├── .gitignore ├── CONTRIBUTING.md ├── LICENSE ├── README.md ├── __init__.py ├── cases └── test.zh.md ├── create_pretraining_data.py ├── data └── lm │ ├── result.zh.json │ ├── test.en.tsv │ └── test.zh.tsv ├── extract_features.py ├── modeling.py ├── modeling_test.py ├── multilingual.md ├── optimization.py ├── optimization_test.py ├── requirements.txt ├── run_classifier.py ├── run_lm_predict.py ├── run_pretraining.py ├── run_squad.py ├── sample_text.txt ├── tokenization.py └── tokenization_test.py /.gitignore: -------------------------------------------------------------------------------- 1 | # Initially taken from Github's Python gitignore file 2 | .idea 3 | 4 | # Byte-compiled / optimized / DLL files 5 | __pycache__/ 6 | *.py[cod] 7 | *$py.class 8 | 9 | # C extensions 10 | *.so 11 | 12 | # Distribution / packaging 13 | .Python 14 | build/ 15 | develop-eggs/ 16 | dist/ 17 | downloads/ 18 | eggs/ 19 | .eggs/ 20 | lib/ 21 | lib64/ 22 | parts/ 23 | sdist/ 24 | var/ 25 | wheels/ 26 | *.egg-info/ 27 | .installed.cfg 28 | *.egg 29 | MANIFEST 30 | 31 | # PyInstaller 32 | # Usually these files are written by a python script from a template 33 | # before PyInstaller builds the exe, so as to inject date/other infos into it. 34 | *.manifest 35 | *.spec 36 | 37 | # Installer logs 38 | pip-log.txt 39 | pip-delete-this-directory.txt 40 | 41 | # Unit test / coverage reports 42 | htmlcov/ 43 | .tox/ 44 | .nox/ 45 | .coverage 46 | .coverage.* 47 | .cache 48 | nosetests.xml 49 | coverage.xml 50 | *.cover 51 | .hypothesis/ 52 | .pytest_cache/ 53 | 54 | # Translations 55 | *.mo 56 | *.pot 57 | 58 | # Django stuff: 59 | *.log 60 | local_settings.py 61 | db.sqlite3 62 | 63 | # Flask stuff: 64 | instance/ 65 | .webassets-cache 66 | 67 | # Scrapy stuff: 68 | .scrapy 69 | 70 | # Sphinx documentation 71 | docs/_build/ 72 | 73 | # PyBuilder 74 | target/ 75 | 76 | # Jupyter Notebook 77 | .ipynb_checkpoints 78 | 79 | # IPython 80 | profile_default/ 81 | ipython_config.py 82 | 83 | # pyenv 84 | .python-version 85 | 86 | # celery beat schedule file 87 | celerybeat-schedule 88 | 89 | # SageMath parsed files 90 | *.sage.py 91 | 92 | # Environments 93 | .env 94 | .venv 95 | env/ 96 | venv/ 97 | ENV/ 98 | env.bak/ 99 | venv.bak/ 100 | 101 | # Spyder project settings 102 | .spyderproject 103 | .spyproject 104 | 105 | # Rope project settings 106 | .ropeproject 107 | 108 | # mkdocs documentation 109 | /site 110 | 111 | # mypy 112 | .mypy_cache/ 113 | .dmypy.json 114 | dmypy.json 115 | 116 | # Pyre type checker 117 | .pyre/ 118 | -------------------------------------------------------------------------------- /CONTRIBUTING.md: -------------------------------------------------------------------------------- 1 | # How to Contribute 2 | 3 | BERT needs to maintain permanent compatibility with the pre-trained model files, 4 | so we do not plan to make any major changes to this library (other than what was 5 | promised in the README). However, we can accept small patches related to 6 | re-factoring and documentation. To submit contributes, there are just a few 7 | small guidelines you need to follow. 8 | 9 | ## Contributor License Agreement 10 | 11 | Contributions to this project must be accompanied by a Contributor License 12 | Agreement. You (or your employer) retain the copyright to your contribution; 13 | this simply gives us permission to use and redistribute your contributions as 14 | part of the project. Head over to to see 15 | your current agreements on file or to sign a new one. 16 | 17 | You generally only need to submit a CLA once, so if you've already submitted one 18 | (even if it was for a different project), you probably don't need to do it 19 | again. 20 | 21 | ## Code reviews 22 | 23 | All submissions, including submissions by project members, require review. We 24 | use GitHub pull requests for this purpose. Consult 25 | [GitHub Help](https://help.github.com/articles/about-pull-requests/) for more 26 | information on using pull requests. 27 | 28 | ## Community Guidelines 29 | 30 | This project follows 31 | [Google's Open Source Community Guidelines](https://opensource.google.com/conduct/). 32 | -------------------------------------------------------------------------------- /LICENSE: -------------------------------------------------------------------------------- 1 | 2 | Apache License 3 | Version 2.0, January 2004 4 | http://www.apache.org/licenses/ 5 | 6 | TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION 7 | 8 | 1. Definitions. 9 | 10 | "License" shall mean the terms and conditions for use, reproduction, 11 | and distribution as defined by Sections 1 through 9 of this document. 12 | 13 | "Licensor" shall mean the copyright owner or entity authorized by 14 | the copyright owner that is granting the License. 15 | 16 | "Legal Entity" shall mean the union of the acting entity and all 17 | other entities that control, are controlled by, or are under common 18 | control with that entity. For the purposes of this definition, 19 | "control" means (i) the power, direct or indirect, to cause the 20 | direction or management of such entity, whether by contract or 21 | otherwise, or (ii) ownership of fifty percent (50%) or more of the 22 | outstanding shares, or (iii) beneficial ownership of such entity. 23 | 24 | "You" (or "Your") shall mean an individual or Legal Entity 25 | exercising permissions granted by this License. 26 | 27 | "Source" form shall mean the preferred form for making modifications, 28 | including but not limited to software source code, documentation 29 | source, and configuration files. 30 | 31 | "Object" form shall mean any form resulting from mechanical 32 | transformation or translation of a Source form, including but 33 | not limited to compiled object code, generated documentation, 34 | and conversions to other media types. 35 | 36 | "Work" shall mean the work of authorship, whether in Source or 37 | Object form, made available under the License, as indicated by a 38 | copyright notice that is included in or attached to the work 39 | (an example is provided in the Appendix below). 40 | 41 | "Derivative Works" shall mean any work, whether in Source or Object 42 | form, that is based on (or derived from) the Work and for which the 43 | editorial revisions, annotations, elaborations, or other modifications 44 | represent, as a whole, an original work of authorship. For the purposes 45 | of this License, Derivative Works shall not include works that remain 46 | separable from, or merely link (or bind by name) to the interfaces of, 47 | the Work and Derivative Works thereof. 48 | 49 | "Contribution" shall mean any work of authorship, including 50 | the original version of the Work and any modifications or additions 51 | to that Work or Derivative Works thereof, that is intentionally 52 | submitted to Licensor for inclusion in the Work by the copyright owner 53 | or by an individual or Legal Entity authorized to submit on behalf of 54 | the copyright owner. For the purposes of this definition, "submitted" 55 | means any form of electronic, verbal, or written communication sent 56 | to the Licensor or its representatives, including but not limited to 57 | communication on electronic mailing lists, source code control systems, 58 | and issue tracking systems that are managed by, or on behalf of, the 59 | Licensor for the purpose of discussing and improving the Work, but 60 | excluding communication that is conspicuously marked or otherwise 61 | designated in writing by the copyright owner as "Not a Contribution." 62 | 63 | "Contributor" shall mean Licensor and any individual or Legal Entity 64 | on behalf of whom a Contribution has been received by Licensor and 65 | subsequently incorporated within the Work. 66 | 67 | 2. Grant of Copyright License. Subject to the terms and conditions of 68 | this License, each Contributor hereby grants to You a perpetual, 69 | worldwide, non-exclusive, no-charge, royalty-free, irrevocable 70 | copyright license to reproduce, prepare Derivative Works of, 71 | publicly display, publicly perform, sublicense, and distribute the 72 | Work and such Derivative Works in Source or Object form. 73 | 74 | 3. Grant of Patent License. Subject to the terms and conditions of 75 | this License, each Contributor hereby grants to You a perpetual, 76 | worldwide, non-exclusive, no-charge, royalty-free, irrevocable 77 | (except as stated in this section) patent license to make, have made, 78 | use, offer to sell, sell, import, and otherwise transfer the Work, 79 | where such license applies only to those patent claims licensable 80 | by such Contributor that are necessarily infringed by their 81 | Contribution(s) alone or by combination of their Contribution(s) 82 | with the Work to which such Contribution(s) was submitted. If You 83 | institute patent litigation against any entity (including a 84 | cross-claim or counterclaim in a lawsuit) alleging that the Work 85 | or a Contribution incorporated within the Work constitutes direct 86 | or contributory patent infringement, then any patent licenses 87 | granted to You under this License for that Work shall terminate 88 | as of the date such litigation is filed. 89 | 90 | 4. Redistribution. You may reproduce and distribute copies of the 91 | Work or Derivative Works thereof in any medium, with or without 92 | modifications, and in Source or Object form, provided that You 93 | meet the following conditions: 94 | 95 | (a) You must give any other recipients of the Work or 96 | Derivative Works a copy of this License; and 97 | 98 | (b) You must cause any modified files to carry prominent notices 99 | stating that You changed the files; and 100 | 101 | (c) You must retain, in the Source form of any Derivative Works 102 | that You distribute, all copyright, patent, trademark, and 103 | attribution notices from the Source form of the Work, 104 | excluding those notices that do not pertain to any part of 105 | the Derivative Works; and 106 | 107 | (d) If the Work includes a "NOTICE" text file as part of its 108 | distribution, then any Derivative Works that You distribute must 109 | include a readable copy of the attribution notices contained 110 | within such NOTICE file, excluding those notices that do not 111 | pertain to any part of the Derivative Works, in at least one 112 | of the following places: within a NOTICE text file distributed 113 | as part of the Derivative Works; within the Source form or 114 | documentation, if provided along with the Derivative Works; or, 115 | within a display generated by the Derivative Works, if and 116 | wherever such third-party notices normally appear. The contents 117 | of the NOTICE file are for informational purposes only and 118 | do not modify the License. You may add Your own attribution 119 | notices within Derivative Works that You distribute, alongside 120 | or as an addendum to the NOTICE text from the Work, provided 121 | that such additional attribution notices cannot be construed 122 | as modifying the License. 123 | 124 | You may add Your own copyright statement to Your modifications and 125 | may provide additional or different license terms and conditions 126 | for use, reproduction, or distribution of Your modifications, or 127 | for any such Derivative Works as a whole, provided Your use, 128 | reproduction, and distribution of the Work otherwise complies with 129 | the conditions stated in this License. 130 | 131 | 5. Submission of Contributions. Unless You explicitly state otherwise, 132 | any Contribution intentionally submitted for inclusion in the Work 133 | by You to the Licensor shall be under the terms and conditions of 134 | this License, without any additional terms or conditions. 135 | Notwithstanding the above, nothing herein shall supersede or modify 136 | the terms of any separate license agreement you may have executed 137 | with Licensor regarding such Contributions. 138 | 139 | 6. Trademarks. This License does not grant permission to use the trade 140 | names, trademarks, service marks, or product names of the Licensor, 141 | except as required for reasonable and customary use in describing the 142 | origin of the Work and reproducing the content of the NOTICE file. 143 | 144 | 7. Disclaimer of Warranty. Unless required by applicable law or 145 | agreed to in writing, Licensor provides the Work (and each 146 | Contributor provides its Contributions) on an "AS IS" BASIS, 147 | WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or 148 | implied, including, without limitation, any warranties or conditions 149 | of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A 150 | PARTICULAR PURPOSE. You are solely responsible for determining the 151 | appropriateness of using or redistributing the Work and assume any 152 | risks associated with Your exercise of permissions under this License. 153 | 154 | 8. Limitation of Liability. In no event and under no legal theory, 155 | whether in tort (including negligence), contract, or otherwise, 156 | unless required by applicable law (such as deliberate and grossly 157 | negligent acts) or agreed to in writing, shall any Contributor be 158 | liable to You for damages, including any direct, indirect, special, 159 | incidental, or consequential damages of any character arising as a 160 | result of this License or out of the use or inability to use the 161 | Work (including but not limited to damages for loss of goodwill, 162 | work stoppage, computer failure or malfunction, or any and all 163 | other commercial damages or losses), even if such Contributor 164 | has been advised of the possibility of such damages. 165 | 166 | 9. Accepting Warranty or Additional Liability. While redistributing 167 | the Work or Derivative Works thereof, You may choose to offer, 168 | and charge a fee for, acceptance of support, warranty, indemnity, 169 | or other liability obligations and/or rights consistent with this 170 | License. However, in accepting such obligations, You may act only 171 | on Your own behalf and on Your sole responsibility, not on behalf 172 | of any other Contributor, and only if You agree to indemnify, 173 | defend, and hold each Contributor harmless for any liability 174 | incurred by, or claims asserted against, such Contributor by reason 175 | of your accepting any such warranty or additional liability. 176 | 177 | END OF TERMS AND CONDITIONS 178 | 179 | APPENDIX: How to apply the Apache License to your work. 180 | 181 | To apply the Apache License to your work, attach the following 182 | boilerplate notice, with the fields enclosed by brackets "[]" 183 | replaced with your own identifying information. (Don't include 184 | the brackets!) The text should be enclosed in the appropriate 185 | comment syntax for the file format. We also recommend that a 186 | file or class name and description of purpose be included on the 187 | same "printed page" as the copyright notice for easier 188 | identification within third-party archives. 189 | 190 | Copyright [yyyy] [name of copyright owner] 191 | 192 | Licensed under the Apache License, Version 2.0 (the "License"); 193 | you may not use this file except in compliance with the License. 194 | You may obtain a copy of the License at 195 | 196 | http://www.apache.org/licenses/LICENSE-2.0 197 | 198 | Unless required by applicable law or agreed to in writing, software 199 | distributed under the License is distributed on an "AS IS" BASIS, 200 | WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 201 | See the License for the specific language governing permissions and 202 | limitations under the License. 203 | -------------------------------------------------------------------------------- /README.md: -------------------------------------------------------------------------------- 1 | **[🤗Demo](#demo)** | 2 | **[📖cases-en](#test-case)** | 3 | **[📖cases-zh](cases/test.zh.md)** | 4 | 5 | 6 | 7 | 8 | ## BERT as Language Model 9 | 10 | For a sentence S = w_1, w_2,..., w_k , we have 11 | 12 | p(S) = \prod_{i=1}^{k} p(w_i | context) 13 | 14 | 15 | In traditional language model, such as RNN, context = w_1, ..., w_{i-1} , 16 | 17 | p(S) = \prod_{i=1}^{k} p(w_i | w_1, ..., w_{i-1}) 18 | 19 | 20 | In bidirectional language model, it has larger context, context = w_1, ..., w_{i-1},w_{i+1},...,w_k. 21 | 22 | In this implementation, we simply adopt the following approximation, 23 | 24 | p(S) \approx \prod_{i=1}^{k} p(w_i | w_1, ..., w_{i-1},w_{i+1}, ...,w_k). 25 | 26 | 27 | 31 | 32 | 37 | 38 | ### Demo 39 | 40 | Try out the [Web Demo](https://huggingface.co/spaces/eson/bert-perplexity) at [![Hugging Face Spaces](https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-Spaces-blue)](https://huggingface.co/spaces/eson/bert-perplexity) 41 | 42 | ### test-case 43 | 44 | > [more cases: 中文](cases/test.zh.md) 45 | 46 | 47 | ```bash 48 | export BERT_BASE_DIR=model/uncased_L-12_H-768_A-12 49 | export INPUT_FILE=data/lm/test.en.tsv 50 | python run_lm_predict.py \ 51 | --input_file=$INPUT_FILE \ 52 | --vocab_file=$BERT_BASE_DIR/vocab.txt \ 53 | --bert_config_file=$BERT_BASE_DIR/bert_config.json \ 54 | --init_checkpoint=$BERT_BASE_DIR/bert_model.ckpt \ 55 | --max_seq_length=128 \ 56 | --output_dir=/tmp/lm_output/ 57 | ``` 58 | 59 | for the following test case 60 | 61 | ```bash 62 | $ cat data/lm/test.en.tsv 63 | there is a book on the desk 64 | there is a plane on the desk 65 | there is a book in the desk 66 | 67 | $ cat /tmp/lm/output/test_result.json 68 | ``` 69 | output: 70 | 71 | ```yml 72 | # prob: probability 73 | # ppl: perplexity 74 | [ 75 | { 76 | "tokens": [ 77 | { 78 | "token": "there", 79 | "prob": 0.9988962411880493 80 | }, 81 | { 82 | "token": "is", 83 | "prob": 0.013578361831605434 84 | }, 85 | { 86 | "token": "a", 87 | "prob": 0.9420605897903442 88 | }, 89 | { 90 | "token": "book", 91 | "prob": 0.07452250272035599 92 | }, 93 | { 94 | "token": "on", 95 | "prob": 0.9607976675033569 96 | }, 97 | { 98 | "token": "the", 99 | "prob": 0.4983428418636322 100 | }, 101 | { 102 | "token": "desk", 103 | "prob": 4.040586190967588e-06 104 | } 105 | ], 106 | "ppl": 17.69329728285426 107 | }, 108 | { 109 | "tokens": [ 110 | { 111 | "token": "there", 112 | "prob": 0.996775209903717 113 | }, 114 | { 115 | "token": "is", 116 | "prob": 0.03194097802042961 117 | }, 118 | { 119 | "token": "a", 120 | "prob": 0.8877727389335632 121 | }, 122 | { 123 | "token": "plane", 124 | "prob": 3.4907534427475184e-05 # low probability 125 | }, 126 | { 127 | "token": "on", 128 | "prob": 0.1902322769165039 129 | }, 130 | { 131 | "token": "the", 132 | "prob": 0.5981084704399109 133 | }, 134 | { 135 | "token": "desk", 136 | "prob": 3.3164762953674654e-06 137 | } 138 | ], 139 | "ppl": 59.646456254851806 140 | }, 141 | { 142 | "tokens": [ 143 | { 144 | "token": "there", 145 | "prob": 0.9969795942306519 146 | }, 147 | { 148 | "token": "is", 149 | "prob": 0.03379646688699722 150 | }, 151 | { 152 | "token": "a", 153 | "prob": 0.9095568060874939 154 | }, 155 | { 156 | "token": "book", 157 | "prob": 0.013939591124653816 158 | }, 159 | { 160 | "token": "in", 161 | "prob": 0.000823647016659379 # low probability 162 | }, 163 | { 164 | "token": "the", 165 | "prob": 0.5844194293022156 166 | }, 167 | { 168 | "token": "desk", 169 | "prob": 3.3361218356731115e-06 170 | } 171 | ], 172 | "ppl": 54.65941516205144 173 | } 174 | ] 175 | ``` 176 | 177 | 178 | 179 | -------------------------------------------------------------------------------- /__init__.py: -------------------------------------------------------------------------------- 1 | # coding=utf-8 2 | # Copyright 2018 The Google AI Language Team Authors. 3 | # 4 | # Licensed under the Apache License, Version 2.0 (the "License"); 5 | # you may not use this file except in compliance with the License. 6 | # You may obtain a copy of the License at 7 | # 8 | # http://www.apache.org/licenses/LICENSE-2.0 9 | # 10 | # Unless required by applicable law or agreed to in writing, software 11 | # distributed under the License is distributed on an "AS IS" BASIS, 12 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 13 | # See the License for the specific language governing permissions and 14 | # limitations under the License. 15 | 16 | -------------------------------------------------------------------------------- /cases/test.zh.md: -------------------------------------------------------------------------------- 1 | 2 | # 总体评价 3 | 4 | 5 | 6 | 7 | 8 | **优势** 9 | 1. bert给句子打分，摆脱了传统auto regressive的局限，可并行。 10 | 1. 得益于双向语言模型的全局感受野，bert给word打分准确度较高 11 | 12 | **缺陷** 13 | 1. 给每个word打分，都要跑一遍inference，计算量较大，且冗余。有优化的空间 14 | 1. 该实现中采用的句子概率是近似概率，不够严谨 15 | 16 | 17 | 另外 18 | 1. char-level的语言模型，由于词组内的高概率，会使整个句子ppl普遍偏高。 19 | 1. 句子间的相对ppl还靠谱。 20 | 21 | > **建议**: 22 | 用分词后的中文重新pretrain，然后进行word-level language model predict。 23 | 24 | 25 | # 中文测试 26 | 27 | ```bash 28 | export BERT_BASE_DIR=model/chinese_L-12_H-768_A-12 29 | export INPUT_FILE=data/lm/test.zh.tsv 30 | python run_lm_predict.py \ 31 | --input_file=$INPUT_FILE \ 32 | --vocab_file=$BERT_BASE_DIR/vocab.txt \ 33 | --bert_config_file=$BERT_BASE_DIR/bert_config.json \ 34 | --init_checkpoint=$BERT_BASE_DIR/bert_model.ckpt \ 35 | --max_seq_length=128 \ 36 | --output_dir=/tmp/lm_output/ 37 | ``` 38 | 39 | 以下是部分结果，更多见[result.zh.json](/data/lm/result.zh.json) 40 | 41 | ```yml 42 | [ 43 | { 44 | "tokens": [ 45 | { 46 | "token": "2016", 47 | "prob": 0.06563900411128998 48 | }, 49 | { 50 | "token": "全", 51 | "prob": 0.4981258511543274 52 | }, 53 | { 54 | "token": "国", 55 | "prob": 0.9088247418403625 56 | }, 57 | { 58 | "token": "低", 59 | "prob": 1.6259804397122934e-05 # 低概率 60 | }, 61 | { 62 | "token": "考", 63 | "prob": 0.4023572504520416 64 | }, 65 | ... 66 | ], 67 | "ppl": 13.400421357093588 68 | }, 69 | { 70 | "tokens": [ 71 | { 72 | "token": "落", 73 | "prob": 0.1483132392168045 74 | }, 75 | { 76 | "token": "霞", 77 | "prob": 0.42232587933540344 78 | }, 79 | { 80 | "token": "与", 81 | "prob": 0.8615185022354126 82 | }, 83 | { 84 | "token": "孤", 85 | "prob": 0.9975666999816895 86 | }, 87 | { 88 | "token": "鹜", 89 | "prob": 0.5613960027694702 90 | }, 91 | { 92 | "token": "齐", 93 | "prob": 0.18012434244155884 94 | }, 95 | { 96 | "token": "跑", 97 | "prob": 1.3388593288254924e-05 # 低概率 98 | }, 99 | ... 100 | ], 101 | "ppl": 11.983086642867598 102 | }, 103 | ``` 104 | 105 | 106 | 中文测试样例来源于[百度云dnnlm](https://cloud.baidu.com/product/nlp/dnnlm_cn) 107 | 108 | 124 | -------------------------------------------------------------------------------- /create_pretraining_data.py: -------------------------------------------------------------------------------- 1 | # coding=utf-8 2 | # Copyright 2018 The Google AI Language Team Authors. 3 | # 4 | # Licensed under the Apache License, Version 2.0 (the "License"); 5 | # you may not use this file except in compliance with the License. 6 | # You may obtain a copy of the License at 7 | # 8 | # http://www.apache.org/licenses/LICENSE-2.0 9 | # 10 | # Unless required by applicable law or agreed to in writing, software 11 | # distributed under the License is distributed on an "AS IS" BASIS, 12 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 13 | # See the License for the specific language governing permissions and 14 | # limitations under the License. 15 | """Create masked LM/next sentence masked_lm TF examples for BERT.""" 16 | 17 | from __future__ import absolute_import 18 | from __future__ import division 19 | from __future__ import print_function 20 | 21 | import collections 22 | import random 23 | 24 | import tokenization 25 | import tensorflow as tf 26 | 27 | flags = tf.flags 28 | 29 | FLAGS = flags.FLAGS 30 | 31 | flags.DEFINE_string("input_file", None, 32 | "Input raw text file (or comma-separated list of files).") 33 | 34 | flags.DEFINE_string( 35 | "output_file", None, 36 | "Output TF example file (or comma-separated list of files).") 37 | 38 | flags.DEFINE_string("vocab_file", None, 39 | "The vocabulary file that the BERT model was trained on.") 40 | 41 | flags.DEFINE_bool( 42 | "do_lower_case", True, 43 | "Whether to lower case the input text. Should be True for uncased " 44 | "models and False for cased models.") 45 | 46 | flags.DEFINE_integer("max_seq_length", 128, "Maximum sequence length.") 47 | 48 | flags.DEFINE_integer("max_predictions_per_seq", 20, 49 | "Maximum number of masked LM predictions per sequence.") 50 | 51 | flags.DEFINE_integer("random_seed", 12345, "Random seed for data generation.") 52 | 53 | flags.DEFINE_integer( 54 | "dupe_factor", 10, 55 | "Number of times to duplicate the input data (with different masks).") 56 | 57 | flags.DEFINE_float("masked_lm_prob", 0.15, "Masked LM probability.") 58 | 59 | flags.DEFINE_float( 60 | "short_seq_prob", 0.1, 61 | "Probability of creating sequences which are shorter than the " 62 | "maximum length.") 63 | 64 | 65 | class TrainingInstance(object): 66 | """A single training instance (sentence pair).""" 67 | 68 | def __init__(self, tokens, segment_ids, masked_lm_positions, masked_lm_labels, 69 | is_random_next): 70 | self.tokens = tokens 71 | self.segment_ids = segment_ids 72 | self.is_random_next = is_random_next 73 | self.masked_lm_positions = masked_lm_positions 74 | self.masked_lm_labels = masked_lm_labels 75 | 76 | def __str__(self): 77 | s = "" 78 | s += "tokens: %s\n" % (" ".join( 79 | [tokenization.printable_text(x) for x in self.tokens])) 80 | s += "segment_ids: %s\n" % (" ".join([str(x) for x in self.segment_ids])) 81 | s += "is_random_next: %s\n" % self.is_random_next 82 | s += "masked_lm_positions: %s\n" % (" ".join( 83 | [str(x) for x in self.masked_lm_positions])) 84 | s += "masked_lm_labels: %s\n" % (" ".join( 85 | [tokenization.printable_text(x) for x in self.masked_lm_labels])) 86 | s += "\n" 87 | return s 88 | 89 | def __repr__(self): 90 | return self.__str__() 91 | 92 | 93 | def write_instance_to_example_files(instances, tokenizer, max_seq_length, 94 | max_predictions_per_seq, output_files): 95 | """Create TF example files from `TrainingInstance`s.""" 96 | writers = [] 97 | for output_file in output_files: 98 | writers.append(tf.python_io.TFRecordWriter(output_file)) 99 | 100 | writer_index = 0 101 | 102 | total_written = 0 103 | for (inst_index, instance) in enumerate(instances): 104 | input_ids = tokenizer.convert_tokens_to_ids(instance.tokens) 105 | input_mask = [1] * len(input_ids) 106 | segment_ids = list(instance.segment_ids) 107 | assert len(input_ids) <= max_seq_length 108 | 109 | while len(input_ids) < max_seq_length: 110 | input_ids.append(0) 111 | input_mask.append(0) 112 | segment_ids.append(0) 113 | 114 | assert len(input_ids) == max_seq_length 115 | assert len(input_mask) == max_seq_length 116 | assert len(segment_ids) == max_seq_length 117 | 118 | masked_lm_positions = list(instance.masked_lm_positions) 119 | masked_lm_ids = tokenizer.convert_tokens_to_ids(instance.masked_lm_labels) 120 | masked_lm_weights = [1.0] * len(masked_lm_ids) 121 | 122 | while len(masked_lm_positions) < max_predictions_per_seq: 123 | masked_lm_positions.append(0) 124 | masked_lm_ids.append(0) 125 | masked_lm_weights.append(0.0) 126 | 127 | next_sentence_label = 1 if instance.is_random_next else 0 128 | 129 | features = collections.OrderedDict() 130 | features["input_ids"] = create_int_feature(input_ids) 131 | features["input_mask"] = create_int_feature(input_mask) 132 | features["segment_ids"] = create_int_feature(segment_ids) 133 | features["masked_lm_positions"] = create_int_feature(masked_lm_positions) 134 | features["masked_lm_ids"] = create_int_feature(masked_lm_ids) 135 | features["masked_lm_weights"] = create_float_feature(masked_lm_weights) 136 | features["next_sentence_labels"] = create_int_feature([next_sentence_label]) 137 | 138 | tf_example = tf.train.Example(features=tf.train.Features(feature=features)) 139 | 140 | writers[writer_index].write(tf_example.SerializeToString()) 141 | writer_index = (writer_index + 1) % len(writers) 142 | 143 | total_written += 1 144 | 145 | if inst_index < 20: 146 | tf.logging.info("*** Example ***") 147 | tf.logging.info("tokens: %s" % " ".join( 148 | [tokenization.printable_text(x) for x in instance.tokens])) 149 | 150 | for feature_name in features.keys(): 151 | feature = features[feature_name] 152 | values = [] 153 | if feature.int64_list.value: 154 | values = feature.int64_list.value 155 | elif feature.float_list.value: 156 | values = feature.float_list.value 157 | tf.logging.info( 158 | "%s: %s" % (feature_name, " ".join([str(x) for x in values]))) 159 | 160 | for writer in writers: 161 | writer.close() 162 | 163 | tf.logging.info("Wrote %d total instances", total_written) 164 | 165 | 166 | def create_int_feature(values): 167 | feature = tf.train.Feature(int64_list=tf.train.Int64List(value=list(values))) 168 | return feature 169 | 170 | 171 | def create_float_feature(values): 172 | feature = tf.train.Feature(float_list=tf.train.FloatList(value=list(values))) 173 | return feature 174 | 175 | 176 | def create_training_instances(input_files, tokenizer, max_seq_length, 177 | dupe_factor, short_seq_prob, masked_lm_prob, 178 | max_predictions_per_seq, rng): 179 | """Create `TrainingInstance`s from raw text.""" 180 | all_documents = [[]] 181 | 182 | # Input file format: 183 | # (1) One sentence per line. These should ideally be actual sentences, not 184 | # entire paragraphs or arbitrary spans of text. (Because we use the 185 | # sentence boundaries for the "next sentence prediction" task). 186 | # (2) Blank lines between documents. Document boundaries are needed so 187 | # that the "next sentence prediction" task doesn't span between documents. 188 | for input_file in input_files: 189 | with tf.gfile.GFile(input_file, "r") as reader: 190 | while True: 191 | line = tokenization.convert_to_unicode(reader.readline()) 192 | if not line: 193 | break 194 | line = line.strip() 195 | 196 | # Empty lines are used as document delimiters 197 | if not line: 198 | all_documents.append([]) 199 | tokens = tokenizer.tokenize(line) 200 | if tokens: 201 | all_documents[-1].append(tokens) 202 | 203 | # Remove empty documents 204 | all_documents = [x for x in all_documents if x] 205 | rng.shuffle(all_documents) 206 | 207 | vocab_words = list(tokenizer.vocab.keys()) 208 | instances = [] 209 | for _ in range(dupe_factor): 210 | for document_index in range(len(all_documents)): 211 | instances.extend( 212 | create_instances_from_document( 213 | all_documents, document_index, max_seq_length, short_seq_prob, 214 | masked_lm_prob, max_predictions_per_seq, vocab_words, rng)) 215 | 216 | rng.shuffle(instances) 217 | return instances 218 | 219 | 220 | def create_instances_from_document( 221 | all_documents, document_index, max_seq_length, short_seq_prob, 222 | masked_lm_prob, max_predictions_per_seq, vocab_words, rng): 223 | """Creates `TrainingInstance`s for a single document.""" 224 | document = all_documents[document_index] 225 | 226 | # Account for [CLS], [SEP], [SEP] 227 | max_num_tokens = max_seq_length - 3 228 | 229 | # We *usually* want to fill up the entire sequence since we are padding 230 | # to `max_seq_length` anyways, so short sequences are generally wasted 231 | # computation. However, we *sometimes* 232 | # (i.e., short_seq_prob == 0.1 == 10% of the time) want to use shorter 233 | # sequences to minimize the mismatch between pre-training and fine-tuning. 234 | # The `target_seq_length` is just a rough target however, whereas 235 | # `max_seq_length` is a hard limit. 236 | target_seq_length = max_num_tokens 237 | if rng.random() < short_seq_prob: 238 | target_seq_length = rng.randint(2, max_num_tokens) 239 | 240 | # We DON'T just concatenate all of the tokens from a document into a long 241 | # sequence and choose an arbitrary split point because this would make the 242 | # next sentence prediction task too easy. Instead, we split the input into 243 | # segments "A" and "B" based on the actual "sentences" provided by the user 244 | # input. 245 | instances = [] 246 | current_chunk = [] 247 | current_length = 0 248 | i = 0 249 | while i < len(document): 250 | segment = document[i] 251 | current_chunk.append(segment) 252 | current_length += len(segment) 253 | if i == len(document) - 1 or current_length >= target_seq_length: 254 | if current_chunk: 255 | # `a_end` is how many segments from `current_chunk` go into the `A` 256 | # (first) sentence. 257 | a_end = 1 258 | if len(current_chunk) >= 2: 259 | a_end = rng.randint(1, len(current_chunk) - 1) 260 | 261 | tokens_a = [] 262 | for j in range(a_end): 263 | tokens_a.extend(current_chunk[j]) 264 | 265 | tokens_b = [] 266 | # Random next 267 | is_random_next = False 268 | if len(current_chunk) == 1 or rng.random() < 0.5: 269 | is_random_next = True 270 | target_b_length = target_seq_length - len(tokens_a) 271 | 272 | # This should rarely go for more than one iteration for large 273 | # corpora. However, just to be careful, we try to make sure that 274 | # the random document is not the same as the document 275 | # we're processing. 276 | for _ in range(10): 277 | random_document_index = rng.randint(0, len(all_documents) - 1) 278 | if random_document_index != document_index: 279 | break 280 | 281 | random_document = all_documents[random_document_index] 282 | random_start = rng.randint(0, len(random_document) - 1) 283 | for j in range(random_start, len(random_document)): 284 | tokens_b.extend(random_document[j]) 285 | if len(tokens_b) >= target_b_length: 286 | break 287 | # We didn't actually use these segments so we "put them back" so 288 | # they don't go to waste. 289 | num_unused_segments = len(current_chunk) - a_end 290 | i -= num_unused_segments 291 | # Actual next 292 | else: 293 | is_random_next = False 294 | for j in range(a_end, len(current_chunk)): 295 | tokens_b.extend(current_chunk[j]) 296 | truncate_seq_pair(tokens_a, tokens_b, max_num_tokens, rng) 297 | 298 | assert len(tokens_a) >= 1 299 | assert len(tokens_b) >= 1 300 | 301 | tokens = [] 302 | segment_ids = [] 303 | tokens.append("[CLS]") 304 | segment_ids.append(0) 305 | for token in tokens_a: 306 | tokens.append(token) 307 | segment_ids.append(0) 308 | 309 | tokens.append("[SEP]") 310 | segment_ids.append(0) 311 | 312 | for token in tokens_b: 313 | tokens.append(token) 314 | segment_ids.append(1) 315 | tokens.append("[SEP]") 316 | segment_ids.append(1) 317 | 318 | (tokens, masked_lm_positions, 319 | masked_lm_labels) = create_masked_lm_predictions( 320 | tokens, masked_lm_prob, max_predictions_per_seq, vocab_words, rng) 321 | instance = TrainingInstance( 322 | tokens=tokens, 323 | segment_ids=segment_ids, 324 | is_random_next=is_random_next, 325 | masked_lm_positions=masked_lm_positions, 326 | masked_lm_labels=masked_lm_labels) 327 | instances.append(instance) 328 | current_chunk = [] 329 | current_length = 0 330 | i += 1 331 | 332 | return instances 333 | 334 | 335 | def create_masked_lm_predictions(tokens, masked_lm_prob, 336 | max_predictions_per_seq, vocab_words, rng): 337 | """Creates the predictions for the masked LM objective.""" 338 | 339 | cand_indexes = [] 340 | for (i, token) in enumerate(tokens): 341 | if token == "[CLS]" or token == "[SEP]": 342 | continue 343 | cand_indexes.append(i) 344 | 345 | rng.shuffle(cand_indexes) 346 | 347 | output_tokens = list(tokens) 348 | 349 | masked_lm = collections.namedtuple("masked_lm", ["index", "label"]) # pylint: disable=invalid-name 350 | 351 | num_to_predict = min(max_predictions_per_seq, 352 | max(1, int(round(len(tokens) * masked_lm_prob)))) 353 | 354 | masked_lms = [] 355 | covered_indexes = set() 356 | for index in cand_indexes: 357 | if len(masked_lms) >= num_to_predict: 358 | break 359 | if index in covered_indexes: 360 | continue 361 | covered_indexes.add(index) 362 | 363 | masked_token = None 364 | # 80% of the time, replace with [MASK] 365 | if rng.random() < 0.8: 366 | masked_token = "[MASK]" 367 | else: 368 | # 10% of the time, keep original 369 | if rng.random() < 0.5: 370 | masked_token = tokens[index] 371 | # 10% of the time, replace with random word 372 | else: 373 | masked_token = vocab_words[rng.randint(0, len(vocab_words) - 1)] 374 | 375 | output_tokens[index] = masked_token 376 | 377 | masked_lms.append(masked_lm(index=index, label=tokens[index])) 378 | 379 | masked_lms = sorted(masked_lms, key=lambda x: x.index) 380 | 381 | masked_lm_positions = [] 382 | masked_lm_labels = [] 383 | for p in masked_lms: 384 | masked_lm_positions.append(p.index) 385 | masked_lm_labels.append(p.label) 386 | 387 | return (output_tokens, masked_lm_positions, masked_lm_labels) 388 | 389 | 390 | def truncate_seq_pair(tokens_a, tokens_b, max_num_tokens, rng): 391 | """Truncates a pair of sequences to a maximum sequence length.""" 392 | while True: 393 | total_length = len(tokens_a) + len(tokens_b) 394 | if total_length <= max_num_tokens: 395 | break 396 | 397 | trunc_tokens = tokens_a if len(tokens_a) > len(tokens_b) else tokens_b 398 | assert len(trunc_tokens) >= 1 399 | 400 | # We want to sometimes truncate from the front and sometimes from the 401 | # back to add more randomness and avoid biases. 402 | if rng.random() < 0.5: 403 | del trunc_tokens[0] 404 | else: 405 | trunc_tokens.pop() 406 | 407 | 408 | def main(_): 409 | tf.logging.set_verbosity(tf.logging.INFO) 410 | 411 | tokenizer = tokenization.FullTokenizer( 412 | vocab_file=FLAGS.vocab_file, do_lower_case=FLAGS.do_lower_case) 413 | 414 | input_files = [] 415 | for input_pattern in FLAGS.input_file.split(","): 416 | input_files.extend(tf.gfile.Glob(input_pattern)) 417 | 418 | tf.logging.info("*** Reading from input files ***") 419 | for input_file in input_files: 420 | tf.logging.info(" %s", input_file) 421 | 422 | rng = random.Random(FLAGS.random_seed) 423 | instances = create_training_instances( 424 | input_files, tokenizer, FLAGS.max_seq_length, FLAGS.dupe_factor, 425 | FLAGS.short_seq_prob, FLAGS.masked_lm_prob, FLAGS.max_predictions_per_seq, 426 | rng) 427 | 428 | output_files = FLAGS.output_file.split(",") 429 | tf.logging.info("*** Writing to output files ***") 430 | for output_file in output_files: 431 | tf.logging.info(" %s", output_file) 432 | 433 | write_instance_to_example_files(instances, tokenizer, FLAGS.max_seq_length, 434 | FLAGS.max_predictions_per_seq, output_files) 435 | 436 | 437 | if __name__ == "__main__": 438 | flags.mark_flag_as_required("input_file") 439 | flags.mark_flag_as_required("output_file") 440 | flags.mark_flag_as_required("vocab_file") 441 | tf.app.run() 442 | -------------------------------------------------------------------------------- /data/lm/result.zh.json: -------------------------------------------------------------------------------- 1 | [ 2 | { 3 | "tokens": [ 4 | { 5 | "token": "2016", 6 | "prob": 0.10201895982027054 7 | }, 8 | { 9 | "token": "全", 10 | "prob": 0.8663827180862427 11 | }, 12 | { 13 | "token": "国", 14 | "prob": 0.9864780902862549 15 | }, 16 | { 17 | "token": "高", 18 | "prob": 0.34686800837516785 19 | }, 20 | { 21 | "token": "考", 22 | "prob": 0.9513261914253235 23 | }, 24 | { 25 | "token": "卷", 26 | "prob": 0.002143713878467679 27 | }, 28 | { 29 | "token": "答", 30 | "prob": 0.04595280811190605 31 | }, 32 | { 33 | "token": "题", 34 | "prob": 0.052179522812366486 35 | }, 36 | { 37 | "token": "模", 38 | "prob": 0.9097260236740112 39 | }, 40 | { 41 | "token": "板", 42 | "prob": 0.5003664493560791 43 | } 44 | ], 45 | "ppl": 5.214168292180062 46 | }, 47 | { 48 | "tokens": [ 49 | { 50 | "token": "2016", 51 | "prob": 0.07322586327791214 52 | }, 53 | { 54 | "token": "全", 55 | "prob": 0.4594285190105438 56 | }, 57 | { 58 | "token": "国", 59 | "prob": 0.9693033695220947 60 | }, 61 | { 62 | "token": "大", 63 | "prob": 0.002398423384875059 64 | }, 65 | { 66 | "token": "考", 67 | "prob": 0.35024359822273254 68 | }, 69 | { 70 | "token": "卷", 71 | "prob": 0.005357644520699978 72 | }, 73 | { 74 | "token": "答", 75 | "prob": 0.05084587633609772 76 | }, 77 | { 78 | "token": "题", 79 | "prob": 0.08499259501695633 80 | }, 81 | { 82 | "token": "模", 83 | "prob": 0.8091200590133667 84 | }, 85 | { 86 | "token": "板", 87 | "prob": 0.4416927099227905 88 | } 89 | ], 90 | "ppl": 9.214288062044073 91 | }, 92 | { 93 | "tokens": [ 94 | { 95 | "token": "2016", 96 | "prob": 0.06563900411128998 97 | }, 98 | { 99 | "token": "全", 100 | "prob": 0.4981258511543274 101 | }, 102 | { 103 | "token": "国", 104 | "prob": 0.9088247418403625 105 | }, 106 | { 107 | "token": "低", 108 | "prob": 1.6259804397122934e-05 109 | }, 110 | { 111 | "token": "考", 112 | "prob": 0.4023572504520416 113 | }, 114 | { 115 | "token": "卷", 116 | "prob": 0.014088313095271587 117 | }, 118 | { 119 | "token": "答", 120 | "prob": 0.0639762282371521 121 | }, 122 | { 123 | "token": "题", 124 | "prob": 0.09024914354085922 125 | }, 126 | { 127 | "token": "模", 128 | "prob": 0.8052080273628235 129 | }, 130 | { 131 | "token": "板", 132 | "prob": 0.4206086993217468 133 | } 134 | ], 135 | "ppl": 13.400421357093588 136 | }, 137 | { 138 | "tokens": [ 139 | { 140 | "token": "床", 141 | "prob": 0.06141529977321625 142 | }, 143 | { 144 | "token": "前", 145 | "prob": 0.6264788508415222 146 | }, 147 | { 148 | "token": "明", 149 | "prob": 0.19100141525268555 150 | }, 151 | { 152 | "token": "月", 153 | "prob": 0.8327909111976624 154 | }, 155 | { 156 | "token": "光", 157 | "prob": 0.0027572319377213717 158 | }, 159 | { 160 | "token": "，", 161 | "prob": 0.9875602126121521 162 | }, 163 | { 164 | "token": "疑", 165 | "prob": 0.26570719480514526 166 | }, 167 | { 168 | "token": "是", 169 | "prob": 0.20073072612285614 170 | }, 171 | { 172 | "token": "地", 173 | "prob": 0.0013851922703906894 174 | }, 175 | { 176 | "token": "上", 177 | "prob": 0.08247583359479904 178 | }, 179 | { 180 | "token": "霜", 181 | "prob": 1.1996443838313553e-09 182 | } 183 | ], 184 | "ppl": 52.41817724710573 185 | }, 186 | { 187 | "tokens": [ 188 | { 189 | "token": "床", 190 | "prob": 0.007852567359805107 191 | }, 192 | { 193 | "token": "前", 194 | "prob": 0.023406021296977997 195 | }, 196 | { 197 | "token": "星", 198 | "prob": 0.0016379707958549261 199 | }, 200 | { 201 | "token": "星", 202 | "prob": 0.08062248677015305 203 | }, 204 | { 205 | "token": "光", 206 | "prob": 0.003109232522547245 207 | }, 208 | { 209 | "token": "，", 210 | "prob": 0.7677633762359619 211 | }, 212 | { 213 | "token": "疑", 214 | "prob": 0.02753453515470028 215 | }, 216 | { 217 | "token": "是", 218 | "prob": 0.18585893511772156 219 | }, 220 | { 221 | "token": "地", 222 | "prob": 0.006168654654175043 223 | }, 224 | { 225 | "token": "上", 226 | "prob": 0.09490019828081131 227 | }, 228 | { 229 | "token": "霜", 230 | "prob": 5.2765480873517845e-09 231 | } 232 | ], 233 | "ppl": 153.20970629080068 234 | }, 235 | { 236 | "tokens": [ 237 | { 238 | "token": "床", 239 | "prob": 0.05734571814537048 240 | }, 241 | { 242 | "token": "前", 243 | "prob": 0.4436746835708618 244 | }, 245 | { 246 | "token": "白", 247 | "prob": 0.0020346769597381353 248 | }, 249 | { 250 | "token": "月", 251 | "prob": 0.23601755499839783 252 | }, 253 | { 254 | "token": "光", 255 | "prob": 0.05537288263440132 256 | }, 257 | { 258 | "token": "，", 259 | "prob": 0.9799799919128418 260 | }, 261 | { 262 | "token": "疑", 263 | "prob": 0.22007052600383759 264 | }, 265 | { 266 | "token": "是", 267 | "prob": 0.21303308010101318 268 | }, 269 | { 270 | "token": "地", 271 | "prob": 0.0022917157039046288 272 | }, 273 | { 274 | "token": "上", 275 | "prob": 0.07564827054738998 276 | }, 277 | { 278 | "token": "霜", 279 | "prob": 6.6843566237650975e-09 280 | } 281 | ], 282 | "ppl": 58.55804131999474 283 | }, 284 | { 285 | "tokens": [ 286 | { 287 | "token": "落", 288 | "prob": 0.14436914026737213 289 | }, 290 | { 291 | "token": "霞", 292 | "prob": 0.4585109353065491 293 | }, 294 | { 295 | "token": "与", 296 | "prob": 0.8038789629936218 297 | }, 298 | { 299 | "token": "孤", 300 | "prob": 0.9961937665939331 301 | }, 302 | { 303 | "token": "鹜", 304 | "prob": 0.7430610060691833 305 | }, 306 | { 307 | "token": "齐", 308 | "prob": 0.5758452415466309 309 | }, 310 | { 311 | "token": "飞", 312 | "prob": 0.8611886501312256 313 | }, 314 | { 315 | "token": "，", 316 | "prob": 0.8926917314529419 317 | }, 318 | { 319 | "token": "秋", 320 | "prob": 0.26929906010627747 321 | }, 322 | { 323 | "token": "水", 324 | "prob": 0.29808318614959717 325 | }, 326 | { 327 | "token": "共", 328 | "prob": 0.16475750505924225 329 | }, 330 | { 331 | "token": "长", 332 | "prob": 0.04301689937710762 333 | }, 334 | { 335 | "token": "天", 336 | "prob": 0.8850820660591125 337 | }, 338 | { 339 | "token": "一", 340 | "prob": 0.675239622592926 341 | }, 342 | { 343 | "token": "色", 344 | "prob": 5.158061867405195e-06 345 | } 346 | ], 347 | "ppl": 5.023060982907458 348 | }, 349 | { 350 | "tokens": [ 351 | { 352 | "token": "落", 353 | "prob": 0.1483132392168045 354 | }, 355 | { 356 | "token": "霞", 357 | "prob": 0.42232587933540344 358 | }, 359 | { 360 | "token": "与", 361 | "prob": 0.8615185022354126 362 | }, 363 | { 364 | "token": "孤", 365 | "prob": 0.9975666999816895 366 | }, 367 | { 368 | "token": "鹜", 369 | "prob": 0.5613960027694702 370 | }, 371 | { 372 | "token": "齐", 373 | "prob": 0.18012434244155884 374 | }, 375 | { 376 | "token": "跑", 377 | "prob": 1.3388593288254924e-05 378 | }, 379 | { 380 | "token": "，", 381 | "prob": 0.8621458411216736 382 | }, 383 | { 384 | "token": "秋", 385 | "prob": 0.24820539355278015 386 | }, 387 | { 388 | "token": "水", 389 | "prob": 0.32748720049858093 390 | }, 391 | { 392 | "token": "共", 393 | "prob": 0.2348739057779312 394 | }, 395 | { 396 | "token": "长", 397 | "prob": 0.040592435747385025 398 | }, 399 | { 400 | "token": "天", 401 | "prob": 0.9231186509132385 402 | }, 403 | { 404 | "token": "一", 405 | "prob": 0.5295999646186829 406 | }, 407 | { 408 | "token": "色", 409 | "prob": 2.7847559067595284e-06 410 | } 411 | ], 412 | "ppl": 11.983086642867598 413 | }, 414 | { 415 | "tokens": [ 416 | { 417 | "token": "落", 418 | "prob": 0.14112578332424164 419 | }, 420 | { 421 | "token": "霞", 422 | "prob": 0.5331157445907593 423 | }, 424 | { 425 | "token": "与", 426 | "prob": 0.41120263934135437 427 | }, 428 | { 429 | "token": "孤", 430 | "prob": 0.9943874478340149 431 | }, 432 | { 433 | "token": "鹜", 434 | "prob": 0.6526917219161987 435 | }, 436 | { 437 | "token": "双", 438 | "prob": 0.03388489782810211 439 | }, 440 | { 441 | "token": "飞", 442 | "prob": 0.4864092767238617 443 | }, 444 | { 445 | "token": "，", 446 | "prob": 0.8635039925575256 447 | }, 448 | { 449 | "token": "秋", 450 | "prob": 0.27505722641944885 451 | }, 452 | { 453 | "token": "水", 454 | "prob": 0.2975866198539734 455 | }, 456 | { 457 | "token": "共", 458 | "prob": 0.1109926775097847 459 | }, 460 | { 461 | "token": "长", 462 | "prob": 0.05403125286102295 463 | }, 464 | { 465 | "token": "天", 466 | "prob": 0.8708866238594055 467 | }, 468 | { 469 | "token": "一", 470 | "prob": 0.6170234680175781 471 | }, 472 | { 473 | "token": "色", 474 | "prob": 1.1969897968810983e-05 475 | } 476 | ], 477 | "ppl": 6.352968507499718 478 | }, 479 | { 480 | "tokens": [ 481 | { 482 | "token": "众", 483 | "prob": 0.9997541308403015 484 | }, 485 | { 486 | "token": "里", 487 | "prob": 0.9997285008430481 488 | }, 489 | { 490 | "token": "寻", 491 | "prob": 0.9988683462142944 492 | }, 493 | { 494 | "token": "他", 495 | "prob": 0.7197673916816711 496 | }, 497 | { 498 | "token": "千", 499 | "prob": 0.9998348951339722 500 | }, 501 | { 502 | "token": "百", 503 | "prob": 0.9883707761764526 504 | }, 505 | { 506 | "token": "度", 507 | "prob": 0.9988011121749878 508 | }, 509 | { 510 | "token": "，", 511 | "prob": 0.9991918206214905 512 | }, 513 | { 514 | "token": "蓦", 515 | "prob": 0.9992905855178833 516 | }, 517 | { 518 | "token": "然", 519 | "prob": 0.999387264251709 520 | }, 521 | { 522 | "token": "回", 523 | "prob": 0.9999867677688599 524 | }, 525 | { 526 | "token": "首", 527 | "prob": 0.9995602965354919 528 | }, 529 | { 530 | "token": "，", 531 | "prob": 0.9979262351989746 532 | }, 533 | { 534 | "token": "那", 535 | "prob": 0.957136869430542 536 | }, 537 | { 538 | "token": "人", 539 | "prob": 0.9965313076972961 540 | }, 541 | { 542 | "token": "却", 543 | "prob": 0.9842029809951782 544 | }, 545 | { 546 | "token": "在", 547 | "prob": 0.9936916828155518 548 | }, 549 | { 550 | "token": "，", 551 | "prob": 0.9937483072280884 552 | }, 553 | { 554 | "token": "灯", 555 | "prob": 0.9998267292976379 556 | }, 557 | { 558 | "token": "火", 559 | "prob": 0.9999244213104248 560 | }, 561 | { 562 | "token": "阑", 563 | "prob": 0.9999798536300659 564 | }, 565 | { 566 | "token": "珊", 567 | "prob": 0.9999972581863403 568 | }, 569 | { 570 | "token": "处", 571 | "prob": 1.918866399108765e-09 572 | } 573 | ], 574 | "ppl": 2.4378636592211476 575 | }, 576 | { 577 | "tokens": [ 578 | { 579 | "token": "众", 580 | "prob": 0.9995400905609131 581 | }, 582 | { 583 | "token": "里", 584 | "prob": 0.9989792108535767 585 | }, 586 | { 587 | "token": "寻", 588 | "prob": 0.9996756315231323 589 | }, 590 | { 591 | "token": "她", 592 | "prob": 0.26533517241477966 593 | }, 594 | { 595 | "token": "千", 596 | "prob": 0.9998635053634644 597 | }, 598 | { 599 | "token": "百", 600 | "prob": 0.9769723415374756 601 | }, 602 | { 603 | "token": "度", 604 | "prob": 0.997245192527771 605 | }, 606 | { 607 | "token": "，", 608 | "prob": 0.9979918003082275 609 | }, 610 | { 611 | "token": "蓦", 612 | "prob": 0.9991554021835327 613 | }, 614 | { 615 | "token": "然", 616 | "prob": 0.9993096590042114 617 | }, 618 | { 619 | "token": "回", 620 | "prob": 0.9999856948852539 621 | }, 622 | { 623 | "token": "首", 624 | "prob": 0.9995132684707642 625 | }, 626 | { 627 | "token": "，", 628 | "prob": 0.9980834722518921 629 | }, 630 | { 631 | "token": "那", 632 | "prob": 0.969828724861145 633 | }, 634 | { 635 | "token": "人", 636 | "prob": 0.9686450958251953 637 | }, 638 | { 639 | "token": "却", 640 | "prob": 0.9804452061653137 641 | }, 642 | { 643 | "token": "在", 644 | "prob": 0.9953302145004272 645 | }, 646 | { 647 | "token": "，", 648 | "prob": 0.9902727007865906 649 | }, 650 | { 651 | "token": "灯", 652 | "prob": 0.999835729598999 653 | }, 654 | { 655 | "token": "火", 656 | "prob": 0.9999294281005859 657 | }, 658 | { 659 | "token": "阑", 660 | "prob": 0.9999833106994629 661 | }, 662 | { 663 | "token": "珊", 664 | "prob": 0.9999969005584717 665 | }, 666 | { 667 | "token": "处", 668 | "prob": 2.7301996219364355e-09 669 | } 670 | ], 671 | "ppl": 2.511098534866279 672 | }, 673 | { 674 | "tokens": [ 675 | { 676 | "token": "众", 677 | "prob": 0.9973291158676147 678 | }, 679 | { 680 | "token": "里", 681 | "prob": 0.9993185997009277 682 | }, 683 | { 684 | "token": "寻", 685 | "prob": 0.9946662187576294 686 | }, 687 | { 688 | "token": "ta", 689 | "prob": 9.90712862858345e-08 690 | }, 691 | { 692 | "token": "千", 693 | "prob": 0.999840259552002 694 | }, 695 | { 696 | "token": "百", 697 | "prob": 0.9573060870170593 698 | }, 699 | { 700 | "token": "度", 701 | "prob": 0.9984676241874695 702 | }, 703 | { 704 | "token": "，", 705 | "prob": 0.998295247554779 706 | }, 707 | { 708 | "token": "蓦", 709 | "prob": 0.9991011619567871 710 | }, 711 | { 712 | "token": "然", 713 | "prob": 0.9991719722747803 714 | }, 715 | { 716 | "token": "回", 717 | "prob": 0.9999833106994629 718 | }, 719 | { 720 | "token": "首", 721 | "prob": 0.999624490737915 722 | }, 723 | { 724 | "token": "，", 725 | "prob": 0.9986977577209473 726 | }, 727 | { 728 | "token": "那", 729 | "prob": 0.9325712323188782 730 | }, 731 | { 732 | "token": "人", 733 | "prob": 0.9955593347549438 734 | }, 735 | { 736 | "token": "却", 737 | "prob": 0.9584773182868958 738 | }, 739 | { 740 | "token": "在", 741 | "prob": 0.9958037734031677 742 | }, 743 | { 744 | "token": "，", 745 | "prob": 0.9910857081413269 746 | }, 747 | { 748 | "token": "灯", 749 | "prob": 0.9998859167098999 750 | }, 751 | { 752 | "token": "火", 753 | "prob": 0.9999510049819946 754 | }, 755 | { 756 | "token": "阑", 757 | "prob": 0.9999864101409912 758 | }, 759 | { 760 | "token": "珊", 761 | "prob": 0.9999982118606567 762 | }, 763 | { 764 | "token": "处", 765 | "prob": 1.1214358330846608e-09 766 | } 767 | ], 768 | "ppl": 4.980078099549084 769 | } -------------------------------------------------------------------------------- /data/lm/test.en.tsv: -------------------------------------------------------------------------------- 1 | there is a book on the desk 2 | there is a plane on the desk 3 | there is a book in the desk 4 | 5 | -------------------------------------------------------------------------------- /data/lm/test.zh.tsv: -------------------------------------------------------------------------------- 1 | 2016全国高考卷答题模板 2 | 2016全国大考卷答题模板 3 | 2016全国低考卷答题模板 4 | 床前明月光，疑是地上霜 5 | 床前星星光，疑是地上霜 6 | 床前白月光，疑是地上霜 7 | 落霞与孤鹜齐飞，秋水共长天一色 8 | 落霞与孤鹜齐跑，秋水共长天一色 9 | 落霞与孤鹜双飞，秋水共长天一色 10 | 众里寻他千百度，蓦然回首，那人却在，灯火阑珊处 11 | 众里寻她千百度，蓦然回首，那人却在，灯火阑珊处 12 | 众里寻ta千百度，蓦然回首，那人却在，灯火阑珊处 -------------------------------------------------------------------------------- /extract_features.py: -------------------------------------------------------------------------------- 1 | # coding=utf-8 2 | # Copyright 2018 The Google AI Language Team Authors. 3 | # 4 | # Licensed under the Apache License, Version 2.0 (the "License"); 5 | # you may not use this file except in compliance with the License. 6 | # You may obtain a copy of the License at 7 | # 8 | # http://www.apache.org/licenses/LICENSE-2.0 9 | # 10 | # Unless required by applicable law or agreed to in writing, software 11 | # distributed under the License is distributed on an "AS IS" BASIS, 12 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 13 | # See the License for the specific language governing permissions and 14 | # limitations under the License. 15 | """Extract pre-computed feature vectors from BERT.""" 16 | 17 | from __future__ import absolute_import 18 | from __future__ import division 19 | from __future__ import print_function 20 | 21 | import codecs 22 | import collections 23 | import json 24 | import re 25 | 26 | import modeling 27 | import tokenization 28 | import tensorflow as tf 29 | 30 | flags = tf.flags 31 | 32 | FLAGS = flags.FLAGS 33 | 34 | flags.DEFINE_string("input_file", None, "") 35 | 36 | flags.DEFINE_string("output_file", None, "") 37 | 38 | flags.DEFINE_string("layers", "-1,-2,-3,-4", "") 39 | 40 | flags.DEFINE_string( 41 | "bert_config_file", None, 42 | "The config json file corresponding to the pre-trained BERT model. " 43 | "This specifies the model architecture.") 44 | 45 | flags.DEFINE_integer( 46 | "max_seq_length", 128, 47 | "The maximum total input sequence length after WordPiece tokenization. " 48 | "Sequences longer than this will be truncated, and sequences shorter " 49 | "than this will be padded.") 50 | 51 | flags.DEFINE_string( 52 | "init_checkpoint", None, 53 | "Initial checkpoint (usually from a pre-trained BERT model).") 54 | 55 | flags.DEFINE_string("vocab_file", None, 56 | "The vocabulary file that the BERT model was trained on.") 57 | 58 | flags.DEFINE_bool( 59 | "do_lower_case", True, 60 | "Whether to lower case the input text. Should be True for uncased " 61 | "models and False for cased models.") 62 | 63 | flags.DEFINE_integer("batch_size", 32, "Batch size for predictions.") 64 | 65 | flags.DEFINE_bool("use_tpu", False, "Whether to use TPU or GPU/CPU.") 66 | 67 | flags.DEFINE_string("master", None, 68 | "If using a TPU, the address of the master.") 69 | 70 | flags.DEFINE_integer( 71 | "num_tpu_cores", 8, 72 | "Only used if `use_tpu` is True. Total number of TPU cores to use.") 73 | 74 | flags.DEFINE_bool( 75 | "use_one_hot_embeddings", False, 76 | "If True, tf.one_hot will be used for embedding lookups, otherwise " 77 | "tf.nn.embedding_lookup will be used. On TPUs, this should be True " 78 | "since it is much faster.") 79 | 80 | 81 | class InputExample(object): 82 | 83 | def __init__(self, unique_id, text_a, text_b): 84 | self.unique_id = unique_id 85 | self.text_a = text_a 86 | self.text_b = text_b 87 | 88 | 89 | class InputFeatures(object): 90 | """A single set of features of data.""" 91 | 92 | def __init__(self, unique_id, tokens, input_ids, input_mask, input_type_ids): 93 | self.unique_id = unique_id 94 | self.tokens = tokens 95 | self.input_ids = input_ids 96 | self.input_mask = input_mask 97 | self.input_type_ids = input_type_ids 98 | 99 | 100 | def input_fn_builder(features, seq_length): 101 | """Creates an `input_fn` closure to be passed to TPUEstimator.""" 102 | 103 | all_unique_ids = [] 104 | all_input_ids = [] 105 | all_input_mask = [] 106 | all_input_type_ids = [] 107 | 108 | for feature in features: 109 | all_unique_ids.append(feature.unique_id) 110 | all_input_ids.append(feature.input_ids) 111 | all_input_mask.append(feature.input_mask) 112 | all_input_type_ids.append(feature.input_type_ids) 113 | 114 | def input_fn(params): 115 | """The actual input function.""" 116 | batch_size = params["batch_size"] 117 | 118 | num_examples = len(features) 119 | 120 | # This is for demo purposes and does NOT scale to large data sets. We do 121 | # not use Dataset.from_generator() because that uses tf.py_func which is 122 | # not TPU compatible. The right way to load data is with TFRecordReader. 123 | d = tf.data.Dataset.from_tensor_slices({ 124 | "unique_ids": 125 | tf.constant(all_unique_ids, shape=[num_examples], dtype=tf.int32), 126 | "input_ids": 127 | tf.constant( 128 | all_input_ids, shape=[num_examples, seq_length], 129 | dtype=tf.int32), 130 | "input_mask": 131 | tf.constant( 132 | all_input_mask, 133 | shape=[num_examples, seq_length], 134 | dtype=tf.int32), 135 | "input_type_ids": 136 | tf.constant( 137 | all_input_type_ids, 138 | shape=[num_examples, seq_length], 139 | dtype=tf.int32), 140 | }) 141 | 142 | d = d.batch(batch_size=batch_size, drop_remainder=False) 143 | return d 144 | 145 | return input_fn 146 | 147 | 148 | def model_fn_builder(bert_config, init_checkpoint, layer_indexes, use_tpu, 149 | use_one_hot_embeddings): 150 | """Returns `model_fn` closure for TPUEstimator.""" 151 | 152 | def model_fn(features, labels, mode, params): # pylint: disable=unused-argument 153 | """The `model_fn` for TPUEstimator.""" 154 | 155 | unique_ids = features["unique_ids"] 156 | input_ids = features["input_ids"] 157 | input_mask = features["input_mask"] 158 | input_type_ids = features["input_type_ids"] 159 | 160 | model = modeling.BertModel( 161 | config=bert_config, 162 | is_training=False, 163 | input_ids=input_ids, 164 | input_mask=input_mask, 165 | token_type_ids=input_type_ids, 166 | use_one_hot_embeddings=use_one_hot_embeddings) 167 | 168 | if mode != tf.estimator.ModeKeys.PREDICT: 169 | raise ValueError("Only PREDICT modes are supported: %s" % (mode)) 170 | 171 | tvars = tf.trainable_variables() 172 | scaffold_fn = None 173 | (assignment_map, 174 | initialized_variable_names) = modeling.get_assignment_map_from_checkpoint( 175 | tvars, init_checkpoint) 176 | if use_tpu: 177 | 178 | def tpu_scaffold(): 179 | tf.train.init_from_checkpoint(init_checkpoint, assignment_map) 180 | return tf.train.Scaffold() 181 | 182 | scaffold_fn = tpu_scaffold 183 | else: 184 | tf.train.init_from_checkpoint(init_checkpoint, assignment_map) 185 | 186 | tf.logging.info("**** Trainable Variables ****") 187 | for var in tvars: 188 | init_string = "" 189 | if var.name in initialized_variable_names: 190 | init_string = ", *INIT_FROM_CKPT*" 191 | tf.logging.info(" name = %s, shape = %s%s", var.name, var.shape, 192 | init_string) 193 | 194 | all_layers = model.get_all_encoder_layers() 195 | 196 | predictions = { 197 | "unique_id": unique_ids, 198 | } 199 | 200 | for (i, layer_index) in enumerate(layer_indexes): 201 | predictions["layer_output_%d" % i] = all_layers[layer_index] 202 | 203 | output_spec = tf.contrib.tpu.TPUEstimatorSpec( 204 | mode=mode, predictions=predictions, scaffold_fn=scaffold_fn) 205 | return output_spec 206 | 207 | return model_fn 208 | 209 | 210 | def convert_examples_to_features(examples, seq_length, tokenizer): 211 | """Loads a data file into a list of `InputBatch`s.""" 212 | 213 | features = [] 214 | for (ex_index, example) in enumerate(examples): 215 | tokens_a = tokenizer.tokenize(example.text_a) 216 | 217 | tokens_b = None 218 | if example.text_b: 219 | tokens_b = tokenizer.tokenize(example.text_b) 220 | 221 | if tokens_b: 222 | # Modifies `tokens_a` and `tokens_b` in place so that the total 223 | # length is less than the specified length. 224 | # Account for [CLS], [SEP], [SEP] with "- 3" 225 | _truncate_seq_pair(tokens_a, tokens_b, seq_length - 3) 226 | else: 227 | # Account for [CLS] and [SEP] with "- 2" 228 | if len(tokens_a) > seq_length - 2: 229 | tokens_a = tokens_a[0:(seq_length - 2)] 230 | 231 | # The convention in BERT is: 232 | # (a) For sequence pairs: 233 | # tokens: [CLS] is this jack ##son ##ville ? [SEP] no it is not . [SEP] 234 | # type_ids: 0 0 0 0 0 0 0 0 1 1 1 1 1 1 235 | # (b) For single sequences: 236 | # tokens: [CLS] the dog is hairy . [SEP] 237 | # type_ids: 0 0 0 0 0 0 0 238 | # 239 | # Where "type_ids" are used to indicate whether this is the first 240 | # sequence or the second sequence. The embedding vectors for `type=0` and 241 | # `type=1` were learned during pre-training and are added to the wordpiece 242 | # embedding vector (and position vector). This is not *strictly* necessary 243 | # since the [SEP] token unambiguously separates the sequences, but it makes 244 | # it easier for the model to learn the concept of sequences. 245 | # 246 | # For classification tasks, the first vector (corresponding to [CLS]) is 247 | # used as as the "sentence vector". Note that this only makes sense because 248 | # the entire model is fine-tuned. 249 | tokens = [] 250 | input_type_ids = [] 251 | tokens.append("[CLS]") 252 | input_type_ids.append(0) 253 | for token in tokens_a: 254 | tokens.append(token) 255 | input_type_ids.append(0) 256 | tokens.append("[SEP]") 257 | input_type_ids.append(0) 258 | 259 | if tokens_b: 260 | for token in tokens_b: 261 | tokens.append(token) 262 | input_type_ids.append(1) 263 | tokens.append("[SEP]") 264 | input_type_ids.append(1) 265 | 266 | input_ids = tokenizer.convert_tokens_to_ids(tokens) 267 | 268 | # The mask has 1 for real tokens and 0 for padding tokens. Only real 269 | # tokens are attended to. 270 | input_mask = [1] * len(input_ids) 271 | 272 | # Zero-pad up to the sequence length. 273 | while len(input_ids) < seq_length: 274 | input_ids.append(0) 275 | input_mask.append(0) 276 | input_type_ids.append(0) 277 | 278 | assert len(input_ids) == seq_length 279 | assert len(input_mask) == seq_length 280 | assert len(input_type_ids) == seq_length 281 | 282 | if ex_index < 5: 283 | tf.logging.info("*** Example ***") 284 | tf.logging.info("unique_id: %s" % (example.unique_id)) 285 | tf.logging.info("tokens: %s" % " ".join( 286 | [tokenization.printable_text(x) for x in tokens])) 287 | tf.logging.info("input_ids: %s" % " ".join([str(x) for x in input_ids])) 288 | tf.logging.info("input_mask: %s" % " ".join([str(x) for x in input_mask])) 289 | tf.logging.info( 290 | "input_type_ids: %s" % " ".join([str(x) for x in input_type_ids])) 291 | 292 | features.append( 293 | InputFeatures( 294 | unique_id=example.unique_id, 295 | tokens=tokens, 296 | input_ids=input_ids, 297 | input_mask=input_mask, 298 | input_type_ids=input_type_ids)) 299 | return features 300 | 301 | 302 | def _truncate_seq_pair(tokens_a, tokens_b, max_length): 303 | """Truncates a sequence pair in place to the maximum length.""" 304 | 305 | # This is a simple heuristic which will always truncate the longer sequence 306 | # one token at a time. This makes more sense than truncating an equal percent 307 | # of tokens from each, since if one sequence is very short then each token 308 | # that's truncated likely contains more information than a longer sequence. 309 | while True: 310 | total_length = len(tokens_a) + len(tokens_b) 311 | if total_length <= max_length: 312 | break 313 | if len(tokens_a) > len(tokens_b): 314 | tokens_a.pop() 315 | else: 316 | tokens_b.pop() 317 | 318 | 319 | def read_examples(input_file): 320 | """Read a list of `InputExample`s from an input file.""" 321 | examples = [] 322 | unique_id = 0 323 | with tf.gfile.GFile(input_file, "r") as reader: 324 | while True: 325 | line = tokenization.convert_to_unicode(reader.readline()) 326 | if not line: 327 | break 328 | line = line.strip() 329 | text_a = None 330 | text_b = None 331 | m = re.match(r"^(.*) \|\|\| (.*)$", line) 332 | if m is None: 333 | text_a = line 334 | else: 335 | text_a = m.group(1) 336 | text_b = m.group(2) 337 | examples.append( 338 | InputExample(unique_id=unique_id, text_a=text_a, text_b=text_b)) 339 | unique_id += 1 340 | return examples 341 | 342 | 343 | def main(_): 344 | tf.logging.set_verbosity(tf.logging.INFO) 345 | 346 | layer_indexes = [int(x) for x in FLAGS.layers.split(",")] 347 | 348 | bert_config = modeling.BertConfig.from_json_file(FLAGS.bert_config_file) 349 | 350 | tokenizer = tokenization.FullTokenizer( 351 | vocab_file=FLAGS.vocab_file, do_lower_case=FLAGS.do_lower_case) 352 | 353 | is_per_host = tf.contrib.tpu.InputPipelineConfig.PER_HOST_V2 354 | run_config = tf.contrib.tpu.RunConfig( 355 | master=FLAGS.master, 356 | tpu_config=tf.contrib.tpu.TPUConfig( 357 | num_shards=FLAGS.num_tpu_cores, 358 | per_host_input_for_training=is_per_host)) 359 | 360 | examples = read_examples(FLAGS.input_file) 361 | 362 | features = convert_examples_to_features( 363 | examples=examples, seq_length=FLAGS.max_seq_length, tokenizer=tokenizer) 364 | 365 | unique_id_to_feature = {} 366 | for feature in features: 367 | unique_id_to_feature[feature.unique_id] = feature 368 | 369 | model_fn = model_fn_builder( 370 | bert_config=bert_config, 371 | init_checkpoint=FLAGS.init_checkpoint, 372 | layer_indexes=layer_indexes, 373 | use_tpu=FLAGS.use_tpu, 374 | use_one_hot_embeddings=FLAGS.use_one_hot_embeddings) 375 | 376 | # If TPU is not available, this will fall back to normal Estimator on CPU 377 | # or GPU. 378 | estimator = tf.contrib.tpu.TPUEstimator( 379 | use_tpu=FLAGS.use_tpu, 380 | model_fn=model_fn, 381 | config=run_config, 382 | predict_batch_size=FLAGS.batch_size) 383 | 384 | input_fn = input_fn_builder( 385 | features=features, seq_length=FLAGS.max_seq_length) 386 | 387 | with codecs.getwriter("utf-8")(tf.gfile.Open(FLAGS.output_file, 388 | "w")) as writer: 389 | for result in estimator.predict(input_fn, yield_single_examples=True): 390 | unique_id = int(result["unique_id"]) 391 | feature = unique_id_to_feature[unique_id] 392 | output_json = collections.OrderedDict() 393 | output_json["linex_index"] = unique_id 394 | all_features = [] 395 | for (i, token) in enumerate(feature.tokens): 396 | all_layers = [] 397 | for (j, layer_index) in enumerate(layer_indexes): 398 | layer_output = result["layer_output_%d" % j] 399 | layers = collections.OrderedDict() 400 | layers["index"] = layer_index 401 | layers["values"] = [ 402 | round(float(x), 6) for x in layer_output[i:(i + 1)].flat 403 | ] 404 | all_layers.append(layers) 405 | features = collections.OrderedDict() 406 | features["token"] = token 407 | features["layers"] = all_layers 408 | all_features.append(features) 409 | output_json["features"] = all_features 410 | writer.write(json.dumps(output_json) + "\n") 411 | 412 | 413 | if __name__ == "__main__": 414 | flags.mark_flag_as_required("input_file") 415 | flags.mark_flag_as_required("vocab_file") 416 | flags.mark_flag_as_required("bert_config_file") 417 | flags.mark_flag_as_required("init_checkpoint") 418 | flags.mark_flag_as_required("output_file") 419 | tf.app.run() 420 | -------------------------------------------------------------------------------- /modeling_test.py: -------------------------------------------------------------------------------- 1 | # coding=utf-8 2 | # Copyright 2018 The Google AI Language Team Authors. 3 | # 4 | # Licensed under the Apache License, Version 2.0 (the "License"); 5 | # you may not use this file except in compliance with the License. 6 | # You may obtain a copy of the License at 7 | # 8 | # http://www.apache.org/licenses/LICENSE-2.0 9 | # 10 | # Unless required by applicable law or agreed to in writing, software 11 | # distributed under the License is distributed on an "AS IS" BASIS, 12 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 13 | # See the License for the specific language governing permissions and 14 | # limitations under the License. 15 | from __future__ import absolute_import 16 | from __future__ import division 17 | from __future__ import print_function 18 | 19 | import collections 20 | import json 21 | import random 22 | import re 23 | 24 | import modeling 25 | import six 26 | import tensorflow as tf 27 | 28 | 29 | class BertModelTest(tf.test.TestCase): 30 | 31 | class BertModelTester(object): 32 | 33 | def __init__(self, 34 | parent, 35 | batch_size=13, 36 | seq_length=7, 37 | is_training=True, 38 | use_input_mask=True, 39 | use_token_type_ids=True, 40 | vocab_size=99, 41 | hidden_size=32, 42 | num_hidden_layers=5, 43 | num_attention_heads=4, 44 | intermediate_size=37, 45 | hidden_act="gelu", 46 | hidden_dropout_prob=0.1, 47 | attention_probs_dropout_prob=0.1, 48 | max_position_embeddings=512, 49 | type_vocab_size=16, 50 | initializer_range=0.02, 51 | scope=None): 52 | self.parent = parent 53 | self.batch_size = batch_size 54 | self.seq_length = seq_length 55 | self.is_training = is_training 56 | self.use_input_mask = use_input_mask 57 | self.use_token_type_ids = use_token_type_ids 58 | self.vocab_size = vocab_size 59 | self.hidden_size = hidden_size 60 | self.num_hidden_layers = num_hidden_layers 61 | self.num_attention_heads = num_attention_heads 62 | self.intermediate_size = intermediate_size 63 | self.hidden_act = hidden_act 64 | self.hidden_dropout_prob = hidden_dropout_prob 65 | self.attention_probs_dropout_prob = attention_probs_dropout_prob 66 | self.max_position_embeddings = max_position_embeddings 67 | self.type_vocab_size = type_vocab_size 68 | self.initializer_range = initializer_range 69 | self.scope = scope 70 | 71 | def create_model(self): 72 | input_ids = BertModelTest.ids_tensor([self.batch_size, self.seq_length], 73 | self.vocab_size) 74 | 75 | input_mask = None 76 | if self.use_input_mask: 77 | input_mask = BertModelTest.ids_tensor( 78 | [self.batch_size, self.seq_length], vocab_size=2) 79 | 80 | token_type_ids = None 81 | if self.use_token_type_ids: 82 | token_type_ids = BertModelTest.ids_tensor( 83 | [self.batch_size, self.seq_length], self.type_vocab_size) 84 | 85 | config = modeling.BertConfig( 86 | vocab_size=self.vocab_size, 87 | hidden_size=self.hidden_size, 88 | num_hidden_layers=self.num_hidden_layers, 89 | num_attention_heads=self.num_attention_heads, 90 | intermediate_size=self.intermediate_size, 91 | hidden_act=self.hidden_act, 92 | hidden_dropout_prob=self.hidden_dropout_prob, 93 | attention_probs_dropout_prob=self.attention_probs_dropout_prob, 94 | max_position_embeddings=self.max_position_embeddings, 95 | type_vocab_size=self.type_vocab_size, 96 | initializer_range=self.initializer_range) 97 | 98 | model = modeling.BertModel( 99 | config=config, 100 | is_training=self.is_training, 101 | input_ids=input_ids, 102 | input_mask=input_mask, 103 | token_type_ids=token_type_ids, 104 | scope=self.scope) 105 | 106 | outputs = { 107 | "embedding_output": model.get_embedding_output(), 108 | "sequence_output": model.get_sequence_output(), 109 | "pooled_output": model.get_pooled_output(), 110 | "all_encoder_layers": model.get_all_encoder_layers(), 111 | } 112 | return outputs 113 | 114 | def check_output(self, result): 115 | self.parent.assertAllEqual( 116 | result["embedding_output"].shape, 117 | [self.batch_size, self.seq_length, self.hidden_size]) 118 | 119 | self.parent.assertAllEqual( 120 | result["sequence_output"].shape, 121 | [self.batch_size, self.seq_length, self.hidden_size]) 122 | 123 | self.parent.assertAllEqual(result["pooled_output"].shape, 124 | [self.batch_size, self.hidden_size]) 125 | 126 | def test_default(self): 127 | self.run_tester(BertModelTest.BertModelTester(self)) 128 | 129 | def test_config_to_json_string(self): 130 | config = modeling.BertConfig(vocab_size=99, hidden_size=37) 131 | obj = json.loads(config.to_json_string()) 132 | self.assertEqual(obj["vocab_size"], 99) 133 | self.assertEqual(obj["hidden_size"], 37) 134 | 135 | def run_tester(self, tester): 136 | with self.test_session() as sess: 137 | ops = tester.create_model() 138 | init_op = tf.group(tf.global_variables_initializer(), 139 | tf.local_variables_initializer()) 140 | sess.run(init_op) 141 | output_result = sess.run(ops) 142 | tester.check_output(output_result) 143 | 144 | self.assert_all_tensors_reachable(sess, [init_op, ops]) 145 | 146 | @classmethod 147 | def ids_tensor(cls, shape, vocab_size, rng=None, name=None): 148 | """Creates a random int32 tensor of the shape within the vocab size.""" 149 | if rng is None: 150 | rng = random.Random() 151 | 152 | total_dims = 1 153 | for dim in shape: 154 | total_dims *= dim 155 | 156 | values = [] 157 | for _ in range(total_dims): 158 | values.append(rng.randint(0, vocab_size - 1)) 159 | 160 | return tf.constant(value=values, dtype=tf.int32, shape=shape, name=name) 161 | 162 | def assert_all_tensors_reachable(self, sess, outputs): 163 | """Checks that all the tensors in the graph are reachable from outputs.""" 164 | graph = sess.graph 165 | 166 | ignore_strings = [ 167 | "^.*/assert_less_equal/.*$", 168 | "^.*/dilation_rate$", 169 | "^.*/Tensordot/concat$", 170 | "^.*/Tensordot/concat/axis$", 171 | "^testing/.*$", 172 | ] 173 | 174 | ignore_regexes = [re.compile(x) for x in ignore_strings] 175 | 176 | unreachable = self.get_unreachable_ops(graph, outputs) 177 | filtered_unreachable = [] 178 | for x in unreachable: 179 | do_ignore = False 180 | for r in ignore_regexes: 181 | m = r.match(x.name) 182 | if m is not None: 183 | do_ignore = True 184 | if do_ignore: 185 | continue 186 | filtered_unreachable.append(x) 187 | unreachable = filtered_unreachable 188 | 189 | self.assertEqual( 190 | len(unreachable), 0, "The following ops are unreachable: %s" % 191 | (" ".join([x.name for x in unreachable]))) 192 | 193 | @classmethod 194 | def get_unreachable_ops(cls, graph, outputs): 195 | """Finds all of the tensors in graph that are unreachable from outputs.""" 196 | outputs = cls.flatten_recursive(outputs) 197 | output_to_op = collections.defaultdict(list) 198 | op_to_all = collections.defaultdict(list) 199 | assign_out_to_in = collections.defaultdict(list) 200 | 201 | for op in graph.get_operations(): 202 | for x in op.inputs: 203 | op_to_all[op.name].append(x.name) 204 | for y in op.outputs: 205 | output_to_op[y.name].append(op.name) 206 | op_to_all[op.name].append(y.name) 207 | if str(op.type) == "Assign": 208 | for y in op.outputs: 209 | for x in op.inputs: 210 | assign_out_to_in[y.name].append(x.name) 211 | 212 | assign_groups = collections.defaultdict(list) 213 | for out_name in assign_out_to_in.keys(): 214 | name_group = assign_out_to_in[out_name] 215 | for n1 in name_group: 216 | assign_groups[n1].append(out_name) 217 | for n2 in name_group: 218 | if n1 != n2: 219 | assign_groups[n1].append(n2) 220 | 221 | seen_tensors = {} 222 | stack = [x.name for x in outputs] 223 | while stack: 224 | name = stack.pop() 225 | if name in seen_tensors: 226 | continue 227 | seen_tensors[name] = True 228 | 229 | if name in output_to_op: 230 | for op_name in output_to_op[name]: 231 | if op_name in op_to_all: 232 | for input_name in op_to_all[op_name]: 233 | if input_name not in stack: 234 | stack.append(input_name) 235 | 236 | expanded_names = [] 237 | if name in assign_groups: 238 | for assign_name in assign_groups[name]: 239 | expanded_names.append(assign_name) 240 | 241 | for expanded_name in expanded_names: 242 | if expanded_name not in stack: 243 | stack.append(expanded_name) 244 | 245 | unreachable_ops = [] 246 | for op in graph.get_operations(): 247 | is_unreachable = False 248 | all_names = [x.name for x in op.inputs] + [x.name for x in op.outputs] 249 | for name in all_names: 250 | if name not in seen_tensors: 251 | is_unreachable = True 252 | if is_unreachable: 253 | unreachable_ops.append(op) 254 | return unreachable_ops 255 | 256 | @classmethod 257 | def flatten_recursive(cls, item): 258 | """Flattens (potentially nested) a tuple/dictionary/list to a list.""" 259 | output = [] 260 | if isinstance(item, list): 261 | output.extend(item) 262 | elif isinstance(item, tuple): 263 | output.extend(list(item)) 264 | elif isinstance(item, dict): 265 | for (_, v) in six.iteritems(item): 266 | output.append(v) 267 | else: 268 | return [item] 269 | 270 | flat_output = [] 271 | for x in output: 272 | flat_output.extend(cls.flatten_recursive(x)) 273 | return flat_output 274 | 275 | 276 | if __name__ == "__main__": 277 | tf.test.main() 278 | -------------------------------------------------------------------------------- /multilingual.md: -------------------------------------------------------------------------------- 1 | ## Models 2 | 3 | There are two multilingual models currently available. We do not plan to release 4 | more single-language models, but we may release `BERT-Large` versions of these 5 | two in the future: 6 | 7 | * **[`BERT-Base, Multilingual Cased (New, recommended)`](https://storage.googleapis.com/bert_models/2018_11_23/multi_cased_L-12_H-768_A-12.zip)**: 8 | 104 languages, 12-layer, 768-hidden, 12-heads, 110M parameters 9 | * **[`BERT-Base, Multilingual Uncased (Orig, not recommended)`](https://storage.googleapis.com/bert_models/2018_11_03/multilingual_L-12_H-768_A-12.zip)**: 10 | 102 languages, 12-layer, 768-hidden, 12-heads, 110M parameters 11 | * **[`BERT-Base, Chinese`](https://storage.googleapis.com/bert_models/2018_11_03/chinese_L-12_H-768_A-12.zip)**: 12 | Chinese Simplified and Traditional, 12-layer, 768-hidden, 12-heads, 110M 13 | parameters 14 | 15 | **The `Multilingual Cased (New)` model also fixes normalization issues in many 16 | languages, so it is recommended in languages with non-Latin alphabets (and is 17 | often better for most languages with Latin alphabets). When using this model, 18 | make sure to pass `--do_lower_case=false` to `run_pretraining.py` and other 19 | scripts.** 20 | 21 | See the [list of languages](#list-of-languages) that the Multilingual model 22 | supports. The Multilingual model does include Chinese (and English), but if your 23 | fine-tuning data is Chinese-only, then the Chinese model will likely produce 24 | better results. 25 | 26 | ## Results 27 | 28 | To evaluate these systems, we use the 29 | [XNLI dataset](https://github.com/facebookresearch/XNLI) dataset, which is a 30 | version of [MultiNLI](https://www.nyu.edu/projects/bowman/multinli/) where the 31 | dev and test sets have been translated (by humans) into 15 languages. Note that 32 | the training set was *machine* translated (we used the translations provided by 33 | XNLI, not Google NMT). For clarity, we only report on 6 languages below: 34 | 35 | 36 | 37 | | System | English | Chinese | Spanish | German | Arabic | Urdu | 38 | | --------------------------------- | -------- | -------- | -------- | -------- | -------- | -------- | 39 | | XNLI Baseline - Translate Train | 73.7 | 67.0 | 68.8 | 66.5 | 65.8 | 56.6 | 40 | | XNLI Baseline - Translate Test | 73.7 | 68.3 | 70.7 | 68.7 | 66.8 | 59.3 | 41 | | BERT - Translate Train Cased | **81.9** | **76.6** | **77.8** | **75.9** | **70.7** | 61.6 | 42 | | BERT - Translate Train Uncased | 81.4 | 74.2 | 77.3 | 75.2 | 70.5 | 61.7 | 43 | | BERT - Translate Test Uncased | 81.4 | 70.1 | 74.9 | 74.4 | 70.4 | **62.1** | 44 | | BERT - Zero Shot Uncased | 81.4 | 63.8 | 74.3 | 70.5 | 62.1 | 58.3 | 45 | 46 | 47 | 48 | The first two rows are baselines from the XNLI paper and the last three rows are 49 | our results with BERT. 50 | 51 | **Translate Train** means that the MultiNLI training set was machine translated 52 | from English into the foreign language. So training and evaluation were both 53 | done in the foreign language. Unfortunately, training was done on 54 | machine-translated data, so it is impossible to quantify how much of the lower 55 | accuracy (compared to English) is due to the quality of the machine translation 56 | vs. the quality of the pre-trained model. 57 | 58 | **Translate Test** means that the XNLI test set was machine translated from the 59 | foreign language into English. So training and evaluation were both done on 60 | English. However, test evaluation was done on machine-translated English, so the 61 | accuracy depends on the quality of the machine translation system. 62 | 63 | **Zero Shot** means that the Multilingual BERT system was fine-tuned on English 64 | MultiNLI, and then evaluated on the foreign language XNLI test. In this case, 65 | machine translation was not involved at all in either the pre-training or 66 | fine-tuning. 67 | 68 | Note that the English result is worse than the 84.2 MultiNLI baseline because 69 | this training used Multilingual BERT rather than English-only BERT. This implies 70 | that for high-resource languages, the Multilingual model is somewhat worse than 71 | a single-language model. However, it is not feasible for us to train and 72 | maintain dozens of single-language model. Therefore, if your goal is to maximize 73 | performance with a language other than English or Chinese, you might find it 74 | beneficial to run pre-training for additional steps starting from our 75 | Multilingual model on data from your language of interest. 76 | 77 | Here is a comparison of training Chinese models with the Multilingual 78 | `BERT-Base` and Chinese-only `BERT-Base`: 79 | 80 | System | Chinese 81 | ----------------------- | ------- 82 | XNLI Baseline | 67.0 83 | BERT Multilingual Model | 74.2 84 | BERT Chinese-only Model | 77.2 85 | 86 | Similar to English, the single-language model does 3% better than the 87 | Multilingual model. 88 | 89 | ## Fine-tuning Example 90 | 91 | The multilingual model does **not** require any special consideration or API 92 | changes. We did update the implementation of `BasicTokenizer` in 93 | `tokenization.py` to support Chinese character tokenization, so please update if 94 | you forked it. However, we did not change the tokenization API. 95 | 96 | To test the new models, we did modify `run_classifier.py` to add support for the 97 | [XNLI dataset](https://github.com/facebookresearch/XNLI). This is a 15-language 98 | version of MultiNLI where the dev/test sets have been human-translated, and the 99 | training set has been machine-translated. 100 | 101 | To run the fine-tuning code, please download the 102 | [XNLI dev/test set](https://s3.amazonaws.com/xnli/XNLI-1.0.zip) and the 103 | [XNLI machine-translated training set](https://s3.amazonaws.com/xnli/XNLI-MT-1.0.zip) 104 | and then unpack both .zip files into some directory `$XNLI_DIR`. 105 | 106 | To run fine-tuning on XNLI. The language is hard-coded into `run_classifier.py` 107 | (Chinese by default), so please modify `XnliProcessor` if you want to run on 108 | another language. 109 | 110 | This is a large dataset, so this will training will take a few hours on a GPU 111 | (or about 30 minutes on a Cloud TPU). To run an experiment quickly for 112 | debugging, just set `num_train_epochs` to a small value like `0.1`. 113 | 114 | ```shell 115 | export BERT_BASE_DIR=/path/to/bert/chinese_L-12_H-768_A-12 # or multilingual_L-12_H-768_A-12 116 | export XNLI_DIR=/path/to/xnli 117 | 118 | python run_classifier.py \ 119 | --task_name=XNLI \ 120 | --do_train=true \ 121 | --do_eval=true \ 122 | --data_dir=$XNLI_DIR \ 123 | --vocab_file=$BERT_BASE_DIR/vocab.txt \ 124 | --bert_config_file=$BERT_BASE_DIR/bert_config.json \ 125 | --init_checkpoint=$BERT_BASE_DIR/bert_model.ckpt \ 126 | --max_seq_length=128 \ 127 | --train_batch_size=32 \ 128 | --learning_rate=5e-5 \ 129 | --num_train_epochs=2.0 \ 130 | --output_dir=/tmp/xnli_output/ 131 | ``` 132 | 133 | With the Chinese-only model, the results should look something like this: 134 | 135 | ``` 136 | ***** Eval results ***** 137 | eval_accuracy = 0.774116 138 | eval_loss = 0.83554 139 | global_step = 24543 140 | loss = 0.74603 141 | ``` 142 | 143 | ## Details 144 | 145 | ### Data Source and Sampling 146 | 147 | The languages chosen were the 148 | [top 100 languages with the largest Wikipedias](https://meta.wikimedia.org/wiki/List_of_Wikipedias). 149 | The entire Wikipedia dump for each language (excluding user and talk pages) was 150 | taken as the training data for each language 151 | 152 | However, the size of the Wikipedia for a given language varies greatly, and 153 | therefore low-resource languages may be "under-represented" in terms of the 154 | neural network model (under the assumption that languages are "competing" for 155 | limited model capacity to some extent). 156 | 157 | However, the size of a Wikipedia also correlates with the number of speakers of 158 | a language, and we also don't want to overfit the model by performing thousands 159 | of epochs over a tiny Wikipedia for a particular language. 160 | 161 | To balance these two factors, we performed exponentially smoothed weighting of 162 | the data during pre-training data creation (and WordPiece vocab creation). In 163 | other words, let's say that the probability of a language is *P(L)*, e.g., 164 | *P(English) = 0.21* means that after concatenating all of the Wikipedias 165 | together, 21% of our data is English. We exponentiate each probability by some 166 | factor *S* and then re-normalize, and sample from that distribution. In our case 167 | we use *S=0.7*. So, high-resource languages like English will be under-sampled, 168 | and low-resource languages like Icelandic will be over-sampled. E.g., in the 169 | original distribution English would be sampled 1000x more than Icelandic, but 170 | after smoothing it's only sampled 100x more. 171 | 172 | ### Tokenization 173 | 174 | For tokenization, we use a 110k shared WordPiece vocabulary. The word counts are 175 | weighted the same way as the data, so low-resource languages are upweighted by 176 | some factor. We intentionally do *not* use any marker to denote the input 177 | language (so that zero-shot training can work). 178 | 179 | Because Chinese (and Japanese Kanji and Korean Hanja) does not have whitespace 180 | characters, we add spaces around every character in the 181 | [CJK Unicode range](https://en.wikipedia.org/wiki/CJK_Unified_Ideographs_\(Unicode_block\)) 182 | before applying WordPiece. This means that Chinese is effectively 183 | character-tokenized. Note that the CJK Unicode block only includes 184 | Chinese-origin characters and does *not* include Hangul Korean or 185 | Katakana/Hiragana Japanese, which are tokenized with whitespace+WordPiece like 186 | all other languages. 187 | 188 | For all other languages, we apply the 189 | [same recipe as English](https://github.com/google-research/bert#tokenization): 190 | (a) lower casing+accent removal, (b) punctuation splitting, (c) whitespace 191 | tokenization. We understand that accent markers have substantial meaning in some 192 | languages, but felt that the benefits of reducing the effective vocabulary make 193 | up for this. Generally the strong contextual models of BERT should make up for 194 | any ambiguity introduced by stripping accent markers. 195 | 196 | ### List of Languages 197 | 198 | The multilingual model supports the following languages. These languages were 199 | chosen because they are the top 100 languages with the largest Wikipedias: 200 | 201 | * Afrikaans 202 | * Albanian 203 | * Arabic 204 | * Aragonese 205 | * Armenian 206 | * Asturian 207 | * Azerbaijani 208 | * Bashkir 209 | * Basque 210 | * Bavarian 211 | * Belarusian 212 | * Bengali 213 | * Bishnupriya Manipuri 214 | * Bosnian 215 | * Breton 216 | * Bulgarian 217 | * Burmese 218 | * Catalan 219 | * Cebuano 220 | * Chechen 221 | * Chinese (Simplified) 222 | * Chinese (Traditional) 223 | * Chuvash 224 | * Croatian 225 | * Czech 226 | * Danish 227 | * Dutch 228 | * English 229 | * Estonian 230 | * Finnish 231 | * French 232 | * Galician 233 | * Georgian 234 | * German 235 | * Greek 236 | * Gujarati 237 | * Haitian 238 | * Hebrew 239 | * Hindi 240 | * Hungarian 241 | * Icelandic 242 | * Ido 243 | * Indonesian 244 | * Irish 245 | * Italian 246 | * Japanese 247 | * Javanese 248 | * Kannada 249 | * Kazakh 250 | * Kirghiz 251 | * Korean 252 | * Latin 253 | * Latvian 254 | * Lithuanian 255 | * Lombard 256 | * Low Saxon 257 | * Luxembourgish 258 | * Macedonian 259 | * Malagasy 260 | * Malay 261 | * Malayalam 262 | * Marathi 263 | * Minangkabau 264 | * Nepali 265 | * Newar 266 | * Norwegian (Bokmal) 267 | * Norwegian (Nynorsk) 268 | * Occitan 269 | * Persian (Farsi) 270 | * Piedmontese 271 | * Polish 272 | * Portuguese 273 | * Punjabi 274 | * Romanian 275 | * Russian 276 | * Scots 277 | * Serbian 278 | * Serbo-Croatian 279 | * Sicilian 280 | * Slovak 281 | * Slovenian 282 | * South Azerbaijani 283 | * Spanish 284 | * Sundanese 285 | * Swahili 286 | * Swedish 287 | * Tagalog 288 | * Tajik 289 | * Tamil 290 | * Tatar 291 | * Telugu 292 | * Turkish 293 | * Ukrainian 294 | * Urdu 295 | * Uzbek 296 | * Vietnamese 297 | * Volapük 298 | * Waray-Waray 299 | * Welsh 300 | * West 301 | * Western Punjabi 302 | * Yoruba 303 | 304 | The **Multilingual Cased (New)** release contains additionally **Thai** and 305 | **Mongolian**, which were not included in the original release. 306 | -------------------------------------------------------------------------------- /optimization.py: -------------------------------------------------------------------------------- 1 | # coding=utf-8 2 | # Copyright 2018 The Google AI Language Team Authors. 3 | # 4 | # Licensed under the Apache License, Version 2.0 (the "License"); 5 | # you may not use this file except in compliance with the License. 6 | # You may obtain a copy of the License at 7 | # 8 | # http://www.apache.org/licenses/LICENSE-2.0 9 | # 10 | # Unless required by applicable law or agreed to in writing, software 11 | # distributed under the License is distributed on an "AS IS" BASIS, 12 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 13 | # See the License for the specific language governing permissions and 14 | # limitations under the License. 15 | """Functions and classes related to optimization (weight updates).""" 16 | 17 | from __future__ import absolute_import 18 | from __future__ import division 19 | from __future__ import print_function 20 | 21 | import re 22 | import tensorflow as tf 23 | 24 | 25 | def create_optimizer(loss, init_lr, num_train_steps, num_warmup_steps, use_tpu): 26 | """Creates an optimizer training op.""" 27 | global_step = tf.train.get_or_create_global_step() 28 | 29 | learning_rate = tf.constant(value=init_lr, shape=[], dtype=tf.float32) 30 | 31 | # Implements linear decay of the learning rate. 32 | learning_rate = tf.train.polynomial_decay( 33 | learning_rate, 34 | global_step, 35 | num_train_steps, 36 | end_learning_rate=0.0, 37 | power=1.0, 38 | cycle=False) 39 | 40 | # Implements linear warmup. I.e., if global_step < num_warmup_steps, the 41 | # learning rate will be `global_step/num_warmup_steps * init_lr`. 42 | if num_warmup_steps: 43 | global_steps_int = tf.cast(global_step, tf.int32) 44 | warmup_steps_int = tf.constant(num_warmup_steps, dtype=tf.int32) 45 | 46 | global_steps_float = tf.cast(global_steps_int, tf.float32) 47 | warmup_steps_float = tf.cast(warmup_steps_int, tf.float32) 48 | 49 | warmup_percent_done = global_steps_float / warmup_steps_float 50 | warmup_learning_rate = init_lr * warmup_percent_done 51 | 52 | is_warmup = tf.cast(global_steps_int < warmup_steps_int, tf.float32) 53 | learning_rate = ( 54 | (1.0 - is_warmup) * learning_rate + is_warmup * warmup_learning_rate) 55 | 56 | # It is recommended that you use this optimizer for fine tuning, since this 57 | # is how the model was trained (note that the Adam m/v variables are NOT 58 | # loaded from init_checkpoint.) 59 | optimizer = AdamWeightDecayOptimizer( 60 | learning_rate=learning_rate, 61 | weight_decay_rate=0.01, 62 | beta_1=0.9, 63 | beta_2=0.999, 64 | epsilon=1e-6, 65 | exclude_from_weight_decay=["LayerNorm", "layer_norm", "bias"]) 66 | 67 | if use_tpu: 68 | optimizer = tf.contrib.tpu.CrossShardOptimizer(optimizer) 69 | 70 | tvars = tf.trainable_variables() 71 | grads = tf.gradients(loss, tvars) 72 | 73 | # This is how the model was pre-trained. 74 | (grads, _) = tf.clip_by_global_norm(grads, clip_norm=1.0) 75 | 76 | train_op = optimizer.apply_gradients( 77 | zip(grads, tvars), global_step=global_step) 78 | 79 | new_global_step = global_step + 1 80 | train_op = tf.group(train_op, [global_step.assign(new_global_step)]) 81 | return train_op 82 | 83 | 84 | class AdamWeightDecayOptimizer(tf.train.Optimizer): 85 | """A basic Adam optimizer that includes "correct" L2 weight decay.""" 86 | 87 | def __init__(self, 88 | learning_rate, 89 | weight_decay_rate=0.0, 90 | beta_1=0.9, 91 | beta_2=0.999, 92 | epsilon=1e-6, 93 | exclude_from_weight_decay=None, 94 | name="AdamWeightDecayOptimizer"): 95 | """Constructs a AdamWeightDecayOptimizer.""" 96 | super(AdamWeightDecayOptimizer, self).__init__(False, name) 97 | 98 | self.learning_rate = learning_rate 99 | self.weight_decay_rate = weight_decay_rate 100 | self.beta_1 = beta_1 101 | self.beta_2 = beta_2 102 | self.epsilon = epsilon 103 | self.exclude_from_weight_decay = exclude_from_weight_decay 104 | 105 | def apply_gradients(self, grads_and_vars, global_step=None, name=None): 106 | """See base class.""" 107 | assignments = [] 108 | for (grad, param) in grads_and_vars: 109 | if grad is None or param is None: 110 | continue 111 | 112 | param_name = self._get_variable_name(param.name) 113 | 114 | m = tf.get_variable( 115 | name=param_name + "/adam_m", 116 | shape=param.shape.as_list(), 117 | dtype=tf.float32, 118 | trainable=False, 119 | initializer=tf.zeros_initializer()) 120 | v = tf.get_variable( 121 | name=param_name + "/adam_v", 122 | shape=param.shape.as_list(), 123 | dtype=tf.float32, 124 | trainable=False, 125 | initializer=tf.zeros_initializer()) 126 | 127 | # Standard Adam update. 128 | next_m = ( 129 | tf.multiply(self.beta_1, m) + tf.multiply(1.0 - self.beta_1, grad)) 130 | next_v = ( 131 | tf.multiply(self.beta_2, v) + tf.multiply(1.0 - self.beta_2, 132 | tf.square(grad))) 133 | 134 | update = next_m / (tf.sqrt(next_v) + self.epsilon) 135 | 136 | # Just adding the square of the weights to the loss function is *not* 137 | # the correct way of using L2 regularization/weight decay with Adam, 138 | # since that will interact with the m and v parameters in strange ways. 139 | # 140 | # Instead we want ot decay the weights in a manner that doesn't interact 141 | # with the m/v parameters. This is equivalent to adding the square 142 | # of the weights to the loss with plain (non-momentum) SGD. 143 | if self._do_use_weight_decay(param_name): 144 | update += self.weight_decay_rate * param 145 | 146 | update_with_lr = self.learning_rate * update 147 | 148 | next_param = param - update_with_lr 149 | 150 | assignments.extend( 151 | [param.assign(next_param), 152 | m.assign(next_m), 153 | v.assign(next_v)]) 154 | return tf.group(*assignments, name=name) 155 | 156 | def _do_use_weight_decay(self, param_name): 157 | """Whether to use L2 weight decay for `param_name`.""" 158 | if not self.weight_decay_rate: 159 | return False 160 | if self.exclude_from_weight_decay: 161 | for r in self.exclude_from_weight_decay: 162 | if re.search(r, param_name) is not None: 163 | return False 164 | return True 165 | 166 | def _get_variable_name(self, param_name): 167 | """Get the variable name from the tensor name.""" 168 | m = re.match("^(.*):\\d+$", param_name) 169 | if m is not None: 170 | param_name = m.group(1) 171 | return param_name 172 | -------------------------------------------------------------------------------- /optimization_test.py: -------------------------------------------------------------------------------- 1 | # coding=utf-8 2 | # Copyright 2018 The Google AI Language Team Authors. 3 | # 4 | # Licensed under the Apache License, Version 2.0 (the "License"); 5 | # you may not use this file except in compliance with the License. 6 | # You may obtain a copy of the License at 7 | # 8 | # http://www.apache.org/licenses/LICENSE-2.0 9 | # 10 | # Unless required by applicable law or agreed to in writing, software 11 | # distributed under the License is distributed on an "AS IS" BASIS, 12 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 13 | # See the License for the specific language governing permissions and 14 | # limitations under the License. 15 | from __future__ import absolute_import 16 | from __future__ import division 17 | from __future__ import print_function 18 | 19 | import optimization 20 | import tensorflow as tf 21 | 22 | 23 | class OptimizationTest(tf.test.TestCase): 24 | 25 | def test_adam(self): 26 | with self.test_session() as sess: 27 | w = tf.get_variable( 28 | "w", 29 | shape=[3], 30 | initializer=tf.constant_initializer([0.1, -0.2, -0.1])) 31 | x = tf.constant([0.4, 0.2, -0.5]) 32 | loss = tf.reduce_mean(tf.square(x - w)) 33 | tvars = tf.trainable_variables() 34 | grads = tf.gradients(loss, tvars) 35 | global_step = tf.train.get_or_create_global_step() 36 | optimizer = optimization.AdamWeightDecayOptimizer(learning_rate=0.2) 37 | train_op = optimizer.apply_gradients(zip(grads, tvars), global_step) 38 | init_op = tf.group(tf.global_variables_initializer(), 39 | tf.local_variables_initializer()) 40 | sess.run(init_op) 41 | for _ in range(100): 42 | sess.run(train_op) 43 | w_np = sess.run(w) 44 | self.assertAllClose(w_np.flat, [0.4, 0.2, -0.5], rtol=1e-2, atol=1e-2) 45 | 46 | 47 | if __name__ == "__main__": 48 | tf.test.main() 49 | -------------------------------------------------------------------------------- /requirements.txt: -------------------------------------------------------------------------------- 1 | tensorflow >= 1.11.0 # CPU Version of TensorFlow. 2 | # tensorflow-gpu >= 1.11.0 # GPU version of TensorFlow. 3 | -------------------------------------------------------------------------------- /run_classifier.py: -------------------------------------------------------------------------------- 1 | # coding=utf-8 2 | # Copyright 2018 The Google AI Language Team Authors. 3 | # 4 | # Licensed under the Apache License, Version 2.0 (the "License"); 5 | # you may not use this file except in compliance with the License. 6 | # You may obtain a copy of the License at 7 | # 8 | # http://www.apache.org/licenses/LICENSE-2.0 9 | # 10 | # Unless required by applicable law or agreed to in writing, software 11 | # distributed under the License is distributed on an "AS IS" BASIS, 12 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 13 | # See the License for the specific language governing permissions and 14 | # limitations under the License. 15 | """BERT finetuning runner.""" 16 | 17 | from __future__ import absolute_import 18 | from __future__ import division 19 | from __future__ import print_function 20 | 21 | import collections 22 | import csv 23 | import os 24 | import modeling 25 | import optimization 26 | import tokenization 27 | import tensorflow as tf 28 | 29 | flags = tf.flags 30 | 31 | FLAGS = flags.FLAGS 32 | 33 | ## Required parameters 34 | flags.DEFINE_string( 35 | "data_dir", None, 36 | "The input data dir. Should contain the .tsv files (or other data files) " 37 | "for the task.") 38 | 39 | flags.DEFINE_string( 40 | "bert_config_file", None, 41 | "The config json file corresponding to the pre-trained BERT model. " 42 | "This specifies the model architecture.") 43 | 44 | flags.DEFINE_string("task_name", None, "The name of the task to train.") 45 | 46 | flags.DEFINE_string("vocab_file", None, 47 | "The vocabulary file that the BERT model was trained on.") 48 | 49 | flags.DEFINE_string( 50 | "output_dir", None, 51 | "The output directory where the model checkpoints will be written.") 52 | 53 | ## Other parameters 54 | 55 | flags.DEFINE_string( 56 | "init_checkpoint", None, 57 | "Initial checkpoint (usually from a pre-trained BERT model).") 58 | 59 | flags.DEFINE_bool( 60 | "do_lower_case", True, 61 | "Whether to lower case the input text. Should be True for uncased " 62 | "models and False for cased models.") 63 | 64 | flags.DEFINE_integer( 65 | "max_seq_length", 128, 66 | "The maximum total input sequence length after WordPiece tokenization. " 67 | "Sequences longer than this will be truncated, and sequences shorter " 68 | "than this will be padded.") 69 | 70 | flags.DEFINE_bool("do_train", False, "Whether to run training.") 71 | 72 | flags.DEFINE_bool("do_eval", False, "Whether to run eval on the dev set.") 73 | 74 | flags.DEFINE_bool( 75 | "do_predict", False, 76 | "Whether to run the model in inference mode on the test set.") 77 | 78 | flags.DEFINE_integer("train_batch_size", 32, "Total batch size for training.") 79 | 80 | flags.DEFINE_integer("eval_batch_size", 8, "Total batch size for eval.") 81 | 82 | flags.DEFINE_integer("predict_batch_size", 8, "Total batch size for predict.") 83 | 84 | flags.DEFINE_float("learning_rate", 5e-5, "The initial learning rate for Adam.") 85 | 86 | flags.DEFINE_float("num_train_epochs", 3.0, 87 | "Total number of training epochs to perform.") 88 | 89 | flags.DEFINE_float( 90 | "warmup_proportion", 0.1, 91 | "Proportion of training to perform linear learning rate warmup for. " 92 | "E.g., 0.1 = 10% of training.") 93 | 94 | flags.DEFINE_integer("save_checkpoints_steps", 1000, 95 | "How often to save the model checkpoint.") 96 | 97 | flags.DEFINE_integer("iterations_per_loop", 1000, 98 | "How many steps to make in each estimator call.") 99 | 100 | flags.DEFINE_bool("use_tpu", False, "Whether to use TPU or GPU/CPU.") 101 | 102 | tf.flags.DEFINE_string( 103 | "tpu_name", None, 104 | "The Cloud TPU to use for training. This should be either the name " 105 | "used when creating the Cloud TPU, or a grpc://ip.address.of.tpu:8470 " 106 | "url.") 107 | 108 | tf.flags.DEFINE_string( 109 | "tpu_zone", None, 110 | "[Optional] GCE zone where the Cloud TPU is located in. If not " 111 | "specified, we will attempt to automatically detect the GCE project from " 112 | "metadata.") 113 | 114 | tf.flags.DEFINE_string( 115 | "gcp_project", None, 116 | "[Optional] Project name for the Cloud TPU-enabled project. If not " 117 | "specified, we will attempt to automatically detect the GCE project from " 118 | "metadata.") 119 | 120 | tf.flags.DEFINE_string("master", None, "[Optional] TensorFlow master URL.") 121 | 122 | flags.DEFINE_integer( 123 | "num_tpu_cores", 8, 124 | "Only used if `use_tpu` is True. Total number of TPU cores to use.") 125 | 126 | 127 | class InputExample(object): 128 | """A single training/test example for simple sequence classification.""" 129 | 130 | def __init__(self, guid, text_a, text_b=None, label=None): 131 | """Constructs a InputExample. 132 | 133 | Args: 134 | guid: Unique id for the example. 135 | text_a: string. The untokenized text of the first sequence. For single 136 | sequence tasks, only this sequence must be specified. 137 | text_b: (Optional) string. The untokenized text of the second sequence. 138 | Only must be specified for sequence pair tasks. 139 | label: (Optional) string. The label of the example. This should be 140 | specified for train and dev examples, but not for test examples. 141 | """ 142 | self.guid = guid 143 | self.text_a = text_a 144 | self.text_b = text_b 145 | self.label = label 146 | 147 | 148 | class InputFeatures(object): 149 | """A single set of features of data.""" 150 | 151 | def __init__(self, input_ids, input_mask, segment_ids, label_id): 152 | self.input_ids = input_ids 153 | self.input_mask = input_mask 154 | self.segment_ids = segment_ids 155 | self.label_id = label_id 156 | 157 | 158 | class DataProcessor(object): 159 | """Base class for data converters for sequence classification data sets.""" 160 | 161 | def get_train_examples(self, data_dir): 162 | """Gets a collection of `InputExample`s for the train set.""" 163 | raise NotImplementedError() 164 | 165 | def get_dev_examples(self, data_dir): 166 | """Gets a collection of `InputExample`s for the dev set.""" 167 | raise NotImplementedError() 168 | 169 | def get_test_examples(self, data_dir): 170 | """Gets a collection of `InputExample`s for prediction.""" 171 | raise NotImplementedError() 172 | 173 | def get_labels(self): 174 | """Gets the list of labels for this data set.""" 175 | raise NotImplementedError() 176 | 177 | @classmethod 178 | def _read_tsv(cls, input_file, quotechar=None): 179 | """Reads a tab separated value file.""" 180 | with tf.gfile.Open(input_file, "r") as f: 181 | reader = csv.reader(f, delimiter="\t", quotechar=quotechar) 182 | lines = [] 183 | for line in reader: 184 | lines.append(line) 185 | return lines 186 | 187 | 188 | class XnliProcessor(DataProcessor): 189 | """Processor for the XNLI data set.""" 190 | 191 | def __init__(self): 192 | self.language = "zh" 193 | 194 | def get_train_examples(self, data_dir): 195 | """See base class.""" 196 | lines = self._read_tsv( 197 | os.path.join(data_dir, "multinli", 198 | "multinli.train.%s.tsv" % self.language)) 199 | examples = [] 200 | for (i, line) in enumerate(lines): 201 | if i == 0: 202 | continue 203 | guid = "train-%d" % (i) 204 | text_a = tokenization.convert_to_unicode(line[0]) 205 | text_b = tokenization.convert_to_unicode(line[1]) 206 | label = tokenization.convert_to_unicode(line[2]) 207 | if label == tokenization.convert_to_unicode("contradictory"): 208 | label = tokenization.convert_to_unicode("contradiction") 209 | examples.append( 210 | InputExample(guid=guid, text_a=text_a, text_b=text_b, label=label)) 211 | return examples 212 | 213 | def get_dev_examples(self, data_dir): 214 | """See base class.""" 215 | lines = self._read_tsv(os.path.join(data_dir, "xnli.dev.tsv")) 216 | examples = [] 217 | for (i, line) in enumerate(lines): 218 | if i == 0: 219 | continue 220 | guid = "dev-%d" % (i) 221 | language = tokenization.convert_to_unicode(line[0]) 222 | if language != tokenization.convert_to_unicode(self.language): 223 | continue 224 | text_a = tokenization.convert_to_unicode(line[6]) 225 | text_b = tokenization.convert_to_unicode(line[7]) 226 | label = tokenization.convert_to_unicode(line[1]) 227 | examples.append( 228 | InputExample(guid=guid, text_a=text_a, text_b=text_b, label=label)) 229 | return examples 230 | 231 | def get_labels(self): 232 | """See base class.""" 233 | return ["contradiction", "entailment", "neutral"] 234 | 235 | 236 | class MnliProcessor(DataProcessor): 237 | """Processor for the MultiNLI data set (GLUE version).""" 238 | 239 | def get_train_examples(self, data_dir): 240 | """See base class.""" 241 | return self._create_examples( 242 | self._read_tsv(os.path.join(data_dir, "train.tsv")), "train") 243 | 244 | def get_dev_examples(self, data_dir): 245 | """See base class.""" 246 | return self._create_examples( 247 | self._read_tsv(os.path.join(data_dir, "dev_matched.tsv")), 248 | "dev_matched") 249 | 250 | def get_test_examples(self, data_dir): 251 | """See base class.""" 252 | return self._create_examples( 253 | self._read_tsv(os.path.join(data_dir, "test_matched.tsv")), "test") 254 | 255 | def get_labels(self): 256 | """See base class.""" 257 | return ["contradiction", "entailment", "neutral"] 258 | 259 | def _create_examples(self, lines, set_type): 260 | """Creates examples for the training and dev sets.""" 261 | examples = [] 262 | for (i, line) in enumerate(lines): 263 | if i == 0: 264 | continue 265 | guid = "%s-%s" % (set_type, tokenization.convert_to_unicode(line[0])) 266 | text_a = tokenization.convert_to_unicode(line[8]) 267 | text_b = tokenization.convert_to_unicode(line[9]) 268 | if set_type == "test": 269 | label = "contradiction" 270 | else: 271 | label = tokenization.convert_to_unicode(line[-1]) 272 | examples.append( 273 | InputExample(guid=guid, text_a=text_a, text_b=text_b, label=label)) 274 | return examples 275 | 276 | 277 | class MrpcProcessor(DataProcessor): 278 | """Processor for the MRPC data set (GLUE version).""" 279 | 280 | def get_train_examples(self, data_dir): 281 | """See base class.""" 282 | return self._create_examples( 283 | self._read_tsv(os.path.join(data_dir, "train.tsv")), "train") 284 | 285 | def get_dev_examples(self, data_dir): 286 | """See base class.""" 287 | return self._create_examples( 288 | self._read_tsv(os.path.join(data_dir, "dev.tsv")), "dev") 289 | 290 | def get_test_examples(self, data_dir): 291 | """See base class.""" 292 | return self._create_examples( 293 | self._read_tsv(os.path.join(data_dir, "test.tsv")), "test") 294 | 295 | def get_labels(self): 296 | """See base class.""" 297 | return ["0", "1"] 298 | 299 | def _create_examples(self, lines, set_type): 300 | """Creates examples for the training and dev sets.""" 301 | examples = [] 302 | for (i, line) in enumerate(lines): 303 | if i == 0: 304 | continue 305 | guid = "%s-%s" % (set_type, i) 306 | text_a = tokenization.convert_to_unicode(line[3]) 307 | text_b = tokenization.convert_to_unicode(line[4]) 308 | if set_type == "test": 309 | label = "0" 310 | else: 311 | label = tokenization.convert_to_unicode(line[0]) 312 | examples.append( 313 | InputExample(guid=guid, text_a=text_a, text_b=text_b, label=label)) 314 | return examples 315 | 316 | 317 | class ColaProcessor(DataProcessor): 318 | """Processor for the CoLA data set (GLUE version).""" 319 | 320 | def get_train_examples(self, data_dir): 321 | """See base class.""" 322 | return self._create_examples( 323 | self._read_tsv(os.path.join(data_dir, "train.tsv")), "train") 324 | 325 | def get_dev_examples(self, data_dir): 326 | """See base class.""" 327 | return self._create_examples( 328 | self._read_tsv(os.path.join(data_dir, "dev.tsv")), "dev") 329 | 330 | def get_test_examples(self, data_dir): 331 | """See base class.""" 332 | return self._create_examples( 333 | self._read_tsv(os.path.join(data_dir, "test.tsv")), "test") 334 | 335 | def get_labels(self): 336 | """See base class.""" 337 | return ["0", "1"] 338 | 339 | def _create_examples(self, lines, set_type): 340 | """Creates examples for the training and dev sets.""" 341 | examples = [] 342 | for (i, line) in enumerate(lines): 343 | # Only the test set has a header 344 | if set_type == "test" and i == 0: 345 | continue 346 | guid = "%s-%s" % (set_type, i) 347 | if set_type == "test": 348 | text_a = tokenization.convert_to_unicode(line[1]) 349 | label = "0" 350 | else: 351 | text_a = tokenization.convert_to_unicode(line[3]) 352 | label = tokenization.convert_to_unicode(line[1]) 353 | examples.append( 354 | InputExample(guid=guid, text_a=text_a, text_b=None, label=label)) 355 | return examples 356 | 357 | 358 | def convert_single_example(ex_index, example, label_list, max_seq_length, 359 | tokenizer): 360 | """Converts a single `InputExample` into a single `InputFeatures`.""" 361 | label_map = {} 362 | for (i, label) in enumerate(label_list): 363 | label_map[label] = i 364 | 365 | tokens_a = tokenizer.tokenize(example.text_a) 366 | tokens_b = None 367 | if example.text_b: 368 | tokens_b = tokenizer.tokenize(example.text_b) 369 | 370 | if tokens_b: 371 | # Modifies `tokens_a` and `tokens_b` in place so that the total 372 | # length is less than the specified length. 373 | # Account for [CLS], [SEP], [SEP] with "- 3" 374 | _truncate_seq_pair(tokens_a, tokens_b, max_seq_length - 3) 375 | else: 376 | # Account for [CLS] and [SEP] with "- 2" 377 | if len(tokens_a) > max_seq_length - 2: 378 | tokens_a = tokens_a[0:(max_seq_length - 2)] 379 | 380 | # The convention in BERT is: 381 | # (a) For sequence pairs: 382 | # tokens: [CLS] is this jack ##son ##ville ? [SEP] no it is not . [SEP] 383 | # type_ids: 0 0 0 0 0 0 0 0 1 1 1 1 1 1 384 | # (b) For single sequences: 385 | # tokens: [CLS] the dog is hairy . [SEP] 386 | # type_ids: 0 0 0 0 0 0 0 387 | # 388 | # Where "type_ids" are used to indicate whether this is the first 389 | # sequence or the second sequence. The embedding vectors for `type=0` and 390 | # `type=1` were learned during pre-training and are added to the wordpiece 391 | # embedding vector (and position vector). This is not *strictly* necessary 392 | # since the [SEP] token unambiguously separates the sequences, but it makes 393 | # it easier for the model to learn the concept of sequences. 394 | # 395 | # For classification tasks, the first vector (corresponding to [CLS]) is 396 | # used as as the "sentence vector". Note that this only makes sense because 397 | # the entire model is fine-tuned. 398 | tokens = [] 399 | segment_ids = [] 400 | tokens.append("[CLS]") 401 | segment_ids.append(0) 402 | for token in tokens_a: 403 | tokens.append(token) 404 | segment_ids.append(0) 405 | tokens.append("[SEP]") 406 | segment_ids.append(0) 407 | 408 | if tokens_b: 409 | for token in tokens_b: 410 | tokens.append(token) 411 | segment_ids.append(1) 412 | tokens.append("[SEP]") 413 | segment_ids.append(1) 414 | 415 | input_ids = tokenizer.convert_tokens_to_ids(tokens) 416 | 417 | # The mask has 1 for real tokens and 0 for padding tokens. Only real 418 | # tokens are attended to. 419 | input_mask = [1] * len(input_ids) 420 | 421 | # Zero-pad up to the sequence length. 422 | while len(input_ids) < max_seq_length: 423 | input_ids.append(0) 424 | input_mask.append(0) 425 | segment_ids.append(0) 426 | 427 | assert len(input_ids) == max_seq_length 428 | assert len(input_mask) == max_seq_length 429 | assert len(segment_ids) == max_seq_length 430 | 431 | label_id = label_map[example.label] 432 | if ex_index < 5: 433 | tf.logging.info("*** Example ***") 434 | tf.logging.info("guid: %s" % (example.guid)) 435 | tf.logging.info("tokens: %s" % " ".join( 436 | [tokenization.printable_text(x) for x in tokens])) 437 | tf.logging.info("input_ids: %s" % " ".join([str(x) for x in input_ids])) 438 | tf.logging.info("input_mask: %s" % " ".join([str(x) for x in input_mask])) 439 | tf.logging.info("segment_ids: %s" % " ".join([str(x) for x in segment_ids])) 440 | tf.logging.info("label: %s (id = %d)" % (example.label, label_id)) 441 | 442 | feature = InputFeatures( 443 | input_ids=input_ids, 444 | input_mask=input_mask, 445 | segment_ids=segment_ids, 446 | label_id=label_id) 447 | return feature 448 | 449 | 450 | def file_based_convert_examples_to_features( 451 | examples, label_list, max_seq_length, tokenizer, output_file): 452 | """Convert a set of `InputExample`s to a TFRecord file.""" 453 | 454 | writer = tf.python_io.TFRecordWriter(output_file) 455 | 456 | for (ex_index, example) in enumerate(examples): 457 | if ex_index % 10000 == 0: 458 | tf.logging.info("Writing example %d of %d" % (ex_index, len(examples))) 459 | 460 | feature = convert_single_example(ex_index, example, label_list, 461 | max_seq_length, tokenizer) 462 | 463 | def create_int_feature(values): 464 | f = tf.train.Feature(int64_list=tf.train.Int64List(value=list(values))) 465 | return f 466 | 467 | features = collections.OrderedDict() 468 | features["input_ids"] = create_int_feature(feature.input_ids) 469 | features["input_mask"] = create_int_feature(feature.input_mask) 470 | features["segment_ids"] = create_int_feature(feature.segment_ids) 471 | features["label_ids"] = create_int_feature([feature.label_id]) 472 | 473 | tf_example = tf.train.Example(features=tf.train.Features(feature=features)) 474 | writer.write(tf_example.SerializeToString()) 475 | 476 | 477 | def file_based_input_fn_builder(input_file, seq_length, is_training, 478 | drop_remainder): 479 | """Creates an `input_fn` closure to be passed to TPUEstimator.""" 480 | 481 | name_to_features = { 482 | "input_ids": tf.FixedLenFeature([seq_length], tf.int64), 483 | "input_mask": tf.FixedLenFeature([seq_length], tf.int64), 484 | "segment_ids": tf.FixedLenFeature([seq_length], tf.int64), 485 | "label_ids": tf.FixedLenFeature([], tf.int64), 486 | } 487 | 488 | def _decode_record(record, name_to_features): 489 | """Decodes a record to a TensorFlow example.""" 490 | example = tf.parse_single_example(record, name_to_features) 491 | 492 | # tf.Example only supports tf.int64, but the TPU only supports tf.int32. 493 | # So cast all int64 to int32. 494 | for name in list(example.keys()): 495 | t = example[name] 496 | if t.dtype == tf.int64: 497 | t = tf.to_int32(t) 498 | example[name] = t 499 | 500 | return example 501 | 502 | def input_fn(params): 503 | """The actual input function.""" 504 | batch_size = params["batch_size"] 505 | 506 | # For training, we want a lot of parallel reading and shuffling. 507 | # For eval, we want no shuffling and parallel reading doesn't matter. 508 | d = tf.data.TFRecordDataset(input_file) 509 | if is_training: 510 | d = d.repeat() 511 | d = d.shuffle(buffer_size=100) 512 | 513 | d = d.apply( 514 | tf.contrib.data.map_and_batch( 515 | lambda record: _decode_record(record, name_to_features), 516 | batch_size=batch_size, 517 | drop_remainder=drop_remainder)) 518 | 519 | return d 520 | 521 | return input_fn 522 | 523 | 524 | def _truncate_seq_pair(tokens_a, tokens_b, max_length): 525 | """Truncates a sequence pair in place to the maximum length.""" 526 | 527 | # This is a simple heuristic which will always truncate the longer sequence 528 | # one token at a time. This makes more sense than truncating an equal percent 529 | # of tokens from each, since if one sequence is very short then each token 530 | # that's truncated likely contains more information than a longer sequence. 531 | while True: 532 | total_length = len(tokens_a) + len(tokens_b) 533 | if total_length <= max_length: 534 | break 535 | if len(tokens_a) > len(tokens_b): 536 | tokens_a.pop() 537 | else: 538 | tokens_b.pop() 539 | 540 | 541 | def create_model(bert_config, is_training, input_ids, input_mask, segment_ids, 542 | labels, num_labels, use_one_hot_embeddings): 543 | """Creates a classification model.""" 544 | model = modeling.BertModel( 545 | config=bert_config, 546 | is_training=is_training, 547 | input_ids=input_ids, 548 | input_mask=input_mask, 549 | token_type_ids=segment_ids, 550 | use_one_hot_embeddings=use_one_hot_embeddings) 551 | 552 | # In the demo, we are doing a simple classification task on the entire 553 | # segment. 554 | # 555 | # If you want to use the token-level output, use model.get_sequence_output() 556 | # instead. 557 | output_layer = model.get_pooled_output() 558 | 559 | hidden_size = output_layer.shape[-1].value 560 | 561 | output_weights = tf.get_variable( 562 | "output_weights", [num_labels, hidden_size], 563 | initializer=tf.truncated_normal_initializer(stddev=0.02)) 564 | 565 | output_bias = tf.get_variable( 566 | "output_bias", [num_labels], initializer=tf.zeros_initializer()) 567 | 568 | with tf.variable_scope("loss"): 569 | if is_training: 570 | # I.e., 0.1 dropout 571 | output_layer = tf.nn.dropout(output_layer, keep_prob=0.9) 572 | 573 | logits = tf.matmul(output_layer, output_weights, transpose_b=True) 574 | logits = tf.nn.bias_add(logits, output_bias) 575 | probabilities = tf.nn.softmax(logits, axis=-1) 576 | log_probs = tf.nn.log_softmax(logits, axis=-1) 577 | 578 | one_hot_labels = tf.one_hot(labels, depth=num_labels, dtype=tf.float32) 579 | 580 | per_example_loss = -tf.reduce_sum(one_hot_labels * log_probs, axis=-1) 581 | loss = tf.reduce_mean(per_example_loss) 582 | 583 | return (loss, per_example_loss, logits, probabilities) 584 | 585 | 586 | def model_fn_builder(bert_config, num_labels, init_checkpoint, learning_rate, 587 | num_train_steps, num_warmup_steps, use_tpu, 588 | use_one_hot_embeddings): 589 | """Returns `model_fn` closure for TPUEstimator.""" 590 | 591 | def model_fn(features, labels, mode, params): # pylint: disable=unused-argument 592 | """The `model_fn` for TPUEstimator.""" 593 | 594 | tf.logging.info("*** Features ***") 595 | for name in sorted(features.keys()): 596 | tf.logging.info(" name = %s, shape = %s" % (name, features[name].shape)) 597 | 598 | input_ids = features["input_ids"] 599 | input_mask = features["input_mask"] 600 | segment_ids = features["segment_ids"] 601 | label_ids = features["label_ids"] 602 | 603 | is_training = (mode == tf.estimator.ModeKeys.TRAIN) 604 | 605 | (total_loss, per_example_loss, logits, probabilities) = create_model( 606 | bert_config, is_training, input_ids, input_mask, segment_ids, label_ids, 607 | num_labels, use_one_hot_embeddings) 608 | 609 | tvars = tf.trainable_variables() 610 | initialized_variable_names = {} 611 | scaffold_fn = None 612 | if init_checkpoint: 613 | (assignment_map, initialized_variable_names 614 | ) = modeling.get_assignment_map_from_checkpoint(tvars, init_checkpoint) 615 | if use_tpu: 616 | 617 | def tpu_scaffold(): 618 | tf.train.init_from_checkpoint(init_checkpoint, assignment_map) 619 | return tf.train.Scaffold() 620 | 621 | scaffold_fn = tpu_scaffold 622 | else: 623 | tf.train.init_from_checkpoint(init_checkpoint, assignment_map) 624 | 625 | tf.logging.info("**** Trainable Variables ****") 626 | for var in tvars: 627 | init_string = "" 628 | if var.name in initialized_variable_names: 629 | init_string = ", *INIT_FROM_CKPT*" 630 | tf.logging.info(" name = %s, shape = %s%s", var.name, var.shape, 631 | init_string) 632 | 633 | output_spec = None 634 | if mode == tf.estimator.ModeKeys.TRAIN: 635 | 636 | train_op = optimization.create_optimizer( 637 | total_loss, learning_rate, num_train_steps, num_warmup_steps, use_tpu) 638 | 639 | output_spec = tf.contrib.tpu.TPUEstimatorSpec( 640 | mode=mode, 641 | loss=total_loss, 642 | train_op=train_op, 643 | scaffold_fn=scaffold_fn) 644 | elif mode == tf.estimator.ModeKeys.EVAL: 645 | 646 | def metric_fn(per_example_loss, label_ids, logits): 647 | predictions = tf.argmax(logits, axis=-1, output_type=tf.int32) 648 | accuracy = tf.metrics.accuracy(label_ids, predictions) 649 | loss = tf.metrics.mean(per_example_loss) 650 | return { 651 | "eval_accuracy": accuracy, 652 | "eval_loss": loss, 653 | } 654 | 655 | eval_metrics = (metric_fn, [per_example_loss, label_ids, logits]) 656 | output_spec = tf.contrib.tpu.TPUEstimatorSpec( 657 | mode=mode, 658 | loss=total_loss, 659 | eval_metrics=eval_metrics, 660 | scaffold_fn=scaffold_fn) 661 | else: 662 | output_spec = tf.contrib.tpu.TPUEstimatorSpec( 663 | mode=mode, predictions=probabilities, scaffold_fn=scaffold_fn) 664 | return output_spec 665 | 666 | return model_fn 667 | 668 | 669 | # This function is not used by this file but is still used by the Colab and 670 | # people who depend on it. 671 | def input_fn_builder(features, seq_length, is_training, drop_remainder): 672 | """Creates an `input_fn` closure to be passed to TPUEstimator.""" 673 | 674 | all_input_ids = [] 675 | all_input_mask = [] 676 | all_segment_ids = [] 677 | all_label_ids = [] 678 | 679 | for feature in features: 680 | all_input_ids.append(feature.input_ids) 681 | all_input_mask.append(feature.input_mask) 682 | all_segment_ids.append(feature.segment_ids) 683 | all_label_ids.append(feature.label_id) 684 | 685 | def input_fn(params): 686 | """The actual input function.""" 687 | batch_size = params["batch_size"] 688 | 689 | num_examples = len(features) 690 | 691 | # This is for demo purposes and does NOT scale to large data sets. We do 692 | # not use Dataset.from_generator() because that uses tf.py_func which is 693 | # not TPU compatible. The right way to load data is with TFRecordReader. 694 | d = tf.data.Dataset.from_tensor_slices({ 695 | "input_ids": 696 | tf.constant( 697 | all_input_ids, shape=[num_examples, seq_length], 698 | dtype=tf.int32), 699 | "input_mask": 700 | tf.constant( 701 | all_input_mask, 702 | shape=[num_examples, seq_length], 703 | dtype=tf.int32), 704 | "segment_ids": 705 | tf.constant( 706 | all_segment_ids, 707 | shape=[num_examples, seq_length], 708 | dtype=tf.int32), 709 | "label_ids": 710 | tf.constant(all_label_ids, shape=[num_examples], dtype=tf.int32), 711 | }) 712 | 713 | if is_training: 714 | d = d.repeat() 715 | d = d.shuffle(buffer_size=100) 716 | 717 | d = d.batch(batch_size=batch_size, drop_remainder=drop_remainder) 718 | return d 719 | 720 | return input_fn 721 | 722 | 723 | # This function is not used by this file but is still used by the Colab and 724 | # people who depend on it. 725 | def convert_examples_to_features(examples, label_list, max_seq_length, 726 | tokenizer): 727 | """Convert a set of `InputExample`s to a list of `InputFeatures`.""" 728 | 729 | features = [] 730 | for (ex_index, example) in enumerate(examples): 731 | if ex_index % 10000 == 0: 732 | tf.logging.info("Writing example %d of %d" % (ex_index, len(examples))) 733 | 734 | feature = convert_single_example(ex_index, example, label_list, 735 | max_seq_length, tokenizer) 736 | 737 | features.append(feature) 738 | return features 739 | 740 | 741 | def main(_): 742 | tf.logging.set_verbosity(tf.logging.INFO) 743 | 744 | processors = { 745 | "cola": ColaProcessor, 746 | "mnli": MnliProcessor, 747 | "mrpc": MrpcProcessor, 748 | "xnli": XnliProcessor, 749 | } 750 | 751 | if not FLAGS.do_train and not FLAGS.do_eval and not FLAGS.do_predict: 752 | raise ValueError( 753 | "At least one of `do_train`, `do_eval` or `do_predict' must be True.") 754 | 755 | bert_config = modeling.BertConfig.from_json_file(FLAGS.bert_config_file) 756 | 757 | if FLAGS.max_seq_length > bert_config.max_position_embeddings: 758 | raise ValueError( 759 | "Cannot use sequence length %d because the BERT model " 760 | "was only trained up to sequence length %d" % 761 | (FLAGS.max_seq_length, bert_config.max_position_embeddings)) 762 | 763 | tf.gfile.MakeDirs(FLAGS.output_dir) 764 | 765 | task_name = FLAGS.task_name.lower() 766 | 767 | if task_name not in processors: 768 | raise ValueError("Task not found: %s" % (task_name)) 769 | 770 | processor = processors[task_name]() 771 | 772 | label_list = processor.get_labels() 773 | 774 | tokenizer = tokenization.FullTokenizer( 775 | vocab_file=FLAGS.vocab_file, do_lower_case=FLAGS.do_lower_case) 776 | 777 | tpu_cluster_resolver = None 778 | if FLAGS.use_tpu and FLAGS.tpu_name: 779 | tpu_cluster_resolver = tf.contrib.cluster_resolver.TPUClusterResolver( 780 | FLAGS.tpu_name, zone=FLAGS.tpu_zone, project=FLAGS.gcp_project) 781 | 782 | is_per_host = tf.contrib.tpu.InputPipelineConfig.PER_HOST_V2 783 | run_config = tf.contrib.tpu.RunConfig( 784 | cluster=tpu_cluster_resolver, 785 | master=FLAGS.master, 786 | model_dir=FLAGS.output_dir, 787 | save_checkpoints_steps=FLAGS.save_checkpoints_steps, 788 | tpu_config=tf.contrib.tpu.TPUConfig( 789 | iterations_per_loop=FLAGS.iterations_per_loop, 790 | num_shards=FLAGS.num_tpu_cores, 791 | per_host_input_for_training=is_per_host)) 792 | 793 | train_examples = None 794 | num_train_steps = None 795 | num_warmup_steps = None 796 | if FLAGS.do_train: 797 | train_examples = processor.get_train_examples(FLAGS.data_dir) 798 | num_train_steps = int( 799 | len(train_examples) / FLAGS.train_batch_size * FLAGS.num_train_epochs) 800 | num_warmup_steps = int(num_train_steps * FLAGS.warmup_proportion) 801 | 802 | model_fn = model_fn_builder( 803 | bert_config=bert_config, 804 | num_labels=len(label_list), 805 | init_checkpoint=FLAGS.init_checkpoint, 806 | learning_rate=FLAGS.learning_rate, 807 | num_train_steps=num_train_steps, 808 | num_warmup_steps=num_warmup_steps, 809 | use_tpu=FLAGS.use_tpu, 810 | use_one_hot_embeddings=FLAGS.use_tpu) 811 | 812 | # If TPU is not available, this will fall back to normal Estimator on CPU 813 | # or GPU. 814 | estimator = tf.contrib.tpu.TPUEstimator( 815 | use_tpu=FLAGS.use_tpu, 816 | model_fn=model_fn, 817 | config=run_config, 818 | train_batch_size=FLAGS.train_batch_size, 819 | eval_batch_size=FLAGS.eval_batch_size, 820 | predict_batch_size=FLAGS.predict_batch_size) 821 | 822 | if FLAGS.do_train: 823 | train_file = os.path.join(FLAGS.output_dir, "train.tf_record") 824 | file_based_convert_examples_to_features( 825 | train_examples, label_list, FLAGS.max_seq_length, tokenizer, train_file) 826 | tf.logging.info("***** Running training *****") 827 | tf.logging.info(" Num examples = %d", len(train_examples)) 828 | tf.logging.info(" Batch size = %d", FLAGS.train_batch_size) 829 | tf.logging.info(" Num steps = %d", num_train_steps) 830 | train_input_fn = file_based_input_fn_builder( 831 | input_file=train_file, 832 | seq_length=FLAGS.max_seq_length, 833 | is_training=True, 834 | drop_remainder=True) 835 | estimator.train(input_fn=train_input_fn, max_steps=num_train_steps) 836 | 837 | if FLAGS.do_eval: 838 | eval_examples = processor.get_dev_examples(FLAGS.data_dir) 839 | eval_file = os.path.join(FLAGS.output_dir, "eval.tf_record") 840 | file_based_convert_examples_to_features( 841 | eval_examples, label_list, FLAGS.max_seq_length, tokenizer, eval_file) 842 | 843 | tf.logging.info("***** Running evaluation *****") 844 | tf.logging.info(" Num examples = %d", len(eval_examples)) 845 | tf.logging.info(" Batch size = %d", FLAGS.eval_batch_size) 846 | 847 | # This tells the estimator to run through the entire set. 848 | eval_steps = None 849 | # However, if running eval on the TPU, you will need to specify the 850 | # number of steps. 851 | if FLAGS.use_tpu: 852 | # Eval will be slightly WRONG on the TPU because it will truncate 853 | # the last batch. 854 | eval_steps = int(len(eval_examples) / FLAGS.eval_batch_size) 855 | 856 | eval_drop_remainder = True if FLAGS.use_tpu else False 857 | eval_input_fn = file_based_input_fn_builder( 858 | input_file=eval_file, 859 | seq_length=FLAGS.max_seq_length, 860 | is_training=False, 861 | drop_remainder=eval_drop_remainder) 862 | 863 | result = estimator.evaluate(input_fn=eval_input_fn, steps=eval_steps) 864 | 865 | output_eval_file = os.path.join(FLAGS.output_dir, "eval_results.txt") 866 | with tf.gfile.GFile(output_eval_file, "w") as writer: 867 | tf.logging.info("***** Eval results *****") 868 | for key in sorted(result.keys()): 869 | tf.logging.info(" %s = %s", key, str(result[key])) 870 | writer.write("%s = %s\n" % (key, str(result[key]))) 871 | 872 | if FLAGS.do_predict: 873 | predict_examples = processor.get_test_examples(FLAGS.data_dir) 874 | predict_file = os.path.join(FLAGS.output_dir, "predict.tf_record") 875 | file_based_convert_examples_to_features(predict_examples, label_list, 876 | FLAGS.max_seq_length, tokenizer, 877 | predict_file) 878 | 879 | tf.logging.info("***** Running prediction*****") 880 | tf.logging.info(" Num examples = %d", len(predict_examples)) 881 | tf.logging.info(" Batch size = %d", FLAGS.predict_batch_size) 882 | 883 | if FLAGS.use_tpu: 884 | # Warning: According to tpu_estimator.py Prediction on TPU is an 885 | # experimental feature and hence not supported here 886 | raise ValueError("Prediction in TPU not supported") 887 | 888 | predict_drop_remainder = True if FLAGS.use_tpu else False 889 | predict_input_fn = file_based_input_fn_builder( 890 | input_file=predict_file, 891 | seq_length=FLAGS.max_seq_length, 892 | is_training=False, 893 | drop_remainder=predict_drop_remainder) 894 | 895 | result = estimator.predict(input_fn=predict_input_fn) 896 | 897 | output_predict_file = os.path.join(FLAGS.output_dir, "test_results.tsv") 898 | with tf.gfile.GFile(output_predict_file, "w") as writer: 899 | tf.logging.info("***** Predict results *****") 900 | for prediction in result: 901 | output_line = "\t".join( 902 | str(class_probability) for class_probability in prediction) + "\n" 903 | writer.write(output_line) 904 | 905 | 906 | if __name__ == "__main__": 907 | flags.mark_flag_as_required("data_dir") 908 | flags.mark_flag_as_required("task_name") 909 | flags.mark_flag_as_required("vocab_file") 910 | flags.mark_flag_as_required("bert_config_file") 911 | flags.mark_flag_as_required("output_dir") 912 | tf.app.run() 913 | -------------------------------------------------------------------------------- /run_lm_predict.py: -------------------------------------------------------------------------------- 1 | # coding=utf-8 2 | # Copyright 2018 The Google AI Language Team Authors. 3 | # 4 | # Licensed under the Apache License, Version 2.0 (the "License"); 5 | # you may not use this file except in compliance with the License. 6 | # You may obtain a copy of the License at 7 | # 8 | # http://www.apache.org/licenses/LICENSE-2.0 9 | # 10 | # Unless required by applicable law or agreed to in writing, software 11 | # distributed under the License is distributed on an "AS IS" BASIS, 12 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 13 | # See the License for the specific language governing permissions and 14 | # limitations under the License. 15 | """BERT language model predict.""" 16 | 17 | from __future__ import absolute_import 18 | from __future__ import division 19 | from __future__ import print_function 20 | 21 | import os 22 | import json 23 | import modeling 24 | import tokenization 25 | import numpy as np 26 | import tensorflow as tf 27 | 28 | flags = tf.flags 29 | FLAGS = flags.FLAGS 30 | 31 | flags.DEFINE_integer("max_predictions_per_seq", 20, 32 | "In this task, it also refers to maximum number of masked tokens per word.") 33 | 34 | flags.DEFINE_string( 35 | "bert_config_file", None, 36 | "The config json file corresponding to the pre-trained BERT model. " 37 | "This specifies the model architecture.") 38 | 39 | flags.DEFINE_string( 40 | "input_file", None, 41 | "The config json file corresponding to the pre-trained BERT model. " 42 | "This specifies the model architecture.") 43 | 44 | flags.DEFINE_string( 45 | "output_dir", None, 46 | "The output directory where the model checkpoints will be written.") 47 | 48 | flags.DEFINE_string("vocab_file", None, 49 | "The vocabulary file that the BERT model was trained on.") 50 | 51 | ## Other parameters 52 | 53 | flags.DEFINE_string( 54 | "init_checkpoint", None, 55 | "Initial checkpoint (usually from a pre-trained BERT model).") 56 | 57 | flags.DEFINE_bool( 58 | "do_lower_case", True, 59 | "Whether to lower case the input text. Should be True for uncased " 60 | "models and False for cased models.") 61 | 62 | flags.DEFINE_integer( 63 | "max_seq_length", 128, 64 | "The maximum total input sequence length after WordPiece tokenization. " 65 | "Sequences longer than this will be truncated, and sequences shorter " 66 | "than this will be padded.") 67 | 68 | flags.DEFINE_integer("predict_batch_size", 8, "Total batch size for predict.") 69 | 70 | flags.DEFINE_bool("use_tpu", False, "Whether to use TPU or GPU/CPU.") 71 | 72 | tf.flags.DEFINE_string( 73 | "tpu_name", None, 74 | "The Cloud TPU to use for training. This should be either the name " 75 | "used when creating the Cloud TPU, or a grpc://ip.address.of.tpu:8470 " 76 | "url.") 77 | 78 | tf.flags.DEFINE_string( 79 | "tpu_zone", None, 80 | "[Optional] GCE zone where the Cloud TPU is located in. If not " 81 | "specified, we will attempt to automatically detect the GCE project from " 82 | "metadata.") 83 | 84 | tf.flags.DEFINE_string( 85 | "gcp_project", None, 86 | "[Optional] Project name for the Cloud TPU-enabled project. If not " 87 | "specified, we will attempt to automatically detect the GCE project from " 88 | "metadata.") 89 | 90 | tf.flags.DEFINE_string("master", None, "[Optional] TensorFlow master URL.") 91 | 92 | flags.DEFINE_integer( 93 | "num_tpu_cores", 8, 94 | "Only used if `use_tpu` is True. Total number of TPU cores to use.") 95 | 96 | 97 | class InputExample(object): 98 | def __init__(self, unique_id, text): 99 | self.unique_id = unique_id 100 | self.text = text 101 | 102 | 103 | def read_examples(input_file): 104 | """Read a list of `InputExample`s from an input file.""" 105 | examples = [] 106 | unique_id = 0 107 | with tf.gfile.GFile(input_file, "r") as reader: 108 | while True: 109 | line = tokenization.convert_to_unicode(reader.readline()) 110 | if not line: 111 | break 112 | line = line.strip() 113 | unique_id += 1 114 | examples.append( 115 | InputExample(unique_id, line)) 116 | unique_id += 1 117 | return examples 118 | 119 | 120 | def model_fn_builder(bert_config, init_checkpoint, use_tpu, 121 | use_one_hot_embeddings): 122 | """Returns `model_fn` closure for TPUEstimator.""" 123 | 124 | def model_fn(features, mode, params): # pylint: disable=unused-argument 125 | """The `model_fn` for TPUEstimator.""" 126 | 127 | tf.logging.info("*** Features ***") 128 | for name in sorted(features.keys()): 129 | tf.logging.info(" name = %s, shape = %s" % (name, features[name].shape)) 130 | 131 | input_ids = features["input_ids"] 132 | input_mask = features["input_mask"] 133 | segment_ids = features["segment_ids"] 134 | masked_lm_positions = features["masked_lm_positions"] 135 | masked_lm_ids = features["masked_lm_ids"] 136 | 137 | 138 | model = modeling.BertModel( 139 | config=bert_config, 140 | is_training=False, 141 | input_ids=input_ids, 142 | input_mask=input_mask, 143 | token_type_ids=segment_ids, 144 | use_one_hot_embeddings=use_one_hot_embeddings) 145 | 146 | masked_lm_example_loss = get_masked_lm_output( 147 | bert_config, model.get_sequence_output(), model.get_embedding_table(), 148 | masked_lm_positions, masked_lm_ids) 149 | 150 | tvars = tf.trainable_variables() 151 | initialized_variable_names = {} 152 | scaffold_fn = None 153 | if init_checkpoint: 154 | (assignment_map, initialized_variable_names 155 | ) = modeling.get_assignment_map_from_checkpoint(tvars, init_checkpoint) 156 | if use_tpu: 157 | 158 | def tpu_scaffold(): 159 | tf.train.init_from_checkpoint(init_checkpoint, assignment_map) 160 | return tf.train.Scaffold() 161 | 162 | scaffold_fn = tpu_scaffold 163 | else: 164 | tf.train.init_from_checkpoint(init_checkpoint, assignment_map) 165 | 166 | tf.logging.info("**** Trainable Variables ****") 167 | for var in tvars: 168 | init_string = "" 169 | if var.name in initialized_variable_names: 170 | init_string = ", *INIT_FROM_CKPT*" 171 | tf.logging.info(" name = %s, shape = %s%s", var.name, var.shape, 172 | init_string) 173 | 174 | output_spec = None 175 | if mode == tf.estimator.ModeKeys.PREDICT: 176 | output_spec = tf.contrib.tpu.TPUEstimatorSpec( 177 | mode=mode, predictions=masked_lm_example_loss, scaffold_fn=scaffold_fn) # 输出mask_word的score 178 | return output_spec 179 | 180 | return model_fn 181 | 182 | 183 | 184 | 185 | def get_masked_lm_output(bert_config, input_tensor, output_weights, positions, 186 | label_ids): 187 | """Get loss and log probs for the masked LM.""" 188 | input_tensor = gather_indexes(input_tensor, positions) 189 | 190 | with tf.variable_scope("cls/predictions"): 191 | # We apply one more non-linear transformation before the output layer. 192 | # This matrix is not used after pre-training. 193 | with tf.variable_scope("transform"): 194 | input_tensor = tf.layers.dense( 195 | input_tensor, 196 | units=bert_config.hidden_size, 197 | activation=modeling.get_activation(bert_config.hidden_act), 198 | kernel_initializer=modeling.create_initializer( 199 | bert_config.initializer_range)) 200 | input_tensor = modeling.layer_norm(input_tensor) 201 | 202 | # The output weights are the same as the input embeddings, but there is 203 | # an output-only bias for each token. 204 | output_bias = tf.get_variable( 205 | "output_bias", 206 | shape=[bert_config.vocab_size], 207 | initializer=tf.zeros_initializer()) 208 | logits = tf.matmul(input_tensor, output_weights, transpose_b=True) 209 | logits = tf.nn.bias_add(logits, output_bias) 210 | log_probs = tf.nn.log_softmax(logits, axis=-1) 211 | 212 | label_ids = tf.reshape(label_ids, [-1]) 213 | 214 | one_hot_labels = tf.one_hot( 215 | label_ids, depth=bert_config.vocab_size, dtype=tf.float32) 216 | per_example_loss = -tf.reduce_sum(log_probs * one_hot_labels, axis=[-1]) 217 | loss = tf.reshape(per_example_loss, [-1, tf.shape(positions)[1]]) 218 | # TODO: dynamic gather from per_example_loss 219 | return loss 220 | 221 | 222 | 223 | def gather_indexes(sequence_tensor, positions): 224 | """Gathers the vectors at the specific positions over a minibatch.""" 225 | sequence_shape = modeling.get_shape_list(sequence_tensor, expected_rank=3) 226 | batch_size = sequence_shape[0] 227 | seq_length = sequence_shape[1] 228 | width = sequence_shape[2] 229 | 230 | flat_offsets = tf.reshape( 231 | tf.range(0, batch_size, dtype=tf.int32) * seq_length, [-1, 1]) 232 | flat_positions = tf.reshape(positions + flat_offsets, [-1]) 233 | flat_sequence_tensor = tf.reshape(sequence_tensor, 234 | [batch_size * seq_length, width]) 235 | output_tensor = tf.gather(flat_sequence_tensor, flat_positions) 236 | return output_tensor 237 | 238 | 239 | def input_fn_builder(features, seq_length, max_predictions_per_seq): 240 | """Creates an `input_fn` closure to be passed to TPUEstimator.""" 241 | 242 | all_input_ids = [] 243 | all_input_mask = [] 244 | all_segment_ids = [] 245 | all_masked_lm_positions = [] 246 | all_masked_lm_ids = [] 247 | 248 | for feature in features: 249 | all_input_ids.append(feature.input_ids) 250 | all_input_mask.append(feature.input_mask) 251 | all_segment_ids.append(feature.segment_ids) 252 | all_masked_lm_positions.append(feature.masked_lm_positions) 253 | all_masked_lm_ids.append(feature.masked_lm_ids) 254 | 255 | def input_fn(params): 256 | """The actual input function.""" 257 | batch_size = params["batch_size"] 258 | num_examples = len(features) 259 | 260 | # This is for demo purposes and does NOT scale to large data sets. We do 261 | # not use Dataset.from_generator() because that uses tf.py_func which is 262 | # not TPU compatible. The right way to load data is with TFRecordReader. 263 | d = tf.data.Dataset.from_tensor_slices({ 264 | "input_ids": 265 | tf.constant( 266 | all_input_ids, shape=[num_examples, seq_length], 267 | dtype=tf.int32), 268 | "input_mask": 269 | tf.constant( 270 | all_input_mask, 271 | shape=[num_examples, seq_length], 272 | dtype=tf.int32), 273 | "segment_ids": 274 | tf.constant( 275 | all_segment_ids, 276 | shape=[num_examples, seq_length], 277 | dtype=tf.int32), 278 | "masked_lm_positions": 279 | tf.constant( 280 | all_masked_lm_positions, 281 | shape=[num_examples, max_predictions_per_seq], 282 | dtype=tf.int32), 283 | "masked_lm_ids": 284 | tf.constant( 285 | all_masked_lm_ids, 286 | shape=[num_examples, max_predictions_per_seq], 287 | dtype=tf.int32) 288 | }) 289 | 290 | d = d.batch(batch_size=batch_size, drop_remainder=False) 291 | return d 292 | 293 | return input_fn 294 | 295 | 296 | 297 | # This function is not used by this file but is still used by the Colab and 298 | # people who depend on it. 299 | def convert_examples_to_features(examples, max_seq_length, tokenizer): 300 | """Convert a set of `InputExample`s to a list of `InputFeatures`.""" 301 | 302 | all_features = [] 303 | all_tokens = [] 304 | 305 | for (ex_index, example) in enumerate(examples): 306 | if ex_index % 10000 == 0: 307 | tf.logging.info("Writing example %d of %d" % (ex_index, len(examples))) 308 | 309 | features, tokens = convert_single_example(ex_index, example, 310 | max_seq_length, tokenizer) 311 | all_features.extend(features) 312 | all_tokens.extend(tokens) 313 | 314 | return all_features, all_tokens 315 | 316 | tokenizer = tokenization.FullTokenizer( 317 | vocab_file=FLAGS.vocab_file, do_lower_case=FLAGS.do_lower_case) 318 | MASKED_TOKEN = "[MASK]" 319 | MASKED_ID = tokenizer.convert_tokens_to_ids([MASKED_TOKEN])[0] 320 | 321 | 322 | def create_masked_lm_prediction(input_ids, mask_position, mask_count=1): 323 | new_input_ids = list(input_ids) 324 | masked_lm_labels = [] 325 | masked_lm_positions = list(range(mask_position, mask_position + mask_count)) 326 | for i in masked_lm_positions: 327 | new_input_ids[i] = MASKED_ID 328 | masked_lm_labels.append(input_ids[i]) 329 | return new_input_ids, masked_lm_positions, masked_lm_labels 330 | 331 | 332 | class InputFeatures(object): 333 | """A single set of features of data.""" 334 | 335 | def __init__(self, input_ids, segment_ids, input_mask, masked_lm_positions, 336 | masked_lm_ids): 337 | self.input_ids = input_ids, 338 | self.segment_ids = segment_ids, 339 | self.input_mask = input_mask, 340 | self.masked_lm_positions = masked_lm_positions, 341 | self.masked_lm_ids = masked_lm_ids, 342 | 343 | 344 | def convert_single_example(ex_index, example, max_seq_length, 345 | tokenizer): 346 | """Converts a single `InputExample` into a single `InputFeatures`.""" 347 | tokens = tokenizer.tokenize(example.text) 348 | 349 | # Account for [CLS] and [SEP] with "- 2" 350 | if len(tokens) > max_seq_length - 2: 351 | tokens = tokens[0:(max_seq_length - 2)] 352 | 353 | input_tokens = [] 354 | segment_ids = [] 355 | input_tokens.append("[CLS]") 356 | segment_ids.append(0) 357 | for token in tokens: 358 | input_tokens.append(token) 359 | segment_ids.append(0) 360 | input_tokens.append("[SEP]") 361 | segment_ids.append(0) 362 | 363 | input_ids = tokenizer.convert_tokens_to_ids(input_tokens) 364 | 365 | # The mask has 1 for real tokens and 0 for padding tokens. Only real 366 | # tokens are attended to. 367 | input_mask = [1] * len(input_ids) 368 | 369 | # Zero-pad up to the sequence length. 370 | while len(input_ids) < max_seq_length: 371 | input_ids.append(0) 372 | input_mask.append(0) 373 | segment_ids.append(0) 374 | 375 | assert len(input_ids) == max_seq_length 376 | assert len(input_mask) == max_seq_length 377 | assert len(segment_ids) == max_seq_length 378 | 379 | if ex_index < 5: 380 | tf.logging.info("*** Example ***") 381 | tf.logging.info("id: %s" % (example.unique_id)) 382 | tf.logging.info("tokens: %s" % " ".join( 383 | [tokenization.printable_text(x) for x in input_tokens])) 384 | tf.logging.info("input_ids: %s" % " ".join([str(x) for x in input_ids])) 385 | tf.logging.info("input_mask: %s" % " ".join([str(x) for x in input_mask])) 386 | tf.logging.info("segment_ids: %s" % " ".join([str(x) for x in segment_ids])) 387 | 388 | features = create_sequential_mask(input_tokens, input_ids, input_mask, segment_ids, 389 | FLAGS.max_predictions_per_seq) 390 | return features, input_tokens 391 | 392 | 393 | def is_subtoken(x): 394 | return x.startswith("##") 395 | 396 | def create_sequential_mask(input_tokens, input_ids, input_mask, segment_ids, 397 | max_predictions_per_seq): 398 | """Mask each token/word sequentially""" 399 | features = [] 400 | i = 1 401 | while i < len(input_tokens) - 1: 402 | mask_count = 1 403 | while is_subtoken(input_tokens[i+mask_count]): 404 | mask_count += 1 405 | 406 | input_ids_new, masked_lm_positions, masked_lm_labels = create_masked_lm_prediction(input_ids, i, mask_count) 407 | while len(masked_lm_positions) < max_predictions_per_seq: 408 | masked_lm_positions.append(0) 409 | masked_lm_labels.append(0) 410 | 411 | feature = InputFeatures( 412 | input_ids=input_ids_new, 413 | input_mask=input_mask, 414 | segment_ids=segment_ids, 415 | masked_lm_positions=masked_lm_positions, 416 | masked_lm_ids=masked_lm_labels) 417 | features.append(feature) 418 | i += mask_count 419 | return features 420 | 421 | 422 | def parse_result(result, all_tokens, output_file=None): 423 | with tf.gfile.GFile(output_file, "w") as writer: 424 | tf.logging.info("***** Predict results *****") 425 | i = 0 426 | sentences = [] 427 | for word_loss in result: 428 | # start of a sentence 429 | if all_tokens[i] == "[CLS]": 430 | sentence = {} 431 | tokens = [] 432 | sentence_loss = 0.0 433 | word_count_per_sent = 0 434 | i += 1 435 | 436 | # add token 437 | tokens.append({"token": tokenization.printable_text(all_tokens[i]), 438 | "prob": float(np.exp(-word_loss[0])) }) 439 | sentence_loss += word_loss[0] 440 | word_count_per_sent += 1 441 | i += 1 442 | 443 | token_count_per_word = 0 444 | while is_subtoken(all_tokens[i]): 445 | token_count_per_word += 1 446 | tokens.append({"token": tokenization.printable_text(all_tokens[i]), 447 | "prob": float(np.exp(-word_loss[token_count_per_word]))}) 448 | sentence_loss += word_loss[token_count_per_word] 449 | i += 1 450 | 451 | # end of a sentence 452 | if all_tokens[i] == "[SEP]": 453 | sentence["tokens"] = tokens 454 | sentence["ppl"] = float(np.exp(sentence_loss / word_count_per_sent)) 455 | sentences.append(sentence) 456 | i += 1 457 | 458 | if output_file is not None: 459 | tf.logging.info("Saving results to %s" % output_file) 460 | writer.write(json.dumps(sentences, indent=2, ensure_ascii=False)) 461 | 462 | def main(_): 463 | tf.logging.set_verbosity(tf.logging.INFO) 464 | 465 | bert_config = modeling.BertConfig.from_json_file(FLAGS.bert_config_file) 466 | 467 | if FLAGS.max_seq_length > bert_config.max_position_embeddings: 468 | raise ValueError( 469 | "Cannot use sequence length %d because the BERT model " 470 | "was only trained up to sequence length %d" % 471 | (FLAGS.max_seq_length, bert_config.max_position_embeddings)) 472 | 473 | tf.gfile.MakeDirs(FLAGS.output_dir) 474 | 475 | tpu_cluster_resolver = None 476 | if FLAGS.use_tpu and FLAGS.tpu_name: 477 | tpu_cluster_resolver = tf.contrib.cluster_resolver.TPUClusterResolver( 478 | FLAGS.tpu_name, zone=FLAGS.tpu_zone, project=FLAGS.gcp_project) 479 | 480 | is_per_host = tf.contrib.tpu.InputPipelineConfig.PER_HOST_V2 481 | run_config = tf.contrib.tpu.RunConfig( 482 | cluster=tpu_cluster_resolver, 483 | master=FLAGS.master, 484 | model_dir=FLAGS.output_dir, 485 | tpu_config=tf.contrib.tpu.TPUConfig( 486 | num_shards=FLAGS.num_tpu_cores, 487 | per_host_input_for_training=is_per_host)) 488 | 489 | 490 | model_fn = model_fn_builder( 491 | bert_config=bert_config, 492 | init_checkpoint=FLAGS.init_checkpoint, 493 | use_tpu=FLAGS.use_tpu, 494 | use_one_hot_embeddings=FLAGS.use_tpu) 495 | 496 | # If TPU is not available, this will fall back to normal Estimator on CPU 497 | # or GPU. 498 | estimator = tf.contrib.tpu.TPUEstimator( 499 | use_tpu=FLAGS.use_tpu, 500 | model_fn=model_fn, 501 | config=run_config, 502 | predict_batch_size=FLAGS.predict_batch_size) 503 | 504 | 505 | predict_examples = read_examples(FLAGS.input_file) 506 | features, all_tokens = convert_examples_to_features(predict_examples, 507 | FLAGS.max_seq_length, tokenizer) 508 | 509 | tf.logging.info("***** Running prediction*****") 510 | tf.logging.info(" Num examples = %d", len(predict_examples)) 511 | tf.logging.info(" Batch size = %d", FLAGS.predict_batch_size) 512 | 513 | if FLAGS.use_tpu: 514 | # Warning: According to tpu_estimator.py Prediction on TPU is an 515 | # experimental feature and hence not supported here 516 | raise ValueError("Prediction in TPU not supported") 517 | 518 | predict_input_fn = input_fn_builder( 519 | features=features, 520 | seq_length=FLAGS.max_seq_length, 521 | max_predictions_per_seq=FLAGS.max_predictions_per_seq) 522 | 523 | result = estimator.predict(input_fn=predict_input_fn) 524 | output_predict_file = os.path.join(FLAGS.output_dir, "test_results.json") 525 | parse_result(result, all_tokens, output_predict_file) 526 | 527 | if __name__ == "__main__": 528 | tf.app.run() 529 | -------------------------------------------------------------------------------- /run_pretraining.py: -------------------------------------------------------------------------------- 1 | # coding=utf-8 2 | # Copyright 2018 The Google AI Language Team Authors. 3 | # 4 | # Licensed under the Apache License, Version 2.0 (the "License"); 5 | # you may not use this file except in compliance with the License. 6 | # You may obtain a copy of the License at 7 | # 8 | # http://www.apache.org/licenses/LICENSE-2.0 9 | # 10 | # Unless required by applicable law or agreed to in writing, software 11 | # distributed under the License is distributed on an "AS IS" BASIS, 12 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 13 | # See the License for the specific language governing permissions and 14 | # limitations under the License. 15 | """Run masked LM/next sentence masked_lm pre-training for BERT.""" 16 | 17 | from __future__ import absolute_import 18 | from __future__ import division 19 | from __future__ import print_function 20 | 21 | import os 22 | import modeling 23 | import optimization 24 | import tensorflow as tf 25 | 26 | flags = tf.flags 27 | 28 | FLAGS = flags.FLAGS 29 | 30 | ## Required parameters 31 | flags.DEFINE_string( 32 | "bert_config_file", None, 33 | "The config json file corresponding to the pre-trained BERT model. " 34 | "This specifies the model architecture.") 35 | 36 | flags.DEFINE_string( 37 | "input_file", None, 38 | "Input TF example files (can be a glob or comma separated).") 39 | 40 | flags.DEFINE_string( 41 | "output_dir", None, 42 | "The output directory where the model checkpoints will be written.") 43 | 44 | ## Other parameters 45 | flags.DEFINE_string( 46 | "init_checkpoint", None, 47 | "Initial checkpoint (usually from a pre-trained BERT model).") 48 | 49 | flags.DEFINE_integer( 50 | "max_seq_length", 128, 51 | "The maximum total input sequence length after WordPiece tokenization. " 52 | "Sequences longer than this will be truncated, and sequences shorter " 53 | "than this will be padded. Must match data generation.") 54 | 55 | flags.DEFINE_integer( 56 | "max_predictions_per_seq", 20, 57 | "Maximum number of masked LM predictions per sequence. " 58 | "Must match data generation.") 59 | 60 | flags.DEFINE_bool("do_train", False, "Whether to run training.") 61 | 62 | flags.DEFINE_bool("do_eval", False, "Whether to run eval on the dev set.") 63 | 64 | flags.DEFINE_integer("train_batch_size", 32, "Total batch size for training.") 65 | 66 | flags.DEFINE_integer("eval_batch_size", 8, "Total batch size for eval.") 67 | 68 | flags.DEFINE_float("learning_rate", 5e-5, "The initial learning rate for Adam.") 69 | 70 | flags.DEFINE_integer("num_train_steps", 100000, "Number of training steps.") 71 | 72 | flags.DEFINE_integer("num_warmup_steps", 10000, "Number of warmup steps.") 73 | 74 | flags.DEFINE_integer("save_checkpoints_steps", 1000, 75 | "How often to save the model checkpoint.") 76 | 77 | flags.DEFINE_integer("iterations_per_loop", 1000, 78 | "How many steps to make in each estimator call.") 79 | 80 | flags.DEFINE_integer("max_eval_steps", 100, "Maximum number of eval steps.") 81 | 82 | flags.DEFINE_bool("use_tpu", False, "Whether to use TPU or GPU/CPU.") 83 | 84 | tf.flags.DEFINE_string( 85 | "tpu_name", None, 86 | "The Cloud TPU to use for training. This should be either the name " 87 | "used when creating the Cloud TPU, or a grpc://ip.address.of.tpu:8470 " 88 | "url.") 89 | 90 | tf.flags.DEFINE_string( 91 | "tpu_zone", None, 92 | "[Optional] GCE zone where the Cloud TPU is located in. If not " 93 | "specified, we will attempt to automatically detect the GCE project from " 94 | "metadata.") 95 | 96 | tf.flags.DEFINE_string( 97 | "gcp_project", None, 98 | "[Optional] Project name for the Cloud TPU-enabled project. If not " 99 | "specified, we will attempt to automatically detect the GCE project from " 100 | "metadata.") 101 | 102 | tf.flags.DEFINE_string("master", None, "[Optional] TensorFlow master URL.") 103 | 104 | flags.DEFINE_integer( 105 | "num_tpu_cores", 8, 106 | "Only used if `use_tpu` is True. Total number of TPU cores to use.") 107 | 108 | 109 | def model_fn_builder(bert_config, init_checkpoint, learning_rate, 110 | num_train_steps, num_warmup_steps, use_tpu, 111 | use_one_hot_embeddings): 112 | """Returns `model_fn` closure for TPUEstimator.""" 113 | 114 | def model_fn(features, labels, mode, params): # pylint: disable=unused-argument 115 | """The `model_fn` for TPUEstimator.""" 116 | 117 | tf.logging.info("*** Features ***") 118 | for name in sorted(features.keys()): 119 | tf.logging.info(" name = %s, shape = %s" % (name, features[name].shape)) 120 | 121 | input_ids = features["input_ids"] 122 | input_mask = features["input_mask"] 123 | segment_ids = features["segment_ids"] 124 | masked_lm_positions = features["masked_lm_positions"] 125 | masked_lm_ids = features["masked_lm_ids"] 126 | masked_lm_weights = features["masked_lm_weights"] 127 | next_sentence_labels = features["next_sentence_labels"] 128 | 129 | is_training = (mode == tf.estimator.ModeKeys.TRAIN) 130 | 131 | model = modeling.BertModel( 132 | config=bert_config, 133 | is_training=is_training, 134 | input_ids=input_ids, 135 | input_mask=input_mask, 136 | token_type_ids=segment_ids, 137 | use_one_hot_embeddings=use_one_hot_embeddings) 138 | 139 | (masked_lm_loss, 140 | masked_lm_example_loss, masked_lm_log_probs) = get_masked_lm_output( 141 | bert_config, model.get_sequence_output(), model.get_embedding_table(), 142 | masked_lm_positions, masked_lm_ids, masked_lm_weights) 143 | 144 | (next_sentence_loss, next_sentence_example_loss, 145 | next_sentence_log_probs) = get_next_sentence_output( 146 | bert_config, model.get_pooled_output(), next_sentence_labels) 147 | 148 | total_loss = masked_lm_loss + next_sentence_loss 149 | 150 | tvars = tf.trainable_variables() 151 | 152 | initialized_variable_names = {} 153 | scaffold_fn = None 154 | if init_checkpoint: 155 | (assignment_map, initialized_variable_names 156 | ) = modeling.get_assignment_map_from_checkpoint(tvars, init_checkpoint) 157 | if use_tpu: 158 | 159 | def tpu_scaffold(): 160 | tf.train.init_from_checkpoint(init_checkpoint, assignment_map) 161 | return tf.train.Scaffold() 162 | 163 | scaffold_fn = tpu_scaffold 164 | else: 165 | tf.train.init_from_checkpoint(init_checkpoint, assignment_map) 166 | 167 | tf.logging.info("**** Trainable Variables ****") 168 | for var in tvars: 169 | init_string = "" 170 | if var.name in initialized_variable_names: 171 | init_string = ", *INIT_FROM_CKPT*" 172 | tf.logging.info(" name = %s, shape = %s%s", var.name, var.shape, 173 | init_string) 174 | 175 | output_spec = None 176 | if mode == tf.estimator.ModeKeys.TRAIN: 177 | train_op = optimization.create_optimizer( 178 | total_loss, learning_rate, num_train_steps, num_warmup_steps, use_tpu) 179 | 180 | output_spec = tf.contrib.tpu.TPUEstimatorSpec( 181 | mode=mode, 182 | loss=total_loss, 183 | train_op=train_op, 184 | scaffold_fn=scaffold_fn) 185 | elif mode == tf.estimator.ModeKeys.EVAL: 186 | 187 | def metric_fn(masked_lm_example_loss, masked_lm_log_probs, masked_lm_ids, 188 | masked_lm_weights, next_sentence_example_loss, 189 | next_sentence_log_probs, next_sentence_labels): 190 | """Computes the loss and accuracy of the model.""" 191 | masked_lm_log_probs = tf.reshape(masked_lm_log_probs, 192 | [-1, masked_lm_log_probs.shape[-1]]) 193 | masked_lm_predictions = tf.argmax( 194 | masked_lm_log_probs, axis=-1, output_type=tf.int32) 195 | masked_lm_example_loss = tf.reshape(masked_lm_example_loss, [-1]) 196 | masked_lm_ids = tf.reshape(masked_lm_ids, [-1]) 197 | masked_lm_weights = tf.reshape(masked_lm_weights, [-1]) 198 | masked_lm_accuracy = tf.metrics.accuracy( 199 | labels=masked_lm_ids, 200 | predictions=masked_lm_predictions, 201 | weights=masked_lm_weights) 202 | masked_lm_mean_loss = tf.metrics.mean( 203 | values=masked_lm_example_loss, weights=masked_lm_weights) 204 | 205 | next_sentence_log_probs = tf.reshape( 206 | next_sentence_log_probs, [-1, next_sentence_log_probs.shape[-1]]) 207 | next_sentence_predictions = tf.argmax( 208 | next_sentence_log_probs, axis=-1, output_type=tf.int32) 209 | next_sentence_labels = tf.reshape(next_sentence_labels, [-1]) 210 | next_sentence_accuracy = tf.metrics.accuracy( 211 | labels=next_sentence_labels, predictions=next_sentence_predictions) 212 | next_sentence_mean_loss = tf.metrics.mean( 213 | values=next_sentence_example_loss) 214 | 215 | return { 216 | "masked_lm_accuracy": masked_lm_accuracy, 217 | "masked_lm_loss": masked_lm_mean_loss, 218 | "next_sentence_accuracy": next_sentence_accuracy, 219 | "next_sentence_loss": next_sentence_mean_loss, 220 | } 221 | 222 | eval_metrics = (metric_fn, [ 223 | masked_lm_example_loss, masked_lm_log_probs, masked_lm_ids, 224 | masked_lm_weights, next_sentence_example_loss, 225 | next_sentence_log_probs, next_sentence_labels 226 | ]) 227 | output_spec = tf.contrib.tpu.TPUEstimatorSpec( 228 | mode=mode, 229 | loss=total_loss, 230 | eval_metrics=eval_metrics, 231 | scaffold_fn=scaffold_fn) 232 | else: 233 | raise ValueError("Only TRAIN and EVAL modes are supported: %s" % (mode)) 234 | 235 | return output_spec 236 | 237 | return model_fn 238 | 239 | 240 | def get_masked_lm_output(bert_config, input_tensor, output_weights, positions, 241 | label_ids, label_weights): 242 | """Get loss and log probs for the masked LM.""" 243 | input_tensor = gather_indexes(input_tensor, positions) 244 | 245 | with tf.variable_scope("cls/predictions"): 246 | # We apply one more non-linear transformation before the output layer. 247 | # This matrix is not used after pre-training. 248 | with tf.variable_scope("transform"): 249 | input_tensor = tf.layers.dense( 250 | input_tensor, 251 | units=bert_config.hidden_size, 252 | activation=modeling.get_activation(bert_config.hidden_act), 253 | kernel_initializer=modeling.create_initializer( 254 | bert_config.initializer_range)) 255 | input_tensor = modeling.layer_norm(input_tensor) 256 | 257 | # The output weights are the same as the input embeddings, but there is 258 | # an output-only bias for each token. 259 | output_bias = tf.get_variable( 260 | "output_bias", 261 | shape=[bert_config.vocab_size], 262 | initializer=tf.zeros_initializer()) 263 | logits = tf.matmul(input_tensor, output_weights, transpose_b=True) 264 | logits = tf.nn.bias_add(logits, output_bias) 265 | log_probs = tf.nn.log_softmax(logits, axis=-1) 266 | 267 | label_ids = tf.reshape(label_ids, [-1]) 268 | label_weights = tf.reshape(label_weights, [-1]) 269 | 270 | one_hot_labels = tf.one_hot( 271 | label_ids, depth=bert_config.vocab_size, dtype=tf.float32) 272 | 273 | # The `positions` tensor might be zero-padded (if the sequence is too 274 | # short to have the maximum number of predictions). The `label_weights` 275 | # tensor has a value of 1.0 for every real prediction and 0.0 for the 276 | # padding predictions. 277 | per_example_loss = -tf.reduce_sum(log_probs * one_hot_labels, axis=[-1]) 278 | numerator = tf.reduce_sum(label_weights * per_example_loss) 279 | denominator = tf.reduce_sum(label_weights) + 1e-5 280 | loss = numerator / denominator 281 | 282 | return (loss, per_example_loss, log_probs) 283 | 284 | 285 | def get_next_sentence_output(bert_config, input_tensor, labels): 286 | """Get loss and log probs for the next sentence prediction.""" 287 | 288 | # Simple binary classification. Note that 0 is "next sentence" and 1 is 289 | # "random sentence". This weight matrix is not used after pre-training. 290 | with tf.variable_scope("cls/seq_relationship"): 291 | output_weights = tf.get_variable( 292 | "output_weights", 293 | shape=[2, bert_config.hidden_size], 294 | initializer=modeling.create_initializer(bert_config.initializer_range)) 295 | output_bias = tf.get_variable( 296 | "output_bias", shape=[2], initializer=tf.zeros_initializer()) 297 | 298 | logits = tf.matmul(input_tensor, output_weights, transpose_b=True) 299 | logits = tf.nn.bias_add(logits, output_bias) 300 | log_probs = tf.nn.log_softmax(logits, axis=-1) 301 | labels = tf.reshape(labels, [-1]) 302 | one_hot_labels = tf.one_hot(labels, depth=2, dtype=tf.float32) 303 | per_example_loss = -tf.reduce_sum(one_hot_labels * log_probs, axis=-1) 304 | loss = tf.reduce_mean(per_example_loss) 305 | return (loss, per_example_loss, log_probs) 306 | 307 | 308 | def gather_indexes(sequence_tensor, positions): 309 | """Gathers the vectors at the specific positions over a minibatch.""" 310 | sequence_shape = modeling.get_shape_list(sequence_tensor, expected_rank=3) 311 | batch_size = sequence_shape[0] 312 | seq_length = sequence_shape[1] 313 | width = sequence_shape[2] 314 | 315 | flat_offsets = tf.reshape( 316 | tf.range(0, batch_size, dtype=tf.int32) * seq_length, [-1, 1]) 317 | flat_positions = tf.reshape(positions + flat_offsets, [-1]) 318 | flat_sequence_tensor = tf.reshape(sequence_tensor, 319 | [batch_size * seq_length, width]) 320 | output_tensor = tf.gather(flat_sequence_tensor, flat_positions) 321 | return output_tensor 322 | 323 | 324 | def input_fn_builder(input_files, 325 | max_seq_length, 326 | max_predictions_per_seq, 327 | is_training, 328 | num_cpu_threads=4): 329 | """Creates an `input_fn` closure to be passed to TPUEstimator.""" 330 | 331 | def input_fn(params): 332 | """The actual input function.""" 333 | batch_size = params["batch_size"] 334 | 335 | name_to_features = { 336 | "input_ids": 337 | tf.FixedLenFeature([max_seq_length], tf.int64), 338 | "input_mask": 339 | tf.FixedLenFeature([max_seq_length], tf.int64), 340 | "segment_ids": 341 | tf.FixedLenFeature([max_seq_length], tf.int64), 342 | "masked_lm_positions": 343 | tf.FixedLenFeature([max_predictions_per_seq], tf.int64), 344 | "masked_lm_ids": 345 | tf.FixedLenFeature([max_predictions_per_seq], tf.int64), 346 | "masked_lm_weights": 347 | tf.FixedLenFeature([max_predictions_per_seq], tf.float32), 348 | "next_sentence_labels": 349 | tf.FixedLenFeature([1], tf.int64), 350 | } 351 | 352 | # For training, we want a lot of parallel reading and shuffling. 353 | # For eval, we want no shuffling and parallel reading doesn't matter. 354 | if is_training: 355 | d = tf.data.Dataset.from_tensor_slices(tf.constant(input_files)) 356 | d = d.repeat() 357 | d = d.shuffle(buffer_size=len(input_files)) 358 | 359 | # `cycle_length` is the number of parallel files that get read. 360 | cycle_length = min(num_cpu_threads, len(input_files)) 361 | 362 | # `sloppy` mode means that the interleaving is not exact. This adds 363 | # even more randomness to the training pipeline. 364 | d = d.apply( 365 | tf.contrib.data.parallel_interleave( 366 | tf.data.TFRecordDataset, 367 | sloppy=is_training, 368 | cycle_length=cycle_length)) 369 | d = d.shuffle(buffer_size=100) 370 | else: 371 | d = tf.data.TFRecordDataset(input_files) 372 | # Since we evaluate for a fixed number of steps we don't want to encounter 373 | # out-of-range exceptions. 374 | d = d.repeat() 375 | 376 | # We must `drop_remainder` on training because the TPU requires fixed 377 | # size dimensions. For eval, we assume we are evaluating on the CPU or GPU 378 | # and we *don't* want to drop the remainder, otherwise we wont cover 379 | # every sample. 380 | d = d.apply( 381 | tf.contrib.data.map_and_batch( 382 | lambda record: _decode_record(record, name_to_features), 383 | batch_size=batch_size, 384 | num_parallel_batches=num_cpu_threads, 385 | drop_remainder=True)) 386 | return d 387 | 388 | return input_fn 389 | 390 | 391 | def _decode_record(record, name_to_features): 392 | """Decodes a record to a TensorFlow example.""" 393 | example = tf.parse_single_example(record, name_to_features) 394 | 395 | # tf.Example only supports tf.int64, but the TPU only supports tf.int32. 396 | # So cast all int64 to int32. 397 | for name in list(example.keys()): 398 | t = example[name] 399 | if t.dtype == tf.int64: 400 | t = tf.to_int32(t) 401 | example[name] = t 402 | 403 | return example 404 | 405 | 406 | def main(_): 407 | tf.logging.set_verbosity(tf.logging.INFO) 408 | 409 | if not FLAGS.do_train and not FLAGS.do_eval: 410 | raise ValueError("At least one of `do_train` or `do_eval` must be True.") 411 | 412 | bert_config = modeling.BertConfig.from_json_file(FLAGS.bert_config_file) 413 | 414 | tf.gfile.MakeDirs(FLAGS.output_dir) 415 | 416 | input_files = [] 417 | for input_pattern in FLAGS.input_file.split(","): 418 | input_files.extend(tf.gfile.Glob(input_pattern)) 419 | 420 | tf.logging.info("*** Input Files ***") 421 | for input_file in input_files: 422 | tf.logging.info(" %s" % input_file) 423 | 424 | tpu_cluster_resolver = None 425 | if FLAGS.use_tpu and FLAGS.tpu_name: 426 | tpu_cluster_resolver = tf.contrib.cluster_resolver.TPUClusterResolver( 427 | FLAGS.tpu_name, zone=FLAGS.tpu_zone, project=FLAGS.gcp_project) 428 | 429 | is_per_host = tf.contrib.tpu.InputPipelineConfig.PER_HOST_V2 430 | run_config = tf.contrib.tpu.RunConfig( 431 | cluster=tpu_cluster_resolver, 432 | master=FLAGS.master, 433 | model_dir=FLAGS.output_dir, 434 | save_checkpoints_steps=FLAGS.save_checkpoints_steps, 435 | tpu_config=tf.contrib.tpu.TPUConfig( 436 | iterations_per_loop=FLAGS.iterations_per_loop, 437 | num_shards=FLAGS.num_tpu_cores, 438 | per_host_input_for_training=is_per_host)) 439 | 440 | model_fn = model_fn_builder( 441 | bert_config=bert_config, 442 | init_checkpoint=FLAGS.init_checkpoint, 443 | learning_rate=FLAGS.learning_rate, 444 | num_train_steps=FLAGS.num_train_steps, 445 | num_warmup_steps=FLAGS.num_warmup_steps, 446 | use_tpu=FLAGS.use_tpu, 447 | use_one_hot_embeddings=FLAGS.use_tpu) 448 | 449 | # If TPU is not available, this will fall back to normal Estimator on CPU 450 | # or GPU. 451 | estimator = tf.contrib.tpu.TPUEstimator( 452 | use_tpu=FLAGS.use_tpu, 453 | model_fn=model_fn, 454 | config=run_config, 455 | train_batch_size=FLAGS.train_batch_size, 456 | eval_batch_size=FLAGS.eval_batch_size) 457 | 458 | if FLAGS.do_train: 459 | tf.logging.info("***** Running training *****") 460 | tf.logging.info(" Batch size = %d", FLAGS.train_batch_size) 461 | train_input_fn = input_fn_builder( 462 | input_files=input_files, 463 | max_seq_length=FLAGS.max_seq_length, 464 | max_predictions_per_seq=FLAGS.max_predictions_per_seq, 465 | is_training=True) 466 | estimator.train(input_fn=train_input_fn, max_steps=FLAGS.num_train_steps) 467 | 468 | if FLAGS.do_eval: 469 | tf.logging.info("***** Running evaluation *****") 470 | tf.logging.info(" Batch size = %d", FLAGS.eval_batch_size) 471 | 472 | eval_input_fn = input_fn_builder( 473 | input_files=input_files, 474 | max_seq_length=FLAGS.max_seq_length, 475 | max_predictions_per_seq=FLAGS.max_predictions_per_seq, 476 | is_training=False) 477 | 478 | result = estimator.evaluate( 479 | input_fn=eval_input_fn, steps=FLAGS.max_eval_steps) 480 | 481 | output_eval_file = os.path.join(FLAGS.output_dir, "eval_results.txt") 482 | with tf.gfile.GFile(output_eval_file, "w") as writer: 483 | tf.logging.info("***** Eval results *****") 484 | for key in sorted(result.keys()): 485 | tf.logging.info(" %s = %s", key, str(result[key])) 486 | writer.write("%s = %s\n" % (key, str(result[key]))) 487 | 488 | 489 | if __name__ == "__main__": 490 | flags.mark_flag_as_required("input_file") 491 | flags.mark_flag_as_required("bert_config_file") 492 | flags.mark_flag_as_required("output_dir") 493 | tf.app.run() 494 | -------------------------------------------------------------------------------- /sample_text.txt: -------------------------------------------------------------------------------- 1 | This text is included to make sure Unicode is handled properly: 力加勝北区ᴵᴺᵀᵃছজটডণত 2 | Text should be one-sentence-per-line, with empty lines between documents. 3 | This sample text is public domain and was randomly selected from Project Guttenberg. 4 | 5 | The rain had only ceased with the gray streaks of morning at Blazing Star, and the settlement awoke to a moral sense of cleanliness, and the finding of forgotten knives, tin cups, and smaller camp utensils, where the heavy showers had washed away the debris and dust heaps before the cabin doors. 6 | Indeed, it was recorded in Blazing Star that a fortunate early riser had once picked up on the highway a solid chunk of gold quartz which the rain had freed from its incumbering soil, and washed into immediate and glittering popularity. 7 | Possibly this may have been the reason why early risers in that locality, during the rainy season, adopted a thoughtful habit of body, and seldom lifted their eyes to the rifted or india-ink washed skies above them. 8 | "Cass" Beard had risen early that morning, but not with a view to discovery. 9 | A leak in his cabin roof,--quite consistent with his careless, improvident habits,--had roused him at 4 A. M., with a flooded "bunk" and wet blankets. 10 | The chips from his wood pile refused to kindle a fire to dry his bed-clothes, and he had recourse to a more provident neighbor's to supply the deficiency. 11 | This was nearly opposite. 12 | Mr. Cassius crossed the highway, and stopped suddenly. 13 | Something glittered in the nearest red pool before him. 14 | Gold, surely! 15 | But, wonderful to relate, not an irregular, shapeless fragment of crude ore, fresh from Nature's crucible, but a bit of jeweler's handicraft in the form of a plain gold ring. 16 | Looking at it more attentively, he saw that it bore the inscription, "May to Cass." 17 | Like most of his fellow gold-seekers, Cass was superstitious. 18 | 19 | The fountain of classic wisdom, Hypatia herself. 20 | As the ancient sage--the name is unimportant to a monk--pumped water nightly that he might study by day, so I, the guardian of cloaks and parasols, at the sacred doors of her lecture-room, imbibe celestial knowledge. 21 | From my youth I felt in me a soul above the matter-entangled herd. 22 | She revealed to me the glorious fact, that I am a spark of Divinity itself. 23 | A fallen star, I am, sir!' continued he, pensively, stroking his lean stomach--'a fallen star!--fallen, if the dignity of philosophy will allow of the simile, among the hogs of the lower world--indeed, even into the hog-bucket itself. Well, after all, I will show you the way to the Archbishop's. 24 | There is a philosophic pleasure in opening one's treasures to the modest young. 25 | Perhaps you will assist me by carrying this basket of fruit?' And the little man jumped up, put his basket on Philammon's head, and trotted off up a neighbouring street. 26 | Philammon followed, half contemptuous, half wondering at what this philosophy might be, which could feed the self-conceit of anything so abject as his ragged little apish guide; 27 | but the novel roar and whirl of the street, the perpetual stream of busy faces, the line of curricles, palanquins, laden asses, camels, elephants, which met and passed him, and squeezed him up steps and into doorways, as they threaded their way through the great Moon-gate into the ample street beyond, drove everything from his mind but wondering curiosity, and a vague, helpless dread of that great living wilderness, more terrible than any dead wilderness of sand which he had left behind. 28 | Already he longed for the repose, the silence of the Laura--for faces which knew him and smiled upon him; but it was too late to turn back now. 29 | His guide held on for more than a mile up the great main street, crossed in the centre of the city, at right angles, by one equally magnificent, at each end of which, miles away, appeared, dim and distant over the heads of the living stream of passengers, the yellow sand-hills of the desert; 30 | while at the end of the vista in front of them gleamed the blue harbour, through a network of countless masts. 31 | At last they reached the quay at the opposite end of the street; 32 | and there burst on Philammon's astonished eyes a vast semicircle of blue sea, ringed with palaces and towers. 33 | He stopped involuntarily; and his little guide stopped also, and looked askance at the young monk, to watch the effect which that grand panorama should produce on him. 34 | -------------------------------------------------------------------------------- /tokenization.py: -------------------------------------------------------------------------------- 1 | # coding=utf-8 2 | # Copyright 2018 The Google AI Language Team Authors. 3 | # 4 | # Licensed under the Apache License, Version 2.0 (the "License"); 5 | # you may not use this file except in compliance with the License. 6 | # You may obtain a copy of the License at 7 | # 8 | # http://www.apache.org/licenses/LICENSE-2.0 9 | # 10 | # Unless required by applicable law or agreed to in writing, software 11 | # distributed under the License is distributed on an "AS IS" BASIS, 12 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 13 | # See the License for the specific language governing permissions and 14 | # limitations under the License. 15 | """Tokenization classes.""" 16 | 17 | from __future__ import absolute_import 18 | from __future__ import division 19 | from __future__ import print_function 20 | 21 | import collections 22 | import unicodedata 23 | import six 24 | import tensorflow as tf 25 | 26 | 27 | def convert_to_unicode(text): 28 | """Converts `text` to Unicode (if it's not already), assuming utf-8 input.""" 29 | if six.PY3: 30 | if isinstance(text, str): 31 | return text 32 | elif isinstance(text, bytes): 33 | return text.decode("utf-8", "ignore") 34 | else: 35 | raise ValueError("Unsupported string type: %s" % (type(text))) 36 | elif six.PY2: 37 | if isinstance(text, str): 38 | return text.decode("utf-8", "ignore") 39 | elif isinstance(text, unicode): 40 | return text 41 | else: 42 | raise ValueError("Unsupported string type: %s" % (type(text))) 43 | else: 44 | raise ValueError("Not running on Python2 or Python 3?") 45 | 46 | 47 | def printable_text(text): 48 | """Returns text encoded in a way suitable for print or `tf.logging`.""" 49 | 50 | # These functions want `str` for both Python2 and Python3, but in one case 51 | # it's a Unicode string and in the other it's a byte string. 52 | if six.PY3: 53 | if isinstance(text, str): 54 | return text 55 | elif isinstance(text, bytes): 56 | return text.decode("utf-8", "ignore") 57 | else: 58 | raise ValueError("Unsupported string type: %s" % (type(text))) 59 | elif six.PY2: 60 | if isinstance(text, str): 61 | return text 62 | elif isinstance(text, unicode): 63 | return text.encode("utf-8") 64 | else: 65 | raise ValueError("Unsupported string type: %s" % (type(text))) 66 | else: 67 | raise ValueError("Not running on Python2 or Python 3?") 68 | 69 | 70 | def load_vocab(vocab_file): 71 | """Loads a vocabulary file into a dictionary.""" 72 | vocab = collections.OrderedDict() 73 | index = 0 74 | with tf.gfile.GFile(vocab_file, "r") as reader: 75 | while True: 76 | token = convert_to_unicode(reader.readline()) 77 | if not token: 78 | break 79 | token = token.strip() 80 | vocab[token] = index 81 | index += 1 82 | return vocab 83 | 84 | 85 | def convert_by_vocab(vocab, items): 86 | """Converts a sequence of [tokens|ids] using the vocab.""" 87 | output = [] 88 | for item in items: 89 | output.append(vocab[item]) 90 | return output 91 | 92 | 93 | def convert_tokens_to_ids(vocab, tokens): 94 | return convert_by_vocab(vocab, tokens) 95 | 96 | 97 | def convert_ids_to_tokens(inv_vocab, ids): 98 | return convert_by_vocab(inv_vocab, ids) 99 | 100 | 101 | def whitespace_tokenize(text): 102 | """Runs basic whitespace cleaning and splitting on a piece of text.""" 103 | text = text.strip() 104 | if not text: 105 | return [] 106 | tokens = text.split() 107 | return tokens 108 | 109 | 110 | class FullTokenizer(object): 111 | """Runs end-to-end tokenziation.""" 112 | 113 | def __init__(self, vocab_file, do_lower_case=True): 114 | self.vocab = load_vocab(vocab_file) 115 | self.inv_vocab = {v: k for k, v in self.vocab.items()} 116 | self.basic_tokenizer = BasicTokenizer(do_lower_case=do_lower_case) 117 | self.wordpiece_tokenizer = WordpieceTokenizer(vocab=self.vocab) 118 | 119 | def tokenize(self, text): 120 | split_tokens = [] 121 | for token in self.basic_tokenizer.tokenize(text): 122 | for sub_token in self.wordpiece_tokenizer.tokenize(token): 123 | split_tokens.append(sub_token) 124 | 125 | return split_tokens 126 | 127 | def convert_tokens_to_ids(self, tokens): 128 | return convert_by_vocab(self.vocab, tokens) 129 | 130 | def convert_ids_to_tokens(self, ids): 131 | return convert_by_vocab(self.inv_vocab, ids) 132 | 133 | 134 | class BasicTokenizer(object): 135 | """Runs basic tokenization (punctuation splitting, lower casing, etc.).""" 136 | 137 | def __init__(self, do_lower_case=True): 138 | """Constructs a BasicTokenizer. 139 | 140 | Args: 141 | do_lower_case: Whether to lower case the input. 142 | """ 143 | self.do_lower_case = do_lower_case 144 | 145 | def tokenize(self, text): 146 | """Tokenizes a piece of text.""" 147 | text = convert_to_unicode(text) 148 | text = self._clean_text(text) 149 | 150 | # This was added on November 1st, 2018 for the multilingual and Chinese 151 | # models. This is also applied to the English models now, but it doesn't 152 | # matter since the English models were not trained on any Chinese data 153 | # and generally don't have any Chinese data in them (there are Chinese 154 | # characters in the vocabulary because Wikipedia does have some Chinese 155 | # words in the English Wikipedia.). 156 | text = self._tokenize_chinese_chars(text) 157 | 158 | orig_tokens = whitespace_tokenize(text) 159 | split_tokens = [] 160 | for token in orig_tokens: 161 | if self.do_lower_case: 162 | token = token.lower() 163 | token = self._run_strip_accents(token) 164 | split_tokens.extend(self._run_split_on_punc(token)) 165 | 166 | output_tokens = whitespace_tokenize(" ".join(split_tokens)) 167 | return output_tokens 168 | 169 | def _run_strip_accents(self, text): 170 | """Strips accents from a piece of text.""" 171 | text = unicodedata.normalize("NFD", text) 172 | output = [] 173 | for char in text: 174 | cat = unicodedata.category(char) 175 | if cat == "Mn": 176 | continue 177 | output.append(char) 178 | return "".join(output) 179 | 180 | def _run_split_on_punc(self, text): 181 | """Splits punctuation on a piece of text.""" 182 | chars = list(text) 183 | i = 0 184 | start_new_word = True 185 | output = [] 186 | while i < len(chars): 187 | char = chars[i] 188 | if _is_punctuation(char): 189 | output.append([char]) 190 | start_new_word = True 191 | else: 192 | if start_new_word: 193 | output.append([]) 194 | start_new_word = False 195 | output[-1].append(char) 196 | i += 1 197 | 198 | return ["".join(x) for x in output] 199 | 200 | def _tokenize_chinese_chars(self, text): 201 | """Adds whitespace around any CJK character.""" 202 | output = [] 203 | for char in text: 204 | cp = ord(char) 205 | if self._is_chinese_char(cp): 206 | output.append(" ") 207 | output.append(char) 208 | output.append(" ") 209 | else: 210 | output.append(char) 211 | return "".join(output) 212 | 213 | def _is_chinese_char(self, cp): 214 | """Checks whether CP is the codepoint of a CJK character.""" 215 | # This defines a "chinese character" as anything in the CJK Unicode block: 216 | # https://en.wikipedia.org/wiki/CJK_Unified_Ideographs_(Unicode_block) 217 | # 218 | # Note that the CJK Unicode block is NOT all Japanese and Korean characters, 219 | # despite its name. The modern Korean Hangul alphabet is a different block, 220 | # as is Japanese Hiragana and Katakana. Those alphabets are used to write 221 | # space-separated words, so they are not treated specially and handled 222 | # like the all of the other languages. 223 | if ((cp >= 0x4E00 and cp <= 0x9FFF) or # 224 | (cp >= 0x3400 and cp <= 0x4DBF) or # 225 | (cp >= 0x20000 and cp <= 0x2A6DF) or # 226 | (cp >= 0x2A700 and cp <= 0x2B73F) or # 227 | (cp >= 0x2B740 and cp <= 0x2B81F) or # 228 | (cp >= 0x2B820 and cp <= 0x2CEAF) or 229 | (cp >= 0xF900 and cp <= 0xFAFF) or # 230 | (cp >= 0x2F800 and cp <= 0x2FA1F)): # 231 | return True 232 | 233 | return False 234 | 235 | def _clean_text(self, text): 236 | """Performs invalid character removal and whitespace cleanup on text.""" 237 | output = [] 238 | for char in text: 239 | cp = ord(char) 240 | if cp == 0 or cp == 0xfffd or _is_control(char): 241 | continue 242 | if _is_whitespace(char): 243 | output.append(" ") 244 | else: 245 | output.append(char) 246 | return "".join(output) 247 | 248 | 249 | class WordpieceTokenizer(object): 250 | """Runs WordPiece tokenziation.""" 251 | 252 | def __init__(self, vocab, unk_token="[UNK]", max_input_chars_per_word=200): 253 | self.vocab = vocab 254 | self.unk_token = unk_token 255 | self.max_input_chars_per_word = max_input_chars_per_word 256 | 257 | def tokenize(self, text): 258 | """Tokenizes a piece of text into its word pieces. 259 | 260 | This uses a greedy longest-match-first algorithm to perform tokenization 261 | using the given vocabulary. 262 | 263 | For example: 264 | input = "unaffable" 265 | output = ["un", "##aff", "##able"] 266 | 267 | Args: 268 | text: A single token or whitespace separated tokens. This should have 269 | already been passed through `BasicTokenizer. 270 | 271 | Returns: 272 | A list of wordpiece tokens. 273 | """ 274 | 275 | text = convert_to_unicode(text) 276 | 277 | output_tokens = [] 278 | for token in whitespace_tokenize(text): 279 | chars = list(token) 280 | if len(chars) > self.max_input_chars_per_word: 281 | output_tokens.append(self.unk_token) 282 | continue 283 | 284 | is_bad = False 285 | start = 0 286 | sub_tokens = [] 287 | while start < len(chars): 288 | end = len(chars) 289 | cur_substr = None 290 | while start < end: 291 | substr = "".join(chars[start:end]) 292 | if start > 0: 293 | substr = "##" + substr 294 | if substr in self.vocab: 295 | cur_substr = substr 296 | break 297 | end -= 1 298 | if cur_substr is None: 299 | is_bad = True 300 | break 301 | sub_tokens.append(cur_substr) 302 | start = end 303 | 304 | if is_bad: 305 | output_tokens.append(self.unk_token) 306 | else: 307 | output_tokens.extend(sub_tokens) 308 | return output_tokens 309 | 310 | 311 | def _is_whitespace(char): 312 | """Checks whether `chars` is a whitespace character.""" 313 | # \t, \n, and \r are technically contorl characters but we treat them 314 | # as whitespace since they are generally considered as such. 315 | if char == " " or char == "\t" or char == "\n" or char == "\r": 316 | return True 317 | cat = unicodedata.category(char) 318 | if cat == "Zs": 319 | return True 320 | return False 321 | 322 | 323 | def _is_control(char): 324 | """Checks whether `chars` is a control character.""" 325 | # These are technically control characters but we count them as whitespace 326 | # characters. 327 | if char == "\t" or char == "\n" or char == "\r": 328 | return False 329 | cat = unicodedata.category(char) 330 | if cat.startswith("C"): 331 | return True 332 | return False 333 | 334 | 335 | def _is_punctuation(char): 336 | """Checks whether `chars` is a punctuation character.""" 337 | cp = ord(char) 338 | # We treat all non-letter/number ASCII as punctuation. 339 | # Characters such as "^", "$", and "`" are not in the Unicode 340 | # Punctuation class but we treat them as punctuation anyways, for 341 | # consistency. 342 | if ((cp >= 33 and cp <= 47) or (cp >= 58 and cp <= 64) or 343 | (cp >= 91 and cp <= 96) or (cp >= 123 and cp <= 126)): 344 | return True 345 | cat = unicodedata.category(char) 346 | if cat.startswith("P"): 347 | return True 348 | return False 349 | -------------------------------------------------------------------------------- /tokenization_test.py: -------------------------------------------------------------------------------- 1 | # coding=utf-8 2 | # Copyright 2018 The Google AI Language Team Authors. 3 | # 4 | # Licensed under the Apache License, Version 2.0 (the "License"); 5 | # you may not use this file except in compliance with the License. 6 | # You may obtain a copy of the License at 7 | # 8 | # http://www.apache.org/licenses/LICENSE-2.0 9 | # 10 | # Unless required by applicable law or agreed to in writing, software 11 | # distributed under the License is distributed on an "AS IS" BASIS, 12 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 13 | # See the License for the specific language governing permissions and 14 | # limitations under the License. 15 | from __future__ import absolute_import 16 | from __future__ import division 17 | from __future__ import print_function 18 | 19 | import os 20 | import tempfile 21 | 22 | import tokenization 23 | import tensorflow as tf 24 | 25 | 26 | class TokenizationTest(tf.test.TestCase): 27 | 28 | def test_full_tokenizer(self): 29 | vocab_tokens = [ 30 | "[UNK]", "[CLS]", "[SEP]", "want", "##want", "##ed", "wa", "un", "runn", 31 | "##ing", "," 32 | ] 33 | with tempfile.NamedTemporaryFile(delete=False) as vocab_writer: 34 | vocab_writer.write("".join([x + "\n" for x in vocab_tokens])) 35 | 36 | vocab_file = vocab_writer.name 37 | 38 | tokenizer = tokenization.FullTokenizer(vocab_file) 39 | os.unlink(vocab_file) 40 | 41 | tokens = tokenizer.tokenize(u"UNwant\u00E9d,running") 42 | self.assertAllEqual(tokens, ["un", "##want", "##ed", ",", "runn", "##ing"]) 43 | 44 | self.assertAllEqual( 45 | tokenizer.convert_tokens_to_ids(tokens), [7, 4, 5, 10, 8, 9]) 46 | 47 | def test_chinese(self): 48 | tokenizer = tokenization.BasicTokenizer() 49 | 50 | self.assertAllEqual( 51 | tokenizer.tokenize(u"ah\u535A\u63A8zz"), 52 | [u"ah", u"\u535A", u"\u63A8", u"zz"]) 53 | 54 | def test_basic_tokenizer_lower(self): 55 | tokenizer = tokenization.BasicTokenizer(do_lower_case=True) 56 | 57 | self.assertAllEqual( 58 | tokenizer.tokenize(u" \tHeLLo!how \n Are yoU? "), 59 | ["hello", "!", "how", "are", "you", "?"]) 60 | self.assertAllEqual(tokenizer.tokenize(u"H\u00E9llo"), ["hello"]) 61 | 62 | def test_basic_tokenizer_no_lower(self): 63 | tokenizer = tokenization.BasicTokenizer(do_lower_case=False) 64 | 65 | self.assertAllEqual( 66 | tokenizer.tokenize(u" \tHeLLo!how \n Are yoU? "), 67 | ["HeLLo", "!", "how", "Are", "yoU", "?"]) 68 | 69 | def test_wordpiece_tokenizer(self): 70 | vocab_tokens = [ 71 | "[UNK]", "[CLS]", "[SEP]", "want", "##want", "##ed", "wa", "un", "runn", 72 | "##ing" 73 | ] 74 | 75 | vocab = {} 76 | for (i, token) in enumerate(vocab_tokens): 77 | vocab[token] = i 78 | tokenizer = tokenization.WordpieceTokenizer(vocab=vocab) 79 | 80 | self.assertAllEqual(tokenizer.tokenize(""), []) 81 | 82 | self.assertAllEqual( 83 | tokenizer.tokenize("unwanted running"), 84 | ["un", "##want", "##ed", "runn", "##ing"]) 85 | 86 | self.assertAllEqual( 87 | tokenizer.tokenize("unwantedX running"), ["[UNK]", "runn", "##ing"]) 88 | 89 | def test_convert_tokens_to_ids(self): 90 | vocab_tokens = [ 91 | "[UNK]", "[CLS]", "[SEP]", "want", "##want", "##ed", "wa", "un", "runn", 92 | "##ing" 93 | ] 94 | 95 | vocab = {} 96 | for (i, token) in enumerate(vocab_tokens): 97 | vocab[token] = i 98 | 99 | self.assertAllEqual( 100 | tokenization.convert_tokens_to_ids( 101 | vocab, ["un", "##want", "##ed", "runn", "##ing"]), [7, 4, 5, 8, 9]) 102 | 103 | def test_is_whitespace(self): 104 | self.assertTrue(tokenization._is_whitespace(u" ")) 105 | self.assertTrue(tokenization._is_whitespace(u"\t")) 106 | self.assertTrue(tokenization._is_whitespace(u"\r")) 107 | self.assertTrue(tokenization._is_whitespace(u"\n")) 108 | self.assertTrue(tokenization._is_whitespace(u"\u00A0")) 109 | 110 | self.assertFalse(tokenization._is_whitespace(u"A")) 111 | self.assertFalse(tokenization._is_whitespace(u"-")) 112 | 113 | def test_is_control(self): 114 | self.assertTrue(tokenization._is_control(u"\u0005")) 115 | 116 | self.assertFalse(tokenization._is_control(u"A")) 117 | self.assertFalse(tokenization._is_control(u" ")) 118 | self.assertFalse(tokenization._is_control(u"\t")) 119 | self.assertFalse(tokenization._is_control(u"\r")) 120 | 121 | def test_is_punctuation(self): 122 | self.assertTrue(tokenization._is_punctuation(u"-")) 123 | self.assertTrue(tokenization._is_punctuation(u"$")) 124 | self.assertTrue(tokenization._is_punctuation(u"`")) 125 | self.assertTrue(tokenization._is_punctuation(u".")) 126 | 127 | self.assertFalse(tokenization._is_punctuation(u"A")) 128 | self.assertFalse(tokenization._is_punctuation(u" ")) 129 | 130 | 131 | if __name__ == "__main__": 132 | tf.test.main() 133 | --------------------------------------------------------------------------------