_logging: print("Length not equal after stripping spaces: '%s' vs '%s'", orig_ns_text, tok_ns_text) return orig_text # We then project the characters in `pred_text` back to `orig_text` using # the character-to-character alignment. tok_s_to_ns_map = {} for (i, tok_index) in six.iteritems(tok_ns_to_s_map): tok_s_to_ns_map[tok_index] = i orig_start_position = None if start_position in tok_s_to_ns_map: ns_start_position = tok_s_to_ns_map[start_position] if ns_start_position in orig_ns_to_s_map: orig_start_position = orig_ns_to_s_map[ns_start_position] if orig_start_position is None: if verbose_logging: print("Couldn't map start position") return orig_text orig_end_position = None if end_position in tok_s_to_ns_map: ns_end_position = tok_s_to_ns_map[end_position] if ns_end_position in orig_ns_to_s_map: orig_end_position = orig_ns_to_s_map[ns_end_position] if orig_end_position is None: if verbose_logging: print("Couldn't map end position") return orig_text output_text = orig_text[orig_start_position:(orig_end_position + 1)] return output_text ================================================ FILE: mdr/retrieval/__init__.py ================================================ # Copyright (c) Facebook, Inc. and its affiliates. # All rights reserved. # # This source code is licensed under the license found in the # LICENSE file in the root directory of this source tree. #!/usr/bin/env python # Copyright 2017-present, Facebook, Inc. # All rights reserved. # # This source code is licensed under the license found in the # LICENSE file in the root directory of this source tree. from . import data from . import models from . import utils ================================================ FILE: mdr/retrieval/config.py ================================================ # Copyright (c) Facebook, Inc. and its affiliates. # All rights reserved. # # This source code is licensed under the license found in the # LICENSE file in the root directory of this source tree. import argparse from ast import parse from typing import NamedTuple class ClusterConfig(NamedTuple): dist_backend: str dist_url: str def common_args(): parser = argparse.ArgumentParser() # task parser.add_argument("--train_file", type=str, default="../data/nq-with-neg-train.txt") parser.add_argument("--predict_file", type=str, default="../data/nq-with-neg-dev.txt") parser.add_argument("--num_workers", default=30, type=int) parser.add_argument("--do_train", default=False, action='store_true', help="Whether to run training.") parser.add_argument("--do_predict", default=False, action='store_true', help="Whether to run eval on the dev set.") # model parser.add_argument("--model_name", default="bert-base-uncased", type=str) parser.add_argument("--init_checkpoint", type=str, help="Initial checkpoint (usually from a pre-trained BERT model).", default="") parser.add_argument("--max_c_len", default=512, type=int, help="The maximum total input sequence length after WordPiece tokenization. Sequences " "longer than this will be truncated, and sequences shorter than this will be padded.") parser.add_argument("--max_q_len", default=50, type=int, help="The maximum number of tokens for the question. Questions longer than this will " "be truncated to this length.") parser.add_argument('--fp16', action='store_true') parser.add_argument('--fp16_opt_level', type=str, default='O1', help="For fp16: Apex AMP optimization level selected in ['O0', 'O1', 'O2', and 'O3']." "See details at https://nvidia.github.io/apex/amp.html") parser.add_argument("--no_cuda", default=False, action='store_true', help="Whether not to use CUDA when available") parser.add_argument("--local_rank", type=int, default=-1, help="local_rank for distributed training on gpus") parser.add_argument("--max_q_sp_len", default=50, type=int) parser.add_argument("--sent-level", action="store_true") parser.add_argument("--rnn-retriever", action="store_true") parser.add_argument("--predict_batch_size", default=512, type=int, help="Total batch size for predictions.") parser.add_argument("--shared-encoder", action="store_true") # multi vector scheme parser.add_argument("--multi-vector", type=int, default=1) parser.add_argument("--scheme", type=str, help="how to get the multivector, layerwise or tokenwise", default="none") # momentum parser.add_argument("--momentum", action="store_true") parser.add_argument("--init-retriever", type=str, default="") parser.add_argument("--k", type=int, default=38400, help="memory bank size") parser.add_argument("--m", type=float, default=0.999, help="momentum") # NQ multihop trial parser.add_argument("--nq-multi", action="store_true", help="train the NQ retrieval model to recover from error cases") return parser def train_args(): parser = common_args() # optimization parser.add_argument('--prefix', type=str, default="eval") parser.add_argument("--weight_decay", default=0.0, type=float, help="Weight decay if we apply some.") parser.add_argument("--temperature", default=1, type=float) parser.add_argument("--output_dir", default="./logs", type=str, help="The output directory where the model checkpoints will be written.") parser.add_argument("--train_batch_size", default=128, type=int, help="Total batch size for training.") parser.add_argument("--learning_rate", default=1e-5, type=float, help="The initial learning rate for Adam.") parser.add_argument("--adam_epsilon", default=1e-8, type=float, help="Epsilon for Adam optimizer.") parser.add_argument("--num_train_epochs", default=50, type=float, help="Total number of training epochs to perform.") parser.add_argument("--save_checkpoints_steps", default=20000, type=int, help="How often to save the model checkpoint.") parser.add_argument("--iterations_per_loop", default=1000, type=int, help="How many steps to make in each estimator call.") parser.add_argument("--accumulate_gradients", type=int, default=1, help="Number of steps to accumulate gradient on (divide the batch_size and accumulate)") parser.add_argument('--seed', type=int, default=3, help="random seed for initialization") parser.add_argument('--gradient_accumulation_steps', type=int, default=1, help="Number of updates steps to accumualte before performing a backward/update pass.") parser.add_argument('--eval-period', type=int, default=2500) parser.add_argument("--max_grad_norm", default=2.0, type=float, help="Max gradient norm.") parser.add_argument("--stop-drop", default=0, type=float) parser.add_argument("--use-adam", action="store_true") parser.add_argument("--warmup-ratio", default=0, type=float, help="Linear warmup over warmup_steps.") return parser.parse_args() def encode_args(): parser = common_args() parser.add_argument('--embed_save_path', type=str, default="") parser.add_argument('--is_query_embed', action="store_true") args = parser.parse_args() return args ================================================ FILE: mdr/retrieval/criterions.py ================================================ # Copyright (c) Facebook, Inc. and its affiliates. # All rights reserved. # # This source code is licensed under the license found in the # LICENSE file in the root directory of this source tree. import torch from torch.nn import CrossEntropyLoss import torch.nn.functional as F # def loss_single(model, batch, momentum=False): # outputs = model(batch) # q = outputs['q'] # c = outputs['c'] # neg_c = outputs['neg_c'] # product_in_batch = torch.mm(q, c.t()) # product_neg = (q * neg_c).sum(-1).unsqueeze(1) # product = torch.cat([product_in_batch, product_neg], dim=-1) # if momentum: # queue_c = model.module.encode_queue_ctx() # product_queue = torch.mm(q, queue_c.t()) # product = torch.cat([product, product_queue], dim=-1) # model.module.dequeue_and_enqueue(batch) # target = torch.arange(product.size(0)).to(product.device) # loss = F.cross_entropy(product, target) # return loss # """ # multi-hop retrieval for NQ, train the model to recover from # """ # def loss_nq_mhop(model, batch, momentum=False): # outputs = model(batch) # product_in_batch = torch.mm(outputs['q'], outputs['c'].t()) # product_neg = (outputs['q'] * outputs['neg']).sum(-1).unsqueeze(1) # # product_neg1 = (outputs['q'] * outputs['dense_neg1']).sum(-1).unsqueeze(1) # # product_neg2 = (outputs['q'] * outputs['dense_neg2']).sum(-1).unsqueeze(1) # scores1 = torch.cat([product_in_batch, product_neg], dim=-1) # product_in_batch_from_error = torch.mm(outputs["q_neg1"], outputs['c'].t()) # dense_neg = torch.cat([outputs["dense_neg1"].unsqueeze(1), outputs["dense_neg2"].unsqueeze(1)], dim=1) # product_neg_from_error = torch.bmm(outputs["q_neg1"].unsqueeze(1), dense_neg.transpose(1,2)).squeeze(1) # scores2 = torch.cat([product_in_batch_from_error, product_neg_from_error], dim=-1) # if momentum: # queue_neg_scores_1 = torch.mm(outputs['q'], model.module.queue.clone().detach().t()) # queue_neg_scores_2 = torch.mm(outputs["q_neg1"], model.module.queue.clone().detach().t()) # scores1 = torch.cat([scores1, queue_neg_scores_1], dim=1) # scores2 = torch.cat([scores2, queue_neg_scores_2], dim=1) # model.module.dequeue_and_enqueue(outputs["c"].detach()) # # model.module.momentum_update_key_encoder() # target = torch.arange(scores1.size(0)).to(scores1.device) # loss = F.cross_entropy(scores1, target) + F.cross_entropy(scores2, target) # # loss = F.cross_entropy(scores1, target) # return loss # def eval_nq_mhop(model, batch): # outputs = model(batch) # product_in_batch = torch.mm(outputs['q'], outputs['c'].t()) # product_neg = (outputs['q'] * outputs['neg']).sum(-1).unsqueeze(1) # # product_neg1 = (outputs['q'] * outputs['dense_neg1']).sum(-1).unsqueeze(1) # # product_neg2 = (outputs['q'] * outputs['dense_neg2']).sum(-1).unsqueeze(1) # scores1 = torch.cat([product_in_batch, product_neg], dim=-1) # product_in_batch_from_error = torch.mm(outputs["q_neg1"], outputs['c'].t()) # dense_neg = torch.cat([outputs["dense_neg1"].unsqueeze(1), outputs["dense_neg2"].unsqueeze(1)], dim=1) # product_neg_from_error = torch.bmm(outputs["q_neg1"].unsqueeze(1), dense_neg.transpose(1,2)).squeeze(1) # scores2 = torch.cat([product_in_batch_from_error, product_neg_from_error], dim=-1) # target = torch.arange(scores1.size(0)).to(scores1.device) # rrs, rrs_2hop = [], [] # ranked = scores1.argsort(dim=1, descending=True) # ranked_2hop = scores2.argsort(dim=1, descending=True) # idx2rank = ranked.argsort(dim=1) # for idx, t in enumerate(target.tolist()): # rrs.append(1 / (idx2rank[idx][t].item() +1)) # idx2rank2hop = ranked_2hop.argsort(dim=1) # for idx, t in enumerate(target.tolist()): # rrs_2hop.append(1 / (idx2rank2hop[idx][t].item() +1)) # return rrs, rrs_2hop # def eval_vanilla(outputs): # """ # view the two sp passages as the same, no multi-hop modeling; # select the passages from all passages in the batch # """ # rrs = [] # q = outputs['q'] # c1 = outputs['c1'] # c2 = outputs['c2'] # c = torch.cat([c1.unsqueeze(1), c2.unsqueeze(1)], dim=1) # B x 2 x D # c = c.view(-1, q.size(-1)) # 2B x D # product_in_batch = torch.mm(q, c.t()) # Bx2B # neg_c = outputs['neg_c'] # product_neg = (q * neg_c).sum(-1).unsqueeze(1) # product = torch.cat([product_in_batch, product_neg], dim=-1) # target = torch.arange(product.size(0)).to(product.device).unsqueeze(1) # target = torch.cat([target*2, target*2+1], dim=1) # ranked = product.argsort(dim=1, descending=True) # # MRR # idx2rank = ranked.argsort(dim=1) # for idx, t in enumerate(target): # correct_idx = t.tolist() # for _ in correct_idx: # rrs.append(1 / (idx2rank[idx][_].item() + 1)) # return rrs def mhop_loss(model, batch, args): outputs = model(batch) loss_fct = CrossEntropyLoss(ignore_index=-1) all_ctx = torch.cat([outputs['c1'], outputs['c2']], dim=0) neg_ctx = torch.cat([outputs["neg_1"].unsqueeze(1), outputs["neg_2"].unsqueeze(1)], dim=1) # B x 2 x M x h scores_1_hop = torch.mm(outputs["q"], all_ctx.t()) neg_scores_1 = torch.bmm(outputs["q"].unsqueeze(1), neg_ctx.transpose(1,2)).squeeze(1) scores_2_hop = torch.mm(outputs["q_sp1"], all_ctx.t()) neg_scores_2 = torch.bmm(outputs["q_sp1"].unsqueeze(1), neg_ctx.transpose(1,2)).squeeze(1) # mask the 1st hop bsize = outputs["q"].size(0) scores_1_mask = torch.cat([torch.zeros(bsize, bsize), torch.eye(bsize)], dim=1).to(outputs["q"].device) scores_1_hop = scores_1_hop.float().masked_fill(scores_1_mask.bool(), float('-inf')).type_as(scores_1_hop) scores_1_hop = torch.cat([scores_1_hop, neg_scores_1], dim=1) scores_2_hop = torch.cat([scores_2_hop, neg_scores_2], dim=1) if args.momentum: queue_neg_scores_1 = torch.mm(outputs["q"], model.module.queue.clone().detach().t()) queue_neg_scores_2 = torch.mm(outputs["q_sp1"], model.module.queue.clone().detach().t()) # queue_neg_scores_1 = queue_neg_scores_1 / args.temperature # queue_neg_scores_2 = queue_neg_scores_2 / args.temperature scores_1_hop = torch.cat([scores_1_hop, queue_neg_scores_1], dim=1) scores_2_hop = torch.cat([scores_2_hop, queue_neg_scores_2], dim=1) model.module.dequeue_and_enqueue(all_ctx.detach()) # model.module.momentum_update_key_encoder() target_1_hop = torch.arange(outputs["q"].size(0)).to(outputs["q"].device) target_2_hop = torch.arange(outputs["q"].size(0)).to(outputs["q"].device) + outputs["q"].size(0) retrieve_loss = loss_fct(scores_1_hop, target_1_hop) + loss_fct(scores_2_hop, target_2_hop) return retrieve_loss def mhop_eval(outputs, args): all_ctx = torch.cat([outputs['c1'], outputs['c2']], dim=0) neg_ctx = torch.cat([outputs["neg_1"].unsqueeze(1), outputs["neg_2"].unsqueeze(1)], dim=1) scores_1_hop = torch.mm(outputs["q"], all_ctx.t()) neg_scores_1 = torch.bmm(outputs["q"].unsqueeze(1), neg_ctx.transpose(1,2)).squeeze(1) scores_2_hop = torch.mm(outputs["q_sp1"], all_ctx.t()) neg_scores_2 = torch.bmm(outputs["q_sp1"].unsqueeze(1), neg_ctx.transpose(1,2)).squeeze(1) bsize = outputs["q"].size(0) scores_1_mask = torch.cat([torch.zeros(bsize, bsize), torch.eye(bsize)], dim=1).to(outputs["q"].device) scores_1_hop = scores_1_hop.float().masked_fill(scores_1_mask.bool(), float('-inf')).type_as(scores_1_hop) scores_1_hop = torch.cat([scores_1_hop, neg_scores_1], dim=1) scores_2_hop = torch.cat([scores_2_hop, neg_scores_2], dim=1) target_1_hop = torch.arange(outputs["q"].size(0)).to(outputs["q"].device) target_2_hop = torch.arange(outputs["q"].size(0)).to(outputs["q"].device) + outputs["q"].size(0) ranked_1_hop = scores_1_hop.argsort(dim=1, descending=True) ranked_2_hop = scores_2_hop.argsort(dim=1, descending=True) idx2ranked_1 = ranked_1_hop.argsort(dim=1) idx2ranked_2 = ranked_2_hop.argsort(dim=1) rrs_1, rrs_2 = [], [] for t, idx2ranked in zip(target_1_hop, idx2ranked_1): rrs_1.append(1 / (idx2ranked[t].item() + 1)) for t, idx2ranked in zip(target_2_hop, idx2ranked_2): rrs_2.append(1 / (idx2ranked[t].item() + 1)) return {"rrs_1": rrs_1, "rrs_2": rrs_2} def unified_loss(model, batch, args): outputs = model(batch) all_ctx = torch.cat([outputs['c1'], outputs['c2']], dim=0) neg_ctx = torch.cat([outputs["neg_1"].unsqueeze(1), outputs["neg_2"].unsqueeze(1)], dim=1) scores_1_hop = torch.mm(outputs["q"], all_ctx.t()) neg_scores_1 = torch.bmm(outputs["q"].unsqueeze(1), neg_ctx.transpose(1,2)).squeeze(1) scores_2_hop = torch.mm(outputs["q_sp1"], all_ctx.t()) neg_scores_2 = torch.bmm(outputs["q_sp1"].unsqueeze(1), neg_ctx.transpose(1,2)).squeeze(1) # mask for 1st hop bsize = outputs["q"].size(0) scores_1_mask = torch.cat([torch.zeros(bsize, bsize), torch.eye(bsize)], dim=1).to(outputs["q"].device) scores_1_hop = scores_1_hop.float().masked_fill(scores_1_mask.bool(), float('-inf')).type_as(scores_1_hop) scores_1_hop = torch.cat([scores_1_hop, neg_scores_1], dim=1) scores_2_hop = torch.cat([scores_2_hop, neg_scores_2], dim=1) stop_loss = F.cross_entropy(outputs["stop_logits"], batch["stop_targets"].view(-1), reduction="sum") target_1_hop = torch.arange(outputs["q"].size(0)).to(outputs["q"].device) target_2_hop = torch.arange(outputs["q"].size(0)).to(outputs["q"].device) + outputs["q"].size(0) 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() return retrieve_loss + stop_loss def unified_eval(outputs, batch): all_ctx = torch.cat([outputs['c1'], outputs['c2']], dim=0) neg_ctx = torch.cat([outputs["neg_1"].unsqueeze(1), outputs["neg_2"].unsqueeze(1)], dim=1) scores_1_hop = torch.mm(outputs["q"], all_ctx.t()) scores_2_hop = torch.mm(outputs["q_sp1"], all_ctx.t()) neg_scores_1 = torch.bmm(outputs["q"].unsqueeze(1), neg_ctx.transpose(1,2)).squeeze(1) neg_scores_2 = torch.bmm(outputs["q_sp1"].unsqueeze(1), neg_ctx.transpose(1,2)).squeeze(1) bsize = outputs["q"].size(0) scores_1_mask = torch.cat([torch.zeros(bsize, bsize), torch.eye(bsize)], dim=1).to(outputs["q"].device) scores_1_hop = scores_1_hop.float().masked_fill(scores_1_mask.bool(), float('-inf')).type_as(scores_1_hop) scores_1_hop = torch.cat([scores_1_hop, neg_scores_1], dim=1) scores_2_hop = torch.cat([scores_2_hop, neg_scores_2], dim=1) target_1_hop = torch.arange(outputs["q"].size(0)).to(outputs["q"].device) target_2_hop = torch.arange(outputs["q"].size(0)).to(outputs["q"].device) + outputs["q"].size(0) # stop accuracy stop_pred = outputs["stop_logits"].argmax(dim=1) stop_targets = batch["stop_targets"].view(-1) stop_acc = (stop_pred == stop_targets).float().tolist() ranked_1_hop = scores_1_hop.argsort(dim=1, descending=True) ranked_2_hop = scores_2_hop.argsort(dim=1, descending=True) idx2ranked_1 = ranked_1_hop.argsort(dim=1) idx2ranked_2 = ranked_2_hop.argsort(dim=1) rrs_1_mhop, rrs_2_mhop, rrs_nq = [], [], [] for t1, idx2ranked1, t2, idx2ranked2, stop in zip(target_1_hop, idx2ranked_1, target_2_hop, idx2ranked_2, stop_targets): if stop: # rrs_1_mhop.append(1 / (idx2ranked1[t1].item() + 1)) rrs_2_mhop.append(1 / (idx2ranked2[t2].item() + 1)) else: rrs_nq.append(1 / (idx2ranked1[t1].item() + 1)) return { "stop_acc": stop_acc, "rrs_1_mhop": rrs_1_mhop, "rrs_2_mhop": rrs_2_mhop, "rrs_nq": rrs_nq } ================================================ FILE: mdr/retrieval/decomposed_analysis.py ================================================ # Copyright (c) Facebook, Inc. and its affiliates. # All rights reserved. # # This source code is licensed under the license found in the # LICENSE file in the root directory of this source tree. import json def decomposed_errors(): top1_pred = [json.loads(l) for l in open("/private/home/xwhan/data/hotpot/dense_val_b1_top1.json").readlines()] analysis_folder = "/private/home/xwhan/data/hotpot/analysis" start_errors, bridge_errors, failed = [], [], [] correct = [] for item in top1_pred: pred_titles = [_[0] for _ in item["candidate_chains"][0]] gold_titles = [_[0] for _ in item["sp"]] if set(pred_titles) == set(gold_titles): if item["type"] == "bridge": correct.append(item) continue if item["type"] == "bridge": start_title = None for t in gold_titles: if t != item["bridge"]: start_title = t assert start_title is not None if item["bridge"] in pred_titles and start_title not in pred_titles: start_errors.append(item) elif item["bridge"] not in pred_titles and start_title in pred_titles: bridge_errors.append(item) else: failed.append(item) with open(analysis_folder + "/correct.json", "w") as g: for _ in correct: _["predicted"] = _.pop("candidate_chains")[0] g.write(json.dumps(_) + "\n") with open(analysis_folder + "/start_errors.json", "w") as g: for _ in start_errors: _["predicted"] = _.pop("candidate_chains")[0] g.write(json.dumps(_) + "\n") with open(analysis_folder + "/bridge_errors.json", "w") as g: for _ in bridge_errors: _["predicted"] = _.pop("candidate_chains")[0] g.write(json.dumps(_) + "\n") with open(analysis_folder + "/total_errors.json", "w") as g: for _ in failed: _["predicted"] = _.pop("candidate_chains")[0] g.write(json.dumps(_) + "\n") print(len(correct)) print(len(start_errors)) print(len(bridge_errors)) print(len(failed)) import random def collect_gold_decomposition(): """ interactively collect """ dev_qdmr = [json.loads(l) for l in open("/private/home/xwhan/data/QDMR/dev.json").readlines()] bridge_dev = [_ for _ in dev_qdmr if _["type"] == "bridge"] random.shuffle(bridge_dev) idx = 0 samples_to_inspect = [] while True: print(f"\n-----{len(samples_to_inspect)} samples collected so far-----") sample = bridge_dev[idx] idx += 1 print(f"Original Q: {sample['q']}") print(f"Decomposed Q: {sample['q_decom']}") print(f"Supporting Passages: {sample['sp']}") subq1 = input("Type SUB Q1:") if subq1 == "bad": continue elif subq1 == "stop": break subq2 = input("Type SUB Q2:") samples_to_inspect.append({ "id": sample["id"], "sp": sample["sp"], "orig_q": sample['q'], "subQ_1": subq1, "subQ_2": subq2 }) print(f"{len(samples_to_inspect)} samples collected in total..") with open("/private/home/xwhan/data/QDMR/inspect.json", "w") as g: for _ in samples_to_inspect: g.write(json.dumps(_) + "\n") def qdmr_utils(): """ change file format for decomposed and end-to-end retrieval """ qdmr_data = [json.loads(l) for l in open("/private/home/xwhan/data/QDMR/inspect.json").readlines()] mhop_data, decomposed_data = [], [] for idx, item in enumerate(qdmr_data): if idx in [65,66,67]: continue sp = [_["title"] for _ in item["sp"]] question = item["orig_q"] mhop_data.append({ "question": question, "sp": sp, "type": "bridge", "_id": item["id"] }) decomposed_data.append(item) # with open("/private/home/xwhan/data/QDMR/qdmr_decomposed.json", "w") as g: # for item in decomposed_data: # g.write(json.dumps(item) + "\n") with open("/private/home/xwhan/data/QDMR/qdmr_e2e.json", "w") as g: for item in mhop_data: g.write(json.dumps(item) + "\n") def analyze_results(): decomposed_results = [json.loads(l) for l in open("/private/home/xwhan/data/QDMR/qdmr_decomposed_results.json")] e2e_results = [json.loads(l) for l in open("/private/home/xwhan/data/QDMR/qdmr_e2e_results.json")] better = 0 worse = 0 both = 0 for res1, res2 in zip(decomposed_results, e2e_results): sp_titles = set([_[0] for _ in res1["sp"]]) res1_top1 = [_[0] for _ in res1["candidate_chains"][0]] res2_top1 = [_[0] for _ in res2["candidate_chains"][0]] assert res1["_id"] == res2["_id"] question = res1["question"] q_pairs = res1["q_pairs"] if set(res2_top1) == sp_titles and set(res1_top1) != sp_titles: # print(sp_titles) # import pdb; pdb.set_trace() better += 1 elif set(res2_top1) != sp_titles and set(res1_top1) == sp_titles: worse += 1 elif set(res2_top1) == sp_titles and set(res1_top1) == sp_titles: both += 1 print(both) print(better) print(worse) print(len(decomposed_results)) if __name__ == "__main__": # collect_gold_decomposition() # qdmr_utils() analyze_results() ================================================ FILE: mdr/retrieval/fever.ipynb ================================================ { "cells": [ { "cell_type": "code", "execution_count": 90, "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "1059\n", "12273\n" ] } ], "source": [ "import json\n", "import numpy as np\n", "import random\n", "\n", "fever_path = \"/private/home/xwhan/data/fever/retrieval/\"\n", "\n", "dev = [json.loads(l) for l in open(fever_path + \"dev.txt\").readlines()]\n", "multi_dev = []\n", "single_dev = []\n", "all_evidence_lens = [] # for multi evidence\n", "random.shuffle(dev)\n", "all_claim_lens = []\n", "for item in dev:\n", " evidence_lens = []\n", " all_claim_lens.append(len(item[\"claim\"].split()))\n", " \n", " for chain in item[\"evidence\"]:\n", " if len(chain) > 1:\n", "# evidence_lens.append(len(chain))\n", " chain_titles = set([p[\"title\"] for p in chain])\n", " evidence_lens.append(len(chain_titles)) \n", "# print(item[\"claim\"])\n", "# print(chain)\n", "# assert False\n", " else:\n", " evidence_lens.append(1)\n", " multi_count = np.sum([int(c > 1) for c in evidence_lens])\n", " \n", " if multi_count == len(evidence_lens):\n", " multi_dev.append(item)\n", " all_evidence_lens += evidence_lens\n", " else:\n", " single_dev.append(item)\n", " \n", "print(len(multi_dev))\n", "print(len(single_dev))\n", "with open(\"/private/home/xwhan/data/fever/retrieval/dev_multi_evidence_compact.txt\", \"w\") as g:\n", " for l in multi_dev:\n", " g.write(json.dumps(l) + \"\\n\")\n", "# with open(\"/private/home/xwhan/data/fever/retrieval/dev_single_evidence.txt\", \"w\") as g:\n", "# for l in single_dev:\n", "# g.write(json.dumps(l) + \"\\n\")\n", "# with open(\"/private/home/xwhan/data/fever/retrieval/dev_all.txt\", \"w\") as g:\n", "# for l in single_dev + multi_dev:\n", "# g.write(json.dumps(l) + \"\\n\")" ] }, { "cell_type": "code", "execution_count": 81, "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "1741\n", "2.0\n", "0.5835726593911545 0.5835726593911545 0.5835726593911545\n" ] } ], "source": [ "# baseline retrieval for single/multihop subsets\n", "\n", "import unicodedata\n", "def normalize(text):\n", " \"\"\"Resolve different type of unicode encodings.\"\"\"\n", " return unicodedata.normalize('NFD', text)\n", "\n", "el_results = [json.loads(l) for l in open(\"/private/home/xwhan/data/fever/retrieval/dev.ensembles.s10.jsonl\").readlines()]\n", "id2el_docs = {_[\"id\"]:_[\"predicted_pages\"] for _ in el_results}\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", "single_gold = {_[\"id\"]:_ for _ in single_dev}\n", "multi_gold = {_[\"id\"]:_ for _ in multi_dev}\n", "all_gold = {_[\"id\"]:_ for _ in multi_dev + single_dev}\n", "\n", "subset = multi_gold\n", "precs, recalls = [], []\n", "doc_count = []\n", "dense_docs = []\n", "for item in dense_multi_results:\n", " if item[\"id\"] in subset:\n", "# pred = set(item[\"predicted_pages\"])\n", " retrieved_chains = item[\"candidate_chains\"] \n", " pred = []\n", " for chain in retrieved_chains:\n", " for p in chain:\n", " if normalize(p[0]) not in pred:\n", " pred.append(normalize(p[0]))\n", " pred = pred[:2]\n", "# pred = [_[\"title\"] for _ in item[\"topk\"][:1]]\n", " \n", " pred = set(pred)\n", "# el_pred = id2el_docs[item[\"id\"]]\n", "# el_count = 0\n", "# for title in el_pred:\n", "# if title not in pred:\n", "# pred.add(title)\n", "# el_count +=1\n", "# if el_count == 2:\n", "# break\n", " pred = list(pred)\n", " \n", " dense_docs.append({\n", " \"claim\": item[\"claim\"],\n", " \"id\": item[\"id\"],\n", " \"predicted_pages\": list(pred)\n", " })\n", " \n", " doc_count.append(len(pred))\n", " \n", " gold_docs = set()\n", " recall = 0\n", " for chain in subset[item[\"id\"]][\"evidence\"]:\n", " chain_titles = set([normalize(p[\"title\"]) for p in chain])\n", " for t in chain_titles: gold_docs.add(t)\n", " chain_covered = [int(t in pred) for t in chain_titles]\n", " if np.sum(chain_covered) == len(chain_titles):\n", " recall = 1\n", " break\n", " \n", " if len(gold_docs) > 0:\n", " if len(pred) == 0:\n", " prec = 0\n", " else:\n", " prec = np.mean([int(doc in gold_docs) for doc in pred])\n", " \n", " precs.append(prec)\n", " recalls.append(recall)\n", " \n", "print(len(precs))\n", "print(np.mean(doc_count))\n", "pr, rec = np.mean(precs), np.mean(recalls)\n", "print(pr, rec, 2.0 * pr * rec / (pr + rec))\n", "\n", "# with open(\"/private/home/xwhan/data/fever/retrieval/dense_wiki_pages_top2.jsonl\", \"w\") as g:\n", "# for _ in dense_docs:\n", "# g.write(json.dumps(_) + \"\\n\")" ] }, { "cell_type": "code", "execution_count": 95, "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "1741\n", "2.9959793222286044\n", "0.6764680633362424 1.0 0.8070157471297632\n" ] } ], "source": [ "# inspect the FEVER results\n", "dev_all_results = [json.loads(l) for l in open(\"/private/home/xwhan/code/Transformer-XH/data/fever_dev_graph.json\").readlines()]\n", "subset = multi_gold\n", "precs, recalls = [], []\n", "doc_count = []\n", "\n", "for item in dev_all_results:\n", " if item[\"qid\"] in subset:\n", " \n", " pred = [_[\"name\"] for _ in item[\"node\"]]\n", " pred = set(pred)\n", " pred = list(pred)\n", " \n", " doc_count.append(len(pred))\n", " \n", " gold_docs = set()\n", " recall = 0\n", " for chain in subset[item[\"qid\"]][\"evidence\"]:\n", " chain_titles = set([normalize(p[\"title\"]) for p in chain])\n", " for t in chain_titles: gold_docs.add(t)\n", " chain_covered = [int(t in pred) for t in chain_titles]\n", " if np.sum(chain_covered) == len(chain_titles):\n", " recall = 1\n", " break\n", " \n", " if len(gold_docs) > 0:\n", " if len(pred) == 0:\n", " prec = 0\n", " else:\n", " prec = np.mean([int(doc in gold_docs) for doc in pred])\n", " \n", " precs.append(prec)\n", " recalls.append(recall)\n", " \n", "print(len(precs))\n", "print(np.mean(doc_count))\n", "pr, rec = np.mean(precs), np.mean(recalls)\n", "print(pr, rec, 2.0 * pr * rec / (pr + rec))" ] }, { "cell_type": "code", "execution_count": 40, "metadata": {}, "outputs": [], "source": [ "# passage retrieval evaluation \n", "def fever_retrieval_eval(results, topk=5):\n", " \n", " precs, recalls = [], []\n", " for item in results:\n", " gold = item[\"correct_normalized\"]\n", " pred = item[\"bm25_topk\"][:topk]\n", " \n", " if len(gold) > 0:\n", " prec = np.mean([int(doc in gold) for doc in pred])\n", " else:\n", " prec = 1\n", " recall = 0\n", " for chain in item[\"evidence\"]:\n", " chain_titles = set([normalize(p[\"title\"]) for p in chain])\n", " chain_covered = [int(t in pred) for t in chain_titles]\n", " if np.sum(chain_covered) == len(chain_titles):\n", " recall = 1\n", " break\n", " precs.append(prec)\n", " recalls.append(recall)\n", " pr, rec = np.mean(precs), np.mean(recalls)\n", " return pr, rec, 2.0 * pr * rec / (pr + rec)" ] }, { "cell_type": "code", "execution_count": 42, "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "(0.12268811028144745, 0.5020103388856979, 0.19718537768662206)\n" ] } ], "source": [ "tfidf_results = [json.loads(l) for l in open(\"/private/home/xwhan/data/fever/retrieval/multi_dev_tfidf.txt\").readlines()]\n", "print(fever_retrieval_eval(tfidf_results, 10))" ] }, { "cell_type": "code", "execution_count": 234, "metadata": {}, "outputs": [], "source": [] }, { "cell_type": "code", "execution_count": 244, "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "1741\n", "1741\n", "(0.6223243346735593, 0.46927053417576103, 0.5350675077296432)\n" ] } ], "source": [ "# phrase_matching_results = [json.loads(l) for l in open(\"/private/home/xwhan/data/fever/retrieval/all_dev.json\").readlines()]\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", "phrase_matching_results = [json.loads(l) for l in open(\"/private/home/xwhan/code/Transformer-XH/data/fever_dev_graph.json\").readlines()]\n", "# for _ in phrase_matching_results:\n", "# _[\"id\"] = _[\"qid\"]\n", "\n", "phrase_matching_results = [_ for _ in phrase_matching_results if _[\"id\"] in multihop_ids]\n", "# phrase_matching_results = [json.loads(l) for l in open(\"/private/home/xwhan/data/fever/retrieval/esim_mhop_dev.json\").readlines()]\n", "\n", "# json.dump(phrase_matching_results, open(\"/private/home/xwhan/data/fever/retrieval/dev_el_wiki_pages.jsonl\", \"w\"))\n", "\n", "print(len(phrase_matching_results))\n", "\n", "tfidf_results = [json.loads(l) for l in open(\"/private/home/xwhan/data/fever/retrieval/multi_dev_tfidf.txt\").readlines()]\n", "print(len(tfidf_results))\n", "pred_lens = []\n", "def fever_retrieval_eval_phrase(tfidf_results, phrase_results, topk=5):\n", " id2gold = {_[\"id\"]:_[\"correct_normalized\"] for _ in tfidf_results}\n", " id2gold_evidence = {_[\"id\"]:_[\"evidence\"] for _ in tfidf_results}\n", " precs, recalls = [], []\n", " for item in phrase_results:\n", " gold = id2gold[item[\"id\"]]\n", "# print(gold)\n", " retrieved_evidence = item[\"evidence\"] \n", " pred = []\n", " for e in retrieved_evidence:\n", " pred.append(normalize(e[0]))\n", " \n", "# pred = item[\"predicted_pages\"] + item[\"wiki_results\"]\n", "# pred = item[\"wiki_results\"]\n", "# pred = item[\"predicted_pages\"]\n", " \n", "# pred = []\n", "# for n in item[\"node\"]:\n", "# pred.append(n[\"name\"])\n", " \n", " pred = list(set(pred))\n", " pred_lens.append(len(pred))\n", " \n", " if len(gold) > 0:\n", " if len(pred) == 0:\n", " prec = 0\n", " else:\n", " prec = np.mean([int(doc in gold) for doc in pred])\n", " else:\n", " prec = 1\n", " recall = 0\n", " for chain in id2gold_evidence[item[\"id\"]]:\n", " chain_titles = set([normalize(p[\"title\"]) for p in chain])\n", " chain_covered = [int(t in pred) for t in chain_titles]\n", " if np.sum(chain_covered) == len(chain_titles):\n", " recall = 1\n", " break\n", " precs.append(prec)\n", " recalls.append(recall)\n", " \n", " pr, rec = np.mean(precs), np.mean(recalls)\n", " return pr, rec, 2.0 * pr * rec / (pr + rec)\n", "print(fever_retrieval_eval_phrase(tfidf_results, phrase_matching_results, topk=5))" ] }, { "cell_type": "code", "execution_count": 181, "metadata": {}, "outputs": [ { "data": { "text/plain": [ "2.1039632395175185" ] }, "execution_count": 181, "metadata": {}, "output_type": "execute_result" } ], "source": [ "np.mean(pred_lens)" ] }, { "cell_type": "code", "execution_count": 153, "metadata": {}, "outputs": [ { "data": { "text/plain": [ "0.12426242624262426" ] }, "execution_count": 153, "metadata": {}, "output_type": "execute_result" } ], "source": [ "test_with_all = [json.loads(l) for l in open(\"/private/home/xwhan/data/fever/retrieval/all_dev.json\").readlines()]\n", "title_count = []\n", "for item in test_with_all:\n", " titles = set()\n", " for e in item[\"evidence\"]:\n", " titles.add(e[0])\n", " title_count.append(len(titles))\n", "np.sum(np.array(title_count) > 7) / len(title_count)" ] }, { "cell_type": "code", "execution_count": 154, "metadata": {}, "outputs": [ { "data": { "text/plain": [ "0.11321132113211321" ] }, "execution_count": 154, "metadata": {}, "output_type": "execute_result" } ], "source": [ "test_with_all = [json.loads(l) for l in open(\"/private/home/xwhan/data/fever/retrieval/all_test.json\").readlines()]\n", "title_count = []\n", "for item in test_with_all:\n", " titles = set()\n", " for e in item[\"evidence\"]:\n", " titles.add(e[0])\n", " title_count.append(len(titles))\n", "np.sum(np.array(title_count) > 7) / len(title_count)" ] }, { "cell_type": "code", "execution_count": 91, "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "1960\n", "1960\n" ] } ], "source": [ "# build dense final prediction for evaluation\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", "final_retrieval = [json.loads(l) for l in open(\"/private/home/xwhan/code/KernelGAT/data/bert_dense_top4_mhop_sents.json\").readlines()]\n", "all_dev_gold = [json.loads(l) for l in open(\"/private/home/xwhan/data/fever/shared_task_dev.jsonl\").readlines()]\n", "id2gold = {_[\"id\"]:_ for _ in all_dev_gold}\n", "\n", "print(len(final_retrieval))\n", "final = []\n", "for pred, retrieval in zip(final_pred, final_retrieval):\n", " assert pred[\"id\"] == retrieval[\"id\"]\n", " final.append({\n", " \"id\": pred[\"id\"],\n", " \"label\": id2gold[pred[\"id\"]][\"label\"],\n", " \"evidence\": id2gold[pred[\"id\"]][\"evidence\"],\n", " \"predicted_label\": pred[\"predicted_label\"],\n", " \"predicted_evidence\": [[normalize(e[0]), int(e[1])] for e in retrieval[\"evidence\"][:5]]\n", " })\n", "\n", "print(len(final))\n", "with open(\"/private/home/xwhan/data/fever/results/dense_top4_mhop_dev.json\", \"w\") as g:\n", " for l in final:\n", " g.write(json.dumps(l) + \"\\n\")" ] }, { "cell_type": "code", "execution_count": 92, "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "1960\n", "1960\n" ] } ], "source": [ "# build EL final prediction for evaluation\n", "final_pred = [json.loads(l) for l in open(\"/private/home/xwhan/code/KernelGAT/kgat/output/el_bert_dev_mhop.json\").readlines()]\n", "final_retrieval = [json.loads(l) for l in open(\"/private/home/xwhan/code/KernelGAT/data/bert_dev_multi_el.json\").readlines()]\n", "all_dev_gold = [json.loads(l) for l in open(\"/private/home/xwhan/data/fever/shared_task_dev.jsonl\").readlines()]\n", "id2gold = {_[\"id\"]:_ for _ in all_dev_gold}\n", "\n", "print(len(final_retrieval))\n", "final = []\n", "for pred, retrieval in zip(final_pred, final_retrieval):\n", " assert pred[\"id\"] == retrieval[\"id\"]\n", " final.append({\n", " \"id\": pred[\"id\"],\n", " \"label\": id2gold[pred[\"id\"]][\"label\"],\n", " \"evidence\": id2gold[pred[\"id\"]][\"evidence\"],\n", " \"predicted_label\": pred[\"predicted_label\"],\n", " \"predicted_evidence\": [[normalize(e[0]), int(e[1])] for e in retrieval[\"evidence\"][:5]]\n", " })\n", "\n", "print(len(final))\n", "with open(\"/private/home/xwhan/data/fever/results/el_mhop_dev.json\", \"w\") as g:\n", " for l in final:\n", " g.write(json.dumps(l) + \"\\n\")" ] }, { "cell_type": "code", "execution_count": 93, "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "1960\n" ] } ], "source": [ "final_pred = [json.loads(l) for l in open(\"/private/home/xwhan/code/KernelGAT/kgat/output/esim_mhop_dev.json\").readlines()]\n", "final_retrieval = [json.loads(l) for l in open(\"/private/home/xwhan/data/fever/retrieval/esim_mhop_dev.json\").readlines()]\n", "all_dev_gold = [json.loads(l) for l in open(\"/private/home/xwhan/data/fever/shared_task_dev.jsonl\").readlines()]\n", "id2gold = {_[\"id\"]:_ for _ in all_dev_gold}\n", "\n", "final = []\n", "for pred, retrieval in zip(final_pred, final_retrieval):\n", " assert pred[\"id\"] == retrieval[\"id\"]\n", " final.append({\n", " \"id\": pred[\"id\"],\n", " \"label\": id2gold[pred[\"id\"]][\"label\"],\n", " \"evidence\": id2gold[pred[\"id\"]][\"evidence\"],\n", " \"predicted_label\": pred[\"predicted_label\"],\n", " \"predicted_evidence\": [[e[0], int(e[1])] for e in retrieval[\"evidence\"][:5]]\n", " })\n", "\n", "print(len(final))\n", "with open(\"/private/home/xwhan/data/fever/results/esim_mhop_dev.json\", \"w\") as g:\n", " for l in final:\n", " g.write(json.dumps(l) + \"\\n\")" ] } ], "metadata": { "kernelspec": { "display_name": "Python 3", "language": "python", "name": "python3" }, "language_info": { "codemirror_mode": { "name": "ipython", "version": 3 }, "file_extension": ".py", "mimetype": "text/x-python", "name": "python", "nbconvert_exporter": "python", "pygments_lexer": "ipython3", "version": "3.6.3" } }, "nbformat": 4, "nbformat_minor": 2 } ================================================ FILE: mdr/retrieval/hotpot.ipynb ================================================ { "cells": [ { "cell_type": "code", "execution_count": 2, "metadata": {}, "outputs": [], "source": [ "import json\n", "\n", "test_qas = json.load(open(\"/private/home/xwhan/data/hotpot/hotpot_test_fullwiki_v1.json\"))\n", "test_results = json.load(open(\"/private/home/xwhan/data/hotpot/results/hotpot_test_b200_k500.json\"))" ] }, { "cell_type": "code", "execution_count": 4, "metadata": {}, "outputs": [ { "data": { "text/plain": [ "True" ] }, "execution_count": 4, "metadata": {}, "output_type": "execute_result" } ], "source": [ "len(test_qas) == len(test_results[\"answer\"])" ] }, { "cell_type": "code", "execution_count": 15, "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "Who has been in more bands, Deron Miller or Steve Marriott?\n", "Deron John Miller\n" ] } ], "source": [ "import random\n", "qid2question = {_[\"_id\"]:_[\"question\"] for _ in test_qas}\n", "qids = list(test_results[\"answer\"].keys())\n", "random.shuffle(qids)\n", "\n", "\n", "\n", "print(qid2question[qids[0]])\n", "print(test_results[\"answer\"][qids[0]])" ] }, { "cell_type": "code", "execution_count": 63, "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "Collecting seaborn\r\n", " Using cached https://files.pythonhosted.org/packages/c7/e6/54aaaafd0b87f51dfba92ba73da94151aa3bc179e5fe88fc5dfb3038e860/seaborn-0.10.1-py3-none-any.whl\r\n", "Collecting matplotlib>=2.1.2 (from seaborn)\r\n", " Using cached https://files.pythonhosted.org/packages/96/a7/b6fa244fd8a8814ef9408c8a5a7e4ed0340e232a6f0ce2046b42e50672c0/matplotlib-3.3.1-cp36-cp36m-manylinux1_x86_64.whl\r\n", "Requirement already satisfied: scipy>=1.0.1 in /public/apps/anaconda3/5.0.1/lib/python3.6/site-packages (from seaborn)\r\n", "Requirement already satisfied: numpy>=1.13.3 in /public/apps/anaconda3/5.0.1/lib/python3.6/site-packages (from seaborn)\r\n", "Requirement already satisfied: pandas>=0.22.0 in /public/apps/anaconda3/5.0.1/lib/python3.6/site-packages (from seaborn)\r\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", "Collecting pillow>=6.2.0 (from matplotlib>=2.1.2->seaborn)\r\n", " Using cached https://files.pythonhosted.org/packages/30/bf/92385b4262178ca22b34f82e0e09c2922eb351fe39f3cc7b8ba9ea555b41/Pillow-7.2.0-cp36-cp36m-manylinux1_x86_64.whl\r\n", "Collecting certifi>=2020.06.20 (from matplotlib>=2.1.2->seaborn)\r\n", " Using cached https://files.pythonhosted.org/packages/5e/c4/6c4fe722df5343c33226f0b4e0bb042e4dc13483228b4718baf286f86d87/certifi-2020.6.20-py2.py3-none-any.whl\r\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", "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", "Collecting kiwisolver>=1.0.1 (from matplotlib>=2.1.2->seaborn)\r\n", " Using cached https://files.pythonhosted.org/packages/ae/23/147de658aabbf968324551ea22c0c13a00284c4ef49a77002e91f79657b7/kiwisolver-1.2.0-cp36-cp36m-manylinux1_x86_64.whl\r\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", "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", "Installing collected packages: pillow, certifi, kiwisolver, matplotlib, seaborn\r\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", "\u001b[33mYou are using pip version 9.0.1, however version 20.2.2 is available.\r\n", "You should consider upgrading via the 'pip install --upgrade pip' command.\u001b[0m\r\n" ] } ], "source": [ "# Install a pip package in the current Jupyter kernel\n", "import sys\n", "!{sys.executable} -m pip install seaborn --user" ] }, { "cell_type": "code", "execution_count": 119, "metadata": {}, "outputs": [ { "data": { "image/png": "iVBORw0KGgoAAAANSUhEUgAAAYsAAAESCAYAAAAMifkAAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\nAABJRU5ErkJggg==\n", "text/plain": [ "<matplotlib.figure.Figure at 0x7f1cd8117748>" ] }, "metadata": {}, "output_type": "display_data" } ], "source": [ "# figure showing the efficieicnty trade off using 16-core CPUs\n", "import matplotlib.pyplot as plt\n", "import seaborn as sns\n", "sns.set(style='ticks')\n", "# Asai et al.2020\n", "plt.xlim(-5, 210)\n", "plt.ylim(35, 70)\n", "\n", "# Asai et al. 500 + 8*3\n", "plt.scatter(133, 61.4, c='#1f77b4')\n", "plt.text(133 - 30, 61.4 - 3, \"Graph Recurrent Retriever\")\n", "\n", "# Gold \n", "plt.scatter(0.5, 39.1, c='#1f77b4')\n", "plt.text(3.5, 39.1 - 0.5, \"GoldEn\")\n", "\n", "# DiKIT\n", "plt.scatter(0.5, 42.9, c='#1f77b4')\n", "plt.text(3.5, 42.9 - 0.5, \"DrKIT\")\n", "\n", "# Semmantic Retrieval\n", "plt.scatter(50*0.3, 47.6, c='#1f77b4')\n", "plt.text(50*0.3 + 3, 47.6 - 0.5, \"Semantic Retrieval\")\n", "\n", "# HGN\n", "plt.scatter(50*0.3, 60.0, c='#1f77b4')\n", "plt.text(50*0.3 + 3, 60 - 0.5, \"HGN\")\n", "\n", "# TransformerXH 115 cross attention 52.9\n", "plt.scatter(115*0.3, 52.9, c='#1f77b4')\n", "plt.text(115*0.3 + 3, 52.9 - 0.5, \"TransformerXH\")\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", "xs = [_[0] for _ in time_f1s]\n", "ys =[_[1] for _ in time_f1s]\n", "plt.plot(xs,ys, linestyle='--', marker='s', c='r', label='Recursive Dense Retriever Topk')\n", "\n", "plt.xlabel('sec/Q', fontsize=12)\n", "plt.ylabel('Joint F1', fontsize=12)\n", "\n", "lg = plt.legend(loc='lower right', fontsize=10)\n", "frame = lg.get_frame()\n", "# frame.set_edgecolor('black')\n", "lg.draw_frame(True)\n", "\n", "\n", "plt.savefig('efficiency.pdf')\n", "plt.show()" ] }, { "cell_type": "code", "execution_count": 431, "metadata": {}, "outputs": [], "source": [ "def autolabel(rects):\n", " \"\"\"Attach a text label above each bar in *rects*, displaying its height.\"\"\"\n", " for rect in rects:\n", " height = rect.get_height()\n", " ax.annotate('{}'.format(height),\n", " xy=(rect.get_x() + rect.get_width() / 2, height),\n", " xytext=(0, 3), # 3 points vertical offset\n", " textcoords=\"offset points\",\n", " ha='center', va='bottom')" ] }, { "cell_type": "code", "execution_count": 450, "metadata": {}, "outputs": [ { "name": "stderr", "output_type": "stream", "text": [ "/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", " \"matplotlib is currently using a non-GUI backend, \"\n" ] }, { "data": { "image/png": "iVBORw0KGgoAAAANSUhEUgAAAb0AAAEtCAYAAACcSL4HAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\nrPQkSZIksyErPUmSJMls/B8DEzApxeEo9gAAAABJRU5ErkJggg==\n", "text/plain": [ "<matplotlib.figure.Figure at 0x7f1c4e731278>" ] }, "metadata": {}, "output_type": "display_data" } ], "source": [ "# lines comparing the performance of bridge and comparison questions, use beam search to get top100\n", "import matplotlib.pyplot as plt\n", "import seaborn as sns\n", "%matplotlib inline\n", "sns.color_palette(\"tab10\")\n", "sns.set(style='ticks')\n", "\n", "k = [1,2,5,10,20,50,80,100]\n", "comparison = [95.8, 98.3, 99.3, 99.5, 99.7, 99.9, 99.9,100]\n", "bridge = [61.2, 66.8, 72.8, 75.7, 78.1, 80.4, 81.7, 82.0]\n", "\n", "f, (a0, a1) = plt.subplots(1, 2, gridspec_kw={'width_ratios': [2.5, 1]})\n", "a0.plot(k, comparison, marker=\"o\", linestyle='--', label=\"Comparison\")\n", "a0.plot(k, bridge, marker=\"s\", linestyle='--', label=\"Bridge\")\n", "a0.set_xlabel('Top-k Chain')\n", "a0.set_ylabel('Recall')\n", "a0.legend()\n", "\n", "a1.yaxis.tick_right()\n", "a1.yaxis.set_label_position(\"right\")\n", "p = a1.bar(np.arange(2), [97.8, 79.0], 0.5, color=('#1f77b4', '#ff7f0e'))\n", "autolabel(p)\n", "a1.set_xlabel('Q type')\n", "a1.set_ylabel('Reranked Top1 EM Accuracy')\n", "plt.sca(a1)\n", "plt.xticks(np.arange(2), [\"Comparison\", \"Bridge\"])\n", "f.tight_layout(pad=0.05)\n", "plt.savefig('retrieva_types.pdf')\n", "f.show()\n" ] }, { "cell_type": "code", "execution_count": 451, "metadata": {}, "outputs": [ { "data": { "image/png": "iVBORw0KGgoAAAANSUhEUgAAAYsAAAESCAYAAAAMifkAAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\nogAAAABJRU5ErkJggg==\n", "text/plain": [ "<matplotlib.figure.Figure at 0x7f1c4e6e3a58>" ] }, "metadata": {}, "output_type": "display_data" } ], "source": [ "# close-book QA diagnosis experiments\n", "import matplotlib.pyplot as plt\n", "import seaborn as sns\n", "import numpy as np\n", "import pandas as pd\n", "sns.set_style(\"white\")\n", "sns.set_palette(\"colorblind\", 10)\n", "# sns.set_color_codes(\"g\")\n", "bart_ems = [26.5, 27.4, 14.5]\n", "datasets = [\"NQ\", \"WebQ\", \"HotpotQA\"]\n", "sota_ems = [51.4, 45.5, 60.0]\n", "\n", "width = 0.2\n", "ind = np.arange(3)\n", "fig, ax = plt.subplots()\n", "p1 = ax.bar(ind - width/2, bart_ems, width, label=\"retrieval-free BART\", color='tab:blue')\n", "p2 = ax.bar(ind + width/2, sota_ems, width, label=\"retrieved-based SoTA\", color='tab:red')\n", "plt.xticks(ind, datasets)\n", "# plt.yticks(np.arange(0, 60, 10))\n", "plt.ylabel('Answer EM', fontsize=12)\n", "plt.xlabel('Datasets', fontsize=12)\n", "\n", "def autolabel(rects):\n", " \"\"\"Attach a text label above each bar in *rects*, displaying its height.\"\"\"\n", " for rect in rects:\n", " height = rect.get_height()\n", " ax.annotate('{}'.format(height),\n", " xy=(rect.get_x() + rect.get_width() / 2, height),\n", " xytext=(0, 3), # 3 points vertical offset\n", " textcoords=\"offset points\",\n", " ha='center', va='bottom')\n", " \n", "autolabel(p1)\n", "autolabel(p2)\n", "\n", "lg = plt.legend(loc='upper left', fontsize=10)\n", "frame = lg.get_frame()\n", "lg.draw_frame(True)\n", "plt.ylim(0, 70)\n", "plt.savefig('retrieval_free.pdf')\n", "plt.show()" ] }, { "cell_type": "code", "execution_count": 177, "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "7405\n", "0.3683997299122215\n", "0.6267386900742742\n" ] } ], "source": [ "# evalute TFIDF hotpotQA retrieval results\n", "import json\n", "val_inputs = [json.loads(l) for l in open(\"/private/home/xwhan/data/hotpot/hotpot_qas_val.json\").readlines()]\n", "tfidf_results = json.load(open(\"/private/home/xwhan/data/hotpot/tfidf/hotpot_dev_tfidf_results.json\"))\n", "k = 20\n", "\n", "tfidf_covered = []\n", "bm25_covered = []\n", "for gold, res in zip(val_inputs, tfidf_results):\n", " assert gold[\"question\"] == res[\"question\"]\n", " \n", " gold_sp = gold[\"sp\"]\n", "\n", " tfidf_topk = res[\"tfidf_topk\"][:k]\n", " bm25_topk = res[\"bm25_topk\"][:k]\n", " \n", " tfidf_covered.append(np.sum([int(_ in tfidf_topk) for _ in gold_sp]) == len(gold_sp))\n", " bm25_covered.append(np.sum([int(_ in bm25_topk) for _ in gold_sp]) == len(gold_sp))\n", "\n", "print(len(tfidf_covered))\n", "print(np.mean(tfidf_covered))\n", "print(np.mean(bm25_covered))" ] }, { "cell_type": "code", "execution_count": 188, "metadata": {}, "outputs": [], "source": [ "# get transformerXH P EM\n", "dev_transformer_xh = [json.loads(l) for l in open(\"/private/home/xwhan/code/Transformer-XH/data/hotpot_dev_graph.json\").readlines()]\n", "qid2goldsp = {_[\"_id\"]:_[\"sp\"] for _ in val_inputs}" ] }, { "cell_type": "code", "execution_count": 192, "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "5904\n", "0.8162262872628726\n" ] } ], "source": [ "len(dev_transformer_xh[2][\"node\"])\n", "p_covered = []\n", "for item in dev_transformer_xh:\n", " node_names = set([_[\"name\"] for _ in item[\"node\"]])\n", " qid = item[\"qid\"]\n", " gold_sp = qid2goldsp[qid]\n", " gold_sp = [_.lower().replace(\" \", \"_\") for _ in gold_sp]\n", " p_covered.append(np.sum([int(_ in node_names) for _ in gold_sp]) == len(gold_sp))\n", "print(len(p_covered))\n", "print(np.mean(p_covered))" ] }, { "cell_type": "code", "execution_count": 294, "metadata": {}, "outputs": [], "source": [ "id2doc = json.load(open(\"/private/home/xwhan/Mhop-Pretrain/retrieval/index/abstracts_id2doc.json\"))\n", "title2text = {v[0]:v[1] for v in id2doc.values()}" ] }, { "cell_type": "code", "execution_count": 332, "metadata": {}, "outputs": [], "source": [ "# pick an intro example to use \n", "val_inputs = [json.loads(l) for l in open(\"/private/home/xwhan/data/hotpot/hotpot_qas_train.json\").readlines()]\n", "bridge_val = [_ for _ in val_inputs if _[\"type\"] == \"bridge\" and len(_[\"question\"].split()) < 10]" ] }, { "cell_type": "code", "execution_count": 371, "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "3826\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" ] } ], "source": [ "print(len(bridge_val))\n", "import random\n", "random.shuffle(bridge_val)\n", "print(bridge_val[0])" ] }, { "cell_type": "code", "execution_count": 328, "metadata": {}, "outputs": [ { "data": { "text/plain": [ "'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\".'" ] }, "execution_count": 328, "metadata": {}, "output_type": "execute_result" } ], "source": [ "title2text['Clark Gable']" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "{'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", "{'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", "{'question': 'Where does the descendant of the Red Setter originate? ', '_id': '5abde4595542991f66106095', 'answer': ['Scotland'], 'sp': ['Irish Setter', 'Scotch Collie'], 'type': 'bridge'}" ] }, { "cell_type": "code", "execution_count": 453, "metadata": {}, "outputs": [], "source": [ "corpus = json.load(open(\"index/hotpotQA_corpus_dict.json\"))\n", "title2text = {v[\"title\"]:v[\"text\"] for v in corpus.values()}" ] }, { "cell_type": "code", "execution_count": 456, "metadata": {}, "outputs": [], "source": [ "# bridge errors after reranking\n", "val_inputs = [json.loads(l) for l in open(\"/private/home/xwhan/data/hotpot/hotpot_qas_val.json\").readlines()]\n", "id2goldsp = {_[\"_id\"]:_[\"sp\"] for _ in val_inputs}\n", "id2goldans = {_[\"_id\"]:_[\"answer\"] for _ in val_inputs}\n", "id2type = {_[\"_id\"]:_[\"type\"] for _ in val_inputs}\n", "id2item = {_[\"_id\"]:_ for _ in val_inputs}\n", "results = json.load(open(\"/private/home/xwhan/data/hotpot/results/hotpot_val_b250_k250.json\"))" ] }, { "cell_type": "code", "execution_count": 469, "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "0.20986819871578236\n", "7405\n" ] } ], "source": [ "bridge_errors = []\n", "bridge_c = 0\n", "for qid in results[\"titles\"].keys():\n", " type_ = id2type[qid]\n", " if type_ != \"bridge\":\n", " continue\n", " chain = results[\"titles\"][qid]\n", " sp = id2goldsp[qid]\n", " sp_covered = int(np.sum([int(_ in chain) for _ in sp]) == len(sp))\n", " if not sp_covered:\n", " bridge_errors.append({\n", " \"id\":qid,\n", " \"question\": id2item[qid][\"question\"],\n", " \"sp\": sp,\n", " \"error\": chain,\n", " \"answer\": id2goldans[qid],\n", " })\n", " bridge_c += 1\n", "print(len(bridge_errors)/ bridge_c)\n", "print(len(results[\"titles\"].keys()))" ] }, { "cell_type": "code", "execution_count": 470, "metadata": {}, "outputs": [], "source": [ "random.shuffle(bridge_errors)" ] }, { "cell_type": "code", "execution_count": 546, "metadata": {}, "outputs": [], "source": [ "id2error = {_[\"id\"]: idx for idx, _ in enumerate(bridge_errors[:50])}\n", "json.dump(id2error, open(\"/private/home/xwhan/data/hotpot/retrieval_errors_50sampled.json\", \"w\"))" ] }, { "cell_type": "code", "execution_count": 564, "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "{'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", "Gold:\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", "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", "Predicted:\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", "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" ] } ], "source": [ "idx = id2error['5abba91e554299642a094b10']\n", "item = bridge_errors[idx]\n", "print(item)\n", "print(\"Gold:\")\n", "for t in item[\"sp\"]:\n", " print(title2text[t])\n", "print(\"Predicted:\")\n", "for t in item[\"error\"]:\n", " print(title2text[t])" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [] } ], "metadata": { "kernelspec": { "display_name": "Python 3", "language": "python", "name": "python3" }, "language_info": { "codemirror_mode": { "name": "ipython", "version": 3 }, "file_extension": ".py", "mimetype": "text/x-python", "name": "python", "nbconvert_exporter": "python", "pygments_lexer": "ipython3", "version": "3.6.3" } }, "nbformat": 4, "nbformat_minor": 2 } ================================================ FILE: mdr/retrieval/interactive_retrieval.py ================================================ # Copyright (c) Facebook, Inc. and its affiliates. # All rights reserved. # # This source code is licensed under the license found in the # LICENSE file in the root directory of this source tree. from models.mhop_retriever import MhopRetriever import faiss import numpy as np import torch from tqdm import tqdm from transformers import AutoConfig, AutoTokenizer import json import logging import argparse from .utils.utils import (load_saved, move_to_cuda) parser = argparse.ArgumentParser() parser.add_argument('--topk', type=int, default=2, help="topk paths") parser.add_argument('--num-workers', type=int, default=10) parser.add_argument('--max-q-len', type=int, default=70) parser.add_argument('--max-c-len', type=int, default=300) parser.add_argument('--max-q-sp-len', type=int, default=350) parser.add_argument('--model-name', type=str, default='bert-base-uncased') parser.add_argument('--gpu', action="store_true") parser.add_argument('--shared-encoder', action="store_true") parser.add_argument("--stop-drop", default=0, type=float) args = parser.parse_args() index_path = "index/abstracts_v0_fixed.npy" corpus_path = "index/abstracts_id2doc.json" model_path = "logs/08-05-2020/baseline_v0_fixed-seed16-bsz150-fp16True-lr2e-05-decay0.0-warm0.1-valbsz3000-sharedTrue-multi1-schemenone/checkpoint_best.pt" print(f"Loading corpus and index...") id2doc = json.load(open(corpus_path)) index_vectors = np.load(index_path).astype('float32') index = faiss.IndexFlatIP(768) index.add(index_vectors) res = faiss.StandardGpuResources() index = faiss.index_cpu_to_gpu(res, 1, index) print(f"Loading retrieval model...") bert_config = AutoConfig.from_pretrained("bert-base-uncased") tokenizer = AutoTokenizer.from_pretrained("bert-base-uncased") model = MhopRetriever(bert_config, args) model = load_saved(model, args.model_path, exact=False) cuda = torch.device('cuda') model.to(cuda) from apex import amp model = amp.initialize(model, opt_level='O1') model.eval() while True: question = input("Type Question:") question = "the Danish musicians who died in 1931" batch_q_encodes = tokenizer.batch_encode_plus(["question"], max_length=args.max_q_len, pad_to_max_length=True, return_tensors="pt") batch_q_encodes = move_to_cuda(dict(batch_q_encodes)) q_embeds = model.encode_q(batch_q_encodes["input_ids"], batch_q_encodes["attention_mask"], batch_q_encodes.get("token_type_ids", None)) q_embeds_numpy = q_embeds.cpu().contiguous().numpy() D, I = index.search(q_embeds_numpy, 1) print(I) ================================================ FILE: mdr/retrieval/mhop_trainer.py ================================================ # Copyright (c) Facebook, Inc. and its affiliates. # All rights reserved. # # This source code is licensed under the license found in the # LICENSE file in the root directory of this source tree. """ submitit trainer for hyperparameter tuning """ import os import os.path as osp from typing import Optional, NamedTuple import torch import torch.distributed import torch.nn as nn import torch.optim as optim import attr import submitit from functools import partial import numpy as np import random from torch.utils.tensorboard import SummaryWriter from pathlib import Path import json from transformers import ( AdamW, AutoConfig, AutoTokenizer, get_linear_schedule_with_warmup) from torch.optim import Adam from .utils.utils import move_to_cuda, AverageMeter from .config import ClusterConfig from .data.mhop_dataset import MhopDataset, mhop_collate from .models.mhop_retriever import (MhopRetriever, RobertaRetriever) from .criterions import (mhop_loss, mhop_eval) from tqdm import tqdm import apex apex.amp.register_half_function(torch, 'einsum') from apex import amp @attr.s(auto_attribs=True) class TrainerState: """ Contains the state of the Trainer. It can be saved to checkpoint the training and loaded to resume it. """ epoch: int model: nn.Module optimizer: optim.Optimizer lr_scheduler: torch.optim.lr_scheduler._LRScheduler global_step: int def save(self, filename: str) -> None: data = attr.asdict(self) # store only the state dict data["model"] = self.model.state_dict() data["optimizer"] = self.optimizer.state_dict() data["lr_scheduler"] = self.lr_scheduler.state_dict() torch.save(data, filename) @classmethod def load(cls, filename: str, default: "TrainerState", gpu: int) -> "TrainerState": data = torch.load(filename, map_location=lambda storage, loc: storage.cuda(gpu)) # We need this default to load the state dict model = default.model model.load_state_dict(data["model"]) data["model"] = model optimizer = default.optimizer optimizer.load_state_dict(data["optimizer"]) data["optimizer"] = optimizer lr_scheduler = default.lr_scheduler lr_scheduler.load_state_dict(data["lr_scheduler"]) data["lr_scheduler"] = lr_scheduler return cls(**data) class Trainer: def __init__(self, train_cfg: NamedTuple, cluster_cfg: ClusterConfig) -> None: self._train_cfg = train_cfg self._cluster_cfg = cluster_cfg def __call__(self) -> Optional[float]: """ Called by submitit for each task. :return: The master task return the final accuracy of the model. """ self._setup_process_group() self._init_state() final_acc = self._train() return final_acc def log(self, log_data: dict): job_env = submitit.JobEnvironment() # z = {**vars(self._train_cfg), **log_data} save_dir = Path(self._train_cfg.output_dir) os.makedirs(save_dir, exist_ok=True) with open(save_dir / 'log.txt', 'a') as f: f.write(json.dumps(log_data) + '\n') def checkpoint(self, rm_init=True) -> submitit.helpers.DelayedSubmission: # will be called by submitit in case of preemption job_env = submitit.JobEnvironment() save_dir = osp.join(self._train_cfg.output_dir, str(job_env.job_id)) os.makedirs(save_dir, exist_ok=True) self._state.save(osp.join(save_dir, "checkpoint.pth")) # Trick here: when the job will be requeue, we will use the same init file # but it must not exist when we initialize the process group # so we delete it, but only when this method is called by submitit for requeue if rm_init and osp.exists(self._cluster_cfg.dist_url[7:]): os.remove(self._cluster_cfg.dist_url[7:]) # remove file:// at the beginning # This allow to remove any non-pickable part of the Trainer instance. empty_trainer = Trainer(self._train_cfg, self._cluster_cfg) return submitit.helpers.DelayedSubmission(empty_trainer) def _setup_process_group(self) -> None: job_env = submitit.JobEnvironment() torch.cuda.set_device(job_env.local_rank) torch.distributed.init_process_group( backend=self._cluster_cfg.dist_backend, init_method=self._cluster_cfg.dist_url, world_size=job_env.num_tasks, rank=job_env.global_rank, ) print(f"Process group: {job_env.num_tasks} tasks, rank: {job_env.global_rank}") def _init_state(self) -> None: """ Initialize the state and load it from an existing checkpoint if any """ job_env = submitit.JobEnvironment() if job_env.global_rank == 0: # config_path = Path(args.save_folder) / str(job_env.job_id) / 'config.json' os.makedirs(self._train_cfg.output_dir, exist_ok=True) config_path = Path(self._train_cfg.output_dir) / 'config.json' with open(config_path, "w") as g: g.write(json.dumps(self._train_cfg._asdict())) print(f"Setting random seed {self._train_cfg.seed}", flush=True) random.seed(self._train_cfg.seed) np.random.seed(self._train_cfg.seed) torch.manual_seed(self._train_cfg.seed) torch.cuda.manual_seed_all(self._train_cfg.seed) print("Create data loaders", flush=True) tokenizer = AutoTokenizer.from_pretrained(self._train_cfg.model_name) collate_fc = partial(mhop_collate, pad_id=tokenizer.pad_token_id) 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) 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) 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) self._test_loader = torch.utils.data.DataLoader( test_set, batch_size=self._train_cfg.predict_batch_size, num_workers=self._train_cfg.num_workers, collate_fn=collate_fc, pin_memory=True ) print("Create model", flush=True) print(f"Local rank {job_env.local_rank}", flush=True) bert_config = AutoConfig.from_pretrained(self._train_cfg.model_name) if "roberta" in self._train_cfg.model_name: model = RobertaRetriever(bert_config, self._train_cfg) else: model = MhopRetriever(bert_config, self._train_cfg) model.cuda(job_env.local_rank) no_decay = ['bias', 'LayerNorm.weight'] optimizer_parameters = [ {'params': [p for n, p in model.named_parameters() if not any( nd in n for nd in no_decay)], 'weight_decay': self._train_cfg.weight_decay}, {'params': [p for n, p in model.named_parameters() if any( nd in n for nd in no_decay)], 'weight_decay': 0.0} ] optimizer = Adam(optimizer_parameters, lr=self._train_cfg.learning_rate, eps=self._train_cfg.adam_epsilon) if self._train_cfg.fp16: model, optimizer = amp.initialize( model, optimizer, opt_level=self._train_cfg.fp16_opt_level) t_total = len(self._train_loader) // self._train_cfg.gradient_accumulation_steps * self._train_cfg.num_train_epochs warmup_steps = t_total * self._train_cfg.warmup_ratio lr_scheduler = get_linear_schedule_with_warmup( optimizer, num_warmup_steps=warmup_steps, num_training_steps=t_total ) model = torch.nn.DataParallel(model) self._state = TrainerState( epoch=0, model=model, optimizer=optimizer, lr_scheduler=lr_scheduler, global_step=0 ) self.tb_logger = SummaryWriter(self._train_cfg.output_dir.replace("logs", "tflogs")) checkpoint_fn = osp.join(self._train_cfg.output_dir, str(job_env.job_id), "checkpoint.pth") # checkpoint_fn = osp.join(self._train_cfg.output_dir, "checkpoint.pth") if os.path.isfile(checkpoint_fn): print(f"Load existing checkpoint from {checkpoint_fn}", flush=True) self._state = TrainerState.load( checkpoint_fn, default=self._state, gpu=job_env.local_rank) def _train(self) -> Optional[float]: job_env = submitit.JobEnvironment() batch_step = 0 # forward batch count best_mrr = 0 train_loss_meter = AverageMeter() print(f"Start training", flush=True) # Start from the loaded epoch start_epoch = self._state.epoch global_step = self._state.global_step for epoch in range(start_epoch, self._train_cfg.num_train_epochs): print(f"Start epoch {epoch}", flush=True) self._state.model.train() self._state.epoch = epoch for batch in self._train_loader: batch_step += 1 batch = move_to_cuda(batch) loss = mhop_loss(self._state.model, batch, self._train_cfg) if self._train_cfg.gradient_accumulation_steps > 1: loss = loss / self._train_cfg.gradient_accumulation_steps if self._train_cfg.fp16: with amp.scale_loss(loss, self._state.optimizer) as scaled_loss: scaled_loss.backward() else: loss.backward() train_loss_meter.update(loss.item()) if (batch_step + 1) % self._train_cfg.gradient_accumulation_steps == 0: if self._train_cfg.fp16: torch.nn.utils.clip_grad_norm_( amp.master_params(self._state.optimizer), self._train_cfg.max_grad_norm) else: torch.nn.utils.clip_grad_norm_( self._state.model.parameters(), self._train_cfg.max_grad_norm) self._state.optimizer.step() self._state.lr_scheduler.step() self._state.model.zero_grad() global_step += 1 self._state.global_step = global_step self.tb_logger.add_scalar('batch_train_loss', loss.item(), global_step) self.tb_logger.add_scalar('smoothed_train_loss', train_loss_meter.avg, global_step) # Checkpoint only on the master # if job_env.global_rank == 0: self.checkpoint(rm_init=False) mrrs = self._eval() mrr = mrrs["mrr_avg"] self.tb_logger.add_scalar('dev_mrr', mrr*100, epoch) self._state.lr_scheduler.step(mrr) if best_mrr < mrr: print("Saving model with best MRR %.2f -> MRR %.2f on epoch=%d" % (best_mrr*100, mrr*100, epoch)) torch.save(self._state.model.state_dict(), os.path.join(self._train_cfg.output_dir, f"checkpoint_best.pt")) best_mrr = mrr self.log({ "best_mrr": best_mrr, "curr_mrr": mrr, "smoothed_loss": train_loss_meter.avg, "epoch": epoch }) return best_mrr def _eval(self) -> float: print("Start evaluation of the model", flush=True) job_env = submitit.JobEnvironment() args = self._train_cfg eval_dataloader = self._test_loader self._state.model.eval() rrs_1, rrs_2 = [], [] # reciprocal rank for batch in tqdm(eval_dataloader): batch_to_feed = move_to_cuda(batch) with torch.no_grad(): outputs = self._state.model(batch_to_feed) eval_results = mhop_eval(outputs, args) _rrs_1, _rrs_2 = eval_results["rrs_1"], eval_results["rrs_2"] rrs_1 += _rrs_1 rrs_2 += _rrs_2 mrr_1 = np.mean(rrs_1) mrr_2 = np.mean(rrs_2) print(f"evaluated {len(rrs_1)} examples...") print(f'MRR-1: {mrr_1}') print(f'MRR-2: {mrr_2}') self._state.model.train() return {"mrr_1": mrr_1, "mrr_2": mrr_2, "mrr_avg": (mrr_1 + mrr_2) / 2} ================================================ FILE: mdr/retrieval/single_trainer.py ================================================ # Copyright (c) Facebook, Inc. and its affiliates. # All rights reserved. # # This source code is licensed under the license found in the # LICENSE file in the root directory of this source tree. """ trainer defined for submitit hyperparameter tuning """ import os import os.path as osp from typing import Optional, NamedTuple import torch import torch.distributed import torch.nn as nn import torch.optim as optim import attr import submitit import argparse from functools import partial from torch.nn import CrossEntropyLoss import numpy as np import random from torch.utils.tensorboard import SummaryWriter from pathlib import Path from .utils import move_to_cuda, convert_to_half, AverageMeter from .config import ClusterConfig from .data.sp_datasets import SPDataset, sp_collate from .models.retriever import BertForRetrieverSP from transformers import AdamW, BertConfig, BertTokenizer import json import apex apex.amp.register_half_function(torch, 'einsum') from apex import amp @attr.s(auto_attribs=True) class TrainerState: """ Contains the state of the Trainer. It can be saved to checkpoint the training and loaded to resume it. """ epoch: int model: nn.Module optimizer: optim.Optimizer lr_scheduler: torch.optim.lr_scheduler._LRScheduler global_step: int def save(self, filename: str) -> None: data = attr.asdict(self) # store only the state dict data["model"] = self.model.state_dict() data["optimizer"] = self.optimizer.state_dict() data["lr_scheduler"] = self.lr_scheduler.state_dict() torch.save(data, filename) @classmethod def load(cls, filename: str, default: "TrainerState", gpu: int) -> "TrainerState": data = torch.load(filename, map_location=lambda storage, loc: storage.cuda(gpu)) # We need this default to load the state dict model = default.model model.load_state_dict(data["model"]) data["model"] = model optimizer = default.optimizer optimizer.load_state_dict(data["optimizer"]) data["optimizer"] = optimizer lr_scheduler = default.lr_scheduler lr_scheduler.load_state_dict(data["lr_scheduler"]) data["lr_scheduler"] = lr_scheduler return cls(**data) class Trainer: def __init__(self, train_cfg: NamedTuple, cluster_cfg: ClusterConfig) -> None: self._train_cfg = train_cfg self._cluster_cfg = cluster_cfg def __call__(self) -> Optional[float]: """ Called by submitit for each task. :return: The master task return the final accuracy of the model. """ self._setup_process_group() self._init_state() final_acc = self._train() return final_acc def log(self, log_data: dict): job_env = submitit.JobEnvironment() # z = {**vars(self._train_cfg), **log_data} save_dir = Path(self._train_cfg.output_dir) os.makedirs(save_dir, exist_ok=True) with open(save_dir / 'log.txt', 'a') as f: f.write(json.dumps(log_data) + '\n') def checkpoint(self, rm_init=True) -> submitit.helpers.DelayedSubmission: # will be called by submitit in case of preemption job_env = submitit.JobEnvironment() save_dir = osp.join(self._train_cfg.output_dir, str(job_env.job_id)) os.makedirs(save_dir, exist_ok=True) self._state.save(osp.join(save_dir, "checkpoint.pth")) # Trick here: when the job will be requeue, we will use the same init file # but it must not exist when we initialize the process group # so we delete it, but only when this method is called by submitit for requeue if rm_init and osp.exists(self._cluster_cfg.dist_url[7:]): os.remove(self._cluster_cfg.dist_url[7:]) # remove file:// at the beginning # This allow to remove any non-pickable part of the Trainer instance. empty_trainer = Trainer(self._train_cfg, self._cluster_cfg) return submitit.helpers.DelayedSubmission(empty_trainer) def _setup_process_group(self) -> None: job_env = submitit.JobEnvironment() torch.cuda.set_device(job_env.local_rank) torch.distributed.init_process_group( backend=self._cluster_cfg.dist_backend, init_method=self._cluster_cfg.dist_url, world_size=job_env.num_tasks, rank=job_env.global_rank, ) print(f"Process group: {job_env.num_tasks} tasks, rank: {job_env.global_rank}") def _init_state(self) -> None: """ Initialize the state and load it from an existing checkpoint if any """ job_env = submitit.JobEnvironment() if job_env.global_rank == 0: # config_path = Path(args.save_folder) / str(job_env.job_id) / 'config.json' os.makedirs(self._train_cfg.output_dir, exist_ok=True) config_path = Path(self._train_cfg.output_dir) / 'config.json' with open(config_path, "w") as g: g.write(json.dumps(self._train_cfg._asdict())) print(f"Setting random seed {self._train_cfg.seed}", flush=True) random.seed(self._train_cfg.seed) np.random.seed(self._train_cfg.seed) torch.manual_seed(self._train_cfg.seed) print("Create data loaders", flush=True) tokenizer = BertTokenizer.from_pretrained(self._train_cfg.bert_model_name) collate_fc = sp_collate train_set = SPDataset(tokenizer, self._train_cfg.train_file, self._train_cfg.max_q_len, self._train_cfg.max_c_len, train=True) # train_sampler = torch.utils.data.distributed.DistributedSampler( # train_set, num_replicas=job_env.num_tasks, rank=job_env.global_rank # ) # self._train_loader = torch.utils.data.DataLoader( # train_set, # batch_size=self._train_cfg.train_batch_size, # num_workers=4, # sampler=train_sampler, collate_fn=collate_fc # ) self._train_loader = torch.utils.data.DataLoader(train_set, batch_size=self._train_cfg.train_batch_size, num_workers=4, collate_fn=collate_fc) test_set = SPDataset(tokenizer, self._train_cfg.predict_file, self._train_cfg.max_q_len, self._train_cfg.max_c_len) self._test_loader = torch.utils.data.DataLoader( test_set, batch_size=self._train_cfg.predict_batch_size, num_workers=4, collate_fn=collate_fc ) print(f"Per Node batch_size: {self._train_cfg.train_batch_size // job_env.num_tasks}", flush=True) print("Create model", flush=True) print(f"Local rank {job_env.local_rank}", flush=True) bert_config = BertConfig.from_pretrained(self._train_cfg.bert_model_name) model = BertForRetrieverSP(bert_config, self._train_cfg) model.cuda(job_env.local_rank) no_decay = ['bias', 'LayerNorm.weight'] optimizer_parameters = [ {'params': [p for n, p in model.named_parameters() if not any( nd in n for nd in no_decay)], 'weight_decay': self._train_cfg.weight_decay}, {'params': [p for n, p in model.named_parameters() if any( nd in n for nd in no_decay)], 'weight_decay': 0.0} ] optimizer = AdamW(optimizer_parameters, lr=self._train_cfg.learning_rate) lr_scheduler = optim.lr_scheduler.ReduceLROnPlateau(optimizer, mode='max', factor=0.5) if self._train_cfg.fp16: model, optimizer = amp.initialize( model, optimizer, opt_level=self._train_cfg.fp16_opt_level) model = torch.nn.DataParallel(model) # self._state = TrainerState( epoch=0, model=model, optimizer=optimizer, lr_scheduler=lr_scheduler, global_step=0 ) self.tb_logger = SummaryWriter(os.path.join(self._train_cfg.output_dir, "tblog")) checkpoint_fn = osp.join(self._train_cfg.output_dir, str(job_env.job_id), "checkpoint.pth") # checkpoint_fn = osp.join(self._train_cfg.output_dir, "checkpoint.pth") if os.path.isfile(checkpoint_fn): print(f"Load existing checkpoint from {checkpoint_fn}", flush=True) self._state = TrainerState.load( checkpoint_fn, default=self._state, gpu=job_env.local_rank) def _train(self) -> Optional[float]: job_env = submitit.JobEnvironment() loss_fct = CrossEntropyLoss() batch_step = 0 # forward batch count best_mrr = 0 train_loss_meter = AverageMeter() print(f"Start training", flush=True) # Start from the loaded epoch start_epoch = self._state.epoch global_step = self._state.global_step for epoch in range(start_epoch, self._train_cfg.num_train_epochs): print(f"Start epoch {epoch}", flush=True) self._state.model.train() self._state.epoch = epoch for batch in self._train_loader: batch_step += 1 batch = move_to_cuda(batch) outputs = self._state.model(batch) q = outputs['q'] c = outputs['c'] neg_c = outputs['neg_c'] product_in_batch = torch.mm(q, c.t()) product_neg = (q * neg_c).sum(-1).unsqueeze(1) product = torch.cat([product_in_batch, product_neg], dim=-1) target = torch.arange(product.size(0)).to(product.device) loss = loss_fct(product, target) if self._train_cfg.gradient_accumulation_steps > 1: loss = loss / self._train_cfg.gradient_accumulation_steps if self._train_cfg.fp16: with amp.scale_loss(loss, self._state.optimizer) as scaled_loss: scaled_loss.backward() else: loss.backward() train_loss_meter.update(loss.item()) self.tb_logger.add_scalar('batch_train_loss', loss.item(), global_step) self.tb_logger.add_scalar('smoothed_train_loss', train_loss_meter.avg, global_step) if (batch_step + 1) % self._train_cfg.gradient_accumulation_steps == 0: if self._train_cfg.fp16: torch.nn.utils.clip_grad_norm_( amp.master_params(self._state.optimizer), self._train_cfg.max_grad_norm) else: torch.nn.utils.clip_grad_norm_( self._state.model.parameters(), self._train_cfg.max_grad_norm) self._state.optimizer.step() # We have accumulated enought gradients self._state.model.zero_grad() global_step += 1 self._state.global_step = global_step # Checkpoint only on the master # if job_env.global_rank == 0: self.checkpoint(rm_init=False) mrr = self._eval() self.tb_logger.add_scalar('dev_mrr', mrr*100, epoch) self._state.lr_scheduler.step(mrr) if best_mrr < mrr: print("Saving model with best MRR %.2f -> MRR %.2f on epoch=%d" % (best_mrr*100, mrr*100, epoch)) torch.save(self._state.model.state_dict(), os.path.join(self._train_cfg.output_dir, f"checkpoint_best.pt")) best_mrr = mrr self.log({ "best_mrr": best_mrr, "curr_mrr": mrr, "smoothed_loss": train_loss_meter.avg, "epoch": epoch }) return best_mrr def _eval(self) -> float: print("Start evaluation of the model", flush=True) job_env = submitit.JobEnvironment() args = self._train_cfg eval_dataloader = self._test_loader num_correct = 0 num_total = 0.0 rrs = [] # reciprocal rank self._state.model.eval() for batch in self._test_loader: batch_to_feed = move_to_cuda(batch) with torch.no_grad(): outputs = self._state.model(batch_to_feed) q = outputs['q'] c = outputs['c'] neg_c = outputs['neg_c'] product_in_batch = torch.mm(q, c.t()) product_neg = (q * neg_c).sum(-1).unsqueeze(1) product = torch.cat([product_in_batch, product_neg], dim=-1) target = torch.arange(product.size(0)).to(product.device) ranked = product.argsort(dim=1, descending=True) # MRR idx2rank = ranked.argsort(dim=1) for idx, t in enumerate(target.tolist()): rrs.append(1 / (idx2rank[idx][t].item() +1)) prediction = product.argmax(-1) pred_res = prediction == target num_total += pred_res.size(0) num_correct += pred_res.sum(0) acc = num_correct/num_total mrr = np.mean(rrs) print(f"evaluated {num_total} examples...", flush=True) print(f"avg. Acc: {acc}", flush=True) print(f'MRR: {mrr}', flush=True) self._state.model.train() return mrr ================================================ FILE: mdr/retrieval/train_single.py ================================================ # Copyright (c) Facebook, Inc. and its affiliates. # All rights reserved. # # This source code is licensed under the license found in the # LICENSE file in the root directory of this source tree. """ # DPR baseline shared encoder CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python train_single.py \ --do_train \ --prefix nq_dpr_shared \ --predict_batch_size 5000 \ --model_name bert-base-uncased \ --train_batch_size 256 \ --gradient_accumulation_steps 1 \ --accumulate_gradients 1 \ --learning_rate 2e-5 \ --fp16 \ --train_file /private/home/xwhan/data/nq-dpr/nq-with-neg-train.txt \ --predict_file /private/home/xwhan/data/nq-dpr/nq-with-neg-dev.txt \ --seed 16 \ --eval-period -1 \ --max_c_len 300 \ --max_q_len 50 \ --warmup-ratio 0.1 \ --shared-encoder \ --num_train_epochs 50 # WebQ single train CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python train_single.py \ --do_train \ --prefix wq_dpr_shared \ --predict_batch_size 5000 \ --model_name bert-base-uncased \ --train_batch_size 256 \ --gradient_accumulation_steps 1 \ --accumulate_gradients 1 \ --learning_rate 2e-5 \ --fp16 \ --train_file /private/home/xwhan/data/WebQ/wq-train-simplified.txt \ --predict_file /private/home/xwhan/data/WebQ/wq-dev-simplified.txt \ --seed 16 \ --eval-period -1 \ --max_c_len 300 \ --max_q_len 50 \ --warmup-ratio 0.1 \ --shared-encoder \ --num_train_epochs 50 # FEVER single-hop retrieval CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python train_single.py \ --do_train \ --prefix fever_single \ --predict_batch_size 5000 \ --model_name bert-base-uncased \ --train_batch_size 256 \ --gradient_accumulation_steps 1 \ --accumulate_gradients 1 \ --learning_rate 2e-5 \ --fp16 \ --train_file /private/home/xwhan/data/fever/retrieval/train_tfidf_neg.txt \ --predict_file /private/home/xwhan/data/fever/retrieval/dev_tfidf_neg.txt \ --seed 16 \ --eval-period -1 \ --max_c_len 400 \ --max_q_len 45 \ --shared-encoder \ --num_train_epochs 40 # HotpotQA single-hop CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python train_single.py \ --do_train \ --prefix hotpot_single \ --predict_batch_size 5000 \ --model_name roberta-base \ --train_batch_size 256 \ --gradient_accumulation_steps 1 \ --accumulate_gradients 1 \ --learning_rate 2e-5 \ --fp16 \ --train_file /private/home/xwhan/data/hotpot/hotpot_train_with_neg_v0.json \ --predict_file /private/home/xwhan/data/hotpot/hotpot_val_with_neg_v0.json \ --seed 16 \ --eval-period -1 \ --max_c_len 300 \ --max_q_len 70 \ --shared-encoder \ --warmup-ratio 0.1 \ --num_train_epochs 50 """ import logging import os import random from tqdm import tqdm import numpy as np import torch from datetime import date from torch.utils.data import DataLoader from models.retriever import BertRetrieverSingle, RobertaRetrieverSingle, MomentumRetriever from transformers import AdamW, AutoConfig, AutoTokenizer, get_linear_schedule_with_warmup from torch.utils.tensorboard import SummaryWriter from data.sp_datasets import SPDataset, sp_collate, NQMhopDataset, FeverSingleDataset from utils.utils import move_to_cuda, AverageMeter, load_saved from config import train_args from criterions import loss_single from torch.optim import Adam from functools import partial import apex def main(): args = train_args() if args.fp16: apex.amp.register_half_function(torch, 'einsum') date_curr = date.today().strftime("%m-%d-%Y") 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}" args.output_dir = os.path.join(args.output_dir, date_curr, model_name) tb_logger = SummaryWriter(os.path.join(args.output_dir.replace("logs","tflogs"))) if os.path.exists(args.output_dir) and os.listdir(args.output_dir): print( f"output directory {args.output_dir} already exists and is not empty.") os.makedirs(args.output_dir, exist_ok=True) logging.basicConfig(format='%(asctime)s - %(levelname)s - %(name)s - %(message)s', datefmt='%m/%d/%Y %H:%M:%S', level=logging.INFO, handlers=[logging.FileHandler(os.path.join(args.output_dir, "log.txt")), logging.StreamHandler()]) logger = logging.getLogger(__name__) logger.info(args) if args.local_rank == -1 or args.no_cuda: device = torch.device( "cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu") n_gpu = torch.cuda.device_count() else: device = torch.device("cuda", args.local_rank) n_gpu = 1 torch.distributed.init_process_group(backend='nccl') logger.info("device %s n_gpu %d distributed training %r", device, n_gpu, bool(args.local_rank != -1)) args.train_batch_size = int( args.train_batch_size / args.accumulate_gradients) random.seed(args.seed) np.random.seed(args.seed) torch.manual_seed(args.seed) if n_gpu > 0: torch.cuda.manual_seed_all(args.seed) if not args.do_train and not args.do_predict: raise ValueError( "At least one of `do_train` or `do_predict` must be True.") bert_config = AutoConfig.from_pretrained(args.model_name) if args.momentum: model = MomentumRetriever(bert_config, args) elif "roberta" in args.model_name: model = RobertaRetrieverSingle(bert_config, args) else: model = BertRetrieverSingle(bert_config, args) tokenizer = AutoTokenizer.from_pretrained(args.model_name) collate_fc = partial(sp_collate, pad_id=tokenizer.pad_token_id) if args.do_train and args.max_c_len > bert_config.max_position_embeddings: raise ValueError( "Cannot use sequence length %d because the BERT model " "was only trained up to sequence length %d" % (args.max_c_len, bert_config.max_position_embeddings)) if "fever" in args.predict_file: eval_dataset = FeverSingleDataset(tokenizer, args.predict_file, args.max_q_len, args.max_c_len) else: eval_dataset = SPDataset(tokenizer, args.predict_file, args.max_q_len, args.max_c_len) eval_dataloader = DataLoader( eval_dataset, batch_size=args.predict_batch_size, collate_fn=collate_fc, pin_memory=True, num_workers=args.num_workers) logger.info(f"Num of dev batches: {len(eval_dataloader)}") if args.init_checkpoint != "": model = load_saved(model, args.init_checkpoint) model.to(device) print(f"number of trainable parameters: {sum(p.numel() for p in model.parameters() if p.requires_grad)}") if args.do_train: no_decay = ['bias', 'LayerNorm.weight'] optimizer_parameters = [ {'params': [p for n, p in model.named_parameters() if not any( nd in n for nd in no_decay)], 'weight_decay': args.weight_decay}, {'params': [p for n, p in model.named_parameters() if any( nd in n for nd in no_decay)], 'weight_decay': 0.0} ] optimizer = Adam(optimizer_parameters, lr=args.learning_rate, eps=args.adam_epsilon) if args.fp16: model, optimizer = apex.amp.initialize(model, optimizer, opt_level=args.fp16_opt_level) else: if args.fp16: model = apex.amp.initialize(model, opt_level=args.fp16_opt_level) if n_gpu > 1: model = torch.nn.DataParallel(model) if args.do_train: global_step = 0 # gradient update step batch_step = 0 # forward batch count best_mrr = 0 train_loss_meter = AverageMeter() model.train() if "fever" in args.predict_file: train_dataset = FeverSingleDataset(tokenizer, args.train_file, args.max_q_len, args.max_c_len, train=True) else: train_dataset = SPDataset(tokenizer, args.train_file, args.max_q_len, args.max_c_len, train=True) 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) t_total = len(train_dataloader) // args.gradient_accumulation_steps * args.num_train_epochs warmup_steps = t_total * args.warmup_ratio scheduler = get_linear_schedule_with_warmup( optimizer, num_warmup_steps=warmup_steps, num_training_steps=t_total ) logger.info('Start training....') for epoch in range(int(args.num_train_epochs)): for batch in tqdm(train_dataloader): batch_step += 1 batch = move_to_cuda(batch) loss = loss_single(model, batch, args.momentum) if args.gradient_accumulation_steps > 1: loss = loss / args.gradient_accumulation_steps if args.fp16: with apex.amp.scale_loss(loss, optimizer) as scaled_loss: scaled_loss.backward() else: loss.backward() train_loss_meter.update(loss.item()) if (batch_step + 1) % args.gradient_accumulation_steps == 0: if args.fp16: torch.nn.utils.clip_grad_norm_( apex.amp.master_params(optimizer), args.max_grad_norm) else: torch.nn.utils.clip_grad_norm_( model.parameters(), args.max_grad_norm) optimizer.step() scheduler.step() model.zero_grad() global_step += 1 tb_logger.add_scalar('batch_train_loss', loss.item(), global_step) tb_logger.add_scalar('smoothed_train_loss', train_loss_meter.avg, global_step) if args.eval_period != -1 and global_step % args.eval_period == 0: mrr = predict(args, model, eval_dataloader, device, logger) logger.info("Step %d Train loss %.2f MRR %.2f on epoch=%d" % (global_step, train_loss_meter.avg, mrr*100, epoch)) if best_mrr < mrr: logger.info("Saving model with best MRR %.2f -> MRR %.2f on epoch=%d" % (best_mrr*100, mrr*100, epoch)) torch.save(model.state_dict(), os.path.join( args.output_dir, f"checkpoint_best.pt")) model = model.to(device) best_mrr = mrr mrr = predict(args, model, eval_dataloader, device, logger) logger.info("Step %d Train loss %.2f MRR %.2f on epoch=%d" % ( global_step, train_loss_meter.avg, mrr*100, epoch)) tb_logger.add_scalar('dev_mrr', mrr*100, epoch) if best_mrr < mrr: torch.save(model.state_dict(), os.path.join( args.output_dir, f"checkpoint_last.pt")) logger.info("Saving model with best MRR %.2f -> MRR %.2f on epoch=%d" % (best_mrr*100, mrr*100, epoch)) torch.save(model.state_dict(), os.path.join( args.output_dir, f"checkpoint_best.pt")) model = model.to(device) best_mrr = mrr logger.info("Training finished!") elif args.do_predict: acc = predict(args, model, eval_dataloader, device, logger) logger.info(f"test performance {acc}") def predict(args, model, eval_dataloader, device, logger): model.eval() num_correct = 0 num_total = 0.0 rrs = [] # reciprocal rank for batch in tqdm(eval_dataloader): batch_to_feed = move_to_cuda(batch) with torch.no_grad(): outputs = model(batch_to_feed) q = outputs['q'] c = outputs['c'] neg_c = outputs['neg_c'] product_in_batch = torch.mm(q, c.t()) product_neg = (q * neg_c).sum(-1).unsqueeze(1) product = torch.cat([product_in_batch, product_neg], dim=-1) target = torch.arange(product.size(0)).to(product.device) ranked = product.argsort(dim=1, descending=True) prediction = product.argmax(-1) # MRR idx2rank = ranked.argsort(dim=1) for idx, t in enumerate(target.tolist()): rrs.append(1 / (idx2rank[idx][t].item() +1)) pred_res = prediction == target num_total += pred_res.size(0) num_correct += pred_res.sum(0) acc = num_correct/num_total mrr = np.mean(rrs) logger.info(f"evaluated {num_total} examples...") logger.info(f"avg. Acc: {acc}") logger.info(f'MRR: {mrr}') model.train() return mrr if __name__ == "__main__": main() ================================================ FILE: mdr/retrieval/utils/basic_tokenizer.py ================================================ # Copyright (c) Facebook, Inc. and its affiliates. # All rights reserved. # # This source code is licensed under the license found in the # LICENSE file in the root directory of this source tree. #!/usr/bin/env python3 # Copyright 2017-present, Facebook, Inc. # All rights reserved. # # This source code is licensed under the license found in the # LICENSE file in the root directory of this source tree. """Base tokenizer/tokens classes and utilities.""" import copy class Tokens(object): """A class to represent a list of tokenized text.""" TEXT = 0 TEXT_WS = 1 SPAN = 2 POS = 3 LEMMA = 4 NER = 5 def __init__(self, data, annotators, opts=None): self.data = data self.annotators = annotators self.opts = opts or {} def __len__(self): """The number of tokens.""" return len(self.data) def slice(self, i=None, j=None): """Return a view of the list of tokens from [i, j).""" new_tokens = copy.copy(self) new_tokens.data = self.data[i: j] return new_tokens def untokenize(self): """Returns the original text (with whitespace reinserted).""" return ''.join([t[self.TEXT_WS] for t in self.data]).strip() def words(self, uncased=False): """Returns a list of the text of each token Args: uncased: lower cases text """ if uncased: return [t[self.TEXT].lower() for t in self.data] else: return [t[self.TEXT] for t in self.data] def offsets(self): """Returns a list of [start, end) character offsets of each token.""" return [t[self.SPAN] for t in self.data] def pos(self): """Returns a list of part-of-speech tags of each token. Returns None if this annotation was not included. """ if 'pos' not in self.annotators: return None return [t[self.POS] for t in self.data] def lemmas(self): """Returns a list of the lemmatized text of each token. Returns None if this annotation was not included. """ if 'lemma' not in self.annotators: return None return [t[self.LEMMA] for t in self.data] def entities(self): """Returns a list of named-entity-recognition tags of each token. Returns None if this annotation was not included. """ if 'ner' not in self.annotators: return None return [t[self.NER] for t in self.data] def ngrams(self, n=1, uncased=False, filter_fn=None, as_strings=True): """Returns a list of all ngrams from length 1 to n. Args: n: upper limit of ngram length uncased: lower cases text filter_fn: user function that takes in an ngram list and returns True or False to keep or not keep the ngram as_string: return the ngram as a string vs list """ def _skip(gram): if not filter_fn: return False return filter_fn(gram) words = self.words(uncased) ngrams = [(s, e + 1) for s in range(len(words)) for e in range(s, min(s + n, len(words))) if not _skip(words[s:e + 1])] # Concatenate into strings if as_strings: ngrams = ['{}'.format(' '.join(words[s:e])) for (s, e) in ngrams] return ngrams def entity_groups(self): """Group consecutive entity tokens with the same NER tag.""" entities = self.entities() if not entities: return None non_ent = self.opts.get('non_ent', 'O') groups = [] idx = 0 while idx < len(entities): ner_tag = entities[idx] # Check for entity tag if ner_tag != non_ent: # Chomp the sequence start = idx while (idx < len(entities) and entities[idx] == ner_tag): idx += 1 groups.append((self.slice(start, idx).untokenize(), ner_tag)) else: idx += 1 return groups class Tokenizer(object): """Base tokenizer class. Tokenizers implement tokenize, which should return a Tokens class. """ def tokenize(self, text): raise NotImplementedError def shutdown(self): pass def __del__(self): self.shutdown() import regex import logging logger = logging.getLogger(__name__) class RegexpTokenizer(Tokenizer): DIGIT = r'\p{Nd}+([:\.\,]\p{Nd}+)*' TITLE = (r'(dr|esq|hon|jr|mr|mrs|ms|prof|rev|sr|st|rt|messrs|mmes|msgr)' r'\.(?=\p{Z})') ABBRV = r'([\p{L}]\.){2,}(?=\p{Z}|$)' ALPHA_NUM = r'[\p{L}\p{N}\p{M}]++' HYPHEN = r'{A}([-\u058A\u2010\u2011]{A})+'.format(A=ALPHA_NUM) NEGATION = r"((?!n't)[\p{L}\p{N}\p{M}])++(?=n't)|n't" CONTRACTION1 = r"can(?=not\b)" CONTRACTION2 = r"'([tsdm]|re|ll|ve)\b" START_DQUOTE = r'(?<=[\p{Z}\(\[{<]|^)(``|["\u0093\u201C\u00AB])(?!\p{Z})' START_SQUOTE = r'(?<=[\p{Z}\(\[{<]|^)[\'\u0091\u2018\u201B\u2039](?!\p{Z})' END_DQUOTE = r'(?<!\p{Z})(\'\'|["\u0094\u201D\u00BB])' END_SQUOTE = r'(?<!\p{Z})[\'\u0092\u2019\u203A]' DASH = r'--|[\u0096\u0097\u2013\u2014\u2015]' ELLIPSES = r'\.\.\.|\u2026' PUNCT = r'\p{P}' NON_WS = r'[^\p{Z}\p{C}]' def __init__(self, **kwargs): """ Args: annotators: None or empty set (only tokenizes). substitutions: if true, normalizes some token types (e.g. quotes). """ self._regexp = regex.compile( '(?P<digit>%s)|(?P<title>%s)|(?P<abbr>%s)|(?P<neg>%s)|(?P<hyph>%s)|' '(?P<contr1>%s)|(?P<alphanum>%s)|(?P<contr2>%s)|(?P<sdquote>%s)|' '(?P<edquote>%s)|(?P<ssquote>%s)|(?P<esquote>%s)|(?P<dash>%s)|' '(?<ellipses>%s)|(?P<punct>%s)|(?P<nonws>%s)' % (self.DIGIT, self.TITLE, self.ABBRV, self.NEGATION, self.HYPHEN, self.CONTRACTION1, self.ALPHA_NUM, self.CONTRACTION2, self.START_DQUOTE, self.END_DQUOTE, self.START_SQUOTE, self.END_SQUOTE, self.DASH, self.ELLIPSES, self.PUNCT, self.NON_WS), flags=regex.IGNORECASE + regex.UNICODE + regex.MULTILINE ) if len(kwargs.get('annotators', {})) > 0: logger.warning('%s only tokenizes! Skipping annotators: %s' % (type(self).__name__, kwargs.get('annotators'))) self.annotators = set() self.substitutions = kwargs.get('substitutions', True) def tokenize(self, text): data = [] matches = [m for m in self._regexp.finditer(text)] for i in range(len(matches)): # Get text token = matches[i].group() # Make normalizations for special token types if self.substitutions: groups = matches[i].groupdict() if groups['sdquote']: token = "``" elif groups['edquote']: token = "''" elif groups['ssquote']: token = "`" elif groups['esquote']: token = "'" elif groups['dash']: token = '--' elif groups['ellipses']: token = '...' # Get whitespace span = matches[i].span() start_ws = span[0] if i + 1 < len(matches): end_ws = matches[i + 1].span()[0] else: end_ws = span[1] # Format data data.append(( token, text[start_ws: end_ws], span, )) return Tokens(data, self.annotators) class SimpleTokenizer(Tokenizer): ALPHA_NUM = r'[\p{L}\p{N}\p{M}]+' NON_WS = r'[^\p{Z}\p{C}]' def __init__(self, **kwargs): """ Args: annotators: None or empty set (only tokenizes). """ self._regexp = regex.compile( '(%s)|(%s)' % (self.ALPHA_NUM, self.NON_WS), flags=regex.IGNORECASE + regex.UNICODE + regex.MULTILINE ) if len(kwargs.get('annotators', {})) > 0: logger.warning('%s only tokenizes! Skipping annotators: %s' % (type(self).__name__, kwargs.get('annotators'))) self.annotators = set() def tokenize(self, text): data = [] matches = [m for m in self._regexp.finditer(text)] for i in range(len(matches)): # Get text token = matches[i].group() # Get whitespace span = matches[i].span() start_ws = span[0] if i + 1 < len(matches): end_ws = matches[i + 1].span()[0] else: end_ws = span[1] # Format data data.append(( token, text[start_ws: end_ws], span, )) return Tokens(data, self.annotators) STOPWORDS = { 'i', 'me', 'my', 'myself', 'we', 'our', 'ours', 'ourselves', 'you', 'your', 'yours', 'yourself', 'yourselves', 'he', 'him', 'his', 'himself', 'she', 'her', 'hers', 'herself', 'it', 'its', 'itself', 'they', 'them', 'their', 'theirs', 'themselves', 'what', 'which', 'who', 'whom', 'this', 'that', 'these', 'those', 'am', 'is', 'are', 'was', 'were', 'be', 'been', 'being', 'have', 'has', 'had', 'having', 'do', 'does', 'did', 'doing', 'a', 'an', 'the', 'and', 'but', 'if', 'or', 'because', 'as', 'until', 'while', 'of', 'at', 'by', 'for', 'with', 'about', 'against', 'between', 'into', 'through', 'during', 'before', 'after', 'above', 'below', 'to', 'from', 'up', 'down', 'in', 'out', 'on', 'off', 'over', 'under', 'again', 'further', 'then', 'once', 'here', 'there', 'when', 'where', 'why', 'how', 'all', 'any', 'both', 'each', 'few', 'more', 'most', 'other', 'some', 'such', 'no', 'nor', 'not', 'only', 'own', 'same', 'so', 'than', 'too', 'very', 's', 't', 'can', 'will', 'just', 'don', 'should', 'now', 'd', 'll', 'm', 'o', 're', 've', 'y', 'ain', 'aren', 'couldn', 'didn', 'doesn', 'hadn', 'hasn', 'haven', 'isn', 'ma', 'mightn', 'mustn', 'needn', 'shan', 'shouldn', 'wasn', 'weren', 'won', 'wouldn', "'ll", "'re", "'ve", "n't", "'s", "'d", "'m", "''", "``" } import unicodedata def normalize(text): """Resolve different type of unicode encodings.""" return unicodedata.normalize('NFD', text) def filter_word(text): """Take out english stopwords, punctuation, and compound endings.""" text = normalize(text) if regex.match(r'^\p{P}+$', text): return True if text.lower() in STOPWORDS: return True return False def filter_ngram(gram, mode='any'): """Decide whether to keep or discard an n-gram. Args: gram: list of tokens (length N) mode: Option to throw out ngram if 'any': any single token passes filter_word 'all': all tokens pass filter_word 'ends': book-ended by filterable tokens """ filtered = [filter_word(w) for w in gram] if mode == 'any': return any(filtered) elif mode == 'all': return all(filtered) elif mode == 'ends': return filtered[0] or filtered[-1] else: raise ValueError('Invalid mode: %s' % mode) ================================================ FILE: mdr/retrieval/utils/gen_index_id_map.py ================================================ # Copyright (c) Facebook, Inc. and its affiliates. # All rights reserved. # # This source code is licensed under the license found in the # LICENSE file in the root directory of this source tree. import json mapping = {} with open('../data/para_doc.db') as f_in: for idx, line in enumerate(f_in): sample = json.loads(line.strip()) mapping[idx] = sample['id'] with open('index_data/idx_id.json', 'w') as f_out: json.dump(mapping, f_out) ================================================ FILE: mdr/retrieval/utils/mhop_utils.py ================================================ # Copyright (c) Facebook, Inc. and its affiliates. # All rights reserved. # # This source code is licensed under the license found in the # LICENSE file in the root directory of this source tree. import sys import json from tqdm import tqdm import collections import re from collections import Counter import string from basic_tokenizer import SimpleTokenizer, filter_ngram import csv def pick_bridge_v0(title2linked, title2doc, titles, q, ans): """ 1. mainly based on if the passage includes the answer (assuming that only the 2nd hop passage has the answer) 2. if 1 fails, then resort the linking structure, if A links to B, then B is the """ # check answer if (ans in titles[0] + " " + title2doc[titles[0]]) and ans not in titles[1] + " " + title2doc[titles[1]]: return titles[0] elif (ans in titles[1] + " " + title2doc[titles[1]]) and (ans not in titles[0] + " " + title2doc[titles[0]]): return titles[1] elif titles[0] in title2linked[titles[1]] and titles[1] not in title2linked[titles[0]]: return titles[0] else: return titles[1] def load_annotated(path="/private/home/xwhan/data/hotpot/tfidf/abstracts.txt"): content = [json.loads(l) for l in open(path).readlines()] title2doc = {item["title"]:item["text"] for item in content} title2linked = {item["title"]:item["linked"] for item in content} return title2doc, title2linked def normalize_answer(s): """Lower text and remove punctuation, articles and extra whitespace.""" def remove_articles(text): return re.sub(r'\b(a|an|the)\b', ' ', text) def white_space_fix(text): return ' '.join(text.split()) def remove_punc(text): exclude = set(string.punctuation) return ''.join(ch for ch in text if ch not in exclude) def lower(text): return text.lower() return white_space_fix(remove_articles(remove_punc(lower(s)))) def hotpot_sp_data(raw_path): train = json.load(open(raw_path + '/hotpot_train_v1.1.json')) dev = json.load(open(raw_path + '/hotpot_dev_distractor_v1.json')) title2doc, title2linked = load_annotated() for split_name, split in {"train": train, "val": dev}.items(): data_to_save = [] for item in tqdm(split): title2passage = {_[0]: _[1] for _ in item["context"]} sp_titles = list(set([_[0] for _ in item["supporting_facts"]])) question = item["question"] if item["type"] == "comparison": pos_paras = [] for title in sp_titles: pos_paras.append({ "title": title, "text": "".join(title2passage[title]) }) data_to_save.append({ "question": question, "pos_paras": pos_paras, "neg_paras": [], "type": item["type"], "answers": item["answer"] }) else: bridge = pick_bridge(title2linked, title2doc, sp_titles, question, item["answer"]) if sp_titles[0] == bridge: sp_titles = sp_titles[::-1] start, bridge = sp_titles[0], sp_titles[1] pos_paras = [] for title in sp_titles: pos_paras.append({ "title": title, "text": "".join(title2passage[title]) }) data_to_save.append({ "question": question, "pos_paras": pos_paras, "neg_paras": [], "type": item["type"], "answers": item["answer"], "bridge": bridge }) with open(raw_path + f"/hotpot_retrieval_{split_name}.json", "w") as g: for line in data_to_save: g.write(json.dumps(line) + "\n") def add_qid(raw_path): title2doc, title2linked = load_annotated() raw_train = json.load(open(raw_path + '/hotpot_train_v1.1.json')) raw_val = json.load(open(raw_path + '/hotpot_dev_distractor_v1.json')) for s, raw_data in zip(["train", "val"], [raw_train, raw_val]): qas_data = [] for item in raw_data: question = item["question"] _id = item["_id"] _type = item["type"] answer = [item["answer"]] sp = list(set([f[0] for f in item["supporting_facts"]])) if _type == "bridge": # make sure the sp order follows reasoning process bridge_title = pick_bridge_v0(title2linked, title2doc, sp, question, answer[0]) if sp[0] == bridge_title: sp = sp[::-1] qas_data.append({ "question": question, "_id": _id, "answer": answer, "sp": sp, "type": _type }) with open(raw_path + f"/hotpot_qas_{s}.json", "w") as g1: for _ in qas_data: g1.write(json.dumps(_) + "\n") def add_bridge_ann(raw_path): title2doc, title2linked = load_annotated() raw_train = json.load(open(raw_path + '/hotpot_train_v1.1.json')) raw_val = json.load(open(raw_path + '/hotpot_dev_distractor_v1.json')) for split, raw in zip(["train", "val"], [raw_train, raw_val]): data = json.load(open(raw_path + f"/hotpot_{split}_with_neg.txt")) for idx, item in enumerate(data): assert item["question"] == raw[idx]["question"] item["_id"] = raw[idx]["_id"] if item["type"] == "bridge": ans = item["answers"][0] sp_titles = [p["title"] for p in item["pos_paras"]] bridge_title = pick_bridge_v0(title2linked, title2doc, sp_titles, item["question"], ans) item["bridge"] = bridge_title with open(raw_path + f"/hotpot_{split}_with_neg_v0.json", "w") as g: for _ in data: g.write(json.dumps(_) + "\n") # data = [json.loads(l) for l in open(raw_path + f"/hotpot_qas_{split}.json").readlines()] # for item in data: # if item["type"] == "bridge": # ans = item["answer"][0] # sp_titles = item["sp"] # bridge_title = pick_bridge(title2linked, title2doc, sp_titles, item["question"], ans) # item["bridge"] = bridge_title # with open(raw_path + f"/hotpot_qas_{split}_bridge_label.json", "w") as g: # for _ in data: # g.write(json.dumps(_) + "\n") import numpy as np def check_2hop(raw_path): data = [json.loads(l) for l in open(raw_path + "/bridge_val.json").readlines()] 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] print(np.mean(target_in_query)) def add_sp_labels(raw_path, input_file, save_path, title2sent_map="data/hotpot_index/title2sents.txt"): """ Add sp sentence supervision for QA model training Inputs: raw_path: original HotpotQA data input_file: from MDR retrieval save_path: retrieved results with sentence level annotation """ # raw_train = json.load(open(raw_path + '/hotpot_train_v1.1.json')) # train = [json.loads(l) for l in open(raw_path + "/dense_train_b100_k100.json").readlines()] raw_data = json.load(open(raw_path)) retrieved = [json.loads(l) for l in open(input_file).readlines()] # title2sents title_and_sents = [json.loads(l) for l in open(title2sent_map).readlines()] title2sents = {_['title']:_['sents'] for _ in title_and_sents} for instance, raw in zip(retrieved, raw_data): assert instance["question"] == raw["question"] if "supporting_facts" in raw: orig_sp = raw["supporting_facts"] sptitle2sentids = collections.defaultdict(list) for _ in orig_sp: sptitle2sentids[_[0]].append(_[1]) instance["sp"] = [] for title in sptitle2sentids.keys(): instance["sp"].append({"title": title, "sents": title2sents[title], "sp_sent_ids": sptitle2sentids[title]}) instance["answer"] = [raw["answer"]] with open(save_path, "w") as out: for l in retrieved: out.write(json.dumps(l) + "\n") def explore_QDMR(path="/private/home/xwhan/data/Break-dataset/QDMR-high-level"): """ question decomposition from the BREAK dataset 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 """ hotpot_retrieval_train = [json.loads(l) for l in open("/private/home/xwhan/data/hotpot/hotpot_train_with_neg_v0.json").readlines()] hotpot_retrieval_val = [json.loads(l) for l in open("/private/home/xwhan/data/hotpot/hotpot_val_with_neg_v0.json").readlines()] qid2data = {} for item in hotpot_retrieval_train + hotpot_retrieval_val: qid2data[item["_id"]] = item for split in ["train", "dev"]: break_data = [] with open(f"{path}/{split}.csv") as csvfile: reader = csv.reader(csvfile, quotechar='"', delimiter=',') for row in reader: if row[0] != "question_id": q_id, orig_q, decom = row[0], row[1], row[2] if q_id.startswith("HOTPOT"): orig_split, orig_id = q_id.split("_")[1], q_id.split("_")[2] sp_paras = qid2data[orig_id]["pos_paras"] break_data.append({ "id": orig_id, "split": orig_split, "q": orig_q, "q_decom": decom, "sp": sp_paras, "type": qid2data[orig_id]["type"] }) with open(f"/private/home/xwhan/data/QDMR/{split}.json", "w") as out: for _ in break_data: out.write(json.dumps(_) + "\n") def add_sents_to_corpus_dict(): id2doc = json.load(open("/private/home/xwhan/Mhop-Pretrain/retrieval/index/abstracts_id2doc.json")) title_and_sents = [json.loads(l) for l in open("/private/home/xwhan/data/hotpot/tfidf/title_sents.txt").readlines()] title2sents = {_['title']:_['sents'] for _ in title_and_sents} for k in id2doc.keys(): title, text = id2doc[k][0], id2doc[k][1] sents = title2sents[title] id2doc[k] = { "title": title, "text": text, "sents": sents } json.dump(id2doc, open("/private/home/xwhan/Mhop-Pretrain/retrieval/index/hotpotQA_corpus_dict.json", "w")) if __name__ == "__main__": original_hotpot_data, retrieved, output_path = sys.argv[1], sys.argv[2], sys.argv[3] add_sp_labels(original_hotpot_data, retrieved, output_path) ================================================ FILE: mdr/retrieval/utils/tokenizer.py ================================================ # Copyright (c) Facebook, Inc. and its affiliates. # All rights reserved. # # This source code is licensed under the license found in the # LICENSE file in the root directory of this source tree. # coding=utf-8 # Copyright 2018 The Google AI Language Team Authors. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """Tokenization classes.""" from __future__ import absolute_import from __future__ import division from __future__ import print_function import collections import unicodedata import six import tensorflow as tf def convert_tokens_to_ids(vocab, tokens): """Converts a sequence of tokens into ids using the vocab.""" ids = [] for token in tokens: ids.append(vocab[token]) return ids def whitespace_tokenize(text): """Runs basic whitespace cleaning and splitting on a peice of text.""" text = text.strip() if not text: return [] tokens = text.split() return tokens def convert_to_unicode(text): """Converts `text` to Unicode (if it's not already), assuming utf-8 input.""" if six.PY3: if isinstance(text, str): return text elif isinstance(text, bytes): return text.decode("utf-8", "ignore") else: raise ValueError("Unsupported string type: %s" % (type(text))) elif six.PY2: if isinstance(text, str): return text.decode("utf-8", "ignore") elif isinstance(text, unicode): return text else: raise ValueError("Unsupported string type: %s" % (type(text))) else: raise ValueError("Not running on Python2 or Python 3?") def _is_whitespace(char): """Checks whether `chars` is a whitespace character.""" # \t, \n, and \r are technically contorl characters but we treat them # as whitespace since they are generally considered as such. if char == " " or char == "\t" or char == "\n" or char == "\r": return True cat = unicodedata.category(char) if cat == "Zs": return True return False def _is_control(char): """Checks whether `chars` is a control character.""" # These are technically control characters but we count them as whitespace # characters. if char == "\t" or char == "\n" or char == "\r": return False cat = unicodedata.category(char) if cat.startswith("C"): return True return False class BasicTokenizer(object): """Runs basic tokenization (punctuation splitting, lower casing, etc.).""" def __init__(self, do_lower_case=True): """Constructs a BasicTokenizer. Args: do_lower_case: Whether to lower case the input. """ self.do_lower_case = do_lower_case def tokenize(self, text): """Tokenizes a piece of text.""" text = convert_to_unicode(text) text = self._clean_text(text) orig_tokens = whitespace_tokenize(text) split_tokens = [] for token in orig_tokens: if self.do_lower_case: token = token.lower() token = self._run_strip_accents(token) split_tokens.extend(self._run_split_on_punc(token)) output_tokens = whitespace_tokenize(" ".join(split_tokens)) return output_tokens def _run_strip_accents(self, text): """Strips accents from a piece of text.""" text = unicodedata.normalize("NFD", text) output = [] for char in text: cat = unicodedata.category(char) if cat == "Mn": continue output.append(char) return "".join(output) def _run_split_on_punc(self, text): """Splits punctuation on a piece of text.""" chars = list(text) i = 0 start_new_word = True output = [] while i < len(chars): char = chars[i] if _is_punctuation(char): output.append([char]) start_new_word = True else: if start_new_word: output.append([]) start_new_word = False output[-1].append(char) i += 1 return ["".join(x) for x in output] def _clean_text(self, text): """Performs invalid character removal and whitespace cleanup on text.""" output = [] for char in text: cp = ord(char) if cp == 0 or cp == 0xfffd or _is_control(char): continue if _is_whitespace(char): output.append(" ") else: output.append(char) return "".join(output) def _is_punctuation(char): """Checks whether `chars` is a punctuation character.""" cp = ord(char) # We treat all non-letter/number ASCII as punctuation. # Characters such as "^", "$", and "`" are not in the Unicode # Punctuation class but we treat them as punctuation anyways, for # consistency. if ((cp >= 33 and cp <= 47) or (cp >= 58 and cp <= 64) or (cp >= 91 and cp <= 96) or (cp >= 123 and cp <= 126)): return True cat = unicodedata.category(char) if cat.startswith("P"): return True return False def process(s, tokenizer): try: return tokenizer.tokenize(s) except: print('failed on', s) raise if __name__ == "__main__": _is_whitespace("a") ================================================ FILE: mdr/retrieval/utils/utils.py ================================================ # Copyright (c) Facebook, Inc. and its affiliates. # All rights reserved. # # This source code is licensed under the license found in the # LICENSE file in the root directory of this source tree. import torch import sqlite3 import unicodedata def load_saved(model, path, exact=True): try: state_dict = torch.load(path) except: state_dict = torch.load(path, map_location=torch.device('cpu')) def filter(x): return x[7:] if x.startswith('module.') else x if exact: state_dict = {filter(k): v for (k, v) in state_dict.items()} else: state_dict = {filter(k): v for (k, v) in state_dict.items() if filter(k) in model.state_dict()} model.load_state_dict(state_dict) return model def move_to_cuda(sample): if len(sample) == 0: return {} def _move_to_cuda(maybe_tensor): if torch.is_tensor(maybe_tensor): return maybe_tensor.cuda() elif isinstance(maybe_tensor, dict): return { key: _move_to_cuda(value) for key, value in maybe_tensor.items() } elif isinstance(maybe_tensor, list): return [_move_to_cuda(x) for x in maybe_tensor] else: return maybe_tensor return _move_to_cuda(sample) def convert_to_half(sample): if len(sample) == 0: return {} def _convert_to_half(maybe_floatTensor): if torch.is_tensor(maybe_floatTensor) and maybe_floatTensor.type() == "torch.FloatTensor": return maybe_floatTensor.half() elif isinstance(maybe_floatTensor, dict): return { key: _convert_to_half(value) for key, value in maybe_floatTensor.items() } elif isinstance(maybe_floatTensor, list): return [_convert_to_half(x) for x in maybe_floatTensor] else: return maybe_floatTensor return _convert_to_half(sample) class AverageMeter(object): """Computes and stores the average and current value""" def __init__(self): self.reset() def reset(self): self.val = 0 self.avg = 0 self.sum = 0 self.count = 0 def update(self, val, n=1): self.val = val self.sum += val * n self.count += n self.avg = self.sum / self.count def normalize(text): """Resolve different type of unicode encodings.""" return unicodedata.normalize('NFD', text) class DocDB(object): """Sqlite backed document storage. Implements get_doc_text(doc_id). """ def __init__(self, db_path=None): self.path = db_path self.connection = sqlite3.connect(self.path, check_same_thread=False) def __enter__(self): return self def __exit__(self, *args): self.close() def close(self): """Close the connection to the database.""" self.connection.close() def get_doc_ids(self): """Fetch all ids of docs stored in the db.""" cursor = self.connection.cursor() cursor.execute("SELECT id FROM documents") results = [r[0] for r in cursor.fetchall()] cursor.close() return results def get_doc_text(self, doc_id): """Fetch the raw text of the doc for 'doc_id'.""" cursor = self.connection.cursor() cursor.execute( "SELECT text FROM documents WHERE id = ?", (normalize(doc_id),) ) result = cursor.fetchone() cursor.close() return result if result is None else result[0] def para_has_answer(answer, para, tokenizer): assert isinstance(answer, list) text = normalize(para) tokens = tokenizer.tokenize(text) text = tokens.words(uncased=True) assert len(text) == len(tokens) for single_answer in answer: single_answer = normalize(single_answer) single_answer = tokenizer.tokenize(single_answer) single_answer = single_answer.words(uncased=True) for i in range(0, len(text) - len(single_answer) + 1): if single_answer == text[i: i + len(single_answer)]: return True return False def complex_ans_recall(): """ calculate retrieval metrics for complexwebQ """ import json import numpy as np from basic_tokenizer import SimpleTokenizer tok = SimpleTokenizer() predictions = json.load(open("/private/home/xwhan/code/learning_to_retrieve_reasoning_paths/results/complexwebq_retrieval_res.json")) raw_dev = [json.loads(l) for l in open("/private/home/xwhan/data/ComplexWebQ/complexwebq_dev_qas.txt").readlines()] id2qas = {_["id"]:_ for _ in raw_dev} assert len(predictions) == len(raw_dev) answer_recalls = [] for item in predictions: qid = item["q_id"] title2passage = item["context"] gold_answers = id2qas[qid]["answer"] chain_coverage = [] for chain in item["topk_titles"]: chain_text = " ".join([title2passage[_] for _ in chain]) chain_coverage.append(para_has_answer(gold_answers, chain_text, tok)) answer_recalls.append(np.sum(chain_coverage) > 0) print(len(answer_recalls)) print(np.mean(answer_recalls)) if __name__ == "__main__": complex_ans_recall() ================================================ FILE: requirements.txt ================================================ transformers==2.11.0 tensorboard>=1.15.0 numpy tqdm ujson streamlit ================================================ FILE: scripts/add_sp_label.sh ================================================ #!/bin/bash # Copyright (c) Facebook, Inc. and its affiliates. # All rights reserved. # # This source code is licensed under the license found in the # LICENSE file in the root directory of this source tree. #!/bin/bash # 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 ORIGINAL_DATA=$1 RETRIEVED_DATA=$2 SAVED_PATH=$3 python mdr/retrieval/utils/mhop_utils.py ${ORIGINAL_DATA} ${RETRIEVED_DATA} ${SASAVED_PATH} ================================================ FILE: scripts/demo.py ================================================ # Copyright (c) Facebook, Inc. and its affiliates. # All rights reserved. # # This source code is licensed under the license found in the # LICENSE file in the root directory of this source tree. import streamlit as st import torch import os import numpy as np from apex import amp import faiss import json import argparse from functools import partial from transformers import AutoConfig, AutoTokenizer from torch.utils.data import DataLoader from mdr.retrieval.models.mhop_retriever import RobertaRetriever from mdr.retrieval.utils.basic_tokenizer import SimpleTokenizer from mdr.retrieval.utils.utils import load_saved, move_to_cuda from mdr.qa.qa_model import QAModel from mdr.qa.qa_dataset import qa_collate, QAEvalDataset from train_qa import eval_final @st.cache(allow_output_mutation=True) def init_retrieval(args): print("Initializing retrieval module...") bert_config = AutoConfig.from_pretrained(args.model_name) tokenizer = AutoTokenizer.from_pretrained(args.model_name) retriever = RobertaRetriever(bert_config, args) retriever = load_saved(retriever, args.model_path, exact=False) cuda = torch.device('cuda') retriever.to(cuda) retriever = amp.initialize(retriever, opt_level='O1') retriever.eval() print("Loading index...") index = faiss.IndexFlatIP(768) xb = np.load(args.indexpath).astype('float32') index.add(xb) if args.index_gpu != -1: res = faiss.StandardGpuResources() index = faiss.index_cpu_to_gpu(res, args.index_gpu, index) print("Loading documents...") id2doc = json.load(open(args.corpus_dict)) print("Index ready...") return retriever, index, id2doc, tokenizer @st.cache(allow_output_mutation=True) def init_reader(args): qa_config = AutoConfig.from_pretrained( 'google/electra-large-discriminator') qa_tokenizer = AutoTokenizer.from_pretrained( 'google/electra-large-discriminator') retriever_name = args.model_name args.model_name = "google/electra-large-discriminator" reader = QAModel(qa_config, args) reader = load_saved(reader, args.reader_path, False) cuda = torch.device('cuda') reader.to(cuda) reader = amp.initialize(reader, opt_level='O1') reader.eval() args.model_name = retriever_name return reader, qa_tokenizer st.markdown( "# Multi-hop Open-domain QA with [MDR](https://github.com/facebookresearch/multihop_dense_retrieval)") parser = argparse.ArgumentParser() parser.add_argument('--indexpath', type=str, default='data/hotpot_index/wiki_index.npy') parser.add_argument('--corpus_dict', type=str, default='data/hotpot_index/wiki_id2doc.json') parser.add_argument('--model_path', type=str, default='models/q_encoder.pt') parser.add_argument('--topk', type=int, default=20, help="topk paths") parser.add_argument('--max-q-len', type=int, default=70) parser.add_argument('--max-c-len', type=int, default=300) parser.add_argument('--max-q-sp-len', type=int, default=350) parser.add_argument('--model-name', type=str, default='roberta-base') parser.add_argument('--reader_path', type=str, default="models/qa_electra.pt") parser.add_argument("--sp-pred", action="store_true", help="whether to predict sentence sp") parser.add_argument("--sp-weight", default=0, type=float, help="weight of the sp loss") parser.add_argument("--max-ans-len", default=30, type=int) parser.add_argument("--save-prediction", default="", type=str) parser.add_argument("--index-gpu", default=-1, type=int) args = parser.parse_args() reader, qa_tokenizer = init_reader(args) retriever, index, id2doc, retriever_tokenizer = init_retrieval(args) st.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.*") query = st.text_input('Enter your question') if query: query = query[:-1] if query.endswith("?") else query with torch.no_grad(): print("Retrieving") q_encodes = retriever_tokenizer.batch_encode_plus( [query], max_length=args.max_q_len, pad_to_max_length=True, return_tensors="pt") q_encodes = move_to_cuda(dict(q_encodes)) q_embeds = retriever.encode_q( q_encodes["input_ids"], q_encodes["attention_mask"], q_encodes.get("token_type_ids", None)).cpu().numpy() scores_1, docid_1 = index.search(q_embeds, args.topk) query_pairs = [] # for 2nd hop for _, doc_id in enumerate(docid_1[0]): doc = id2doc[str(doc_id)]["text"] if doc.strip() == "": # roberta tokenizer does not accept empty string as segment B doc = id2doc[str(doc_id)]["title"] scores_1[b_idx][_] = float("-inf") query_pairs.append((query, doc)) q_sp_encodes = retriever_tokenizer.batch_encode_plus( query_pairs, max_length=args.max_q_sp_len, pad_to_max_length=True, return_tensors="pt") q_sp_encodes = move_to_cuda(dict(q_sp_encodes)) q_sp_embeds = retriever.encode_q( q_sp_encodes["input_ids"], q_sp_encodes["attention_mask"],q_sp_encodes.get("token_type_ids", None)).cpu().numpy() scores_2, docid_2 = index.search(q_sp_embeds, args.topk) scores_2 = scores_2.reshape(1, args.topk, args.topk) docid_2 = docid_2.reshape(1, args.topk, args.topk) path_scores = np.expand_dims(scores_1, axis=2) + scores_2 search_scores = path_scores[0] ranked_pairs = np.vstack(np.unravel_index(np.argsort(search_scores.ravel())[::-1], (args.topk, args.topk))).transpose() chains = [] topk_docs = {} for _ in range(args.topk): path_ids = ranked_pairs[_] doc1_id = str(docid_1[0, path_ids[0]]) doc2_id = str(docid_2[0, path_ids[0], path_ids[1]]) chains.append([id2doc[doc1_id], id2doc[doc2_id]]) topk_docs[id2doc[doc1_id]['title']] = id2doc[doc1_id]['text'] topk_docs[id2doc[doc2_id]['title']] = id2doc[doc2_id]['text'] reader_input = [{ "_id": 0, "question": query, "candidate_chains": chains }] print(f"Reading {len(chains)} chains...") collate_fc = partial(qa_collate, pad_id=qa_tokenizer.pad_token_id) qa_eval_dataset = QAEvalDataset( qa_tokenizer, reader_input, max_seq_len=512, max_q_len=64) qa_eval_dataloader = DataLoader( qa_eval_dataset, batch_size=args.topk, collate_fn=collate_fc, pin_memory=True, num_workers=0) qa_results = eval_final(args, reader, qa_eval_dataloader, gpu=True) answer_pred = qa_results['answer'][0] sp_pred = qa_results['sp'][0] titles_pred = qa_results['titles'][0] st.markdown(f'**Answer**: {answer_pred}') st.markdown(f'**Supporting passages**:') st.markdown(f'> **{titles_pred[0]}**: {topk_docs[titles_pred[0]].replace(answer_pred, "**" + answer_pred + "**")}') st.markdown( f'> **{titles_pred[1]}**: {topk_docs[titles_pred[1]].replace(answer_pred, "**" + answer_pred + "**")}') # st.write(qa_results) ================================================ FILE: scripts/download_hotpot.sh ================================================ #!/bin/bash # Copyright (c) Facebook, Inc. and its affiliates. # All rights reserved. # # This source code is licensed under the license found in the # LICENSE file in the root directory of this source tree. #!/bin/bash # Make data and model folder. mkdir data mkdir models # Download data cd data mkdir hotpot cd hotpot wget https://dl.fbaipublicfiles.com/mdpr/data/hotpot/hotpot_train_with_neg_v0.json wget https://dl.fbaipublicfiles.com/mdpr/data/hotpot/hotpot_dev_with_neg_v0.json wget https://dl.fbaipublicfiles.com/mdpr/data/hotpot/hotpot_qas_val.json wget https://dl.fbaipublicfiles.com/mdpr/data/hotpot/train_retrieval_b100_k100_sp.json wget https://dl.fbaipublicfiles.com/mdpr/data/hotpot/dev_retrieval_b50_k50_sp.json wget https://dl.fbaipublicfiles.com/mdpr/data/hotpot/dev_retrieval_top100_sp.json cd .. mkdir hotpot_index cd hotpot_index wget https://dl.fbaipublicfiles.com/mdpr/data/hotpot_index/wiki_id2doc.json wget https://dl.fbaipublicfiles.com/mdpr/data/hotpot_index/wiki_index.npy echo "Finished downloading data!" # Download models cd ../../models wget https://dl.fbaipublicfiles.com/mdpr/models/doc_encoder.pt wget https://dl.fbaipublicfiles.com/mdpr/models/q_encoder.pt wget https://dl.fbaipublicfiles.com/mdpr/models/qa_electra.pt ================================================ FILE: scripts/encode_corpus.py ================================================ # Copyright (c) Facebook, Inc. and its affiliates. # All rights reserved. # # This source code is licensed under the license found in the # LICENSE file in the root directory of this source tree. """ Description: encode text corpus into a store of dense vectors. Usage (adjust the batch size according to your GPU memory): CUDA_VISIBLE_DEVICES=0,1,2,3 python scripts/encode_corpus.py \ --do_predict \ --predict_batch_size 1000 \ --model_name roberta-base \ --predict_file ${CORPUS_PATH} \ --init_checkpoint ${MODEL_CHECKPOINT} \ --embed_save_path ${SAVE_PATH} \ --fp16 \ --max_c_len 300 \ --num_workers 20 """ import collections import logging import json import os import random from tqdm import tqdm import numpy as np import torch from transformers import AutoConfig, AutoTokenizer from torch.utils.data import DataLoader from mdr.retrieval.data.encode_datasets import EmDataset, em_collate from mdr.retrieval.models.retriever import CtxEncoder, RobertaCtxEncoder from mdr.retrieval.config import encode_args from mdr.retrieval.utils.utils import move_to_cuda, load_saved def main(): args = encode_args() if args.fp16: import apex apex.amp.register_half_function(torch, 'einsum') if args.local_rank == -1 or args.no_cuda: device = torch.device( "cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu") n_gpu = torch.cuda.device_count() else: device = torch.device("cuda", args.local_rank) n_gpu = 1 torch.distributed.init_process_group(backend='nccl') if not args.predict_file: raise ValueError( "If `do_predict` is True, then `predict_file` must be specified.") bert_config = AutoConfig.from_pretrained(args.model_name) if "roberta" in args.model_name: model = RobertaCtxEncoder(bert_config, args) else: model = CtxEncoder(bert_config, args) tokenizer = AutoTokenizer.from_pretrained(args.model_name) eval_dataset = EmDataset( tokenizer, args.predict_file, args.max_q_len, args.max_c_len, args.is_query_embed, args.embed_save_path) eval_dataloader = DataLoader( eval_dataset, batch_size=args.predict_batch_size, collate_fn=em_collate, pin_memory=True, num_workers=args.num_workers) assert args.init_checkpoint != "" model = load_saved(model, args.init_checkpoint, exact=False) model.to(device) if args.fp16: try: from apex import amp except ImportError: raise ImportError( "Please install apex from https://www.github.com/nvidia/apex to use fp16 training.") model = amp.initialize(model, opt_level=args.fp16_opt_level) if args.local_rank != -1: model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[args.local_rank], output_device=args.local_rank) elif n_gpu > 1: model = torch.nn.DataParallel(model) embeds = predict(model, eval_dataloader) print(embeds.size()) np.save(args.embed_save_path, embeds.cpu().numpy()) def predict(model, eval_dataloader): if type(model) == list: model = [m.eval() for m in model] else: model.eval() embed_array = [] for batch in tqdm(eval_dataloader): batch_to_feed = move_to_cuda(batch) with torch.no_grad(): results = model(batch_to_feed) embed = results['embed'].cpu() embed_array.append(embed) ## linear combination tuning on dev data embed_array = torch.cat(embed_array) model.train() return embed_array if __name__ == "__main__": main() ================================================ FILE: scripts/end2end.py ================================================ #!/usr/bin/env python # Copyright (c) Facebook, Inc. and its affiliates. # All rights reserved. # # This source code is licensed under the license found in the # LICENSE file in the root directory of this source tree. """ Efficient end2end QA with HNSW index taskset --cpu-list 0-15 python end2end.py ../data/hotpot/hotpot_qas_val.json """ import argparse import json import logging from functools import partial import time import argparse import collections import json import logging from torch.utils.data import DataLoader import faiss import numpy as np import torch from tqdm import tqdm from transformers import AutoConfig, AutoTokenizer from retrieval.models.mhop_retriever import RobertaRetriever from retrieval.utils.utils import load_saved from qa.qa_model import QAModel from mdr.qa.qa_dataset import qa_collate, QAEvalDataset from .train_qa import eval_final from qa.hotpot_evaluate_v1 import f1_score, exact_match_score from qa.utils import set_global_logging_level logger = logging.getLogger() logger.setLevel(logging.INFO) if (logger.hasHandlers()): logger.handlers.clear() console = logging.StreamHandler() logger.addHandler(console) set_global_logging_level(logging.ERROR, ["transformers", "nlp", "torch", "tensorflow", "tensorboard", "wandb"]) def convert_hnsw_query(query_vectors): aux_dim = np.zeros(len(query_vectors), dtype='float32') query_nhsw_vectors = np.hstack((query_vectors, aux_dim.reshape(-1, 1))) return query_nhsw_vectors if __name__ == '__main__': parser = argparse.ArgumentParser() parser.add_argument('raw_data', type=str, default=None) parser.add_argument('--indexpath', type=str, default="retrieval/index/wiki_index_hnsw_roberta") parser.add_argument('--corpus_dict', type=str, default='retrieval/index/hotpotQA_corpus_dict.json') 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") 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") parser.add_argument('--topk', type=int, default=1, help="topk paths") parser.add_argument('--num-workers', type=int, default=10) parser.add_argument('--max-q-len', type=int, default=70) parser.add_argument('--max-q-sp-len', type=int, default=350) parser.add_argument('--batch-size', type=int, default=1) parser.add_argument("--max-ans-len", default=35, type=int) parser.add_argument("--save-prediction", default="", type=str) parser.add_argument("--model-name", type=str, default="") parser.add_argument("--sp-pred", action="store_true", help="whether to predict sentence sp") parser.add_argument("--sp-weight", default=0, type=float, help="weight of the sp loss") # parser.add_argument('--hnsw', action="store_true") args = parser.parse_args() logger.info("Loading trained models...") retrieval_config = AutoConfig.from_pretrained('roberta-base') retrieval_tokenizer = AutoTokenizer.from_pretrained('roberta-base') args.model_name = "roberta-base" retriever = RobertaRetriever(retrieval_config, args) retriever = load_saved(retriever, args.retriever_path) retriever.eval() qa_config = AutoConfig.from_pretrained('google/electra-large-discriminator') qa_tokenizer = AutoTokenizer.from_pretrained('google/electra-large-discriminator') args.model_name = "google/electra-large-discriminator" reader = QAModel(qa_config, args) reader = load_saved(reader, args.reader_path, False) reader.eval() logger.info("Loading index...") index = faiss.read_index(args.indexpath) logger.info(f"Loading corpus...") id2doc = json.load(open(args.corpus_dict)) logger.info(f"Corpus size {len(id2doc)}") logger.info("Loading queries...") qas_items = [json.loads(_) for _ in open(args.raw_data).readlines()[:5]] questions = [_["question"][:-1] if _["question"].endswith("?") else _["question"] for _ in qas_items] id2gold_ans = {_["_id"]: _["answer"][0] for _ in qas_items} start = time.time() logger.info("Retrieving...") retrieval_results = [] encode_times = [] search_times = [] with torch.no_grad(): for b_start in tqdm(range(0, len(questions), args.batch_size)): # 1-hop retrieval batch_q = questions[b_start:b_start + args.batch_size] batch_qas = qas_items[b_start:b_start + args.batch_size] batch_q_encodes = retrieval_tokenizer.batch_encode_plus(batch_q, max_length=args.max_q_len, pad_to_max_length=True, return_tensors="pt") q_embeds = retriever.encode_q(batch_q_encodes["input_ids"], batch_q_encodes["attention_mask"], batch_q_encodes.get("token_type_ids", None)) q_embeds_numpy = q_embeds.numpy() q_embeds_numpy = convert_hnsw_query(q_embeds_numpy) scores_1, docid_1 = index.search(q_embeds_numpy, args.topk) # construct 2hop queries bsize = len(batch_q) query_pairs = [] for b_idx in range(bsize): for _, doc_id in enumerate(docid_1[b_idx]): doc = id2doc[str(doc_id)]["text"] if doc.strip() == "": # roberta tokenizer does not accept empty string as segment B doc = id2doc[str(doc_id)]["title"] scores_1[b_idx][_] = float("-inf") query_pairs.append((batch_q[b_idx], doc)) # 2-hop retrieval s1 = time.time() 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") 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)) encode_times.append(time.time() - s1) s2 = time.time() q_sp_embeds = q_sp_embeds.numpy() q_sp_embeds = convert_hnsw_query(q_sp_embeds) scores_2, docid_2 = index.search(q_sp_embeds, args.topk) search_times.append(time.time() - s2) # aggregate chain scores scores_2 = scores_2.reshape(bsize, args.topk, args.topk) docid_2 = docid_2.reshape(bsize, args.topk, args.topk) path_scores = - (np.expand_dims(scores_1, axis=2) + scores_2) for idx in range(bsize): search_scores = path_scores[idx] ranked_pairs = np.vstack(np.unravel_index(np.argsort(search_scores.ravel())[::-1], (args.topk, args.topk))).transpose() chains = [] for _ in range(args.topk): path_ids = ranked_pairs[_] doc1_id = str(docid_1[idx, path_ids[0]]) doc2_id = str(docid_2[idx, path_ids[0], path_ids[1]]) chains.append([id2doc[doc1_id], id2doc[doc2_id]]) retrieval_results.append({ "_id": batch_qas[idx]["_id"], "question": batch_qas[idx]["question"], "candidate_chains": chains }) logger.info("Reading...") collate_fc = partial(qa_collate, pad_id=qa_tokenizer.pad_token_id) qa_eval_dataset = QAEvalDataset(qa_tokenizer, retrieval_results, max_seq_len=512, max_q_len=64) qa_eval_dataloader = DataLoader(qa_eval_dataset, batch_size=args.topk, collate_fn=collate_fc, pin_memory=True, num_workers=0) qa_results = eval_final(args, reader, qa_eval_dataloader, gpu=False) print(f"Finishing evaluation in {time.time() - start}s") ems = [exact_match_score(qa_results["answer"][k], id2gold_ans[k]) for k in qa_results["answer"].keys()] f1s = [f1_score(qa_results["answer"][k], id2gold_ans[k]) for k in qa_results["answer"].keys()] logger.info(f"Answer EM {np.mean(ems)}, F1 {np.mean(f1s)}") ================================================ FILE: scripts/end2end.sh ================================================ #!/bin/bash # Copyright (c) Facebook, Inc. and its affiliates. # All rights reserved. # # This source code is licensed under the license found in the # LICENSE file in the root directory of this source tree. #SBATCH --cpus-per-task=16 #SBATCH --nodes=1 #SBATCH --ntasks-per-node=1 #SBATCH --time=12:00:00 #SBATCH --job-name=hotpot_eval #SBATCH --output=/private/home/xwhan/Mhop-Pretrain/eval_logs #SBATCH --partition=dev #SBATCH --error=/checkpoint/%u/hotpot-jobs/sample-%j.err #SBATCH --mem=500GB #SBATCH --signal=USR1@140 #SBATCH --open-mode=append --wrap="srun python end2end.py \ ../data/hotpot/hotpot_qas_val.json" sbatch --job-name=hotpot_eval \ --error=/checkpoint/hotpot-jobs/hotpot-%j.err \ --output=/checkpoint/hotpot-jobs/hotpot-%j.out \ --partition=dev --nodes=1 --ntasks-per-node=1 \ --cpus-per-task=16 \ --gpus-per-node=1 --open-mode=append \ --time=12:00:00 \ --wrap="srun python end2end.py ../data/hotpot/hotpot_qas_val.json" ================================================ FILE: scripts/eval/eval_mhop_fever.py ================================================ # Copyright (c) Facebook, Inc. and its affiliates. # All rights reserved. # # This source code is licensed under the license found in the # LICENSE file in the root directory of this source tree. """ python 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 # unified retrieval python 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 python 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 # fix parenthesis python 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 """ import argparse import json import logging from os import path import faiss import numpy as np import torch from tqdm import tqdm from transformers import AutoConfig, AutoTokenizer from models.mhop_retriever import RobertaRetriever from models.unified_retriever import UnifiedRetriever from utils.basic_tokenizer import SimpleTokenizer from utils.utils import (load_saved, move_to_cuda, para_has_answer) logger = logging.getLogger() logger.setLevel(logging.INFO) if (logger.hasHandlers()): logger.handlers.clear() console = logging.StreamHandler() logger.addHandler(console) if __name__ == '__main__': parser = argparse.ArgumentParser() parser.add_argument('raw_data', type=str, default=None) parser.add_argument('indexpath', type=str, default=None) parser.add_argument('corpus_dict', type=str, default=None) parser.add_argument('model_path', type=str, default=None) parser.add_argument('--topk', type=int, default=2, help="topk paths") parser.add_argument('--num-workers', type=int, default=10) parser.add_argument('--max-q-len', type=int, default=45) parser.add_argument('--max-c-len', type=int, default=350) parser.add_argument('--max-q-sp-len', type=int, default=400) parser.add_argument('--batch-size', type=int, default=100) parser.add_argument('--beam-size-1', type=int, default=5) parser.add_argument('--beam-size-2', type=int, default=5) parser.add_argument('--model-name', type=str, default='bert-base-uncased') parser.add_argument('--gpu', action="store_true") parser.add_argument('--shared-encoder', action="store_true") parser.add_argument("--save-path", type=str, default="") parser.add_argument("--stop-drop", default=0, type=float) args = parser.parse_args() logger.info("Loading data...") ds_items = [json.loads(_) for _ in open(args.raw_data).readlines()] logger.info("Building index...") d = 768 xb = np.load(args.indexpath).astype('float32') print(xb.shape) index = faiss.IndexFlatIP(d) index.add(xb) if args.gpu: res = faiss.StandardGpuResources() index = faiss.index_cpu_to_gpu(res, 1, index) logger.info(f"Loading corpus...") id2doc = json.load(open(args.corpus_dict)) title2doc = {item[0]:item[1] for item in id2doc.values()} logger.info(f"Corpus size {len(id2doc)}") logger.info("Loading trained model...") bert_config = AutoConfig.from_pretrained(args.model_name) tokenizer = AutoTokenizer.from_pretrained(args.model_name) model = RobertaRetriever(bert_config, args) # model = UnifiedRetriever(bert_config, args) model = load_saved(model, args.model_path, exact=False) simple_tokenizer = SimpleTokenizer() cuda = torch.device('cuda') model.to(cuda) from apex import amp model = amp.initialize(model, opt_level='O1') model.eval() logger.info("Encoding claims and searching") questions = [_["claim"] for _ in ds_items] metrics = [] retrieval_outputs = [] for b_start in tqdm(range(0, len(questions), args.batch_size)): with torch.no_grad(): batch_q = questions[b_start:b_start + args.batch_size] batch_ann = ds_items[b_start:b_start + args.batch_size] bsize = len(batch_q) batch_q_encodes = tokenizer.batch_encode_plus(batch_q, max_length=args.max_q_len, pad_to_max_length=True, return_tensors="pt") batch_q_encodes = move_to_cuda(dict(batch_q_encodes)) q_embeds = model.encode_q(batch_q_encodes["input_ids"], batch_q_encodes["attention_mask"], batch_q_encodes.get("token_type_ids", None)) q_embeds_numpy = q_embeds.cpu().contiguous().numpy() D, I = index.search(q_embeds_numpy, args.beam_size_1) # 2hop search query_pairs = [] for b_idx in range(bsize): for _, doc_id in enumerate(I[b_idx]): doc = id2doc[str(doc_id)][1] if "roberta" in args.model_name and doc.strip() == "": # doc = "fadeaxsaa" * 100 doc = id2doc[str(doc_id)][0] D[b_idx][_] = float("-inf") query_pairs.append((batch_q[b_idx], doc)) 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") batch_q_sp_encodes = move_to_cuda(dict(batch_q_sp_encodes)) 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)) q_sp_embeds = q_sp_embeds.contiguous().cpu().numpy() # search_start = time.time() D_, I_ = index.search(q_sp_embeds, args.beam_size_2) # logger.info(f"MIPS searching: {time.time() - search_start}") D_ = D_.reshape(bsize, args.beam_size_1, args.beam_size_2) I_ = I_.reshape(bsize, args.beam_size_1, args.beam_size_2) # aggregate path scores path_scores = np.expand_dims(D, axis=2) + D_ # path_scores = D_ # eval for idx in range(bsize): search_scores = path_scores[idx] ranked_pairs = np.vstack(np.unravel_index(np.argsort(search_scores.ravel())[::-1], (args.beam_size_1, args.beam_size_2))).transpose() retrieved_titles = [] hop1_titles = [] paths, path_titles = [], [] paths_both_are_intro = [] for _ in range(args.topk): path_ids = ranked_pairs[_] hop_1_id = I[idx, path_ids[0]] hop_2_id = I_[idx, path_ids[0], path_ids[1]] retrieved_titles.append(id2doc[str(hop_1_id)][0]) retrieved_titles.append(id2doc[str(hop_2_id)][0]) paths.append([str(hop_1_id), str(hop_2_id)]) path_titles.append([id2doc[str(hop_1_id)][0], id2doc[str(hop_2_id)][0]]) paths_both_are_intro.append(id2doc[str(hop_1_id)][2] and id2doc[str(hop_2_id)][2]) hop1_titles.append(id2doc[str(hop_1_id)][0]) # saving when there's no annotations if args.save_path != "": candidaite_chains = [] for path in path_titles: candidaite_chains.append([(path[0], title2doc[path[0]]), (path[1], title2doc[path[1]])]) retrieval_outputs.append({ "id": batch_ann[idx]["id"], "claim": batch_ann[idx]["claim"], "candidate_chains": candidaite_chains, }) if args.save_path != "": with open(f"/private/home/xwhan/data/fever/retrieval/{args.save_path}", "w") as out: for l in retrieval_outputs: out.write(json.dumps(l) + "\n") ================================================ FILE: scripts/eval/eval_mhop_retrieval.py ================================================ # Copyright (c) Facebook, Inc. and its affiliates. # All rights reserved. # # This source code is licensed under the license found in the # LICENSE file in the root directory of this source tree. """ Evaluating trained retrieval model. Usage: python eval_mhop_retrieval.py ${EVAL_DATA} ${CORPUS_VECTOR_PATH} ${CORPUS_DICT} ${MODEL_CHECKPOINT} \ --batch-size 50 \ --beam-size-1 20 \ --beam-size-2 5 \ --topk 20 \ --shared-encoder \ --gpu \ --save-path ${PATH_TO_SAVE_RETRIEVAL} """ import argparse import collections import json import logging from os import path import time import faiss import numpy as np import torch from tqdm import tqdm from transformers import AutoConfig, AutoTokenizer from mdr.retrieval.models.mhop_retriever import RobertaRetriever from mdr.retrieval.utils.basic_tokenizer import SimpleTokenizer from mdr.retrieval.utils.utils import (load_saved, move_to_cuda, para_has_answer) logger = logging.getLogger() logger.setLevel(logging.INFO) if (logger.hasHandlers()): logger.handlers.clear() console = logging.StreamHandler() logger.addHandler(console) def convert_hnsw_query(query_vectors): aux_dim = np.zeros(len(query_vectors), dtype='float32') query_nhsw_vectors = np.hstack((query_vectors, aux_dim.reshape(-1, 1))) return query_nhsw_vectors if __name__ == '__main__': parser = argparse.ArgumentParser() parser.add_argument('raw_data', type=str, default=None) parser.add_argument('indexpath', type=str, default=None) parser.add_argument('corpus_dict', type=str, default=None) parser.add_argument('model_path', type=str, default=None) parser.add_argument('--topk', type=int, default=2, help="topk paths") parser.add_argument('--num-workers', type=int, default=10) parser.add_argument('--max-q-len', type=int, default=70) parser.add_argument('--max-c-len', type=int, default=300) parser.add_argument('--max-q-sp-len', type=int, default=350) parser.add_argument('--batch-size', type=int, default=100) parser.add_argument('--beam-size', type=int, default=5) parser.add_argument('--model-name', type=str, default='roberta-base') parser.add_argument('--gpu', action="store_true") parser.add_argument('--save-index', action="store_true") parser.add_argument('--only-eval-ans', action="store_true") parser.add_argument('--shared-encoder', action="store_true") parser.add_argument("--save-path", type=str, default="") parser.add_argument("--stop-drop", default=0, type=float) parser.add_argument('--hnsw', action="store_true") args = parser.parse_args() logger.info("Loading data...") ds_items = [json.loads(_) for _ in open(args.raw_data).readlines()] # filter if args.only_eval_ans: ds_items = [_ for _ in ds_items if _["answer"][0] not in ["yes", "no"]] logger.info("Loading trained model...") bert_config = AutoConfig.from_pretrained(args.model_name) tokenizer = AutoTokenizer.from_pretrained(args.model_name) model = RobertaRetriever(bert_config, args) model = load_saved(model, args.model_path, exact=False) simple_tokenizer = SimpleTokenizer() cuda = torch.device('cuda') model.to(cuda) from apex import amp model = amp.initialize(model, opt_level='O1') model.eval() logger.info("Building index...") d = 768 xb = np.load(args.indexpath).astype('float32') if args.hnsw: if path.exists("data/hotpot_index/wiki_index_hnsw.index"): index = faiss.read_index("index/wiki_index_hnsw.index") else: index = faiss.IndexHNSWFlat(d + 1, 512) index.hnsw.efSearch = 128 index.hnsw.efConstruction = 200 phi = 0 for i, vector in enumerate(xb): norms = (vector ** 2).sum() phi = max(phi, norms) logger.info('HNSWF DotProduct -> L2 space phi={}'.format(phi)) data = xb buffer_size = 50000 n = len(data) print(n) for i in tqdm(range(0, n, buffer_size)): vectors = [np.reshape(t, (1, -1)) for t in data[i:i + buffer_size]] norms = [(doc_vector ** 2).sum() for doc_vector in vectors] aux_dims = [np.sqrt(phi - norm) for norm in norms] hnsw_vectors = [np.hstack((doc_vector, aux_dims[idx].reshape(-1, 1))) for idx, doc_vector in enumerate(vectors)] hnsw_vectors = np.concatenate(hnsw_vectors, axis=0) index.add(hnsw_vectors) else: index = faiss.IndexFlatIP(d) index.add(xb) if args.gpu: res = faiss.StandardGpuResources() index = faiss.index_cpu_to_gpu(res, 6, index) if args.save_index: faiss.write_index(index, "data/hotpot_index/wiki_index_hnsw_roberta") logger.info(f"Loading corpus...") id2doc = json.load(open(args.corpus_dict)) if isinstance(id2doc["0"], list): id2doc = {k: {"title":v[0], "text": v[1]} for k, v in id2doc.items()} # title2text = {v[0]:v[1] for v in id2doc.values()} logger.info(f"Corpus size {len(id2doc)}") logger.info("Encoding questions and searching") questions = [_["question"][:-1] if _["question"].endswith("?") else _["question"] for _ in ds_items] metrics = [] retrieval_outputs = [] for b_start in tqdm(range(0, len(questions), args.batch_size)): with torch.no_grad(): batch_q = questions[b_start:b_start + args.batch_size] batch_ann = ds_items[b_start:b_start + args.batch_size] bsize = len(batch_q) batch_q_encodes = tokenizer.batch_encode_plus(batch_q, max_length=args.max_q_len, pad_to_max_length=True, return_tensors="pt") batch_q_encodes = move_to_cuda(dict(batch_q_encodes)) q_embeds = model.encode_q(batch_q_encodes["input_ids"], batch_q_encodes["attention_mask"], batch_q_encodes.get("token_type_ids", None)) q_embeds_numpy = q_embeds.cpu().contiguous().numpy() if args.hnsw: q_embeds_numpy = convert_hnsw_query(q_embeds_numpy) D, I = index.search(q_embeds_numpy, args.beam_size) # 2hop search query_pairs = [] for b_idx in range(bsize): for _, doc_id in enumerate(I[b_idx]): doc = id2doc[str(doc_id)]["text"] if "roberta" in args.model_name and doc.strip() == "": # doc = "fadeaxsaa" * 100 doc = id2doc[str(doc_id)]["title"] D[b_idx][_] = float("-inf") query_pairs.append((batch_q[b_idx], doc)) 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") batch_q_sp_encodes = move_to_cuda(dict(batch_q_sp_encodes)) s1 = time.time() 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)) # print("Encoding time:", time.time() - s1) q_sp_embeds = q_sp_embeds.contiguous().cpu().numpy() s2 = time.time() if args.hnsw: q_sp_embeds = convert_hnsw_query(q_sp_embeds) D_, I_ = index.search(q_sp_embeds, args.beam_size) D_ = D_.reshape(bsize, args.beam_size, args.beam_size) I_ = I_.reshape(bsize, args.beam_size, args.beam_size) # aggregate path scores path_scores = np.expand_dims(D, axis=2) + D_ if args.hnsw: path_scores = - path_scores for idx in range(bsize): search_scores = path_scores[idx] ranked_pairs = np.vstack(np.unravel_index(np.argsort(search_scores.ravel())[::-1], (args.beam_size, args.beam_size))).transpose() retrieved_titles = [] hop1_titles = [] paths, path_titles = [], [] for _ in range(args.topk): path_ids = ranked_pairs[_] hop_1_id = I[idx, path_ids[0]] hop_2_id = I_[idx, path_ids[0], path_ids[1]] retrieved_titles.append(id2doc[str(hop_1_id)]["title"]) retrieved_titles.append(id2doc[str(hop_2_id)]["title"]) paths.append([str(hop_1_id), str(hop_2_id)]) path_titles.append([id2doc[str(hop_1_id)]["title"], id2doc[str(hop_2_id)]["title"]]) hop1_titles.append(id2doc[str(hop_1_id)]["title"]) if args.only_eval_ans: gold_answers = batch_ann[idx]["answer"] concat_p = "yes no " for p in paths: concat_p += " ".join([id2doc[doc_id]["title"] + " " + id2doc[doc_id]["text"] for doc_id in p]) metrics.append({ "question": batch_ann[idx]["question"], "ans_recall": int(para_has_answer(gold_answers, concat_p, simple_tokenizer)), "type": batch_ann[idx].get("type", "single") }) else: sp = batch_ann[idx]["sp"] assert len(set(sp)) == 2 type_ = batch_ann[idx]["type"] question = batch_ann[idx]["question"] p_recall, p_em = 0, 0 sp_covered = [sp_title in retrieved_titles for sp_title in sp] if np.sum(sp_covered) > 0: p_recall = 1 if np.sum(sp_covered) == len(sp_covered): p_em = 1 path_covered = [int(set(p) == set(sp)) for p in path_titles] path_covered = np.sum(path_covered) > 0 recall_1 = 0 covered_1 = [sp_title in hop1_titles for sp_title in sp] if np.sum(covered_1) > 0: recall_1 = 1 metrics.append({ "question": question, "p_recall": p_recall, "p_em": p_em, "type": type_, 'recall_1': recall_1, 'path_covered': int(path_covered) }) # saving when there's no annotations candidaite_chains = [] for path in paths: candidaite_chains.append([id2doc[path[0]], id2doc[path[1]]]) retrieval_outputs.append({ "_id": batch_ann[idx]["_id"], "question": batch_ann[idx]["question"], "candidate_chains": candidaite_chains, # "sp": sp_chain, # "answer": gold_answers, # "type": type_, # "coverd_k": covered_k }) if args.save_path != "": with open(args.save_path, "w") as out: for l in retrieval_outputs: out.write(json.dumps(l) + "\n") logger.info(f"Evaluating {len(metrics)} samples...") type2items = collections.defaultdict(list) for item in metrics: type2items[item["type"]].append(item) if args.only_eval_ans: logger.info(f'Ans Recall: {np.mean([m["ans_recall"] for m in metrics])}') for t in type2items.keys(): logger.info(f"{t} Questions num: {len(type2items[t])}") logger.info(f'Ans Recall: {np.mean([m["ans_recall"] for m in type2items[t]])}') else: logger.info(f'\tAvg PR: {np.mean([m["p_recall"] for m in metrics])}') logger.info(f'\tAvg P-EM: {np.mean([m["p_em"] for m in metrics])}') logger.info(f'\tAvg 1-Recall: {np.mean([m["recall_1"] for m in metrics])}') logger.info(f'\tPath Recall: {np.mean([m["path_covered"] for m in metrics])}') for t in type2items.keys(): logger.info(f"{t} Questions num: {len(type2items[t])}") logger.info(f'\tAvg PR: {np.mean([m["p_recall"] for m in type2items[t]])}') logger.info(f'\tAvg P-EM: {np.mean([m["p_em"] for m in type2items[t]])}') logger.info(f'\tAvg 1-Recall: {np.mean([m["recall_1"] for m in type2items[t]])}') logger.info(f'\tPath Recall: {np.mean([m["path_covered"] for m in type2items[t]])}') ================================================ FILE: scripts/eval/eval_reranked.py ================================================ # Copyright (c) Facebook, Inc. and its affiliates. # All rights reserved. # # This source code is licensed under the license found in the # LICENSE file in the root directory of this source tree. import json import numpy as np from utils.utils import para_has_answer, normalize from utils.basic_tokenizer import SimpleTokenizer from tqdm import tqdm corpus = json.load(open("../data/hotpot_index/wiki_id2doc.json")) title2text = {v["title"]:v["text"] for v in corpus.values()} val_inputs = [json.loads(l) for l in open("../data/hotpot/hotpot_qas_val.json").readlines()] id2goldsp = {_["_id"]:_["sp"] for _ in val_inputs} id2goldans = {_["_id"]:_["answer"] for _ in val_inputs} id2type = {_["_id"]:_["type"] for _ in val_inputs} simple_tokenizer = SimpleTokenizer() # out best results results = json.load(open("../data/hotpot/results/hotpot_val_top100.json")) # # asai results # results = json.load(open("/private/home/xwhan/code/learning_to_retrieve_reasoning_paths/results/hotpot_dev_reader_titles.json")) # for k in results.keys(): # v = results[k] # v = [normalize(_[:-2]) for _ in v] # # import pdb; pdb.set_trace() # results[k] = v # results = {"titles":results} sp_ems = [] ans_recalls = [] bridge_ems = [] compare_ems = [] for qid in tqdm(results["titles"].keys()): chain = results["titles"][qid] sp = id2goldsp[qid] answer = id2goldans[qid] type_ = id2type[qid] # if answer[0].strip() in ["yes", "no"]: # continue sp_covered = int(np.sum([int(_ in chain) for _ in sp]) == len(sp)) concat_p = "yes no " + " ".join([t + " " + title2text.get(t, "") for t in chain]) ans_covered = para_has_answer(answer, concat_p, simple_tokenizer) ans_recalls.append(ans_covered) sp_ems.append(sp_covered) if type_ == "bridge": bridge_ems.append(sp_covered) else: compare_ems.append(sp_covered) print(len(sp_ems)) print(np.mean(sp_ems)) print(f"Answer Recall: {np.mean(ans_recalls)}, count: {len(ans_recalls)}") print("Bridge P EM:", np.mean(bridge_ems)) print("Comparison P EM:", np.mean(compare_ems)) ================================================ FILE: scripts/eval/eval_retrieval.py ================================================ # Copyright (c) Facebook, Inc. and its affiliates. # All rights reserved. # # This source code is licensed under the license found in the # LICENSE file in the root directory of this source tree. """ Single-hop retrieval evaluation ## Use the unified model (trained with both hotpotQA and NQ) python 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 # DPR shared-encoder baseline bsz256 python 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 # shared encoder on merged corpus python 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 # to get negatives from DPR shared baseline python 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 python 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 python 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 """ import numpy as np import json import faiss import argparse import logging import torch from tqdm import tqdm from multiprocessing import Pool as ProcessPool from multiprocessing.util import Finalize from functools import partial from collections import defaultdict from utils.utils import load_saved, move_to_cuda, para_has_answer from utils.basic_tokenizer import SimpleTokenizer from transformers import AutoConfig, AutoTokenizer from models.retriever import BertRetrieverSingle, RobertaRetrieverSingle from models.unified_retriever import UnifiedRetriever, BertNQRetriever logger = logging.getLogger() logger.setLevel(logging.INFO) if (logger.hasHandlers()): logger.handlers.clear() console = logging.StreamHandler() logger.addHandler(console) PROCESS_TOK = None def init(): global PROCESS_TOK PROCESS_TOK = SimpleTokenizer() Finalize(PROCESS_TOK, PROCESS_TOK.shutdown, exitpriority=100) def get_score(answer_doc, topk=20): """Search through all the top docs to see if they have the answer.""" question, answer, docs = answer_doc top5doc_covered = 0 global PROCESS_TOK topkpara_covered = [] for p in docs: topkpara_covered.append(int(para_has_answer(answer, p["title"] + " " + p["text"], PROCESS_TOK))) return { "5": int(np.sum(topkpara_covered[:5]) > 0), "10": int(np.sum(topkpara_covered[:10]) > 0), "20": int(np.sum(topkpara_covered[:20]) > 0), "50": int(np.sum(topkpara_covered[:50]) > 0), "100": int(np.sum(topkpara_covered[:100]) > 0), "covered": topkpara_covered } def add_marker_q(tokenizer, q): q_toks = tokenizer.tokenize(q) return ['[unused0]'] + q_toks if __name__ == '__main__': parser = argparse.ArgumentParser() parser.add_argument('raw_data', type=str, default=None) parser.add_argument('indexpath', type=str, default=None) parser.add_argument('corpus_dict', type=str, default=None) parser.add_argument('model_path', type=str, default=None) parser.add_argument('--batch-size', type=int, default=100) parser.add_argument('--topk', type=int, default=100) parser.add_argument('--max-q-len', type=int, default=100) parser.add_argument('--num-workers', type=int, default=10) parser.add_argument('--shared-encoder', action="store_true") parser.add_argument('--model-name', type=str, default='bert-base-uncased') parser.add_argument("--stop-drop", default=0, type=float) parser.add_argument("--gpu", action="store_true") parser.add_argument("--save-pred", default="", type=str) parser.add_argument("--unified", action="store_true", help="test with unified trained model") args = parser.parse_args() logger.info(f"Loading questions") qas = [json.loads(line) for line in open(args.raw_data).readlines()] questions = [_["question"][:-1] if _["question"].endswith("?") else _["question"] for _ in qas] answers = [item["answer"] for item in qas] logger.info(f"Loading index") d = 768 xb = np.load(args.indexpath).astype('float32') index = faiss.IndexFlatIP(d) index.add(xb) if args.gpu: res = faiss.StandardGpuResources() index = faiss.index_cpu_to_gpu(res, 1, index) # logger.info(f"Building GPU index") # co = faiss.GpuMultipleClonerOptions() # co.useFloat16 = True # co.shards = True # index = faiss.index_cpu_to_gpus_list(index, co, [1,2,3,4,5,6,7]) # index.add(xb) logger.info("Loading trained model...") bert_config = AutoConfig.from_pretrained(args.model_name) tokenizer = AutoTokenizer.from_pretrained(args.model_name) if args.unified: model = UnifiedRetriever(bert_config, args) elif "roberta" in args.model_name: model = RobertaRetrieverSingle(bert_config, args) else: model = BertRetrieverSingle(bert_config, args) model = load_saved(model, args.model_path, exact=False) cuda = torch.device('cuda') model.to(cuda) from apex import amp model = amp.initialize(model, opt_level='O1') model.eval() logger.info(f"Loading corpus") id2doc = json.load(open(args.corpus_dict)) logger.info(f"Corpus size {len(id2doc)}") retrieved_results = [] retrieved_docids = [] for b_start in tqdm(range(0, len(questions), args.batch_size)): with torch.no_grad(): batch_q = questions[b_start:b_start + args.batch_size] batch_ans = answers[b_start:b_start + args.batch_size] # test retrieval model with marker # batch_q_toks = [add_marker_q(tokenizer, q) for q in batch_q] # 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) 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) batch_q_encodes = move_to_cuda(dict(batch_q_encodes)) q_embeds = model.encode_q(batch_q_encodes["input_ids"], batch_q_encodes["attention_mask"], batch_q_encodes.get("token_type_ids", None)) q_embeds_numpy = q_embeds.cpu().contiguous().numpy() D, I = index.search(q_embeds_numpy, args.topk) for b_idx in range(len(batch_q)): topk_docs = [{"title": id2doc[str(doc_id)][0],"text": id2doc[str(doc_id)][1]} for doc_id in I[b_idx]] retrieved_results.append(topk_docs) retrieved_docids.append([str(doc_id) for doc_id in I[b_idx]]) answers_docs = list(zip(questions, answers, retrieved_results)) processes = ProcessPool( processes=args.num_workers, initializer=init ) get_score_partial = partial( get_score, topk=args.topk) results = processes.map(get_score_partial, answers_docs) if args.save_pred != "": to_save = [] for inputs, metrics, topk_ids in zip(answers_docs, results, retrieved_docids): q, ans, topk_doc = inputs topk_covered = metrics["covered"] assert len(topk_doc) == len(topk_covered) assert len(topk_doc) == len(topk_ids) to_save.append({ "question": q, "ans": ans, "topk": list(zip(topk_doc, topk_covered)), "topkdocs": topk_doc, "metrics": metrics, "topk_ids": topk_ids }) print(f"Saving {len(to_save)} instances...") with open("/private/home/xwhan/data/nq-dpr/results/" + args.save_pred, "w") as g: for l in to_save: g.write(json.dumps(l) + "\n") aggregate = defaultdict(list) for r in results: for k, v in r.items(): aggregate[k].append(v) for k in aggregate: results = aggregate[k] print('Top {} Recall for {} QA pairs: {} ...'.format( k, len(results), np.mean(results))) ================================================ FILE: scripts/eval/eval_single_fever.py ================================================ # Copyright (c) Facebook, Inc. and its affiliates. # All rights reserved. # # This source code is licensed under the license found in the # LICENSE file in the root directory of this source tree. """ python 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 python 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 """ import argparse import json import logging import faiss import numpy as np import torch from tqdm import tqdm from transformers import AutoConfig, AutoTokenizer from models.retriever import BertRetrieverSingle from models.unified_retriever import UnifiedRetriever from utils.basic_tokenizer import SimpleTokenizer from utils.utils import (load_saved, move_to_cuda) logger = logging.getLogger() logger.setLevel(logging.INFO) if (logger.hasHandlers()): logger.handlers.clear() console = logging.StreamHandler() logger.addHandler(console) if __name__ == '__main__': parser = argparse.ArgumentParser() parser.add_argument('raw_data', type=str, default=None) parser.add_argument('indexpath', type=str, default=None) parser.add_argument('corpus_dict', type=str, default=None) parser.add_argument('model_path', type=str, default=None) parser.add_argument('--topk', type=int, default=2, help="topk paths") parser.add_argument('--num-workers', type=int, default=10) parser.add_argument('--max-q-len', type=int, default=45) parser.add_argument('--batch-size', type=int, default=100) parser.add_argument('--model-name', type=str, default='bert-base-uncased') parser.add_argument('--gpu', action="store_true") parser.add_argument('--shared-encoder', action="store_true") parser.add_argument("--save-path", type=str, default="") parser.add_argument("--stop-drop", default=0, type=float) args = parser.parse_args() logger.info("Loading data...") ds_items = [json.loads(_) for _ in open(args.raw_data).readlines()] logger.info("Building index...") d = 768 xb = np.load(args.indexpath).astype('float32') print(xb.shape) index = faiss.IndexFlatIP(d) index.add(xb) if args.gpu: res = faiss.StandardGpuResources() index = faiss.index_cpu_to_gpu(res, 1, index) logger.info(f"Loading corpus...") id2doc = json.load(open(args.corpus_dict)) title2doc = {item[0]:item[1] for item in id2doc.values()} logger.info(f"Corpus size {len(id2doc)}") logger.info("Loading trained model...") bert_config = AutoConfig.from_pretrained(args.model_name) tokenizer = AutoTokenizer.from_pretrained(args.model_name) # model = BertRetrieverSingle(bert_config, args) model = UnifiedRetriever(bert_config, args) model = load_saved(model, args.model_path, exact=False) simple_tokenizer = SimpleTokenizer() cuda = torch.device('cuda') model.to(cuda) from apex import amp model = amp.initialize(model, opt_level='O1') model.eval() logger.info("Encoding claims and searching") questions = [_["claim"] for _ in ds_items] metrics = [] retrieval_outputs = [] for b_start in tqdm(range(0, len(questions), args.batch_size)): with torch.no_grad(): batch_q = questions[b_start:b_start + args.batch_size] batch_ann = ds_items[b_start:b_start + args.batch_size] bsize = len(batch_q) batch_q_encodes = tokenizer.batch_encode_plus(batch_q, max_length=args.max_q_len, pad_to_max_length=True, return_tensors="pt") batch_q_encodes = move_to_cuda(dict(batch_q_encodes)) q_embeds = model.encode_q(batch_q_encodes["input_ids"], batch_q_encodes["attention_mask"], batch_q_encodes.get("token_type_ids", None)) q_embeds_numpy = q_embeds.cpu().contiguous().numpy() D, I = index.search(q_embeds_numpy, args.topk) for b_idx in range(bsize): topk_docs = [] for _, doc_id in enumerate(I[b_idx]): doc = id2doc[str(doc_id)] topk_docs.append({"title": doc[0], "text": doc[1]}) # saving when there's no annotations if args.save_path != "": candidaite_chains = [] retrieval_outputs.append({ "id": batch_ann[b_idx]["id"], "claim": batch_ann[b_idx]["claim"], "topk": topk_docs, }) if args.save_path != "": with open(f"/private/home/xwhan/data/fever/retrieval/{args.save_path}", "w") as out: for l in retrieval_outputs: out.write(json.dumps(l) + "\n") ================================================ FILE: scripts/train_mhop.py ================================================ # Copyright (c) Facebook, Inc. and its affiliates. # All rights reserved. # # This source code is licensed under the license found in the # LICENSE file in the root directory of this source tree. """ Description: train a multi-hop dense retrieval from pretrained BERT/RoBERTa encoder Usage: CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python scripts/train_mhop.py \ --do_train \ --prefix ${RUN_ID} \ --predict_batch_size 3000 \ --model_name roberta-base \ --train_batch_size 150 \ --learning_rate 2e-5 \ --fp16 \ --train_file ${TRAIN_DATA_PATH} \ --predict_file ${DEV_DATA_PATH} \ --seed 16 \ --eval-period -1 \ --max_c_len 300 \ --max_q_len 70 \ --max_q_sp_len 350 \ --shared-encoder \ --warmup-ratio 0.1 """ import logging import os import random from datetime import date from functools import partial import numpy as np import torch from torch.optim import Adam from torch.utils.data import DataLoader from torch.utils.tensorboard import SummaryWriter from tqdm import tqdm from transformers import (AdamW, AutoConfig, AutoTokenizer, get_linear_schedule_with_warmup) from mdr.retrieval.config import train_args from mdr.retrieval.criterions import (mhop_eval, mhop_loss) from mdr.retrieval.data.mhop_dataset import MhopDataset, mhop_collate from mdr.retrieval.models.mhop_retriever import RobertaRetriever from mdr.retrieval.utils.utils import AverageMeter, move_to_cuda, load_saved def main(): args = train_args() if args.fp16: import apex apex.amp.register_half_function(torch, 'einsum') date_curr = date.today().strftime("%m-%d-%Y") 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}" args.output_dir = os.path.join(args.output_dir, date_curr, model_name) tb_logger = SummaryWriter(os.path.join(args.output_dir.replace("logs","tflogs"))) if os.path.exists(args.output_dir) and os.listdir(args.output_dir): print( f"output directory {args.output_dir} already exists and is not empty.") if not os.path.exists(args.output_dir): os.makedirs(args.output_dir, exist_ok=True) logging.basicConfig(format='%(asctime)s - %(levelname)s - %(name)s - %(message)s', datefmt='%m/%d/%Y %H:%M:%S', level=logging.INFO, handlers=[logging.FileHandler(os.path.join(args.output_dir, "log.txt")), logging.StreamHandler()]) logger = logging.getLogger(__name__) logger.info(args) if args.local_rank == -1 or args.no_cuda: device = torch.device( "cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu") n_gpu = torch.cuda.device_count() else: device = torch.device("cuda", args.local_rank) n_gpu = 1 torch.distributed.init_process_group(backend='nccl') logger.info("device %s n_gpu %d distributed training %r", device, n_gpu, bool(args.local_rank != -1)) if args.accumulate_gradients < 1: raise ValueError("Invalid accumulate_gradients parameter: {}, should be >= 1".format( args.accumulate_gradients)) args.train_batch_size = int( args.train_batch_size / args.accumulate_gradients) random.seed(args.seed) np.random.seed(args.seed) torch.manual_seed(args.seed) if n_gpu > 0: torch.cuda.manual_seed_all(args.seed) bert_config = AutoConfig.from_pretrained(args.model_name) model = RobertaRetriever(bert_config, args) tokenizer = AutoTokenizer.from_pretrained(args.model_name) collate_fc = partial(mhop_collate, pad_id=tokenizer.pad_token_id) if args.do_train and args.max_c_len > bert_config.max_position_embeddings: raise ValueError( "Cannot use sequence length %d because the BERT model " "was only trained up to sequence length %d" % (args.max_c_len, bert_config.max_position_embeddings)) eval_dataset = MhopDataset( tokenizer, args.predict_file, args.max_q_len, args.max_q_sp_len, args.max_c_len) eval_dataloader = DataLoader( eval_dataset, batch_size=args.predict_batch_size, collate_fn=collate_fc, pin_memory=True, num_workers=args.num_workers) logger.info(f"Num of dev batches: {len(eval_dataloader)}") if args.init_checkpoint != "": model = load_saved(model, args.init_checkpoint) model.to(device) print(f"number of trainable parameters: {sum(p.numel() for p in model.parameters() if p.requires_grad)}") if args.do_train: no_decay = ['bias', 'LayerNorm.weight'] optimizer_parameters = [ {'params': [p for n, p in model.named_parameters() if not any( nd in n for nd in no_decay)], 'weight_decay': args.weight_decay}, {'params': [p for n, p in model.named_parameters() if any( nd in n for nd in no_decay)], 'weight_decay': 0.0} ] optimizer = Adam(optimizer_parameters, lr=args.learning_rate, eps=args.adam_epsilon) if args.fp16: from apex import amp model, optimizer = amp.initialize( model, optimizer, opt_level=args.fp16_opt_level) else: if args.fp16: from apex import amp model = amp.initialize(model, opt_level=args.fp16_opt_level) if args.local_rank != -1: model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[args.local_rank], output_device=args.local_rank) elif n_gpu > 1: model = torch.nn.DataParallel(model) if args.do_train: global_step = 0 # gradient update step batch_step = 0 # forward batch count best_mrr = 0 train_loss_meter = AverageMeter() model.train() train_dataset = MhopDataset(tokenizer, args.train_file, args.max_q_len, args.max_q_sp_len, args.max_c_len, train=True) 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) t_total = len(train_dataloader) // args.gradient_accumulation_steps * args.num_train_epochs warmup_steps = t_total * args.warmup_ratio scheduler = get_linear_schedule_with_warmup( optimizer, num_warmup_steps=warmup_steps, num_training_steps=t_total ) logger.info('Start training....') for epoch in range(int(args.num_train_epochs)): for batch in tqdm(train_dataloader): batch_step += 1 batch = move_to_cuda(batch) loss = mhop_loss(model, batch, args) if args.gradient_accumulation_steps > 1: loss = loss / args.gradient_accumulation_steps if args.fp16: with amp.scale_loss(loss, optimizer) as scaled_loss: scaled_loss.backward() else: loss.backward() train_loss_meter.update(loss.item()) if (batch_step + 1) % args.gradient_accumulation_steps == 0: if args.fp16: torch.nn.utils.clip_grad_norm_( amp.master_params(optimizer), args.max_grad_norm) else: torch.nn.utils.clip_grad_norm_( model.parameters(), args.max_grad_norm) optimizer.step() scheduler.step() model.zero_grad() global_step += 1 tb_logger.add_scalar('batch_train_loss', loss.item(), global_step) tb_logger.add_scalar('smoothed_train_loss', train_loss_meter.avg, global_step) if args.eval_period != -1 and global_step % args.eval_period == 0: mrrs = predict(args, model, eval_dataloader, device, logger) mrr = mrrs["mrr_avg"] logger.info("Step %d Train loss %.2f MRR %.2f on epoch=%d" % (global_step, train_loss_meter.avg, mrr*100, epoch)) if best_mrr < mrr: logger.info("Saving model with best MRR %.2f -> MRR %.2f on epoch=%d" % (best_mrr*100, mrr*100, epoch)) torch.save(model.state_dict(), os.path.join( args.output_dir, f"checkpoint_best.pt")) model = model.to(device) best_mrr = mrr mrrs = predict(args, model, eval_dataloader, device, logger) mrr = mrrs["mrr_avg"] logger.info("Step %d Train loss %.2f MRR-AVG %.2f on epoch=%d" % ( global_step, train_loss_meter.avg, mrr*100, epoch)) for k, v in mrrs.items(): tb_logger.add_scalar(k, v*100, epoch) torch.save(model.state_dict(), os.path.join( args.output_dir, f"checkpoint_last.pt")) if best_mrr < mrr: logger.info("Saving model with best MRR %.2f -> MRR %.2f on epoch=%d" % (best_mrr*100, mrr*100, epoch)) torch.save(model.state_dict(), os.path.join( args.output_dir, f"checkpoint_best.pt")) best_mrr = mrr logger.info("Training finished!") elif args.do_predict: acc = predict(args, model, eval_dataloader, device, logger) logger.info(f"test performance {acc}") def predict(args, model, eval_dataloader, device, logger): model.eval() rrs_1, rrs_2 = [], [] # reciprocal rank for batch in tqdm(eval_dataloader): batch_to_feed = move_to_cuda(batch) with torch.no_grad(): outputs = model(batch_to_feed) eval_results = mhop_eval(outputs, args) _rrs_1, _rrs_2 = eval_results["rrs_1"], eval_results["rrs_2"] rrs_1 += _rrs_1 rrs_2 += _rrs_2 mrr_1 = np.mean(rrs_1) mrr_2 = np.mean(rrs_2) logger.info(f"evaluated {len(rrs_1)} examples...") logger.info(f'MRR-1: {mrr_1}') logger.info(f'MRR-2: {mrr_2}') model.train() return {"mrr_1": mrr_1, "mrr_2": mrr_2, "mrr_avg": (mrr_1 + mrr_2) / 2} if __name__ == "__main__": main() ================================================ FILE: scripts/train_momentum.py ================================================ # Copyright (c) Facebook, Inc. and its affiliates. # All rights reserved. # # This source code is licensed under the license found in the # LICENSE file in the root directory of this source tree. import logging import os import random from datetime import date from functools import partial import numpy as np import torch from torch.optim import Adam from torch.utils.data import DataLoader from torch.utils.tensorboard import SummaryWriter from tqdm import tqdm from transformers import (AdamW, AutoConfig, AutoTokenizer, get_linear_schedule_with_warmup) from mdr.retrieval.config import train_args from mdr.retrieval.criterions import (mhop_eval, mhop_loss) from mdr.retrieval.data.mhop_dataset import MhopDataset, mhop_collate from mdr.retrieval.models.mhop_retriever import RobertaMomentumRetriever from mdr.retrieval.utils.utils import AverageMeter, move_to_cuda from mdr.retrieval.data.fever_dataset import FeverDataset def main(): args = train_args() if args.fp16: import apex apex.amp.register_half_function(torch, 'einsum') date_curr = date.today().strftime("%m-%d-%Y") 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}" args.output_dir = os.path.join(args.output_dir, date_curr, model_name) tb_logger = SummaryWriter(os.path.join(args.output_dir.replace("logs","tflogs"))) if os.path.exists(args.output_dir) and os.listdir(args.output_dir): print( f"output directory {args.output_dir} already exists and is not empty.") if not os.path.exists(args.output_dir): os.makedirs(args.output_dir, exist_ok=True) logging.basicConfig(format='%(asctime)s - %(levelname)s - %(name)s - %(message)s', datefmt='%m/%d/%Y %H:%M:%S', level=logging.INFO, handlers=[logging.FileHandler(os.path.join(args.output_dir, "log.txt")), logging.StreamHandler()]) logger = logging.getLogger(__name__) logger.info(args) if args.local_rank == -1 or args.no_cuda: device = torch.device( "cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu") n_gpu = torch.cuda.device_count() else: device = torch.device("cuda", args.local_rank) n_gpu = 1 torch.distributed.init_process_group(backend='nccl') logger.info("device %s n_gpu %d distributed training %r", device, n_gpu, bool(args.local_rank != -1)) if args.accumulate_gradients < 1: raise ValueError("Invalid accumulate_gradients parameter: {}, should be >= 1".format( args.accumulate_gradients)) args.train_batch_size = int( args.train_batch_size / args.accumulate_gradients) random.seed(args.seed) np.random.seed(args.seed) torch.manual_seed(args.seed) if n_gpu > 0: torch.cuda.manual_seed_all(args.seed) bert_config = AutoConfig.from_pretrained(args.model_name) model = RobertaMomentumRetriever(bert_config, args) tokenizer = AutoTokenizer.from_pretrained(args.model_name) collate_fc = partial(mhop_collate, pad_id=tokenizer.pad_token_id) if args.do_train and args.max_c_len > bert_config.max_position_embeddings: raise ValueError( "Cannot use sequence length %d because the BERT model " "was only trained up to sequence length %d" % (args.max_c_len, bert_config.max_position_embeddings)) if "fever" in args.predict_file: eval_dataset = FeverDataset( tokenizer, args.predict_file, args.max_q_len, args.max_q_sp_len, args.max_c_len) else: eval_dataset = MhopDataset( tokenizer, args.predict_file, args.max_q_len, args.max_q_sp_len, args.max_c_len) eval_dataloader = DataLoader( eval_dataset, batch_size=args.predict_batch_size, collate_fn=collate_fc, pin_memory=True, num_workers=args.num_workers) logger.info(f"Num of dev batches: {len(eval_dataloader)}") model.to(device) print(f"number of trainable parameters: {sum(p.numel() for p in model.parameters() if p.requires_grad)}") if args.do_train: no_decay = ['bias', 'LayerNorm.weight'] optimizer_parameters = [ {'params': [p for n, p in model.named_parameters() if not any( nd in n for nd in no_decay)], 'weight_decay': args.weight_decay}, {'params': [p for n, p in model.named_parameters() if any( nd in n for nd in no_decay)], 'weight_decay': 0.0} ] optimizer = Adam(optimizer_parameters, lr=args.learning_rate, eps=args.adam_epsilon) if args.fp16: from apex import amp model, optimizer = amp.initialize( model, optimizer, opt_level=args.fp16_opt_level) else: if args.fp16: from apex import amp model = amp.initialize(model, opt_level=args.fp16_opt_level) if args.local_rank != -1: model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[args.local_rank], output_device=args.local_rank) elif n_gpu > 1: model = torch.nn.DataParallel(model) if args.do_train: global_step = 0 # gradient update step batch_step = 0 # forward batch count best_mrr = 0 train_loss_meter = AverageMeter() model.train() if "fever" in args.train_file: train_dataset = FeverDataset(tokenizer, args.train_file, args.max_q_len, args.max_q_sp_len, args.max_c_len, train=True) else: train_dataset = MhopDataset(tokenizer, args.train_file, args.max_q_len, args.max_q_sp_len, args.max_c_len, train=True) 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) t_total = len(train_dataloader) // args.gradient_accumulation_steps * args.num_train_epochs warmup_steps = t_total * args.warmup_ratio scheduler = get_linear_schedule_with_warmup( optimizer, num_warmup_steps=warmup_steps, num_training_steps=t_total ) logger.info('Start training....') for epoch in range(int(args.num_train_epochs)): for batch in tqdm(train_dataloader): batch_step += 1 batch = move_to_cuda(batch) loss = mhop_loss(model, batch, args) if args.gradient_accumulation_steps > 1: loss = loss / args.gradient_accumulation_steps if args.fp16: with amp.scale_loss(loss, optimizer) as scaled_loss: scaled_loss.backward() else: loss.backward() train_loss_meter.update(loss.item()) if (batch_step + 1) % args.gradient_accumulation_steps == 0: if args.fp16: torch.nn.utils.clip_grad_norm_( amp.master_params(optimizer), args.max_grad_norm) else: torch.nn.utils.clip_grad_norm_( model.parameters(), args.max_grad_norm) optimizer.step() scheduler.step() model.zero_grad() global_step += 1 tb_logger.add_scalar('batch_train_loss', loss.item(), global_step) tb_logger.add_scalar('smoothed_train_loss', train_loss_meter.avg, global_step) if args.eval_period != -1 and global_step % args.eval_period == 0: mrrs = predict(args, model, eval_dataloader, device, logger) mrr = mrrs["mrr_avg"] logger.info("Step %d Train loss %.2f MRR %.2f on epoch=%d" % (global_step, train_loss_meter.avg, mrr*100, epoch)) if best_mrr < mrr: logger.info("Saving model with best MRR %.2f -> MRR %.2f on epoch=%d" % (best_mrr*100, mrr*100, epoch)) torch.save(model.module.encoder_q.state_dict(), os.path.join( args.output_dir, f"checkpoint_q_best.pt")) torch.save(model.module.encoder_q.state_dict(), os.path.join( args.output_dir, f"checkpoint_k_best.pt")) model = model.to(device) best_mrr = mrr mrrs = predict(args, model, eval_dataloader, device, logger) mrr = mrrs["mrr_avg"] logger.info("Step %d Train loss %.2f MRR-AVG %.2f on epoch=%d" % ( global_step, train_loss_meter.avg, mrr*100, epoch)) for k, v in mrrs.items(): tb_logger.add_scalar(k, v*100, epoch) if best_mrr < mrr: logger.info("Saving model with best MRR %.2f -> MRR %.2f on epoch=%d" % (best_mrr*100, mrr*100, epoch)) torch.save(model.module.encoder_q.state_dict(), os.path.join( args.output_dir, f"checkpoint_q_best.pt")) torch.save(model.module.encoder_q.state_dict(), os.path.join( args.output_dir, f"checkpoint_k_best.pt")) best_mrr = mrr logger.info("Training finished!") elif args.do_predict: acc = predict(args, model, eval_dataloader, device, logger) logger.info(f"test performance {acc}") def predict(args, model, eval_dataloader, device, logger): model.eval() rrs_1, rrs_2 = [], [] # reciprocal rank for batch in tqdm(eval_dataloader): batch_to_feed = move_to_cuda(batch) with torch.no_grad(): outputs = model(batch_to_feed) eval_results = mhop_eval(outputs, args) _rrs_1, _rrs_2 = eval_results["rrs_1"], eval_results["rrs_2"] rrs_1 += _rrs_1 rrs_2 += _rrs_2 mrr_1 = np.mean(rrs_1) mrr_2 = np.mean(rrs_2) logger.info(f"evaluated {len(rrs_1)} examples...") logger.info(f'MRR-1: {mrr_1}') logger.info(f'MRR-2: {mrr_2}') model.train() return {"mrr_1": mrr_1, "mrr_2": mrr_2, "mrr_avg": (mrr_1 + mrr_2) / 2} if __name__ == "__main__": main() ================================================ FILE: scripts/train_qa.py ================================================ # Copyright (c) Facebook, Inc. and its affiliates. # All rights reserved. # # This source code is licensed under the license found in the # LICENSE file in the root directory of this source tree. import collections import json import logging import os import random from datetime import date from functools import partial import numpy as np from numpy.core.defchararray import encode import torch from torch import sparse_coo_tensor import torch.nn.functional as F from torch.utils.data import DataLoader from torch.utils.tensorboard import SummaryWriter from torch.optim import Adam from tqdm import tqdm from transformers import (AdamW, AutoConfig, AutoTokenizer, get_linear_schedule_with_warmup) from mdr.qa.config import train_args from mdr.qa.qa_dataset import QADataset, qa_collate, MhopSampler from mdr.qa.qa_model import QAModel from mdr.qa.utils import AverageMeter, move_to_cuda, get_final_text from mdr.qa.hotpot_evaluate_v1 import f1_score, exact_match_score, update_sp def load_saved(model, path): state_dict = torch.load(path) def filter(x): return x[7:] if x.startswith('module.') else x state_dict = {filter(k): v for (k, v) in state_dict.items()} model.load_state_dict(state_dict) return model def main(): args = train_args() if args.fp16: import apex apex.amp.register_half_function(torch, 'einsum') date_curr = date.today().strftime("%m-%d-%Y") 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}" args.output_dir = os.path.join(args.output_dir, date_curr, model_name) tb_logger = SummaryWriter(os.path.join(args.output_dir.replace("logs","tflogs"))) if os.path.exists(args.output_dir) and os.listdir(args.output_dir): print( f"output directory {args.output_dir} already exists and is not empty.") if not os.path.exists(args.output_dir): os.makedirs(args.output_dir, exist_ok=True) logging.basicConfig(format='%(asctime)s - %(levelname)s - %(name)s - %(message)s', datefmt='%m/%d/%Y %H:%M:%S', level=logging.INFO, handlers=[logging.FileHandler(os.path.join(args.output_dir, "log.txt")), logging.StreamHandler()]) logger = logging.getLogger(__name__) logger.setLevel(logging.INFO) logger.info(args) if args.local_rank == -1 or args.no_cuda: device = torch.device( "cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu") n_gpu = torch.cuda.device_count() else: device = torch.device("cuda", args.local_rank) n_gpu = 1 # Initializes the distributed backend which will take care of sychronizing nodes/GPUs torch.distributed.init_process_group(backend='nccl') logger.info("device %s n_gpu %d distributed training %r", device, n_gpu, bool(args.local_rank != -1)) if args.shared_norm: # chains of each question are on the same gpu assert (args.train_batch_size // n_gpu) == args.neg_num + 1 random.seed(args.seed) np.random.seed(args.seed) torch.manual_seed(args.seed) if n_gpu > 0: torch.cuda.manual_seed_all(args.seed) # define model if args.model_name == "spanbert": bert_config = AutoConfig.from_pretrained("/private/home/span-bert") tokenizer = AutoTokenizer.from_pretrained('bert-large-cased') else: bert_config = AutoConfig.from_pretrained(args.model_name) tokenizer = AutoTokenizer.from_pretrained(args.model_name) model = QAModel(bert_config, args) collate_fc = partial(qa_collate, pad_id=tokenizer.pad_token_id) eval_dataset = QADataset(tokenizer, args.predict_file, args.max_seq_len, args.max_q_len) eval_dataloader = DataLoader(eval_dataset, batch_size=args.predict_batch_size, collate_fn=collate_fc, pin_memory=True, num_workers=args.num_workers) logger.info(f"Num of dev batches: {len(eval_dataloader)}") if args.init_checkpoint != "": logger.info(f"Loading model from {args.init_checkpoint}") model = load_saved(model, args.init_checkpoint) model.to(device) print(f"number of trainable parameters: {sum(p.numel() for p in model.parameters() if p.requires_grad)}") if args.do_train: no_decay = ['bias', 'LayerNorm.weight'] optimizer_parameters = [ {'params': [p for n, p in model.named_parameters() if not any( nd in n for nd in no_decay)], 'weight_decay': args.weight_decay}, {'params': [p for n, p in model.named_parameters() if any( nd in n for nd in no_decay)], 'weight_decay': 0.0} ] if args.use_adam: optimizer = Adam(optimizer_parameters, lr=args.learning_rate, eps=args.adam_epsilon) else: optimizer = AdamW(optimizer_parameters, lr=args.learning_rate, eps=args.adam_epsilon) if args.fp16: from apex import amp model, optimizer = amp.initialize( model, optimizer, opt_level=args.fp16_opt_level) else: if args.fp16: from apex import amp model = amp.initialize(model, opt_level=args.fp16_opt_level) if args.local_rank != -1: model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[args.local_rank], output_device=args.local_rank) elif n_gpu > 1: model = torch.nn.DataParallel(model) if args.do_train: global_step = 0 # gradient update step batch_step = 0 # forward batch count best_em = 0 train_loss_meter = AverageMeter() model.train() train_dataset = QADataset(tokenizer, args.train_file, args.max_seq_len, args.max_q_len, train=True) train_sampler = MhopSampler(train_dataset, num_neg=args.neg_num) 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) t_total = len(train_dataloader) // args.gradient_accumulation_steps * args.num_train_epochs warmup_steps = t_total * args.warmup_ratio scheduler = get_linear_schedule_with_warmup( optimizer, num_warmup_steps=warmup_steps, num_training_steps=t_total ) logger.info('Start training....') for epoch in range(int(args.num_train_epochs)): for batch in tqdm(train_dataloader): batch_step += 1 batch_inputs = move_to_cuda(batch["net_inputs"]) loss = model(batch_inputs) if n_gpu > 1: loss = loss.mean() if args.gradient_accumulation_steps > 1: loss = loss / args.gradient_accumulation_steps if args.fp16: with amp.scale_loss(loss, optimizer) as scaled_loss: scaled_loss.backward() else: loss.backward() train_loss_meter.update(loss.item()) if (batch_step + 1) % args.gradient_accumulation_steps == 0: if args.fp16: torch.nn.utils.clip_grad_norm_( amp.master_params(optimizer), args.max_grad_norm) else: torch.nn.utils.clip_grad_norm_( model.parameters(), args.max_grad_norm) optimizer.step() scheduler.step() model.zero_grad() global_step += 1 # logger.info(f"current batch loss: {loss.item()}") tb_logger.add_scalar('batch_train_loss', loss.item(), global_step) tb_logger.add_scalar('smoothed_train_loss', train_loss_meter.avg, global_step) if args.eval_period != -1 and global_step % args.eval_period == 0: metrics = predict(args, model, eval_dataloader, logger) em = metrics["em"] logger.info("Step %d Train loss %.2f em %.2f on epoch=%d" % (global_step, train_loss_meter.avg, em*100, epoch)) if best_em < em: logger.info("Saving model with best em %.2f -> em %.2f on step=%d" % (best_em*100, em*100, global_step)) torch.save(model.state_dict(), os.path.join( args.output_dir, f"checkpoint_best.pt")) model = model.to(device) best_em = em metrics = predict(args, model, eval_dataloader, logger) em = metrics["em"] logger.info("Step %d Train loss %.2f em %.2f" % ( global_step, train_loss_meter.avg, em*100)) tb_logger.add_scalar('dev_em', em*100, global_step) if best_em < em: logger.info("Saving model with best em %.2f -> em %.2f on epoch=%d" % (best_em*100, em*100, epoch)) torch.save(model.state_dict(), os.path.join( args.output_dir, f"checkpoint_best.pt")) best_em = em logger.info("Training finished!") elif args.do_predict: metrics = predict(args, model, eval_dataloader, logger, fixed_thresh=0.8) logger.info(f"test performance {metrics}") elif args.do_test: eval_final(args, model, eval_dataloader, weight=0.8) def predict(args, model, eval_dataloader, logger, fixed_thresh=None): model.eval() id2result = collections.defaultdict(list) id2answer = collections.defaultdict(list) id2gold = {} id2goldsp = {} for batch in tqdm(eval_dataloader): batch_to_feed = move_to_cuda(batch["net_inputs"]) batch_qids = batch["qids"] batch_labels = batch["net_inputs"]["label"].view(-1).tolist() with torch.no_grad(): outputs = model(batch_to_feed) scores = outputs["rank_score"] scores = scores.view(-1).tolist() if args.sp_pred: sp_scores = outputs["sp_score"] sp_scores = sp_scores.float().masked_fill(batch_to_feed["sent_offsets"].eq(0), float("-inf")).type_as(sp_scores) batch_sp_scores = sp_scores.sigmoid() # ans_type_predicted = torch.argmax(outputs["ans_type_logits"], dim=1).view(-1).tolist() outs = [outputs["start_logits"], outputs["end_logits"]] for qid, label, score in zip(batch_qids, batch_labels, scores): id2result[qid].append((label, score)) # answer prediction span_scores = outs[0][:, :, None] + outs[1][:, None] max_seq_len = span_scores.size(1) span_mask = np.tril(np.triu(np.ones((max_seq_len, max_seq_len)), 0), args.max_ans_len) span_mask = span_scores.data.new(max_seq_len, max_seq_len).copy_(torch.from_numpy(span_mask)) span_scores_masked = span_scores.float().masked_fill((1 - span_mask[None].expand_as(span_scores)).bool(), -1e10).type_as(span_scores) start_position = span_scores_masked.max(dim=2)[0].max(dim=1)[1] end_position = span_scores_masked.max(dim=2)[1].gather( 1, start_position.unsqueeze(1)).squeeze(1) answer_scores = span_scores_masked.max(dim=2)[0].max(dim=1)[0].tolist() para_offset = batch['para_offsets'] start_position_ = list( np.array(start_position.tolist()) - np.array(para_offset)) end_position_ = list( np.array(end_position.tolist()) - np.array(para_offset)) for idx, qid in enumerate(batch_qids): id2gold[qid] = batch["gold_answer"][idx] id2goldsp[qid] = batch["sp_gold"][idx] rank_score = scores[idx] start = start_position_[idx] end = end_position_[idx] span_score = answer_scores[idx] tok_to_orig_index = batch['tok_to_orig_index'][idx] doc_tokens = batch['doc_tokens'][idx] wp_tokens = batch['wp_tokens'][idx] orig_doc_start = tok_to_orig_index[start] orig_doc_end = tok_to_orig_index[end] orig_tokens = doc_tokens[orig_doc_start:(orig_doc_end + 1)] tok_tokens = wp_tokens[start:end+1] tok_text = " ".join(tok_tokens) tok_text = tok_text.replace(" ##", "") tok_text = tok_text.replace("##", "") tok_text = tok_text.strip() tok_text = " ".join(tok_text.split()) orig_text = " ".join(orig_tokens) pred_str = get_final_text(tok_text, orig_text, do_lower_case=True, verbose_logging=False) # get the sp sentences pred_sp = [] if args.sp_pred: sp_score = batch_sp_scores[idx].tolist() passages = batch["passages"][idx] for passage, sent_offset in zip(passages, [0, len(passages[0]["sents"])]): for idx, _ in enumerate(passage["sents"]): try: if sp_score[idx + sent_offset] >= 0.5: pred_sp.append([passage["title"], idx]) except: # logger.info(f"sentence exceeds max lengths") continue id2answer[qid].append({ "pred_str": pred_str.strip(), "rank_score": rank_score, "span_score": span_score, "pred_sp": pred_sp }) acc = [] for qid, res in id2result.items(): res.sort(key=lambda x: x[1], reverse=True) acc.append(res[0][0] == 1) logger.info(f"evaluated {len(id2result)} questions...") logger.info(f'chain ranking em: {np.mean(acc)}') best_em, best_f1, best_joint_em, best_joint_f1, best_sp_em, best_sp_f1 = 0, 0, 0, 0, 0, 0 best_res = None if fixed_thresh: lambdas = [fixed_thresh] else: # selecting threshhold on the dev data lambdas = [0, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1] for lambda_ in lambdas: ems, f1s, sp_ems, sp_f1s, joint_ems, joint_f1s = [], [], [], [], [], [] results = collections.defaultdict(dict) for qid, res in id2result.items(): ans_res = id2answer[qid] ans_res.sort(key=lambda x: lambda_ * x["rank_score"] + (1 - lambda_) * x["span_score"], reverse=True) top_pred = ans_res[0]["pred_str"] top_pred_sp = ans_res[0]["pred_sp"] results["answer"][qid] = top_pred results["sp"][qid] = top_pred_sp ems.append(exact_match_score(top_pred, id2gold[qid][0])) f1, prec, recall = f1_score(top_pred, id2gold[qid][0]) f1s.append(f1) if args.sp_pred: metrics = {'sp_em': 0, 'sp_f1': 0, 'sp_prec': 0, 'sp_recall': 0} update_sp(metrics, top_pred_sp, id2goldsp[qid]) sp_ems.append(metrics['sp_em']) sp_f1s.append(metrics['sp_f1']) # joint metrics joint_prec = prec * metrics["sp_prec"] joint_recall = recall * metrics["sp_recall"] if joint_prec + joint_recall > 0: joint_f1 = 2 * joint_prec * joint_recall / (joint_prec + joint_recall) else: joint_f1 = 0. joint_em = ems[-1] * sp_ems[-1] joint_ems.append(joint_em) joint_f1s.append(joint_f1) if args.sp_pred: if best_joint_f1 < np.mean(joint_f1s): best_joint_f1 = np.mean(joint_f1s) best_joint_em = np.mean(joint_ems) best_sp_f1 = np.mean(sp_f1s) best_sp_em = np.mean(sp_ems) best_f1 = np.mean(f1s) best_em = np.mean(ems) best_res = results else: if best_f1 < np.mean(f1s): best_f1 = np.mean(f1s) best_em = np.mean(ems) logger.info(f".......Using combination factor {lambda_}......") logger.info(f'answer em: {np.mean(ems)}, count: {len(ems)}') logger.info(f'answer f1: {np.mean(f1s)}, count: {len(f1s)}') logger.info(f'sp em: {np.mean(sp_ems)}, count: {len(sp_ems)}') logger.info(f'sp f1: {np.mean(sp_f1s)}, count: {len(sp_f1s)}') logger.info(f'joint em: {np.mean(joint_ems)}, count: {len(joint_ems)}') logger.info(f'joint f1: {np.mean(joint_f1s)}, count: {len(joint_f1s)}') logger.info(f"Best joint F1 from combination {best_f1}") if args.save_prediction != "": json.dump(best_res, open(f"{args.save_prediction}", "w")) model.train() 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} import time def eval_final(args, model, eval_dataloader, weight=0.8, gpu=True): """ for final submission """ model.eval() id2answer = collections.defaultdict(list) encode_times = [] for batch in tqdm(eval_dataloader): batch_to_feed = move_to_cuda(batch["net_inputs"]) if gpu else batch["net_inputs"] batch_qids = batch["qids"] with torch.no_grad(): start = time.time() outputs = model(batch_to_feed) encode_times.append(time.time() - start) scores = outputs["rank_score"] scores = scores.view(-1).tolist() if args.sp_pred: sp_scores = outputs["sp_score"] sp_scores = sp_scores.float().masked_fill(batch_to_feed["sent_offsets"].eq(0), float("-inf")).type_as(sp_scores) batch_sp_scores = sp_scores.sigmoid() # ans_type_predicted = torch.argmax(outputs["ans_type_logits"], dim=1).view(-1).tolist() outs = [outputs["start_logits"], outputs["end_logits"]] # answer prediction span_scores = outs[0][:, :, None] + outs[1][:, None] max_seq_len = span_scores.size(1) span_mask = np.tril(np.triu(np.ones((max_seq_len, max_seq_len)), 0), args.max_ans_len) span_mask = span_scores.data.new(max_seq_len, max_seq_len).copy_(torch.from_numpy(span_mask)) span_scores_masked = span_scores.float().masked_fill((1 - span_mask[None].expand_as(span_scores)).bool(), -1e10).type_as(span_scores) start_position = span_scores_masked.max(dim=2)[0].max(dim=1)[1] end_position = span_scores_masked.max(dim=2)[1].gather( 1, start_position.unsqueeze(1)).squeeze(1) answer_scores = span_scores_masked.max(dim=2)[0].max(dim=1)[0].tolist() para_offset = batch['para_offsets'] start_position_ = list( np.array(start_position.tolist()) - np.array(para_offset)) end_position_ = list( np.array(end_position.tolist()) - np.array(para_offset)) for idx, qid in enumerate(batch_qids): rank_score = scores[idx] start = start_position_[idx] end = end_position_[idx] span_score = answer_scores[idx] tok_to_orig_index = batch['tok_to_orig_index'][idx] doc_tokens = batch['doc_tokens'][idx] wp_tokens = batch['wp_tokens'][idx] orig_doc_start = tok_to_orig_index[start] orig_doc_end = tok_to_orig_index[end] orig_tokens = doc_tokens[orig_doc_start:(orig_doc_end + 1)] tok_tokens = wp_tokens[start:end+1] tok_text = " ".join(tok_tokens) tok_text = tok_text.replace(" ##", "") tok_text = tok_text.replace("##", "") tok_text = tok_text.strip() tok_text = " ".join(tok_text.split()) orig_text = " ".join(orig_tokens) pred_str = get_final_text(tok_text, orig_text, do_lower_case=True, verbose_logging=False) chain_titles = [_["title"] for _ in batch["passages"][idx]] # get the sp sentences pred_sp = [] if args.sp_pred: sp_score = batch_sp_scores[idx].tolist() passages = batch["passages"][idx] for passage, sent_offset in zip(passages, [0, len(passages[0]["sents"])]): for idx, _ in enumerate(passage["sents"]): try: if sp_score[idx + sent_offset] > 0.5: pred_sp.append([passage["title"], idx]) except: # logger.info(f"sentence exceeds max lengths") continue id2answer[qid].append({ "pred_str": pred_str.strip(), "rank_score": rank_score, "span_score": span_score, "pred_sp": pred_sp, "chain_titles": chain_titles }) lambda_ = weight results = collections.defaultdict(dict) for qid in id2answer.keys(): ans_res = id2answer[qid] ans_res.sort(key=lambda x: lambda_ * x["rank_score"] + (1 - lambda_) * x["span_score"], reverse=True) top_pred = ans_res[0]["pred_str"] top_pred_sp = ans_res[0]["pred_sp"] results["answer"][qid] = top_pred results["sp"][qid] = top_pred_sp results["titles"][qid] = ans_res[0]["chain_titles"] if args.save_prediction != "": json.dump(results, open(f"{args.save_prediction}", "w")) return results if __name__ == "__main__": main() ================================================ FILE: setup.py ================================================ #!/usr/bin/env python # Copyright (c) Facebook, Inc. and its affiliates. # All rights reserved. # # This source code is licensed under the license found in the # LICENSE file in the root directory of this source tree. from setuptools import setup, find_packages import sys import subprocess with open('README.md') as f: readme = f.read() # with open('LICENSE') as f: # license = f.read() with open('requirements.txt') as f: reqs = f.read() setup( name='mdr', version='0.0.1', description='Multi-hop dense retrieval for complex open-domain question answering', long_description='text/markdown', # license=license, python_requires='>=3.6', packages=find_packages(exclude=('data')), install_requires=reqs.strip().split('\n'), ) ================================================ FILE: setup.sh ================================================ #!/bin/bash # Copyright (c) Facebook, Inc. and its affiliates. # All rights reserved. # # This source code is licensed under the license found in the # LICENSE file in the root directory of this source tree. pip install -r requirements.txt conda install faiss-gpu cudatoolkit=10.2 -c pytorch conda install pytorch cudatoolkit=10.2 -c pytorch git clone https://github.com/NVIDIA/apex cd apex pip install -v --disable-pip-version-check --no-cache-dir ./ python setup.py develop ================================================ FILE: submitit/submit_retrieval.sh ================================================ #!/bin/bash # Copyright (c) Facebook, Inc. and its affiliates. # All rights reserved. # # This source code is licensed under the license found in the # LICENSE file in the root directory of this source tree. #!/bin/bash MKL_THREADING_LAYER=GNU python submitit_train.py --prefix mhop_retrieval_roberta \ --train_file /private/home/xwhan/data/hotpot/hotpot_train_with_neg_v0.json \ --predict_file /private/home/xwhan/data/hotpot/hotpot_val_with_neg_v0.json \ --model_name roberta-base \ --max_c_len 300 \ --max_q_len 70 \ --max_q_sp_len 350 \ --fp16 ================================================ FILE: submitit/submitit_qa.sh ================================================ #!/bin/bash # Copyright (c) Facebook, Inc. and its affiliates. # All rights reserved. # # This source code is licensed under the license found in the # LICENSE file in the root directory of this source tree. #!/bin/bash PREFIX=wwm_val_top50 MODEL_BACKEND=bert-large-uncased-whole-word-masking MKL_THREADING_LAYER=GNU python submitit_train_qa.py --prefix ${PREFIX} \ --train_file /private/home/xwhan/data/hotpot/dense_train_b100_k100_sents.json \ --predict_file /private/home/xwhan/data/hotpot/dense_val_b50_k50_roberta_sents.json \ --model_name ${MODEL_BACKEND} \ --fp16 \ --sp-pred \ ================================================ FILE: submitit/submitit_train.py ================================================ # Copyright (c) Facebook, Inc. and its affiliates. # All rights reserved. # # This source code is licensed under the license found in the # LICENSE file in the root directory of this source tree. import os import numpy as np import uuid import itertools from typing import Dict import submitit from collections import Iterable, namedtuple from pathlib import Path from datetime import date # from trainer import Trainer from mhop_trainer import Trainer from config import ClusterConfig, train_args def get_shared_folder() -> Path: return Path("/checkpoint/xwhan/mhop-dense-retrieval") def get_init_file() -> Path: # Init file must not exist, but it's parent dir must exist. os.makedirs(str(get_shared_folder()), exist_ok=True) init_file = get_shared_folder() / f"{uuid.uuid4().hex}_init" if init_file.exists(): os.remove(str(init_file)) return init_file def grid_parameters(grid: Dict): """ Yield all combinations of parameters in the grid (as a dict) """ grid_copy = dict(grid) # Turn single value in an Iterable for k in grid_copy: if not isinstance(grid_copy[k], Iterable): grid_copy[k] = [grid_copy[k]] for p in itertools.product(*grid_copy.values()): yield dict(zip(grid.keys(), p)) def grid_search(args): cluster_cfg = ClusterConfig(dist_backend="nccl", dist_url="") date_curr = date.today().strftime("%m-%d-%Y") log_dir = os.path.join(args.output_dir, date_curr) TrainerConfig = namedtuple("TrainerConfig", sorted(vars(args))) train_cfg = TrainerConfig(**vars(args)) # Create the executor print("Create the submitit Executor (can take time on FB cluster)") # Note that the folder will depend on the job_id, to easily track experiments executor = submitit.AutoExecutor(folder=get_shared_folder() / "%j") num_gpus_per_node = 8 executor.update_parameters( mem_gb=500, gpus_per_node=num_gpus_per_node, tasks_per_node=1, cpus_per_task=80, nodes=1, timeout_min=60*48, slurm_partition="learnfair", slurm_signal_delay_s=120, slurm_constraint='volta32gb' ) # Launch one job per grid position grid_meta = { "num_train_epochs": (50, lambda val: f'epoch{val}'), "learning_rate": ([2e-5, 1e-5, 3e-5], lambda val: f'lr{val}'), "seed": (16, lambda val: f'seed{val}'), "predict_batch_size": (3000, lambda val: f'evalbsize{val}'), "train_batch_size": (150, lambda val: f'trainbsize{val}'), "temperature": ([1, 0.5, 0.07], lambda val: f'tem{val}'), "warmup_ratio": ([0.1, 0.15], lambda val: f'warmup{val}'), } grid = {k:v[0] for k, v in grid_meta.items()} save_key = {k:v[1] for k, v in grid_meta.items()} hyper_parameters = list(grid_parameters(grid)) jobs = [] for i, grid_data in enumerate(hyper_parameters): cluster_cfg = cluster_cfg._replace(dist_url=get_init_file().as_uri()) train_cfg = train_cfg._replace(**grid_data) run_name = f"{train_cfg.prefix}" for k, v in grid_data.items(): run_name += "-" + save_key[k](v) train_cfg = train_cfg._replace(output_dir=os.path.join(log_dir, run_name)) # Chronos needs a different job name each time executor.update_parameters(name=f"sweep_{i:02d}_{uuid.uuid4().hex}") trainer = Trainer(train_cfg, cluster_cfg) job = executor.submit(trainer) jobs.append(job) print(f"Run {i:02d} submitted with train cfg: {train_cfg}, cluster cfg: {cluster_cfg}") print(f"Submitted jobs ids: {','.join([str(job.job_id) for job in jobs])}") # Wait for the master's results of each job results = [job.task(0).result() for job in jobs] print(f"Jobs results: {results}") best_job = np.argmax(results) print(f"Best configuration: {hyper_parameters[best_job]} (val acc = {results[best_job]:.1%})") if __name__ == "__main__": args = train_args() grid_search(args) ================================================ FILE: submitit/submitit_train_qa.py ================================================ # Copyright (c) Facebook, Inc. and its affiliates. # All rights reserved. # # This source code is licensed under the license found in the # LICENSE file in the root directory of this source tree. import os import numpy as np import uuid import itertools from typing import Dict import submitit from collections import Iterable, namedtuple from pathlib import Path from datetime import date from qa_trainer import Trainer from config import ClusterConfig, train_args def get_shared_folder() -> Path: return Path("/checkpoint/xwhan/mhop-qa") def get_init_file() -> Path: # Init file must not exist, but it's parent dir must exist. os.makedirs(str(get_shared_folder()), exist_ok=True) init_file = get_shared_folder() / f"{uuid.uuid4().hex}_init" if init_file.exists(): os.remove(str(init_file)) return init_file def grid_parameters(grid: Dict): """ Yield all combinations of parameters in the grid (as a dict) """ grid_copy = dict(grid) # Turn single value in an Iterable for k in grid_copy: if not isinstance(grid_copy[k], Iterable): grid_copy[k] = [grid_copy[k]] for p in itertools.product(*grid_copy.values()): yield dict(zip(grid.keys(), p)) def grid_search(args): cluster_cfg = ClusterConfig(dist_backend="nccl", dist_url="") date_curr = date.today().strftime("%m-%d-%Y") log_dir = os.path.join(args.output_dir, date_curr) TrainerConfig = namedtuple("TrainerConfig", sorted(vars(args))) train_cfg = TrainerConfig(**vars(args)) # Create the executor print("Create the submitit Executor (can take time on FB cluster)") # Note that the folder will depend on the job_id, to easily track experiments executor = submitit.AutoExecutor(folder=get_shared_folder() / "%j") num_gpus_per_node = 8 executor.update_parameters( mem_gb=400, gpus_per_node=num_gpus_per_node, tasks_per_node=1, # one task per GPU cpus_per_task=10, nodes=1, timeout_min=60*72, slurm_partition="learnfair", slurm_signal_delay_s=120, slurm_constraint='volta32gb' ) # Launch one job per grid position grid_meta = { "num_train_epochs": (7, lambda val: f'epoch{val}'), "learning_rate": ([2e-5, 5e-5, 3e-5], lambda val: f'lr{val}'), "seed": ([42,5], lambda val: f'seed{val}'), "rank_drop": (0, lambda val: f'rdrop{val}'), "qa_drop": (0, lambda val: f'qadrop{val}'), # "max_seq_len": (512, lambda val: f'c_len{val}'), # "max_q_len": (100, lambda val: f'q_len{val}'), "weight_decay": (0, lambda val: f'decay{val}'), "num_q_per_gpu": (2, lambda val: f'qpergpu{val}'), # how many questions per gpu "gradient_accumulation_steps": (8, lambda val: f'aggstep{val}'), "max_grad_norm": (2, lambda val: f'clip{val}'), "eval_period": (250, lambda val: f'evalper{val}'), "predict_batch_size": (1024, lambda val: f'evalbsize{val}'), "neg_num": (5, lambda val: f'negnum{val}'), "warmup_ratio": ([0.1, 0.2], lambda val: f'warmup{val}'), "use_adam": (True, lambda val: f'adam{val}'), "sp_weight": ([0.05, 0.025], lambda val: f'spweight{val}'), "shared_norm": (False, lambda val: f'sn{val}'), } grid = {k:v[0] for k, v in grid_meta.items()} save_key = {k:v[1] for k, v in grid_meta.items()} hyper_parameters = list(grid_parameters(grid)) jobs = [] for i, grid_data in enumerate(hyper_parameters): cluster_cfg = cluster_cfg._replace(dist_url=get_init_file().as_uri()) train_cfg = train_cfg._replace(**grid_data) run_name = f"{train_cfg.prefix}" for k, v in grid_data.items(): run_name += "-" + save_key[k](v) train_cfg = train_cfg._replace(output_dir=os.path.join(log_dir, run_name)) # Chronos needs a different job name each time executor.update_parameters(name=f"sweep_{i:02d}_{uuid.uuid4().hex}") trainer = Trainer(train_cfg, cluster_cfg) job = executor.submit(trainer) jobs.append(job) print(f"Run {i:02d} submitted with train cfg: {train_cfg}, cluster cfg: {cluster_cfg}") print(f"Submitted jobs ids: {','.join([str(job.job_id) for job in jobs])}") # Wait for the master's results of each job results = [job.task(0).result() for job in jobs] print(f"Jobs results: {results}") best_job = np.argmax(results) print(f"Best configuration: {hyper_parameters[best_job]} (val acc = {results[best_job]:.1%})") if __name__ == "__main__": args = train_args() grid_search(args)