├── .github ├── ISSUE_TEMPLATE │ ├── bug-report.md │ ├── feature-request.md │ └── usage-question.md └── stale.yml ├── .gitignore ├── CITATION.cff ├── CONTRIBUTING.md ├── DISCLAIMER ├── LICENSE ├── README.md ├── README_EN.md ├── codeassist ├── __init__.py ├── create_dataset.py ├── gpt2_coder.py └── wizard_coder.py ├── docs ├── 36-text-rep-examples.md ├── api.png ├── codeassist.png ├── hf_model.png ├── source_code_datasets │ ├── source_code_dataset.md │ └── upload_dataset.py └── wechat.jpeg ├── examples ├── data │ ├── code_alpaca_20k_50.jsonl │ └── train_code_5k.txt ├── distilgpt2_demo.py ├── gpt2_demo.py ├── gradio_demo.py ├── inference_demo.py ├── original_gpt2_demo.py ├── prepare_code_data.py ├── server.py ├── training_gpt2_mydata.py ├── training_wizardcoder_mydata.py ├── use_transformers_gpt2.py └── wizardcoder_demo.py ├── requirements.txt ├── setup.cfg ├── setup.py └── tests ├── test.txt ├── test_issue.py └── test_qps.py /.github/ISSUE_TEMPLATE/bug-report.md: -------------------------------------------------------------------------------- 1 | --- 2 | name: Bug Report 3 | about: Create a report to help us improve 4 | title: '' 5 | labels: bug 6 | assignees: '' 7 | 8 | --- 9 | 10 | ### Describe the bug 11 | Please provide a clear and concise description of what the bug is. If applicable, add screenshots to help explain your problem, especially for visualization related problems. 12 | -------------------------------------------------------------------------------- /.github/ISSUE_TEMPLATE/feature-request.md: -------------------------------------------------------------------------------- 1 | --- 2 | name: Feature Request 3 | about: Suggest an idea for this project 4 | title: '' 5 | labels: enhancement 6 | assignees: '' 7 | 8 | --- 9 | 10 | - [ ] I checked to make sure that this is not a duplicate issue 11 | 12 | ### Describe the solution you'd like 13 | A clear and concise description of what you want to happen. 14 | -------------------------------------------------------------------------------- /.github/ISSUE_TEMPLATE/usage-question.md: -------------------------------------------------------------------------------- 1 | --- 2 | name: Usage Question 3 | about: Ask a question about usage 4 | title: '' 5 | labels: question 6 | assignees: '' 7 | 8 | --- 9 | 10 | ### Describe the Question 11 | Please provide a clear and concise description of what the question is. 12 | 13 | -------------------------------------------------------------------------------- /.github/stale.yml: -------------------------------------------------------------------------------- 1 | # Number of days of inactivity before an issue becomes stale 2 | daysUntilStale: 60 3 | # Number of days of inactivity before a stale issue is closed 4 | daysUntilClose: 7 5 | # Issues with these labels will never be considered stale 6 | exemptLabels: 7 | - pinned 8 | - security 9 | # Label to use when marking an issue as stale 10 | staleLabel: wontfix 11 | # Comment to post when marking an issue as stale. Set to `false` to disable 12 | markComment: > 13 | This issue has been automatically marked as stale because it has not had 14 | recent activity. It will be closed if no further activity occurs. Thank you 15 | for your contributions.(由于长期不活动，机器人自动关闭此问题，如果需要欢迎提问) 16 | # Comment to post when closing a stale issue. Set to `false` to disable 17 | closeComment: false -------------------------------------------------------------------------------- /.gitignore: -------------------------------------------------------------------------------- 1 | # Byte-compiled / optimized / DLL files 2 | __pycache__/ 3 | *.py[cod] 4 | *$py.class 5 | 6 | # C extensions 7 | *.so 8 | 9 | # Distribution / packaging 10 | .Python 11 | build/ 12 | develop-eggs/ 13 | dist/ 14 | downloads/ 15 | eggs/ 16 | .eggs/ 17 | lib/ 18 | lib64/ 19 | parts/ 20 | sdist/ 21 | var/ 22 | wheels/ 23 | *.egg-info/ 24 | .installed.cfg 25 | *.egg 26 | MANIFEST 27 | 28 | # PyInstaller 29 | # Usually these files are written by a python script from a template 30 | # before PyInstaller builds the exe, so as to inject date/other infos into it. 31 | *.manifest 32 | *.spec 33 | 34 | # Installer logs 35 | pip-log.txt 36 | pip-delete-this-directory.txt 37 | 38 | # Unit test / coverage reports 39 | htmlcov/ 40 | .tox/ 41 | .coverage 42 | .coverage.* 43 | .cache 44 | nosetests.xml 45 | coverage.xml 46 | *.cover 47 | .hypothesis/ 48 | .pytest_cache/ 49 | 50 | # Translations 51 | *.mo 52 | *.pot 53 | 54 | # Django stuff: 55 | *.log 56 | local_settings.py 57 | db.sqlite3 58 | 59 | # Flask stuff: 60 | instance/ 61 | .webassets-cache 62 | 63 | # Scrapy stuff: 64 | .scrapy 65 | 66 | # Sphinx documentation 67 | docs/_build/ 68 | 69 | # PyBuilder 70 | target/ 71 | 72 | # Jupyter Notebook 73 | .ipynb_checkpoints 74 | 75 | # pyenv 76 | .python-version 77 | 78 | # celery beat schedule file 79 | celerybeat-schedule 80 | 81 | # SageMath parsed files 82 | *.sage.py 83 | 84 | # Environments 85 | .env 86 | .venv 87 | env/ 88 | venv/ 89 | ENV/ 90 | env.bak/ 91 | venv.bak/ 92 | 93 | # Spyder project settings 94 | .spyderproject 95 | .spyproject 96 | 97 | # Rope project settings 98 | .ropeproject 99 | 100 | # mkdocs documentation 101 | /site 102 | 103 | # mypy 104 | .mypy_cache/ 105 | .idea 106 | .github 107 | *.zip 108 | download/ -------------------------------------------------------------------------------- /CITATION.cff: -------------------------------------------------------------------------------- 1 | cff-version: 1.2.0 2 | message: "If you use this software, please cite it as below." 3 | authors: 4 | - family-names: "Xu" 5 | given-names: "Ming" 6 | title: "code-autocomplete: Code AutoComplete with GPT2 model" 7 | url: "https://github.com/shibing624/code-autocomplete" 8 | data-released: 2022-03-01 9 | version: 0.0.4 -------------------------------------------------------------------------------- /CONTRIBUTING.md: -------------------------------------------------------------------------------- 1 | # Contributing 2 | 3 | We are happy to accept your contributions to make `codeassist` better and more awesome! To avoid unnecessary work on either 4 | side, please stick to the following process: 5 | 6 | 1. Check if there is already [an issue](https://github.com/shibing624/codeassist/issues) for your concern. 7 | 2. If there is not, open a new one to start a discussion. We hate to close finished PRs! 8 | 3. If we decide your concern needs code changes, we would be happy to accept a pull request. Please consider the 9 | commit guidelines below. -------------------------------------------------------------------------------- /DISCLAIMER: -------------------------------------------------------------------------------- 1 | The software project, data, and models provided by our GitHub project are provided "as is," without warranty of any kind, express or implied, including but not limited to the warranties of merchantability, fitness for a particular purpose, and non-infringement. 2 | 3 | In no event shall the project owners or contributors be liable for any direct, indirect, incidental, special, exemplary, or consequential damages (including, but not limited to, procurement of substitute goods or services; loss of use, data, or profits; or business interruption) however caused and on any theory of liability, whether in contract, strict liability, or tort (including negligence or otherwise) arising in any way out of the use of this software project, data, or models, even if advised of the possibility of such damage. 4 | 5 | Users of this software project, data, and models are solely responsible for any consequences of their use. The project owners and contributors shall not be held responsible for any subsequent or potential harm caused by the use of this software project, data, or models. 6 | 7 | By using this software project, data, or models, users accept and agree to this disclaimer. If users do not agree to the terms of this disclaimer, they should not use this software project, data, or models. 8 | 9 | It is important to note that this software project, data, and models are still in the research phase and are provided for experimental purposes only. As such, the project owners and contributors do not guarantee the accuracy, completeness, or usefulness of the software project, data, or models. 10 | 11 | Furthermore, due to the experimental nature of this software project, data, and models, it is possible that they may contain or generate inappropriate responses, errors, or inconsistencies. Users should exercise caution when using this software project, data, or models, and should not rely solely on them for any critical or sensitive tasks. 12 | 13 | The project owners and contributors shall not be held responsible for any damages, losses, or liabilities arising from the use of this software project, data, or models, including but not limited to, any inappropriate responses generated by the software project, data, or models. 14 | 15 | By using this software project, data, or models, users acknowledge and accept the experimental nature of the software project, data, and models, and understand the potential risks and limitations associated with their use. If users do not agree to the terms of this disclaimer, they should not use this software project, data, or models. 16 | 17 | The software project, data, and models provided by our GitHub project are intended for research purposes only. They should not be used for any commercial, business, or legal purposes, and should not be relied upon as a substitute for professional advice or judgment. 18 | 19 | Users of this software project, data, and models are strictly prohibited from using them for any commercial purposes, including but not limited to, selling, licensing, or distributing the software project, data, or models to third parties. 20 | 21 | The project owners and contributors shall not be held responsible for any damages, losses, or liabilities arising from the use of this software project, data, or models for any commercial or business purposes. 22 | 23 | By using this software project, data, or models, users agree to use them for research purposes only, and not for any commercial or business purposes. If users do not agree to the terms of this disclaimer, they should not use this software project, data, or models. -------------------------------------------------------------------------------- /LICENSE: -------------------------------------------------------------------------------- 1 | Apache License 2 | Version 2.0, January 2004 3 | http://www.apache.org/licenses/ 4 | 5 | TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION 6 | 7 | 1. Definitions. 8 | 9 | "License" shall mean the terms and conditions for use, reproduction, 10 | and distribution as defined by Sections 1 through 9 of this document. 11 | 12 | "Licensor" shall mean the copyright owner or entity authorized by 13 | the copyright owner that is granting the License. 14 | 15 | "Legal Entity" shall mean the union of the acting entity and all 16 | other entities that control, are controlled by, or are under common 17 | control with that entity. For the purposes of this definition, 18 | "control" means (i) the power, direct or indirect, to cause the 19 | direction or management of such entity, whether by contract or 20 | otherwise, or (ii) ownership of fifty percent (50%) or more of the 21 | outstanding shares, or (iii) beneficial ownership of such entity. 22 | 23 | "You" (or "Your") shall mean an individual or Legal Entity 24 | exercising permissions granted by this License. 25 | 26 | "Source" form shall mean the preferred form for making modifications, 27 | including but not limited to software source code, documentation 28 | source, and configuration files. 29 | 30 | "Object" form shall mean any form resulting from mechanical 31 | transformation or translation of a Source form, including but 32 | not limited to compiled object code, generated documentation, 33 | and conversions to other media types. 34 | 35 | "Work" shall mean the work of authorship, whether in Source or 36 | Object form, made available under the License, as indicated by a 37 | copyright notice that is included in or attached to the work 38 | (an example is provided in the Appendix below). 39 | 40 | "Derivative Works" shall mean any work, whether in Source or Object 41 | form, that is based on (or derived from) the Work and for which the 42 | editorial revisions, annotations, elaborations, or other modifications 43 | represent, as a whole, an original work of authorship. For the purposes 44 | of this License, Derivative Works shall not include works that remain 45 | separable from, or merely link (or bind by name) to the interfaces of, 46 | the Work and Derivative Works thereof. 47 | 48 | "Contribution" shall mean any work of authorship, including 49 | the original version of the Work and any modifications or additions 50 | to that Work or Derivative Works thereof, that is intentionally 51 | submitted to Licensor for inclusion in the Work by the copyright owner 52 | or by an individual or Legal Entity authorized to submit on behalf of 53 | the copyright owner. For the purposes of this definition, "submitted" 54 | means any form of electronic, verbal, or written communication sent 55 | to the Licensor or its representatives, including but not limited to 56 | communication on electronic mailing lists, source code control systems, 57 | and issue tracking systems that are managed by, or on behalf of, the 58 | Licensor for the purpose of discussing and improving the Work, but 59 | excluding communication that is conspicuously marked or otherwise 60 | designated in writing by the copyright owner as "Not a Contribution." 61 | 62 | "Contributor" shall mean Licensor and any individual or Legal Entity 63 | on behalf of whom a Contribution has been received by Licensor and 64 | subsequently incorporated within the Work. 65 | 66 | 2. Grant of Copyright License. Subject to the terms and conditions of 67 | this License, each Contributor hereby grants to You a perpetual, 68 | worldwide, non-exclusive, no-charge, royalty-free, irrevocable 69 | copyright license to reproduce, prepare Derivative Works of, 70 | publicly display, publicly perform, sublicense, and distribute the 71 | Work and such Derivative Works in Source or Object form. 72 | 73 | 3. Grant of Patent License. Subject to the terms and conditions of 74 | this License, each Contributor hereby grants to You a perpetual, 75 | worldwide, non-exclusive, no-charge, royalty-free, irrevocable 76 | (except as stated in this section) patent license to make, have made, 77 | use, offer to sell, sell, import, and otherwise transfer the Work, 78 | where such license applies only to those patent claims licensable 79 | by such Contributor that are necessarily infringed by their 80 | Contribution(s) alone or by combination of their Contribution(s) 81 | with the Work to which such Contribution(s) was submitted. If You 82 | institute patent litigation against any entity (including a 83 | cross-claim or counterclaim in a lawsuit) alleging that the Work 84 | or a Contribution incorporated within the Work constitutes direct 85 | or contributory patent infringement, then any patent licenses 86 | granted to You under this License for that Work shall terminate 87 | as of the date such litigation is filed. 88 | 89 | 4. Redistribution. You may reproduce and distribute copies of the 90 | Work or Derivative Works thereof in any medium, with or without 91 | modifications, and in Source or Object form, provided that You 92 | meet the following conditions: 93 | 94 | (a) You must give any other recipients of the Work or 95 | Derivative Works a copy of this License; and 96 | 97 | (b) You must cause any modified files to carry prominent notices 98 | stating that You changed the files; and 99 | 100 | (c) You must retain, in the Source form of any Derivative Works 101 | that You distribute, all copyright, patent, trademark, and 102 | attribution notices from the Source form of the Work, 103 | excluding those notices that do not pertain to any part of 104 | the Derivative Works; and 105 | 106 | (d) If the Work includes a "NOTICE" text file as part of its 107 | distribution, then any Derivative Works that You distribute must 108 | include a readable copy of the attribution notices contained 109 | within such NOTICE file, excluding those notices that do not 110 | pertain to any part of the Derivative Works, in at least one 111 | of the following places: within a NOTICE text file distributed 112 | as part of the Derivative Works; within the Source form or 113 | documentation, if provided along with the Derivative Works; or, 114 | within a display generated by the Derivative Works, if and 115 | wherever such third-party notices normally appear. The contents 116 | of the NOTICE file are for informational purposes only and 117 | do not modify the License. You may add Your own attribution 118 | notices within Derivative Works that You distribute, alongside 119 | or as an addendum to the NOTICE text from the Work, provided 120 | that such additional attribution notices cannot be construed 121 | as modifying the License. 122 | 123 | You may add Your own copyright statement to Your modifications and 124 | may provide additional or different license terms and conditions 125 | for use, reproduction, or distribution of Your modifications, or 126 | for any such Derivative Works as a whole, provided Your use, 127 | reproduction, and distribution of the Work otherwise complies with 128 | the conditions stated in this License. 129 | 130 | 5. Submission of Contributions. Unless You explicitly state otherwise, 131 | any Contribution intentionally submitted for inclusion in the Work 132 | by You to the Licensor shall be under the terms and conditions of 133 | this License, without any additional terms or conditions. 134 | Notwithstanding the above, nothing herein shall supersede or modify 135 | the terms of any separate license agreement you may have executed 136 | with Licensor regarding such Contributions. 137 | 138 | 6. Trademarks. This License does not grant permission to use the trade 139 | names, trademarks, service marks, or product names of the Licensor, 140 | except as required for reasonable and customary use in describing the 141 | origin of the Work and reproducing the content of the NOTICE file. 142 | 143 | 7. Disclaimer of Warranty. Unless required by applicable law or 144 | agreed to in writing, Licensor provides the Work (and each 145 | Contributor provides its Contributions) on an "AS IS" BASIS, 146 | WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or 147 | implied, including, without limitation, any warranties or conditions 148 | of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A 149 | PARTICULAR PURPOSE. You are solely responsible for determining the 150 | appropriateness of using or redistributing the Work and assume any 151 | risks associated with Your exercise of permissions under this License. 152 | 153 | 8. Limitation of Liability. In no event and under no legal theory, 154 | whether in tort (including negligence), contract, or otherwise, 155 | unless required by applicable law (such as deliberate and grossly 156 | negligent acts) or agreed to in writing, shall any Contributor be 157 | liable to You for damages, including any direct, indirect, special, 158 | incidental, or consequential damages of any character arising as a 159 | result of this License or out of the use or inability to use the 160 | Work (including but not limited to damages for loss of goodwill, 161 | work stoppage, computer failure or malfunction, or any and all 162 | other commercial damages or losses), even if such Contributor 163 | has been advised of the possibility of such damages. 164 | 165 | 9. Accepting Warranty or Additional Liability. While redistributing 166 | the Work or Derivative Works thereof, You may choose to offer, 167 | and charge a fee for, acceptance of support, warranty, indemnity, 168 | or other liability obligations and/or rights consistent with this 169 | License. However, in accepting such obligations, You may act only 170 | on Your own behalf and on Your sole responsibility, not on behalf 171 | of any other Contributor, and only if You agree to indemnify, 172 | defend, and hold each Contributor harmless for any liability 173 | incurred by, or claims asserted against, such Contributor by reason 174 | of your accepting any such warranty or additional liability. 175 | 176 | END OF TERMS AND CONDITIONS 177 | 178 | APPENDIX: How to apply the Apache License to your work. 179 | 180 | To apply the Apache License to your work, attach the following 181 | boilerplate notice, with the fields enclosed by brackets "[]" 182 | replaced with your own identifying information. (Don't include 183 | the brackets!) The text should be enclosed in the appropriate 184 | comment syntax for the file format. We also recommend that a 185 | file or class name and description of purpose be included on the 186 | same "printed page" as the copyright notice for easier 187 | identification within third-party archives. 188 | 189 | Copyright [yyyy] [name of copyright owner] 190 | 191 | Licensed under the Apache License, Version 2.0 (the "License"); 192 | you may not use this file except in compliance with the License. 193 | You may obtain a copy of the License at 194 | 195 | http://www.apache.org/licenses/LICENSE-2.0 196 | 197 | Unless required by applicable law or agreed to in writing, software 198 | distributed under the License is distributed on an "AS IS" BASIS, 199 | WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 200 | See the License for the specific language governing permissions and 201 | limitations under the License. 202 | -------------------------------------------------------------------------------- /README.md: -------------------------------------------------------------------------------- 1 | [**🇨🇳中文**](https://github.com/shibing624/codeassist/blob/main/README.md) | [**🌐English**](https://github.com/shibing624/codeassist/blob/main/README_EN.md) | [**📖文档/Docs**](https://github.com/shibing624/codeassist/wiki) | [**🤖模型/Models**](https://huggingface.co/shibing624) 2 | 3 |
4 | 5 | Logo 6 | 7 |
8 | 9 | ----------------- 10 | 11 | # CodeAssist: Advanced Code Completion Tool 12 | [![PyPI version](https://badge.fury.io/py/CodeAssist.svg)](https://badge.fury.io/py/CodeAssist) 13 | [![Contributions welcome](https://img.shields.io/badge/contributions-welcome-brightgreen.svg)](CONTRIBUTING.md) 14 | [![GitHub contributors](https://img.shields.io/github/contributors/shibing624/CodeAssist.svg)](https://github.com/shibing624/CodeAssist/graphs/contributors) 15 | [![License Apache 2.0](https://img.shields.io/badge/license-Apache%202.0-blue.svg)](LICENSE) 16 | [![python_vesion](https://img.shields.io/badge/Python-3.5%2B-green.svg)](requirements.txt) 17 | [![GitHub issues](https://img.shields.io/github/issues/shibing624/CodeAssist.svg)](https://github.com/shibing624/CodeAssist/issues) 18 | [![Wechat Group](https://img.shields.io/badge/wechat-group-green.svg?logo=wechat)](#Contact) 19 | 20 | ## Introduction 21 | 22 | **CodeAssist** is an advanced code completion tool that intelligently provides high-quality code completions for Python, Java, and C++ and so on. 23 | 24 | CodeAssist 是一个高级代码补全工具，高质量为 Python、Java 和 C++ 等编程语言补全代码 25 | 26 | 27 | ## Features 28 | 29 | - GPT based code completion 30 | - Code completion for `Python`, `Java`, `C++`, `javascript` and so on 31 | - Line and block code completion 32 | - Train(Fine-tuning) and predict model with your own data 33 | 34 | ### Release Models 35 | 36 | | Arch | BaseModel | Model | Model Size | 37 | |:-------|:------------------|:------------------------------------------------------------------------------------------------------------------------|:----------:| 38 | | GPT | gpt2 | [shibing624/code-autocomplete-gpt2-base](https://huggingface.co/shibing624/code-autocomplete-gpt2-base) | 487MB | 39 | | GPT | distilgpt2 | [shibing624/code-autocomplete-distilgpt2-python](https://huggingface.co/shibing624/code-autocomplete-distilgpt2-python) | 319MB | 40 | | GPT | bigcode/starcoder | [WizardLM/WizardCoder-15B-V1.0](https://huggingface.co/WizardLM/WizardCoder-15B-V1.0) | 29GB | 41 | 42 | 43 | 44 | ## Install 45 | 46 | ```shell 47 | pip install torch # conda install pytorch 48 | pip install -U codeassist 49 | ``` 50 | 51 | or 52 | 53 | ```shell 54 | git clone https://github.com/shibing624/codeassist.git 55 | cd CodeAssist 56 | python setup.py install 57 | ``` 58 | 59 | ## Usage 60 | 61 | ### WizardCoder model 62 | 63 | WizardCoder-15b is fine-tuned `bigcode/starcoder` with alpaca code data, you can use the following code to generate code: 64 | 65 | example: [examples/wizardcoder_demo.py](https://github.com/shibing624/CodeAssist/blob/main/examples/wizardcoder_demo.py) 66 | 67 | ```python 68 | import sys 69 | 70 | sys.path.append('..') 71 | from codeassist import WizardCoder 72 | 73 | m = WizardCoder("WizardLM/WizardCoder-15B-V1.0") 74 | print(m.generate('def load_csv_file(file_path):')[0]) 75 | ``` 76 | 77 | output: 78 | 79 | 80 | ```python 81 | import csv 82 | 83 | def load_csv_file(file_path): 84 | """ 85 | Load data from a CSV file and return a list of dictionaries. 86 | """ 87 | # Open the file in read mode 88 | with open(file_path, 'r') as file: 89 | # Create a CSV reader object 90 | csv_reader = csv.DictReader(file) 91 | # Initialize an empty list to store the data 92 | data = [] 93 | # Iterate over each row of data 94 | for row in csv_reader: 95 | # Append the row of data to the list 96 | data.append(row) 97 | # Return the list of data 98 | return data 99 | ``` 100 | 101 | model output is impressively effective, it currently supports English and Chinese input, you can enter instructions or code prefixes as required. 102 | 103 | ### distilgpt2 model 104 | 105 | 106 | distilgpt2 fine-tuned code autocomplete model, you can use the following code: 107 | 108 | example: [examples/distilgpt2_demo.py](https://github.com/shibing624/CodeAssist/blob/main/examples/distilgpt2_demo.py) 109 | 110 | ```python 111 | import sys 112 | 113 | sys.path.append('..') 114 | from codeassist import GPT2Coder 115 | 116 | m = GPT2Coder("shibing624/code-autocomplete-distilgpt2-python") 117 | print(m.generate('import torch.nn as')[0]) 118 | ``` 119 | 120 | output: 121 | 122 | ```shell 123 | import torch.nn as nn 124 | import torch.nn.functional as F 125 | ``` 126 | 127 | ### Use with huggingface/transformers： 128 | 129 | example: [examples/use_transformers_gpt2.py](https://github.com/shibing624/CodeAssist/blob/main/examples/use_transformers_gpt2.py) 130 | 131 | ### Train Model 132 | #### Train WizardCoder model 133 | example: [examples/training_wizardcoder_mydata.py](https://github.com/shibing624/CodeAssist/blob/main/examples/training_wizardcoder_mydata.py) 134 | 135 | ```shell 136 | cd examples 137 | CUDA_VISIBLE_DEVICES=0,1 python training_wizardcoder_mydata.py --do_train --do_predict --num_epochs 1 --output_dir outputs-wizard --model_name WizardLM/WizardCoder-15B-V1.0 138 | ``` 139 | 140 | - GPU memory: 31GB 141 | - finetune need 2*V100(32GB) 142 | - inference need 1*V100(32GB) 143 | 144 | #### Train distilgpt2 model 145 | example: [examples/training_gpt2_mydata.py](https://github.com/shibing624/CodeAssist/blob/main/examples/training_gpt2_mydata.py) 146 | 147 | ```shell 148 | cd examples 149 | python training_gpt2_mydata.py --do_train --do_predict --num_epochs 15 --output_dir outputs-gpt2 --model_name gpt2 150 | ``` 151 | 152 | PS: fine-tuned result model is GPT2-python: [shibing624/code-autocomplete-gpt2-base](https://huggingface.co/shibing624/code-autocomplete-gpt2-base), 153 | I spent about 24 hours with V100 to fine-tune it. 154 | 155 | 156 | ### Server 157 | 158 | start FastAPI server: 159 | 160 | example: [examples/server.py](https://github.com/shibing624/CodeAssist/blob/main/examples/server.py) 161 | 162 | ```shell 163 | cd examples 164 | python server.py 165 | ``` 166 | 167 | open url: http://0.0.0.0:8001/docs 168 | 169 | ![api](https://github.com/shibing624/CodeAssist/blob/main/docs/api.png) 170 | 171 | 172 | 173 | ## Dataset 174 | 175 | This allows to customize dataset building. Below is an example of the building process. 176 | 177 | Let's use Python codes from [Awesome-pytorch-list](https://github.com/bharathgs/Awesome-pytorch-list) 178 | 179 | 1. We want the model to help auto-complete codes at a general level. The codes of The Algorithms suits the need. 180 | 2. This code from this project is well written (high-quality codes). 181 | 182 | dataset tree: 183 | 184 | ```shell 185 | examples/download/python 186 | ├── train.txt 187 | └── valid.txt 188 | └── test.txt 189 | ``` 190 | 191 | There are three ways to build dataset: 192 | 1. Use the huggingface/datasets library load the dataset 193 | huggingface datasets [https://huggingface.co/datasets/shibing624/source_code](https://huggingface.co/datasets/shibing624/source_code) 194 | 195 | ```python 196 | from datasets import load_dataset 197 | dataset = load_dataset("shibing624/source_code", "python") # python or java or cpp 198 | print(dataset) 199 | print(dataset['test'][0:10]) 200 | ``` 201 | 202 | output: 203 | ```shell 204 | DatasetDict({ 205 | train: Dataset({ 206 | features: ['text'], 207 | num_rows: 5215412 208 | }) 209 | validation: Dataset({ 210 | features: ['text'], 211 | num_rows: 10000 212 | }) 213 | test: Dataset({ 214 | features: ['text'], 215 | num_rows: 10000 216 | }) 217 | }) 218 | {'text': [ 219 | " {'max_epochs': [1, 2]},\n", 220 | ' refit=False,\n', ' cv=3,\n', 221 | " scoring='roc_auc',\n", ' )\n', 222 | ' search.fit(*data)\n', 223 | '', 224 | ' def test_module_output_not_1d(self, net_cls, data):\n', 225 | ' from skorch.toy import make_classifier\n', 226 | ' module = make_classifier(\n' 227 | ]} 228 | ``` 229 | 230 | 2. Download dataset from Cloud 231 | 232 | | Name | Source | Download | Size | 233 | | :------- | :--------- | :---------: | :---------: | 234 | | Python+Java+CPP source code | Awesome-pytorch-list(5.22 Million lines) | [github_source_code.zip](https://github.com/shibing624/codeassist/releases/download/0.0.4/source_code.zip) | 105M | 235 | 236 | download dataset and unzip it, put to `examples/`. 237 | 238 | 3. Get source code from scratch and build dataset 239 | 240 | [prepare_code_data.py](https://github.com/shibing624/CodeAssist/blob/main/examples/prepare_code_data.py) 241 | 242 | ```shell 243 | cd examples 244 | python prepare_code_data.py --num_repos 260 245 | ``` 246 | 247 | 248 | ## Contact 249 | 250 | - Issue(建议) 251 | ：[![GitHub issues](https://img.shields.io/github/issues/shibing624/CodeAssist.svg)](https://github.com/shibing624/CodeAssist/issues) 252 | - 邮件我：xuming: xuming624@qq.com 253 | - 微信我： 加我*微信号：xuming624, 备注：个人名称-公司-NLP* 进NLP交流群。 254 | 255 | 256 | 257 | ## Citation 258 | 259 | 如果你在研究中使用了codeassist，请按如下格式引用： 260 | 261 | APA: 262 | ```latex 263 | Xu, M. codeassist: Code AutoComplete with GPT model (Version 1.0.0) [Computer software]. https://github.com/shibing624/codeassist 264 | ``` 265 | 266 | BibTeX: 267 | ```latex 268 | @software{Xu_codeassist, 269 | author = {Ming Xu}, 270 | title = {CodeAssist: Code AutoComplete with Generation model}, 271 | url = {https://github.com/shibing624/codeassist}, 272 | version = {1.0.0} 273 | } 274 | ``` 275 | 276 | ## License 277 | This repository is licensed under the [The Apache License 2.0](LICENSE). 278 | 279 | Please follow the [Attribution-NonCommercial 4.0 International](https://github.com/nlpxucan/WizardLM/blob/main/WizardCoder/MODEL_WEIGHTS_LICENSE) to use the WizardCoder model. 280 | 281 | 282 | ## Contribute 283 | 284 | 项目代码还很粗糙，如果大家对代码有所改进，欢迎提交回本项目，在提交之前，注意以下两点： 285 | 286 | - 在`tests`添加相应的单元测试 287 | - 使用`python setup.py test`来运行所有单元测试，确保所有单测都是通过的 288 | 289 | 之后即可提交PR。 290 | 291 | ## Reference 292 | 293 | - [gpt-2-simple](https://github.com/minimaxir/gpt-2-simple) 294 | - [galois-autocompleter](https://github.com/galois-autocompleter/galois-autocompleter) 295 | - [WizardLM/WizardCoder-15B-V1.0](https://huggingface.co/WizardLM/WizardCoder-15B-V1.0) 296 | -------------------------------------------------------------------------------- /README_EN.md: -------------------------------------------------------------------------------- 1 | [**🇨🇳中文**](https://github.com/shibing624/CodeAssist/blob/main/README.md) | [**🌐English**](https://github.com/shibing624/CodeAssist/blob/main/README_EN.md) | [**📖文档/Docs**](https://github.com/shibing624/CodeAssist/wiki) | [**🤖模型/Models**](https://huggingface.co/shibing624) 2 | 3 |
4 | 5 | Logo 6 | 7 |
8 | 9 | ----------------- 10 | 11 | # CodeAssist: Advanced Code Completion Tool 12 | [![PyPI version](https://badge.fury.io/py/CodeAssist.svg)](https://badge.fury.io/py/CodeAssist) 13 | [![Contributions welcome](https://img.shields.io/badge/contributions-welcome-brightgreen.svg)](CONTRIBUTING.md) 14 | [![GitHub contributors](https://img.shields.io/github/contributors/shibing624/CodeAssist.svg)](https://github.com/shibing624/CodeAssist/graphs/contributors) 15 | [![License Apache 2.0](https://img.shields.io/badge/license-Apache%202.0-blue.svg)](LICENSE) 16 | [![python_vesion](https://img.shields.io/badge/Python-3.5%2B-green.svg)](requirements.txt) 17 | [![GitHub issues](https://img.shields.io/github/issues/shibing624/CodeAssist.svg)](https://github.com/shibing624/CodeAssist/issues) 18 | [![Wechat Group](https://img.shields.io/badge/wechat-group-green.svg?logo=wechat)](#Contact) 19 | 20 | ## Introduction 21 | 22 | **CodeAssist** is an advanced code completion tool that intelligently provides high-quality code completions for Python, Java, and C++ and so on. 23 | 24 | CodeAssist 是一个高级代码补全工具，高质量为 Python、Java 和 C++ 等编程语言补全代码 25 | 26 | 27 | **Guide** 28 | 29 | - [Feature](#Feature) 30 | - [Install](#install) 31 | - [Usage](#usage) 32 | - [Contact](#Contact) 33 | - [Citation](#Citation) 34 | - [Reference](#reference) 35 | 36 | ## Feature 37 | 38 | - GPT based code completion 39 | - Code completion for `Python`, `Java`, `C++`, `javascript` and so on 40 | - Line and block code completion 41 | - Train(Fine-tuning) and predict model with your own data 42 | 43 | ### Release Models 44 | 45 | | Arch | BaseModel | Model | Model Size | 46 | |:-------|:------------------|:------------------------------------------------------------------------------------------------------------------------|:----------:| 47 | | GPT | gpt2 | [shibing624/code-autocomplete-gpt2-base](https://huggingface.co/shibing624/code-autocomplete-gpt2-base) | 487MB | 48 | | GPT | distilgpt2 | [shibing624/code-autocomplete-distilgpt2-python](https://huggingface.co/shibing624/code-autocomplete-distilgpt2-python) | 319MB | 49 | | GPT | bigcode/starcoder | [WizardLM/WizardCoder-15B-V1.0](https://huggingface.co/WizardLM/WizardCoder-15B-V1.0) | 29GB | 50 | 51 | 52 | ### Demo 53 | 54 | HuggingFace Demo: https://huggingface.co/spaces/shibing624/code-autocomplete 55 | 56 | backend model: `shibing624/code-autocomplete-gpt2-base` 57 | 58 | ## Install 59 | 60 | ``` 61 | pip install torch # conda install pytorch 62 | pip install -U codeassist 63 | ``` 64 | 65 | or 66 | 67 | ``` 68 | git clone https://github.com/shibing624/codeassist.git 69 | cd CodeAssist 70 | python setup.py install 71 | ``` 72 | 73 | ## Usage 74 | 75 | ### WizardCoder model 76 | 77 | WizardCoder-15b is fine-tuned `bigcode/starcoder` with alpaca code data, you can use the following code to generate code: 78 | 79 | example: [examples/wizardcoder_demo.py](https://github.com/shibing624/CodeAssist/blob/main/examples/wizardcoder_demo.py) 80 | 81 | ```python 82 | import sys 83 | 84 | sys.path.append('..') 85 | from codeassist import WizardCoder 86 | 87 | m = WizardCoder("WizardLM/WizardCoder-15B-V1.0") 88 | print(m.generate('def load_csv_file(file_path):')[0]) 89 | ``` 90 | 91 | output: 92 | 93 | 94 | ```python 95 | import csv 96 | 97 | def load_csv_file(file_path): 98 | """ 99 | Load data from a CSV file and return a list of dictionaries. 100 | """ 101 | # Open the file in read mode 102 | with open(file_path, 'r') as file: 103 | # Create a CSV reader object 104 | csv_reader = csv.DictReader(file) 105 | # Initialize an empty list to store the data 106 | data = [] 107 | # Iterate over each row of data 108 | for row in csv_reader: 109 | # Append the row of data to the list 110 | data.append(row) 111 | # Return the list of data 112 | return data 113 | ``` 114 | 115 | model output is impressively effective, it currently supports English input only, and you can enter instructions or code prefixes as required. 116 | 117 | ### distilgpt2 model 118 | 119 | 120 | distilgpt2 fine-tuned code autocomplete model, you can use the following code: 121 | 122 | example: [examples/distilgpt2_demo.py](https://github.com/shibing624/CodeAssist/blob/main/examples/distilgpt2_demo.py) 123 | 124 | ```python 125 | import sys 126 | 127 | sys.path.append('..') 128 | from codeassist import GPT2Coder 129 | 130 | m = GPT2Coder("shibing624/code-autocomplete-distilgpt2-python") 131 | print(m.generate('import torch.nn as')[0]) 132 | ``` 133 | 134 | output: 135 | 136 | ```shell 137 | import torch.nn as nn 138 | import torch.nn.functional as F 139 | ``` 140 | 141 | ### Use with huggingface/transformers： 142 | 143 | example: [examples/use_transformers_gpt2.py](https://github.com/shibing624/CodeAssist/blob/main/examples/use_transformers_gpt2.py) 144 | 145 | ### Train Model 146 | #### Train WizardCoder model 147 | example: [examples/training_wizardcoder_mydata.py](https://github.com/shibing624/CodeAssist/blob/main/examples/training_wizardcoder_mydata.py) 148 | 149 | ```shell 150 | cd examples 151 | python training_wizardcoder_mydata.py --do_train --do_predict --num_epochs 1 --output_dir outputs-wizard --model_name WizardLM/WizardCoder-15B-V1.0 152 | ``` 153 | 154 | #### Train distilgpt2 model 155 | example: [examples/training_gpt2_mydata.py](https://github.com/shibing624/CodeAssist/blob/main/examples/training_gpt2_mydata.py) 156 | 157 | ```shell 158 | cd examples 159 | python training_gpt2_mydata.py --do_train --do_predict --num_epochs 15 --output_dir outputs-gpt2 --model_name gpt2 160 | ``` 161 | 162 | PS: fine-tuned result model is GPT2-python: [shibing624/code-autocomplete-gpt2-base](https://huggingface.co/shibing624/code-autocomplete-gpt2-base), 163 | I spent about 24 hours with V100 to fine-tune it. 164 | 165 | 166 | ### Server 167 | 168 | start FastAPI server: 169 | 170 | example: [examples/server.py](https://github.com/shibing624/CodeAssist/blob/main/examples/server.py) 171 | 172 | ```shell 173 | cd examples 174 | python server.py 175 | ``` 176 | 177 | open url: http://0.0.0.0:8001/docs 178 | 179 | ![api](https://github.com/shibing624/CodeAssist/blob/main/docs/api.png) 180 | 181 | 182 | 183 | ## Dataset 184 | 185 | This allows to customize dataset building. Below is an example of the building process. 186 | 187 | Let's use Python codes from [Awesome-pytorch-list](https://github.com/bharathgs/Awesome-pytorch-list) 188 | 189 | 1. We want the model to help auto-complete codes at a general level. The codes of The Algorithms suits the need. 190 | 2. This code from this project is well written (high-quality codes). 191 | 192 | dataset tree: 193 | 194 | ```shell 195 | examples/download/python 196 | ├── train.txt 197 | └── valid.txt 198 | └── test.txt 199 | ``` 200 | 201 | There are three ways to build dataset: 202 | 1. Use the huggingface/datasets library load the dataset 203 | huggingface datasets [https://huggingface.co/datasets/shibing624/source_code](https://huggingface.co/datasets/shibing624/source_code) 204 | 205 | ```python 206 | from datasets import load_dataset 207 | dataset = load_dataset("shibing624/source_code", "python") # python or java or cpp 208 | print(dataset) 209 | print(dataset['test'][0:10]) 210 | ``` 211 | 212 | output: 213 | ```shell 214 | DatasetDict({ 215 | train: Dataset({ 216 | features: ['text'], 217 | num_rows: 5215412 218 | }) 219 | validation: Dataset({ 220 | features: ['text'], 221 | num_rows: 10000 222 | }) 223 | test: Dataset({ 224 | features: ['text'], 225 | num_rows: 10000 226 | }) 227 | }) 228 | {'text': [ 229 | " {'max_epochs': [1, 2]},\n", 230 | ' refit=False,\n', ' cv=3,\n', 231 | " scoring='roc_auc',\n", ' )\n', 232 | ' search.fit(*data)\n', 233 | '', 234 | ' def test_module_output_not_1d(self, net_cls, data):\n', 235 | ' from skorch.toy import make_classifier\n', 236 | ' module = make_classifier(\n' 237 | ]} 238 | ``` 239 | 240 | 2. Download dataset from Cloud 241 | 242 | | Name | Source | Download | Size | 243 | | :------- | :--------- | :---------: | :---------: | 244 | | Python+Java+CPP source code | Awesome-pytorch-list(5.22 Million lines) | [github_source_code.zip](https://github.com/shibing624/CodeAssist/releases/download/0.0.4/source_code.zip) | 105M | 245 | 246 | download dataset and unzip it, put to `examples/`. 247 | 248 | 3. Get source code from scratch and build dataset 249 | 250 | [prepare_code_data.py](https://github.com/shibing624/CodeAssist/blob/main/examples/prepare_code_data.py) 251 | 252 | ```shell 253 | cd examples 254 | python prepare_code_data.py --num_repos 260 255 | ``` 256 | 257 | 258 | ## Contact 259 | 260 | - Issue(建议) 261 | ：[![GitHub issues](https://img.shields.io/github/issues/shibing624/CodeAssist.svg)](https://github.com/shibing624/CodeAssist/issues) 262 | - 邮件我：xuming: xuming624@qq.com 263 | - 微信我： 加我*微信号：xuming624, 备注：个人名称-公司-NLP* 进NLP交流群。 264 | 265 | 266 | 267 | ## Citation 268 | 269 | 如果你在研究中使用了CodeAssist，请按如下格式引用： 270 | 271 | APA: 272 | ```latex 273 | Xu, M. CodeAssist: Code AutoComplete with GPT model (Version 1.0.0) [Computer software]. https://github.com/shibing624/CodeAssist 274 | ``` 275 | 276 | BibTeX: 277 | ```latex 278 | @software{Xu_CodeAssist, 279 | author = {Ming Xu}, 280 | title = {CodeAssist: Code AutoComplete with Generation model}, 281 | url = {https://github.com/shibing624/CodeAssist}, 282 | version = {1.0.0} 283 | } 284 | ``` 285 | 286 | ## License 287 | 288 | 授权协议为 [The Apache License 2.0](/LICENSE)，可免费用做商业用途。请在产品说明中附加CodeAssist的链接和授权协议。 289 | 290 | ## Contribute 291 | 292 | 项目代码还很粗糙，如果大家对代码有所改进，欢迎提交回本项目，在提交之前，注意以下两点： 293 | 294 | - 在`tests`添加相应的单元测试 295 | - 使用`python setup.py test`来运行所有单元测试，确保所有单测都是通过的 296 | 297 | 之后即可提交PR。 298 | 299 | ## Reference 300 | 301 | - [gpt-2-simple](https://github.com/minimaxir/gpt-2-simple) 302 | - [galois-autocompleter](https://github.com/galois-autocompleter/galois-autocompleter) 303 | - [WizardLM/WizardCoder-15B-V1.0](https://huggingface.co/WizardLM/WizardCoder-15B-V1.0) 304 | -------------------------------------------------------------------------------- /codeassist/__init__.py: -------------------------------------------------------------------------------- 1 | # -*- coding: utf-8 -*- 2 | """ 3 | @author:XuMing(xuming624@qq.com) 4 | @description: 5 | """ 6 | 7 | __version__ = "1.0.0" 8 | 9 | from codeassist.gpt2_coder import GPT2Coder 10 | from codeassist.wizard_coder import WizardCoder 11 | -------------------------------------------------------------------------------- /codeassist/create_dataset.py: -------------------------------------------------------------------------------- 1 | # -*- coding: utf-8 -*- 2 | """ 3 | @author:https://github.com/labmlai 4 | @description: Parse all files and write to a single file 5 | 6 | part of code refer: https://github.com/labmlai/python_autocomplete/blob/master/python_autocomplete/create_dataset.py 7 | """ 8 | import glob 9 | import os 10 | import re 11 | import ssl 12 | import string 13 | import urllib.error 14 | import urllib.request 15 | import zipfile 16 | from pathlib import Path 17 | from typing import Optional 18 | from typing import Set 19 | 20 | from loguru import logger 21 | 22 | ssl._create_default_https_context = ssl._create_unverified_context 23 | PRINTABLE = set(string.printable) 24 | 25 | 26 | def read_file(path: str) -> str: 27 | """ 28 | Read a file 29 | """ 30 | with open(path, 'r', encoding='utf8') as f: 31 | try: 32 | content = f.read() 33 | except UnicodeDecodeError as e: 34 | logger.warning(f"UnicodeDecodeError: {path}, file pass") 35 | content = "" 36 | content = ''.join(filter(lambda x: x in PRINTABLE, content)) 37 | 38 | return content 39 | 40 | 41 | def save_file(content, file_path: str): 42 | with open(file_path, 'w', encoding='utf8') as f: 43 | f.write(str(content)) 44 | 45 | 46 | def merge_and_save(source_files, path): 47 | with open(path, 'w', encoding='utf8') as f: 48 | for src in source_files: 49 | f.write(read_file(src) + "\n\n") 50 | 51 | 52 | def get_repos_from_readme(readme_content): 53 | link_pattern = re.compile(r""" 54 | \[(?P[^\]]*)\] # title 55 | \((?P<utl>[^\)]*)\) # url 56 | """, re.VERBOSE) 57 | 58 | res = link_pattern.findall(readme_content) 59 | 60 | github_repos = [] 61 | repo_pattern = re.compile(r'https://github.com/(?P<user>[^/]*)/(?P<repo>[^/#]*)$') 62 | for title, url in res: 63 | repos = repo_pattern.findall(url) 64 | for r in repos: 65 | github_repos.append((r[0], r[1])) 66 | 67 | return github_repos 68 | 69 | 70 | def download_repo(save_dir: str, org: str, repo: str, idx: Optional[int]): 71 | zip_file = Path(f'{save_dir}/{org}_{repo}.zip') 72 | 73 | if zip_file.exists(): 74 | return zip_file 75 | 76 | if idx is not None: 77 | idx_str = f"{idx:03}: " 78 | else: 79 | idx_str = "" 80 | 81 | try: 82 | zip = urllib.request.urlopen(f'https://github.com/{org}/{repo}/archive/master.zip') 83 | except urllib.error.HTTPError as e: 84 | print(e) 85 | return 86 | content = zip.read() 87 | 88 | size = len(content) // 1024 89 | logger.debug(f"{idx_str} {org}/{repo} {size :,}KB") 90 | 91 | with open(str(zip_file), 'wb') as f: 92 | f.write(content) 93 | 94 | return zip_file 95 | 96 | 97 | def extract_zip(source_dir, file_path: Path): 98 | source = Path(source_dir) 99 | logger.debug(f"Extract {file_path}") 100 | repo_source = source / file_path.stem 101 | if repo_source.exists(): 102 | logger.debug(f"Exists: {repo_source}") 103 | return repo_source 104 | try: 105 | with zipfile.ZipFile(file_path, 'r') as repo_zip: 106 | repo_zip.extractall(repo_source) 107 | except zipfile.BadZipfile as e: 108 | print(file_path, e) 109 | 110 | return repo_source 111 | 112 | 113 | def remove_files(path: Path, keep: Set[str]): 114 | """ 115 | Remove files 116 | """ 117 | for p in path.iterdir(): 118 | if p.is_symlink(): 119 | p.unlink() 120 | continue 121 | if p.is_dir(): 122 | remove_files(p, keep) 123 | else: 124 | if p.suffix not in keep: 125 | p.unlink() 126 | 127 | 128 | def get_source_code_by_language(code_languages=("python", "java", "cpp"), 129 | save_dir='download/', 130 | each_limit_repos=3): 131 | sources = dict() 132 | Path(save_dir).mkdir(parents=True, exist_ok=True) 133 | if isinstance(code_languages, str): 134 | code_languages = [code_languages] 135 | logger.info(f"Get source code by language: {code_languages}") 136 | 137 | def get_source_files_by_readme(readme_content, sub_save_dir, limit_size): 138 | zip_dir = sub_save_dir + "/zip" 139 | src_dir = sub_save_dir + "/src" 140 | Path(zip_dir).mkdir(parents=True, exist_ok=True) 141 | Path(src_dir).mkdir(parents=True, exist_ok=True) 142 | repos = get_repos_from_readme(readme_content) 143 | if limit_size: 144 | repos = repos[:limit_size] 145 | # Download repos 146 | for i, r in enumerate(repos): 147 | zip_file = download_repo(zip_dir, r[0], r[1], i) 148 | if not zip_file: 149 | continue 150 | extracted = extract_zip(src_dir, zip_file) 151 | remove_files(extracted, keep={suffix}) 152 | source_files = glob.glob(f"{src_dir}/**/*{suffix}", recursive=True) 153 | logger.info(f"Path: {src_dir}/**/*{suffix}, file size: {len(source_files)}") 154 | return source_files 155 | 156 | if "python" in code_languages: 157 | logger.debug('Get awesome-python') 158 | suffix = '.py' 159 | sub_save_dir = os.path.join(save_dir, 'python') 160 | readme_file = sub_save_dir + '/README.md' 161 | if os.path.exists(readme_file): 162 | readme_content = read_file(readme_file) 163 | else: 164 | readme_content = urllib.request.urlopen( 165 | 'https://raw.githubusercontent.com/bharathgs/Awesome-pytorch-list/master/README.md').read() 166 | readme_content = str(readme_content) 167 | save_file(readme_content, readme_file) 168 | sources['python'] = get_source_files_by_readme(readme_content, sub_save_dir, each_limit_repos) 169 | logger.info(f"Get source code by language: python done") 170 | if "java" in code_languages: 171 | logger.debug('Get awesome-java') 172 | suffix = '.java' 173 | sub_save_dir = os.path.join(save_dir, 'java') 174 | readme_file = sub_save_dir + '/README.md' 175 | if os.path.exists(readme_file): 176 | readme_content = read_file(readme_file) 177 | else: 178 | readme_content = urllib.request.urlopen( 179 | 'https://raw.githubusercontent.com/akullpp/awesome-java/master/README.md').read() 180 | readme_content = str(readme_content) 181 | save_file(readme_content, readme_file) 182 | sources['java'] = get_source_files_by_readme(readme_content, sub_save_dir, each_limit_repos) 183 | logger.info(f"Get source code by language: java done") 184 | if "cpp" in code_languages: 185 | logger.debug('Get awesome-cpp') 186 | suffix = '.cpp' 187 | sub_save_dir = os.path.join(save_dir, 'cpp') 188 | readme_file = sub_save_dir + '/README.md' 189 | if os.path.exists(readme_file): 190 | readme_content = read_file(readme_file) 191 | else: 192 | readme_content = urllib.request.urlopen( 193 | 'https://raw.githubusercontent.com/fffaraz/awesome-cpp/master/README.md').read() 194 | readme_content = str(readme_content) 195 | save_file(readme_content, readme_file) 196 | sources['cpp'] = get_source_files_by_readme(readme_content, sub_save_dir, each_limit_repos) 197 | logger.info(f"Get source code by language: cpp done") 198 | 199 | return sources 200 | -------------------------------------------------------------------------------- /codeassist/gpt2_coder.py: -------------------------------------------------------------------------------- 1 | # -*- coding: utf-8 -*- 2 | """ 3 | @author:XuMing(xuming624@qq.com) 4 | @description: Rewrite the original gpt2 model to support the autocomplete 5 | """ 6 | 7 | import math 8 | import os 9 | import random 10 | from typing import Dict, List 11 | 12 | import numpy as np 13 | import pandas as pd 14 | import torch 15 | from loguru import logger 16 | from torch.nn.utils.rnn import pad_sequence 17 | from torch.utils.data import DataLoader, Dataset, RandomSampler, SequentialSampler 18 | from torch.utils.data.distributed import DistributedSampler 19 | from tqdm import tqdm, trange 20 | from transformers import GPT2LMHeadModel, GPT2Tokenizer 21 | from transformers.data.datasets.language_modeling import TextDataset 22 | from transformers.optimization import AdamW, get_linear_schedule_with_warmup 23 | 24 | os.environ["KMP_DUPLICATE_LIB_OK"] = "TRUE" 25 | device = torch.device('cuda' if torch.cuda.is_available() else 'cpu') 26 | 27 | 28 | class GPT2Coder: 29 | def __init__( 30 | self, 31 | model_name_or_path: str = "shibing624/code-autocomplete-gpt2-base", 32 | max_length: int = 128, 33 | do_lower_case: bool = False, 34 | special_words_dict: Dict = None 35 | ): 36 | """ 37 | Initializes a GPT2 LanguageModelingModel. 38 | 39 | Args: 40 | model_name_or_path: Default Transformer model name or path to a directory containing Transformer model file (pytorch_nodel.bin). 41 | max_length: The maximum total input sequence length after tokenization. 42 | do_lower_case: Set this flag if you are using an uncased model. 43 | special_words_dict: A dictionary of special words and their token ids. 44 | """ 45 | self.model_name_or_path = model_name_or_path 46 | self.do_lower_case = do_lower_case 47 | if max_length > 1024: 48 | logger.warning("GPT only allows a max_length of 1024. Value will be set to 1024") 49 | max_length = 1024 50 | self.max_length = max_length 51 | self.model = GPT2LMHeadModel.from_pretrained(model_name_or_path) 52 | self.model.to(device) 53 | self.tokenizer = GPT2Tokenizer.from_pretrained(model_name_or_path, do_lower_case=do_lower_case) 54 | if special_words_dict is not None: 55 | self.add_special_words(special_words_dict) 56 | self.results = {} 57 | 58 | def set_seed(self, seed): 59 | logger.debug(f"Set seed for random, numpy and torch: {seed}") 60 | random.seed(seed) 61 | np.random.seed(seed) 62 | torch.manual_seed(seed) 63 | if torch.cuda.is_available(): 64 | torch.cuda.manual_seed_all(seed) 65 | 66 | def add_special_words(self, special_words_dict): 67 | origin_num_tokens = len(self.tokenizer) 68 | num_added_tokens = self.tokenizer.add_special_tokens(special_words_dict) 69 | if num_added_tokens > 0: 70 | self.model.resize_token_embeddings(new_num_tokens=origin_num_tokens + num_added_tokens) 71 | 72 | def train_model( 73 | self, 74 | train_file: str, 75 | output_dir: str, 76 | eval_file: str = None, 77 | verbose: bool = True, 78 | batch_size: int = 8, 79 | num_epochs: int = 1, 80 | weight_decay: float = 0.01, 81 | seed: int = 42, 82 | warmup_ratio: float = 0.1, 83 | lr: float = 5e-5, 84 | eps: float = 1e-6, 85 | gradient_accumulation_steps: int = 1, 86 | max_grad_norm: float = 1.0, 87 | max_steps: int = -1 88 | ): 89 | """ 90 | Trains the model on 'train_file' 91 | 92 | Args: 93 | train_file: Path to text file containing the text to train the language model on. 94 | output_dir: The directory where model files will be saved. If not given, self.args.output_dir will be used. 95 | eval_file (optional): Path to eval file containing the text to evaluate the language model on. 96 | verbose (optional): Print logger or not. 97 | batch_size (optional): Batch size for training. 98 | num_epochs (optional): Number of epochs for training. 99 | weight_decay (optional): Weight decay for optimization. 100 | seed (optional): Seed for initialization. 101 | warmup_ratio (optional): Warmup ratio for learning rate. 102 | lr (optional): Learning rate. 103 | eps (optional): Adam epsilon. 104 | gradient_accumulation_steps (optional): Number of updates steps to accumulate before performing a backward/update pass. 105 | max_grad_norm (optional): Max gradient norm. 106 | max_steps (optional): If > 0: set total number of training steps to perform. Override num_epochs. 107 | Returns: 108 | global_step: Number of global steps trained 109 | training_details: Average training loss if evaluate_during_training is False or full training progress scores if evaluate_during_training is True 110 | """ 111 | os.makedirs(output_dir, exist_ok=True) 112 | self.model.to(device) 113 | train_dataset = TextDataset(self.tokenizer, train_file, self.max_length, overwrite_cache=True, 114 | cache_dir=output_dir) 115 | 116 | global_step, training_details = self.train( 117 | train_dataset, 118 | output_dir, 119 | eval_file=eval_file, 120 | verbose=verbose, 121 | batch_size=batch_size, 122 | num_epochs=num_epochs, 123 | weight_decay=weight_decay, 124 | seed=seed, 125 | warmup_ratio=warmup_ratio, 126 | lr=lr, 127 | eps=eps, 128 | gradient_accumulation_steps=gradient_accumulation_steps, 129 | max_grad_norm=max_grad_norm, 130 | max_steps=max_steps 131 | ) 132 | logger.info(f" Training model done. Saved to {output_dir}.") 133 | 134 | return global_step, training_details 135 | 136 | def train( 137 | self, 138 | train_dataset: Dataset, 139 | output_dir: str, 140 | eval_file: str = None, 141 | verbose: bool = True, 142 | batch_size: int = 8, 143 | num_epochs: int = 1, 144 | weight_decay: float = 0.01, 145 | seed: int = 42, 146 | warmup_ratio: float = 0.1, 147 | lr: float = 5e-5, 148 | eps: float = 1e-6, 149 | gradient_accumulation_steps: int = 1, 150 | max_grad_norm: float = 1.0, 151 | max_steps: int = -1 152 | ): 153 | """ 154 | Trains the model on train_dataset. 155 | 156 | Utility function to be used by the train_model() method. Not intended to be used directly. 157 | """ 158 | self.set_seed(seed) 159 | 160 | def collate(examples: List[torch.Tensor]): 161 | if self.tokenizer._pad_token is None: 162 | return pad_sequence(examples, batch_first=True) 163 | return pad_sequence(examples, batch_first=True, padding_value=self.tokenizer.pad_token_id) 164 | 165 | train_sampler = RandomSampler(train_dataset) 166 | train_dataloader = DataLoader( 167 | train_dataset, 168 | batch_size=batch_size, 169 | sampler=train_sampler, 170 | collate_fn=collate, 171 | ) 172 | 173 | total_steps = len(train_dataloader) * num_epochs 174 | param_optimizer = list(self.model.named_parameters()) 175 | no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight'] 176 | optimizer_grouped_parameters = [ 177 | {'params': [p for n, p in param_optimizer if not any(nd in n for nd in no_decay)], 178 | 'weight_decay': weight_decay}, 179 | {'params': [p for n, p in param_optimizer if any(nd in n for nd in no_decay)], 'weight_decay': 0.0} 180 | ] 181 | 182 | warmup_steps = math.ceil(total_steps * warmup_ratio) # by default 10% of train data for warm-up 183 | optimizer = AdamW(optimizer_grouped_parameters, lr=lr, eps=eps, correct_bias=False) 184 | scheduler = get_linear_schedule_with_warmup(optimizer, num_warmup_steps=warmup_steps, 185 | num_training_steps=total_steps) 186 | logger.info("***** Running training *****") 187 | logger.info(f" Num examples = {len(train_dataset)}") 188 | logger.info(f" Batch size = {batch_size}") 189 | logger.info(f" Num steps = {total_steps}") 190 | logger.info(f" Warmup-steps: {warmup_steps}") 191 | 192 | logger.info(" Training started") 193 | global_step = 0 194 | tr_loss, logging_loss = 0.0, 0.0 195 | self.model.zero_grad() 196 | epoch_number = 0 197 | best_eval_metric = 1e3 198 | steps_trained_in_current_epoch = 0 199 | epochs_trained = 0 200 | 201 | if self.model_name_or_path and os.path.exists(self.model_name_or_path): 202 | try: 203 | # set global_step to global_step of last saved checkpoint from model path 204 | checkpoint_suffix = self.model_name_or_path.split("/")[-1].split("-") 205 | if len(checkpoint_suffix) > 2: 206 | checkpoint_suffix = checkpoint_suffix[1] 207 | else: 208 | checkpoint_suffix = checkpoint_suffix[-1] 209 | global_step = int(checkpoint_suffix) 210 | epochs_trained = global_step // (len(train_dataloader) // gradient_accumulation_steps) 211 | steps_trained_in_current_epoch = global_step % (len(train_dataloader) // gradient_accumulation_steps) 212 | logger.info(" Continuing training from checkpoint, will skip to saved global_step") 213 | logger.info(" Continuing training from epoch %d" % epochs_trained) 214 | logger.info(" Continuing training from global step %d" % global_step) 215 | logger.info(" Will skip the first %d steps in the current epoch" % steps_trained_in_current_epoch) 216 | except ValueError: 217 | logger.info(" Starting fine-tuning.") 218 | 219 | training_progress_scores = { 220 | "global_step": [], 221 | "perplexity": [], 222 | "eval_loss": [], 223 | "train_loss": [], 224 | } 225 | train_iterator = trange(int(num_epochs), desc="Epoch", disable=False, mininterval=0) 226 | for current_epoch in train_iterator: 227 | self.model.train() 228 | current_loss = 0 229 | if isinstance(train_dataloader, DataLoader) and isinstance(train_dataloader.sampler, DistributedSampler): 230 | train_dataloader.sampler.set_epoch(current_epoch) 231 | if epochs_trained > 0: 232 | epochs_trained -= 1 233 | continue 234 | train_iterator.set_description(f"Epoch {epoch_number + 1} of {num_epochs}") 235 | batch_iterator = tqdm(train_dataloader, 236 | desc=f"Running Epoch {epoch_number} of {num_epochs}", 237 | disable=False, 238 | mininterval=0) 239 | for step, batch in enumerate(batch_iterator): 240 | if steps_trained_in_current_epoch > 0: 241 | steps_trained_in_current_epoch -= 1 242 | continue 243 | 244 | inputs = batch.to(device) 245 | outputs = self.model(inputs, labels=inputs) 246 | loss = outputs[0] 247 | current_loss = loss.item() 248 | if verbose: 249 | batch_iterator.set_description( 250 | f"Epochs {epoch_number}/{num_epochs}. Running Loss: {current_loss:9.4f}") 251 | 252 | if gradient_accumulation_steps > 1: 253 | loss = loss / gradient_accumulation_steps 254 | 255 | loss.backward() 256 | tr_loss += loss.item() 257 | if (step + 1) % gradient_accumulation_steps == 0: 258 | torch.nn.utils.clip_grad_norm_(self.model.parameters(), max_grad_norm) 259 | optimizer.step() 260 | scheduler.step() # Update learning rate schedule 261 | self.model.zero_grad() 262 | global_step += 1 263 | epoch_number += 1 264 | output_dir_current = os.path.join(output_dir, "checkpoint-{}-epoch-{}".format(global_step, epoch_number)) 265 | results = self.eval_model(eval_file, output_dir_current, verbose=verbose, batch_size=batch_size) 266 | self.save_model(output_dir_current, model=self.model, results=results) 267 | training_progress_scores["global_step"].append(global_step) 268 | training_progress_scores["train_loss"].append(current_loss) 269 | for key in results: 270 | training_progress_scores[key].append(results[key]) 271 | report = pd.DataFrame(training_progress_scores) 272 | report.to_csv(os.path.join(output_dir, "training_progress_scores.csv"), index=False) 273 | 274 | eval_loss = results["eval_loss"] 275 | if eval_loss < best_eval_metric: 276 | best_eval_metric = eval_loss 277 | self.save_model(output_dir, model=self.model, results=results) 278 | 279 | if 0 < max_steps < global_step: 280 | return global_step, training_progress_scores 281 | 282 | return global_step, training_progress_scores 283 | 284 | def eval_model(self, eval_file: str, output_dir: str = None, verbose: bool = True, batch_size: int = 16): 285 | """ 286 | Evaluates the model on eval_df. Saves results to args.output_dir 287 | result: Dictionary containing evaluation results. 288 | """ 289 | self.model.to(device) 290 | eval_dataset = TextDataset(self.tokenizer, eval_file, self.max_length, overwrite_cache=True) 291 | result = self.evaluate(eval_dataset, output_dir, batch_size=batch_size) 292 | self.results.update(result) 293 | 294 | if verbose: 295 | logger.info(self.results) 296 | 297 | return result 298 | 299 | def evaluate(self, eval_dataset, output_dir: str = None, batch_size: int = 16): 300 | """ 301 | Evaluates the model on eval_dataset. 302 | 303 | Utility function to be used by the eval_model() method. Not intended to be used directly. 304 | """ 305 | results = {} 306 | 307 | def collate(examples: List[torch.Tensor]): 308 | if self.tokenizer._pad_token is None: 309 | return pad_sequence(examples, batch_first=True) 310 | return pad_sequence(examples, batch_first=True, padding_value=self.tokenizer.pad_token_id) 311 | 312 | eval_sampler = SequentialSampler(eval_dataset) 313 | eval_dataloader = DataLoader(eval_dataset, 314 | batch_size=batch_size, 315 | sampler=eval_sampler, 316 | collate_fn=collate) 317 | eval_loss = 0.0 318 | nb_eval_steps = 0 319 | self.model.eval() 320 | 321 | for batch in tqdm(eval_dataloader, disable=False, desc="Running Evaluation"): 322 | inputs = batch.to(device) 323 | with torch.no_grad(): 324 | outputs = self.model(inputs, labels=inputs) 325 | lm_loss = outputs[0] 326 | eval_loss += lm_loss.item() 327 | nb_eval_steps += 1 328 | 329 | eval_loss = eval_loss / nb_eval_steps 330 | perplexity = torch.exp(torch.tensor(eval_loss)) 331 | 332 | results["eval_loss"] = eval_loss 333 | results["perplexity"] = perplexity 334 | if output_dir: 335 | os.makedirs(output_dir, exist_ok=True) 336 | with open(os.path.join(output_dir, "eval_results.txt"), "w") as writer: 337 | for key in sorted(results.keys()): 338 | writer.write("{} = {}\n".format(key, str(results[key]))) 339 | 340 | return results 341 | 342 | def save_model(self, output_dir, model, results=None): 343 | """ 344 | Saves the model to output_dir. 345 | :param output_dir: 346 | :param model: 347 | :param results: 348 | :return: 349 | """ 350 | logger.info("Saving model checkpoint to %s", output_dir) 351 | os.makedirs(output_dir, exist_ok=True) 352 | model_to_save = model.module if hasattr(model, "module") else model 353 | model_to_save.save_pretrained(output_dir) 354 | self.tokenizer.save_pretrained(output_dir) 355 | if results: 356 | output_eval_file = os.path.join(output_dir, "eval_results.txt") 357 | with open(output_eval_file, "w") as writer: 358 | for key in sorted(results.keys()): 359 | writer.write("{} = {}\n".format(key, str(results[key]))) 360 | 361 | def generate( 362 | self, 363 | prompt: str, 364 | is_add_prompt: bool = True, 365 | max_length: int = 128, 366 | temperature: int = 1.0, 367 | top_k: int = 50, 368 | top_p: float = 0.95, 369 | repetition_penalty: float = 1.0, 370 | do_sample: bool = True, 371 | num_return_sequences: int = 1, 372 | length_penalty: float = 2.0, 373 | early_stopping: bool = True, 374 | stop_word: str = "\n\n", 375 | bad_words: list = None, 376 | **kwargs, 377 | ): 378 | """ 379 | Generate text using a GPT2 LanguageGenerationModel 380 | 381 | Args: 382 | prompt: A prompt text for the model. 383 | is_add_prompt: Whether to add the prompt to the returned text. 384 | max_length: The maximum length of the sequence to be generated. 385 | temperature: The sampling temperature. 386 | top_k: The number of top k tokens to be considered by sampling. 387 | top_p: The sampling probability for top p tokens. 388 | repetition_penalty: The repetition penalty parameter. 389 | do_sample: Boolean value indicating whether to sample or greedy generate. 390 | num_return_sequences: The number of samples to return. 391 | length_penalty: The length penalty parameter. 392 | early_stopping: Boolean value indicating whether to do early stopping or not. 393 | stop_word: A stop word to stop generation. 394 | bad_words: A list of bad words to be ignored. 395 | Returns: 396 | generated_sequences: list, Sequences of text generated by the model. 397 | """ 398 | encoded_prompt = self.tokenizer(prompt, return_tensors="pt").to(device) 399 | encoded_prompt_ids = encoded_prompt.input_ids 400 | # Get tokens of words that should not be generated 401 | bad_words_ids = [self.tokenizer(bad_word, add_prefix_space=True).input_ids for bad_word in 402 | bad_words] if bad_words else None 403 | output_sequences = self.model.generate( 404 | input_ids=encoded_prompt_ids, 405 | max_length=max_length if max_length is not None else self.max_length, 406 | temperature=temperature, 407 | top_k=top_k, 408 | top_p=top_p, 409 | repetition_penalty=repetition_penalty, 410 | do_sample=do_sample, 411 | num_return_sequences=num_return_sequences, 412 | length_penalty=length_penalty, 413 | early_stopping=early_stopping, 414 | bad_words_id=bad_words_ids, 415 | bos_token_id=self.tokenizer.bos_token_id, 416 | pad_token_id=self.tokenizer.eos_token_id, # tokenizer.pad_token_ids is None 417 | eos_token_id=self.tokenizer.eos_token_id, 418 | **kwargs 419 | ) 420 | 421 | # Remove the batch dimension when returning multiple sequences 422 | if len(output_sequences.shape) > 2: 423 | output_sequences.squeeze_() 424 | 425 | generated_sequences = [] 426 | for generated_sequence in output_sequences: 427 | generated_sequence = generated_sequence.tolist() 428 | # Decode text 429 | text = self.tokenizer.decode(generated_sequence, clean_up_tokenization_spaces=True) 430 | # Remove all text after the stop word token 431 | text = text[: text.find(stop_word) if stop_word else None] 432 | # Remove the excess text that was used for pre-processing 433 | total_sequence = text[len(self.tokenizer.decode(encoded_prompt_ids[0], clean_up_tokenization_spaces=True)):] 434 | # Add the prompt at the beginning of the sequence. 435 | if is_add_prompt: 436 | total_sequence = prompt + total_sequence 437 | generated_sequences.append(total_sequence) 438 | 439 | return generated_sequences 440 | -------------------------------------------------------------------------------- /codeassist/wizard_coder.py: -------------------------------------------------------------------------------- 1 | # -*- coding: utf-8 -*- 2 | """ 3 | @author:XuMing(xuming624@qq.com) 4 | @description: Rewrite the original WizardCoder to support code completion. 5 | """ 6 | 7 | import math 8 | import os 9 | import random 10 | from copy import deepcopy 11 | from dataclasses import dataclass 12 | from typing import List, Optional, Dict, Sequence, Union 13 | 14 | import numpy as np 15 | import torch 16 | from datasets import load_dataset 17 | from loguru import logger 18 | from peft import ( 19 | PeftModel, 20 | LoraConfig, 21 | get_peft_model, 22 | prepare_model_for_int8_training, 23 | ) 24 | from tqdm import tqdm 25 | from transformers import ( 26 | PreTrainedTokenizer, 27 | PreTrainedModel, 28 | AutoModelForCausalLM, 29 | AutoTokenizer, 30 | Trainer, 31 | TrainingArguments, 32 | ) 33 | 34 | os.environ["KMP_DUPLICATE_LIB_OK"] = "TRUE" 35 | has_cuda = torch.cuda.is_available() 36 | IGNORE_INDEX = -100 37 | PROMPT_DICT = { 38 | "prompt_input": ( 39 | "Below is an instruction that describes a task, paired with an input that provides further context. " 40 | "Write a response that appropriately completes the request.\n\n" 41 | "### Instruction:\n{instruction}\n\n### Input:\n{input}\n\n### Response:" 42 | ), 43 | "prompt_no_input": ( 44 | "Below is an instruction that describes a task. " 45 | "Write a response that appropriately completes the request.\n\n" 46 | "### Instruction:\n{instruction}\n\n### Response:" 47 | ), 48 | } 49 | 50 | 51 | class WizardCoder: 52 | def __init__( 53 | self, 54 | model_name_or_path: str = "WizardLM/WizardCoder-15B-V1.0", 55 | peft_name: Optional[str] = None, 56 | special_words_dict: Dict = None, 57 | use_cuda: Optional[bool] = has_cuda, 58 | cuda_device: Optional[int] = -1, 59 | fp16: bool = True, 60 | bf16: bool = False, 61 | **kwargs, 62 | ): 63 | """ 64 | Initializes a AutoModelForCausalLM 65 | 66 | Args: 67 | model_name_or_path: Default Transformer model name or path to a directory containing Transformer model file (pytorch_nodel.bin). 68 | peft_name: The name of the PEFT model to use. 69 | special_words_dict: A dictionary of special words and their token ids. 70 | use_cuda: Use GPU if available. 71 | cuda_device: Which cuda device to use. 72 | fp16: Use 16-bit (mixed) precision (through NVIDIA apex) instead of 32-bit. 73 | bf16: Use 16-bit (mixed) precision (through NVIDIA apex) instead of 32-bit. 74 | **kwargs: Additional kwargs for the Transformers `PreTrainedModel` and `PreTrainedTokenizer` classes. 75 | """ 76 | self.model_name_or_path = model_name_or_path 77 | self.fp16 = fp16 78 | self.bf16 = bf16 79 | self.device_map = "auto" 80 | if use_cuda: 81 | if torch.cuda.is_available(): 82 | if cuda_device == -1: 83 | self.device = torch.device("cuda") 84 | else: 85 | self.device = torch.device(f"cuda:{cuda_device}") 86 | self.device_map = {"": int(cuda_device)} 87 | else: 88 | raise ValueError( 89 | "'use_cuda' set to True when cuda is unavailable." 90 | "Make sure CUDA is available or set `use_cuda=False`." 91 | ) 92 | else: 93 | if torch.backends.mps.is_available(): 94 | self.device = torch.device("mps") 95 | self.device_map = {"": "mps"} 96 | else: 97 | self.device = "cpu" 98 | self.device_map = {"": "cpu"} 99 | logger.debug(f"Device: {self.device}") 100 | if not use_cuda: 101 | self.fp16 = False 102 | self.bf16 = False 103 | world_size = int(os.environ.get("WORLD_SIZE", 1)) 104 | self.ddp = world_size != 1 105 | if self.ddp: 106 | self.device_map = {"": int(os.environ.get("LOCAL_RANK") or 0)} 107 | if torch.cuda.is_bf16_supported() and not self.bf16: 108 | logger.warning("GPU supports bf16, you can enable bf16.") 109 | self.torch_dtype = torch.bfloat16 if self.bf16 else (torch.float16 if self.fp16 else torch.float32) 110 | self.model = AutoModelForCausalLM.from_pretrained( 111 | model_name_or_path, 112 | torch_dtype=self.torch_dtype, 113 | device_map=self.device_map, 114 | **kwargs, 115 | ) 116 | if peft_name: 117 | # Load PEFT model for inference default, if you want to continue training, please set is_trainable=True 118 | self.model = PeftModel.from_pretrained( 119 | self.model, 120 | peft_name, 121 | torch_dtype=self.torch_dtype, 122 | device_map=self.device_map, 123 | is_trainable=False, 124 | ) 125 | logger.info(f"Loaded peft model from {peft_name}") 126 | 127 | self.tokenizer = AutoTokenizer.from_pretrained(model_name_or_path) 128 | if self.tokenizer.pad_token_id is None: 129 | self.tokenizer.pad_token_id = 0 130 | # Set padding side equal to Collator padding side 131 | self.tokenizer.padding_side = "left" 132 | 133 | if "starcoder" in model_name_or_path: 134 | self.tokenizer.add_special_tokens({ 135 | "eos_token": "<|endoftext|>", 136 | "bos_token": "<|endoftext|>", 137 | "unk_token": "<|endoftext|>", 138 | }) 139 | if special_words_dict is not None: 140 | self.tokenizer.add_special_tokens(special_words_dict) 141 | self.results = {} 142 | 143 | def set_seed(self, seed): 144 | logger.debug(f"Set seed for random, numpy and torch: {seed}") 145 | random.seed(seed) 146 | np.random.seed(seed) 147 | torch.manual_seed(seed) 148 | if torch.cuda.is_available(): 149 | torch.cuda.manual_seed_all(seed) 150 | 151 | def train_model( 152 | self, 153 | train_file: str, 154 | output_dir: str, 155 | eval_file: str = None, 156 | batch_size: int = 8, 157 | num_epochs: int = 1, 158 | lr: float = 5e-5, 159 | gradient_accumulation_steps: int = 1, 160 | max_steps: int = -1, 161 | logging_steps: int = 50, 162 | gradient_checkpointing: bool = True, 163 | torch_compile: bool = False, 164 | warmup_steps: int = 200, 165 | save_steps: int = 400, 166 | eval_steps: int = 200, 167 | optimizer: str = "adamw_torch", 168 | save_strategy: str = "steps", 169 | save_total_limit: int = 10, 170 | report_to: Optional[List[str]] = "tensorboard", 171 | overwrite_output_dir: bool = True, 172 | use_peft: bool = True, 173 | int8: bool = False, 174 | max_eval_samples: int = 20, 175 | **kwargs, 176 | ): 177 | """ 178 | Trains the model on 'train_file' 179 | 180 | Args: 181 | train_file: Path to text file containing the text to train the language model on. 182 | output_dir: The directory where model files will be saved. If not given, self.args.output_dir will be used. 183 | eval_file (optional): Path to eval file containing the text to evaluate the language model on. 184 | batch_size (optional): Batch size for training. 185 | num_epochs (optional): Number of epochs for training. 186 | lr (optional): Learning rate. 187 | gradient_accumulation_steps (optional): Number of updates steps to accumulate before performing a backward/update pass. 188 | max_steps (optional): If > 0: set total number of training steps to perform. Override num_epochs. 189 | logging_steps (optional): Number of steps between logging. 190 | gradient_checkpointing (optional): If True, use gradient checkpointing to save memory at the expense of slower backward pass. 191 | torch_compile (optional): If True, use torch's new experimental just-in-time (JIT) compiler to compile the model for faster runtime. 192 | warmup_steps (optional): Number of steps for the warmup in the lr scheduler. 193 | save_steps (optional): Number of steps between saving. 194 | eval_steps (optional): Number of steps between evaluations. 195 | optimizer (optional): Optimizer to use. Can be 'adamw_torch', 'adamw_deepspeed', 'adam', 'sgd', 'lamb' or 'lamb_wd'. 196 | save_strategy (optional): Strategy to save checkpoints. Can be 'steps' or 'epoch'. 197 | save_total_limit (optional): Maximum number of checkpoints to keep. 198 | report_to (optional): The list of integrations to report the results and logs to. 199 | overwrite_output_dir (optional): Overwrite the content of the output directory. 200 | use_peft (optional): If True, use the PEFT scheduler to schedule the training. 201 | int8 (optional): If True, use int8 quantization for the model. 202 | max_eval_samples (optional): Maximum number of samples to use for evaluation. 203 | kwargs (optional): Optional model specific arguments. 204 | 205 | Returns: 206 | global_step: Number of global steps trained 207 | metrics: Dictionary containing the evaluation results. 208 | """ 209 | os.makedirs(output_dir, exist_ok=True) 210 | self.set_seed(42) 211 | logger.debug(f"Tokenizer: {self.tokenizer}") 212 | logger.debug(f"Model: {self.model}") 213 | 214 | training_args = TrainingArguments( 215 | output_dir=output_dir, 216 | dataloader_drop_last=True, 217 | learning_rate=lr, 218 | num_train_epochs=num_epochs, 219 | max_steps=max_steps, 220 | logging_dir=f"{output_dir}/logs", 221 | logging_steps=logging_steps, 222 | per_device_train_batch_size=batch_size, 223 | per_device_eval_batch_size=batch_size, 224 | gradient_checkpointing=gradient_checkpointing, 225 | torch_compile=torch_compile, 226 | gradient_accumulation_steps=gradient_accumulation_steps, 227 | warmup_steps=warmup_steps, 228 | save_steps=save_steps, 229 | optim=optimizer, 230 | save_strategy=save_strategy, 231 | evaluation_strategy='steps' if eval_file is not None else 'no', 232 | eval_steps=eval_steps if eval_file is not None else None, 233 | load_best_model_at_end=True if eval_file is not None else False, 234 | ddp_find_unused_parameters=False, 235 | save_total_limit=save_total_limit, 236 | fp16=self.fp16, 237 | bf16=self.bf16, 238 | report_to=report_to, 239 | overwrite_output_dir=overwrite_output_dir, 240 | no_cuda=True if self.device == "cpu" else False, 241 | **kwargs 242 | ) 243 | # update model train config 244 | if training_args.gradient_checkpointing: 245 | self.model.gradient_checkpointing_enable() 246 | self.model.config.use_cache = False 247 | else: 248 | self.model.config.use_cache = True 249 | self.model.enable_input_require_grads() 250 | 251 | # Tell Trainer not to attempt DataParallel 252 | self.model.is_parallelizable = True 253 | self.model.model_parallel = True 254 | # Setup peft 255 | if use_peft: 256 | peft_config = LoraConfig( 257 | task_type="CAUSAL_LM", 258 | inference_mode=False, 259 | r=16, 260 | lora_alpha=32, 261 | lora_dropout=0.05, 262 | target_modules=["c_proj", "c_attn", "q_attn"], 263 | bias="none", 264 | ) 265 | if int8: 266 | self.model = prepare_model_for_int8_training(self.model) 267 | self.model = get_peft_model(self.model, peft_config) 268 | self.model.print_trainable_parameters() # Be more transparent about the % of trainable params. 269 | else: 270 | logger.warning("Now full model params fine-tune, which is slow, set `use_peft=True` for lora fine-tune.") 271 | logger.debug(f"Tokenizer: {self.tokenizer}") 272 | logger.debug(f"Model: {self.model}") 273 | 274 | # load dataset 275 | raw_train_datasets = load_dataset('json', data_files=train_file, split="train") 276 | logger.debug(f"Example train_dataset[0]: {raw_train_datasets[0]}") 277 | with training_args.main_process_first(desc="Train dataset tokenization"): 278 | train_dataset = raw_train_datasets.map( 279 | self.train_tokenize_function, 280 | batched=True, 281 | num_proc=1, 282 | remove_columns=raw_train_datasets.column_names, 283 | desc="Running tokenizer on train dataset", 284 | fn_kwargs={"tokenizer": self.tokenizer} 285 | ) 286 | logger.debug(f"Train dataset size: {len(train_dataset)}") 287 | logger.debug(f"First sample of the training set: {train_dataset[0]}.") 288 | eval_dataset = None 289 | if eval_file is not None: 290 | raw_eval_datasets = load_dataset('json', data_files=eval_file, split="train") 291 | if max_eval_samples is not None and max_eval_samples > 0: 292 | max_eval_samples = min(len(raw_eval_datasets), max_eval_samples) 293 | raw_eval_datasets = raw_eval_datasets.select(range(max_eval_samples)) 294 | with training_args.main_process_first(desc="Eval dataset tokenization"): 295 | eval_dataset = raw_eval_datasets.map( 296 | self.train_tokenize_function, 297 | batched=True, 298 | num_proc=1, 299 | remove_columns=raw_eval_datasets.column_names, 300 | desc="Running tokenizer on train dataset", 301 | fn_kwargs={"tokenizer": self.tokenizer} 302 | ) 303 | logger.debug(f"Eval dataset size: {len(eval_dataset)}") 304 | logger.debug(f"First sample of the eval set: {eval_dataset[0]}.") 305 | 306 | # Log on each process the small summary: 307 | logger.warning( 308 | f"Process rank: {training_args.local_rank}, device: {training_args.device}, n_gpu: {training_args.n_gpu}, " 309 | + f"distributed training: {bool(training_args.local_rank != -1)}, 16-bits training: {training_args.fp16}" 310 | ) 311 | if training_args.local_rank <= 0: 312 | logger.info(f"Training/evaluation parameters {training_args}") 313 | 314 | data_collator = DataCollatorForSupervisedDataset(tokenizer=self.tokenizer) 315 | trainer = Trainer( 316 | model=self.model, 317 | tokenizer=self.tokenizer, 318 | args=training_args, 319 | train_dataset=train_dataset, 320 | eval_dataset=eval_dataset, 321 | data_collator=data_collator, 322 | ) 323 | logger.info("*** Train ***") 324 | logger.debug(f"Train dataloader example: {next(iter(trainer.get_train_dataloader()))}") 325 | (global_step, training_loss, metrics) = trainer.train() 326 | self.results.update(metrics) 327 | trainer.log_metrics("train", metrics) 328 | trainer.save_metrics("train", metrics) 329 | trainer.save_state() 330 | self.save_model(output_dir=output_dir) 331 | logger.info(f" Training model done. Saved to {output_dir}.") 332 | 333 | if eval_dataset is not None: 334 | logger.info("*** Evaluate ***") 335 | metrics = trainer.evaluate(metric_key_prefix="eval") 336 | metrics['eval_samples'] = len(eval_dataset) 337 | try: 338 | perplexity = math.exp(metrics["eval_loss"]) 339 | except OverflowError: 340 | perplexity = float("inf") 341 | metrics["perplexity"] = perplexity 342 | logger.debug(f"eval metrics: {metrics}") 343 | self.results.update(metrics) 344 | trainer.log_metrics("eval", metrics) 345 | trainer.save_metrics("eval", metrics) 346 | 347 | if training_args.local_rank <= 0: 348 | logger.debug(f"metrics: {self.results}") 349 | logger.info( 350 | " Training of {} model complete. Saved to {}.".format( 351 | self.model_name_or_path, output_dir 352 | ) 353 | ) 354 | 355 | return global_step, metrics 356 | 357 | def save_model(self, output_dir): 358 | """Save the model and the tokenizer.""" 359 | os.makedirs(output_dir, exist_ok=True) 360 | model = self.model 361 | # Take care of distributed/parallel training 362 | model_to_save = model.module if hasattr(model, "module") else model 363 | model_to_save.save_pretrained(output_dir) 364 | self.tokenizer.save_pretrained(output_dir) 365 | 366 | def smart_tokenizer_and_embedding_resize( 367 | self, 368 | special_tokens_dict: Dict, 369 | tokenizer: PreTrainedTokenizer, 370 | model: PreTrainedModel, 371 | ): 372 | """Resize tokenizer and embedding. 373 | 374 | Note: This is the unoptimized version that may make your embedding size not be divisible by 64. 375 | """ 376 | num_new_tokens = tokenizer.add_special_tokens(special_tokens_dict) 377 | model.resize_token_embeddings(len(tokenizer)) 378 | 379 | if num_new_tokens > 0: 380 | input_embeddings = model.get_input_embeddings().weight.data 381 | output_embeddings = model.get_output_embeddings().weight.data 382 | 383 | input_embeddings_avg = input_embeddings[:-num_new_tokens].mean(dim=0, keepdim=True) 384 | output_embeddings_avg = output_embeddings[:-num_new_tokens].mean(dim=0, keepdim=True) 385 | 386 | input_embeddings[-num_new_tokens:] = input_embeddings_avg 387 | output_embeddings[-num_new_tokens:] = output_embeddings_avg 388 | 389 | def _tokenize_fn(self, strings: Sequence[str], tokenizer: PreTrainedTokenizer) -> Dict: 390 | """Tokenize a list of strings.""" 391 | tokenized_list = [ 392 | tokenizer( 393 | text, 394 | return_tensors="pt", 395 | padding="longest", 396 | max_length=tokenizer.model_max_length, 397 | truncation=True, 398 | ) 399 | for text in strings 400 | ] 401 | input_ids = [tokenized.input_ids[0] for tokenized in tokenized_list] 402 | input_ids_lens = [ 403 | tokenized.input_ids.ne(tokenizer.pad_token_id).sum().item() for tokenized in tokenized_list 404 | ] 405 | return dict( 406 | input_ids=input_ids, 407 | input_ids_lens=input_ids_lens, 408 | ) 409 | 410 | def preprocess( 411 | self, 412 | sources: Sequence[str], 413 | targets: Sequence[str], 414 | tokenizer: PreTrainedTokenizer, 415 | ) -> Dict: 416 | """Preprocess the data by tokenizing.""" 417 | examples = [s + t for s, t in zip(sources, targets)] 418 | examples_tokenized, sources_tokenized = [ 419 | self._tokenize_fn(strings, tokenizer) for strings in (examples, sources) 420 | ] 421 | input_ids = examples_tokenized["input_ids"] 422 | labels = deepcopy(input_ids) 423 | for label, source_len in zip(labels, sources_tokenized["input_ids_lens"]): 424 | label[:source_len] = IGNORE_INDEX 425 | return dict(input_ids=input_ids, labels=labels) 426 | 427 | def train_tokenize_function(self, examples, tokenizer): 428 | prompt_input, prompt_no_input = PROMPT_DICT["prompt_input"], PROMPT_DICT["prompt_no_input"] 429 | if 'input' in examples: 430 | sources = [ 431 | prompt_input.format_map(dict(instruction=instruction, input=input)) if input != "" 432 | else prompt_no_input.format_map(dict(instruction=instruction)) 433 | for instruction, input in zip(examples['instruction'], examples['input']) 434 | ] 435 | else: 436 | sources = [ 437 | prompt_no_input.format_map(dict(instruction=instruction)) 438 | for instruction in examples['instruction'] 439 | ] 440 | targets = [f"{output}{tokenizer.eos_token}" for output in examples['output']] 441 | data_dict = self.preprocess(sources, targets, tokenizer) 442 | return data_dict 443 | 444 | @torch.inference_mode() 445 | def generate( 446 | self, 447 | sentences: Union[List[str], str], 448 | keep_prompt: bool = False, 449 | add_system_prompt: bool = True, 450 | eval_batch_size: int = 4, 451 | max_length: int = 256, 452 | temperature: int = 1.0, 453 | top_k: int = 50, 454 | top_p: float = 0.95, 455 | do_sample: bool = True, 456 | num_beams: int = 1, 457 | repetition_penalty: float = 1.0, 458 | length_penalty: float = 1.2, 459 | early_stopping: bool = True, 460 | verbose: bool = True, 461 | **kwargs 462 | ) -> List[str]: 463 | """ 464 | Performs predictions on a list of text. 465 | 466 | Args: 467 | sentences: A prompt text for the model. 468 | keep_prompt: Whether to keep the prompt in the generated text. 469 | add_system_prompt: Whether to add the system prompt to the prompt text. 470 | eval_batch_size: The batch size for evaluation. 471 | max_length: The maximum length of the generated text. 472 | temperature: The sampling temperature. 473 | top_k: The number of top k tokens to be considered by sampling. 474 | top_p: The sampling probability for top p tokens. 475 | num_beams: The number of beams for beam search. 476 | repetition_penalty: The repetition penalty parameter. 477 | do_sample: Boolean value indicating whether to sample or greedy generate. 478 | length_penalty: The length penalty parameter. 479 | early_stopping: Boolean value indicating whether to do early stopping or not. 480 | verbose: Boolean value indicating whether to print the progress bar. 481 | **kwargs: Additional arguments for the generate method of the model. 482 | Returns: 483 | generated_sequences: list, Sequences of text generated by the model. 484 | """ 485 | all_outputs = [] 486 | if isinstance(sentences, str): 487 | sentences = [sentences] 488 | for batch in tqdm( 489 | [sentences[i: i + eval_batch_size] for i in range(0, len(sentences), eval_batch_size)], 490 | desc="Generating outputs", 491 | disable=not verbose, 492 | ): 493 | if add_system_prompt: 494 | batch = [PROMPT_DICT['prompt_no_input'].format(instruction=s) for s in batch] 495 | inputs = self.tokenizer( 496 | batch, 497 | return_tensors="pt", 498 | max_length=max_length, 499 | truncation=True, 500 | padding=True, 501 | ) 502 | outputs = self.model.generate( 503 | input_ids=inputs['input_ids'].to(self.device), 504 | max_new_tokens=max_length, 505 | temperature=temperature, 506 | top_k=top_k, 507 | top_p=top_p, 508 | num_beams=num_beams, 509 | repetition_penalty=repetition_penalty, 510 | do_sample=do_sample, 511 | length_penalty=length_penalty, 512 | early_stopping=early_stopping, 513 | bos_token_id=self.tokenizer.bos_token_id, 514 | pad_token_id=self.tokenizer.pad_token_id, 515 | eos_token_id=self.tokenizer.eos_token_id, 516 | return_dict_in_generate=True, 517 | output_scores=True, 518 | **kwargs 519 | ) 520 | for idx, (prompt_text, generated_sequence) in enumerate(zip(batch, outputs.sequences)): 521 | # Decode text 522 | text = self.tokenizer.decode(generated_sequence, skip_special_tokens=True) 523 | prompt_len = len(prompt_text) 524 | gen_text = text[prompt_len:] 525 | if keep_prompt: 526 | total_sequence = prompt_text + gen_text 527 | else: 528 | total_sequence = gen_text 529 | all_outputs.append(total_sequence) 530 | return all_outputs 531 | 532 | 533 | @dataclass 534 | class DataCollatorForSupervisedDataset(object): 535 | """Collate examples for supervised fine-tuning.""" 536 | 537 | tokenizer: PreTrainedTokenizer 538 | 539 | def __call__(self, instances: Sequence[Dict]) -> Dict[str, torch.Tensor]: 540 | input_ids, labels = tuple([instance[key] for instance in instances] for key in ("input_ids", "labels")) 541 | input_ids = [torch.tensor(x) for x in input_ids] 542 | input_ids = torch.nn.utils.rnn.pad_sequence( 543 | input_ids, batch_first=True, padding_value=self.tokenizer.pad_token_id 544 | ) 545 | labels = [torch.tensor(x) for x in labels] 546 | labels = torch.nn.utils.rnn.pad_sequence(labels, batch_first=True, padding_value=IGNORE_INDEX) 547 | return dict( 548 | input_ids=input_ids, 549 | labels=labels, 550 | attention_mask=input_ids.ne(self.tokenizer.pad_token_id), 551 | ) 552 | -------------------------------------------------------------------------------- /docs/36-text-rep-examples.md: -------------------------------------------------------------------------------- 1 | --- 2 | title: Text Representation Examples 3 | permalink: /docs/text-rep-examples/ 4 | excerpt: "Text Representation Examples" 5 | last_modified_at: 2020/07/26 23:16:38 6 | toc: true 7 | --- 8 | 9 | ### Minimal example for generating word embeddings 10 | Generate a list of contextual word embeddings for every sentence in a list 11 | ```python 12 | from simpletransformers.language_representation import RepresentationModel 13 | 14 | sentences = ["Example sentence 1", "Example sentence 2"] 15 | model = RepresentationModel( 16 | model_type="bert", 17 | model_name="bert-base-uncased", 18 | use_cuda=False 19 | ) 20 | word_vectors = model.encode_sentences(sentences, combine_strategy=None) 21 | assert word_vectors.shape === (2, 5, 768) # token vector for every token in each sentence, bert based models add 2 tokens per sentence by default([CLS] & [SEP]) 22 | ``` 23 | 24 | ### Minimal example for generating sentence embeddings 25 | Same code as for generating word embeddings, the only difference is that we pass `combine_strategy="mean"` parameter 26 | ```python 27 | from simpletransformers.language_representation import RepresentationModel 28 | sentences = ["Example sentence 1", "Example sentence 2"] 29 | model = RepresentationModel( 30 | model_type="bert", 31 | model_name="bert-base-uncased", 32 | use_cuda=False 33 | ) 34 | word_vectors = model.encode_sentences(sentences, combine_strategy="mean") 35 | assert word_vectors.shape === (2, 768) # one sentence embedding per sentence 36 | ``` 37 | -------------------------------------------------------------------------------- /docs/api.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/shibing624/CodeAssist/c5885d2d20e3dbea7d653da17e2a635951f522a5/docs/api.png -------------------------------------------------------------------------------- /docs/codeassist.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/shibing624/CodeAssist/c5885d2d20e3dbea7d653da17e2a635951f522a5/docs/codeassist.png -------------------------------------------------------------------------------- /docs/hf_model.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/shibing624/CodeAssist/c5885d2d20e3dbea7d653da17e2a635951f522a5/docs/hf_model.png -------------------------------------------------------------------------------- /docs/source_code_datasets/source_code_dataset.md: -------------------------------------------------------------------------------- 1 | --- 2 | annotations_creators: 3 | - no-annotation 4 | language_creators: 5 | - crowdsourced 6 | languages: 7 | - en 8 | licenses: 9 | - cc-by-4-0 10 | - gfdl-1-3-or-later 11 | multilinguality: 12 | - monolingual 13 | size_categories: 14 | - 100M<n<200M 15 | source_datasets: 16 | - https://github.com/shibing624/code-autocomplete 17 | - https://github.com/bharathgs/Awesome-pytorch-list 18 | - https://github.com/akullpp/awesome-java 19 | - https://github.com/fffaraz/awesome-cpp 20 | task_categories: 21 | - sequence-modeling 22 | task_ids: 23 | - language-modeling 24 | --- 25 | # Dataset Card for "SourceCode" 26 | ## Table of Contents 27 | - [Dataset Description](#dataset-description) 28 | - [Dataset Summary](#dataset-summary) 29 | - [Supported Tasks and Leaderboards](#supported-tasks-and-leaderboards) 30 | - [Languages](#languages) 31 | - [Dataset Structure](#dataset-structure) 32 | - [Data Instances](#data-instances) 33 | - [Data Fields](#data-fields) 34 | - [Data Splits](#data-splits) 35 | - [Dataset Creation](#dataset-creation) 36 | - [Curation Rationale](#curation-rationale) 37 | - [Source Data](#source-data) 38 | - [Annotations](#annotations) 39 | - [Personal and Sensitive Information](#personal-and-sensitive-information) 40 | - [Considerations for Using the Data](#considerations-for-using-the-data) 41 | - [Social Impact of Dataset](#social-impact-of-dataset) 42 | - [Discussion of Biases](#discussion-of-biases) 43 | - [Other Known Limitations](#other-known-limitations) 44 | - [Additional Information](#additional-information) 45 | - [Dataset Curators](#dataset-curators) 46 | - [Licensing Information](#licensing-information) 47 | - [Citation Information](#citation-information) 48 | - [Contributions](#contributions) 49 | 50 | ## Dataset Description 51 | - **Repository:** [code-autocomplete](https://github.com/shibing624/code-autocomplete) 52 | - **Leaderboard:** [leaderboard](https://github.com/shibing624/code-autocomplete) (located on the homepage) 53 | - **Size of downloaded dataset files:** 105 MB 54 | - **Total amount of disk used:** 570 MB 55 | 56 | ### Dataset Summary 57 | 58 | Source code dataset is a collection of Github awesome repos, it contains Python, Java, C++, and other programming languages. 59 | This dataset can be used in different NLP tasks like language modeling and text generation tasks. 60 | 61 | data source: 62 | 63 | - PYTHON_CODE: https://github.com/bharathgs/Awesome-pytorch-list 64 | - JAVA_CODE: https://github.com/akullpp/awesome-java 65 | - CPP_CODE: https://github.com/fffaraz/awesome-cpp 66 | 67 | 68 | ### Supported Tasks and Leaderboards 69 | - language modeling 70 | - code generation tasks, **Leaderboard:** [code-autocomplete](https://github.com/shibing624/code-autocomplete) 71 | 72 | ### Languages 73 | 74 | - programming languages: Python, Java, C++ 75 | - natural language: English 76 | 77 | ## Dataset Structure 78 | ### Data Instances 79 | An example of 'train' looks as follows. 80 | ``` 81 | This example was too long and was cropped: 82 | 83 | { 84 | "text": """ 85 | import json 86 | import argparse 87 | 88 | 89 | def _parse_args(): 90 | parser = argparse.ArgumentParser( 91 | description=__doc__, 92 | formatter_class=argparse.RawTextHelpFormatter, 93 | ) 94 | parser.add_argument( 95 | '--model-file', 96 | required=True, 97 | help=( 98 | 'A pt file from ' 99 | 'https://github.com/pytorch/fairseq/tree/main/examples/hubert' 100 | ) 101 | ) 102 | return parser.parse_args() 103 | """ 104 | } 105 | ``` 106 | ### Data Fields 107 | The data fields are the same among all splits. 108 | - `text`: a `string` feature. 109 | ### Data Splits 110 | #### python 111 | ```shell 112 | $ wc -l python/* 113 | 10000 python/test.txt 114 | 5215412 python/train.txt 115 | 10000 python/valid.txt 116 | 5235412 total 117 | ``` 118 | #### java 119 | ```shell 120 | $ wc -l java/* 121 | 950083 java/test.txt 122 | 2802880 java/train.txt 123 | 940803 java/valid.txt 124 | 4693766 total 125 | ``` 126 | #### cpp 127 | ```shell 128 | $ wc -l cpp/* 129 | 1060014 cpp/test.txt 130 | 3119241 cpp/train.txt 131 | 1099124 cpp/valid.txt 132 | 5278379 total 133 | ``` 134 | ## Dataset Creation 135 | ### Curation Rationale 136 | As code generation dataset, I upload it to huggingface datasets. 137 | ### Source Data 138 | #### Initial Data Collection and Normalization 139 | #### Who are the source language producers? 140 | Citation: 141 | 142 | APA: 143 | ```latex 144 | Xu, M. code-autocomplete: Code AutoComplete with GPT2 model (Version 0.0.4) [Computer software]. https://github.com/shibing624/code-autocomplete 145 | ``` 146 | 147 | BibTeX: 148 | ```latex 149 | @software{Xu_code-autocomplete_Code_AutoComplete, 150 | author = {Xu, Ming}, 151 | title = {code-autocomplete: Code AutoComplete with GPT2 model}, 152 | url = {https://github.com/shibing624/code-autocomplete}, 153 | version = {0.0.4} 154 | } 155 | ``` 156 | 157 | ### Annotations 158 | #### Annotation process 159 | #### Who are the annotators? 160 | nobody 161 | ### Personal and Sensitive Information 162 | ## Considerations for Using the Data 163 | ### Social Impact of Dataset 164 | This dataset was developed as a benchmark for evaluating code generation model. 165 | ### Discussion of Biases 166 | ### Other Known Limitations 167 | ## Additional Information 168 | ### Dataset Curators 169 | 170 | Github awesome programing code repos. 171 | 172 | ### Licensing Information 173 | 174 | For research use only. 175 | 176 | ### Contributions 177 | Thanks to [@shibing624](https://github.com/shibing624) add this dataset. -------------------------------------------------------------------------------- /docs/source_code_datasets/upload_dataset.py: -------------------------------------------------------------------------------- 1 | # -*- coding: utf-8 -*- 2 | """ 3 | @author:XuMing(xuming624@qq.com) 4 | @description: 5 | """ 6 | 7 | """Code AutoComplete Python dataset Corpus.(code_autocomplete)""" 8 | 9 | import os 10 | 11 | import datasets 12 | 13 | _DESCRIPTION = """纯文本数据，内容：高质量编程源代码，包括Python，Java，CPP源代码""" 14 | 15 | PYTHON_HOME = "https://github.com/bharathgs/Awesome-pytorch-list" 16 | JAVA_HOME = "https://github.com/akullpp/awesome-java" 17 | CPP_HOME = "https://github.com/fffaraz/awesome-cpp" 18 | 19 | _CITATION = "https://github.com/shibing624/code-autocomplete" 20 | 21 | _DATA_URL = "https://github.com/shibing624/code-autocomplete/releases/download/0.0.4/source_code.zip" 22 | 23 | 24 | class SourceCodeConfig(datasets.BuilderConfig): 25 | """BuilderConfig for NLI_zh""" 26 | 27 | def __init__(self, features, data_url, citation, url, **kwargs): 28 | """BuilderConfig for NLI_zh 29 | Args: 30 | features: `list[string]`, list of the features that will appear in the 31 | feature dict. Should not include "label". 32 | data_url: `string`, url to download the zip file from. 33 | citation: `string`, citation for the data set. 34 | url: `string`, url for information about the data set. 35 | **kwargs: keyword arguments forwarded to super. 36 | """ 37 | super().__init__(version=datasets.Version("1.0.0"), **kwargs) 38 | self.features = features 39 | self.data_url = data_url 40 | self.citation = citation 41 | self.url = url 42 | 43 | 44 | class SourceCode(datasets.GeneratorBasedBuilder): 45 | """The Natural Language Inference Chinese(NLI_zh) Corpus.""" 46 | 47 | BUILDER_CONFIGS = [ 48 | SourceCodeConfig( 49 | name="python", 50 | description=_DESCRIPTION, 51 | features=["text"], 52 | data_url=_DATA_URL, 53 | citation=_CITATION, 54 | url=PYTHON_HOME, 55 | ), 56 | SourceCodeConfig( 57 | name="java", 58 | description=_DESCRIPTION, 59 | features=["text"], 60 | data_url=_DATA_URL, 61 | citation=_CITATION, 62 | url=JAVA_HOME, 63 | ), 64 | SourceCodeConfig( 65 | name="cpp", 66 | description=_DESCRIPTION, 67 | features=["text"], 68 | data_url=_DATA_URL, 69 | citation=_CITATION, 70 | url=CPP_HOME, 71 | ), 72 | ] 73 | 74 | def _info(self): 75 | return datasets.DatasetInfo( 76 | description=self.config.description, 77 | features=datasets.Features( 78 | { 79 | "text": datasets.Value("string"), 80 | } 81 | ), 82 | homepage=self.config.url, 83 | citation=self.config.citation, 84 | ) 85 | 86 | def _split_generators(self, dl_manager): 87 | dl_dir = dl_manager.download_and_extract(self.config.data_url) or "" 88 | dl_dir = os.path.join(dl_dir, f"source_code/{self.config.name}") 89 | return [ 90 | datasets.SplitGenerator( 91 | name=datasets.Split.TRAIN, 92 | gen_kwargs={ 93 | "filepath": os.path.join(dl_dir, f"train.txt"), 94 | }, 95 | ), 96 | datasets.SplitGenerator( 97 | name=datasets.Split.VALIDATION, 98 | gen_kwargs={ 99 | "filepath": os.path.join(dl_dir, f"valid.txt"), 100 | }, 101 | ), 102 | datasets.SplitGenerator( 103 | name=datasets.Split.TEST, 104 | gen_kwargs={ 105 | "filepath": os.path.join(dl_dir, f"test.txt"), 106 | }, 107 | ), 108 | ] 109 | 110 | def _generate_examples(self, filepath): 111 | """This function returns the examples in the raw (text) form.""" 112 | with open(filepath, 'r', encoding="utf-8") as f: 113 | for idx, row in enumerate(f): 114 | if row.strip(): 115 | yield idx, {"text": row} 116 | else: 117 | yield idx, {"text": ""} 118 | -------------------------------------------------------------------------------- /docs/wechat.jpeg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/shibing624/CodeAssist/c5885d2d20e3dbea7d653da17e2a635951f522a5/docs/wechat.jpeg -------------------------------------------------------------------------------- /examples/data/code_alpaca_20k_50.jsonl: -------------------------------------------------------------------------------- 1 | {"instruction": "Create an array of length 5 which contains all even numbers between 1 and 10.", "input": "", "output": "arr = [2, 4, 6, 8, 10]"} 2 | {"instruction": "Formulate an equation to calculate the height of a triangle given the angle, side lengths and opposite side length.", "input": "", "output": "Height of triangle = opposite side length * sin (angle) / side length"} 3 | {"instruction": "Write a replace method for a string class which replaces the given string with a given set of characters.", "input": "string = \"Hello World!\"\nreplace_with = \"Greetings!\"", "output": "def replace(self, replace_with):\n new_string = \"\"\n for char in self:\n if char == \" \":\n new_string += replace_with\n else:\n new_string += char\n return new_string"} 4 | {"instruction": "Create an array of length 15 containing numbers divisible by 3 up to 45.", "input": "", "output": "arr = [3, 6, 9, 12, 15, 18, 21, 24, 27, 30, 33, 36, 39, 42, 45]"} 5 | {"instruction": "Write a function to find the number of distinct states in a given matrix.", "input": "matrix = [[1, 0, 0],\n [1, 0, 1],\n [1, 1, 1]]", "output": "def find_num_distinct_states(matrix):\n states = set()\n for row in matrix:\n state = \"\".join([str(x) for x in row])\n states.add(state)\n return len(states)"} 6 | {"instruction": "Create a nested loop to print every combination of numbers between 0-9", "input": "", "output": "for i in range(10):\n for j in range(10):\n print(i, j)"} 7 | {"instruction": "Write a function to find the maximum difference between two numbers in a given array.", "input": "arr = [5, 3, 17, 11, 9]", "output": "def max_difference(arr):\n max_diff = 0\n for i in range(len(arr)-1):\n diff = arr[i+1] - arr[i]\n if diff > max_diff:\n max_diff = diff\n return max_diff"} 8 | {"instruction": "Write a function to generate the nth Fibonacci number.", "input": "< noinput >", "output": "def fib(n):\n if n == 0:\n return 0\n elif n == 1:\n return 1\n else:\n return fib(n-2) + fib(n-1)"} 9 | {"instruction": "Write a class to represent a 2D point with x and y coordinates.", "input": "", "output": "class Point:\n def __init__(self, x, y):\n self.x = x \n self.y = y \n \n def __str__(self): \n return \"({0},{1})\".format(self.x, self.y)"} 10 | {"instruction": "Write code that removes spaces from a given string.", "input": "string = \" A B C D \"", "output": "def remove_spaces(string): \n return \"\".join(string.split())"} 11 | {"instruction": "Write a function that checks if a give number is even or odd.", "input": "", "output": "def is_even(n):\n if n%2 == 0:\n return True \n else:\n return False"} 12 | {"instruction": "Write code to create a list of all prime numbers between 2 and 100.", "input": "", "output": "def find_primes(n): \n prime_list = [2] \n for number in range(2, n + 1): \n is_prime = True\n for k in range(2, number): \n if number % k == 0: \n is_prime = False \n if is_prime: \n prime_list.append(number) \n return prime_list"} 13 | {"instruction": "Edit the following code to print out all even numbers from 1 to 10.", "input": "for i in range(1,11):\n print(i)", "output": "for i in range(1,11):\n if i % 2 == 0:\n print(i)"} 14 | {"instruction": "Write code to find the sum of all numbers between 1 and 10.", "input": "", "output": "sum = 0\nfor i in range(1,11):\n sum += i\nprint(sum)"} 15 | {"instruction": "Create a function to calculate the area of a given circle.", "input": "", "output": "def area_circle(radius):\n return 3.14 * (radius**2)"} 16 | {"instruction": "Follow the comments in the code to complete the program which prints a conversion table from Fahrenheit to Celsius.", "input": "# This program prints a conversion table from Fahrenheit to Celsius\n\n# Print a heading for the table\n# TODO\n\n# Print the column headings\n# TODO\n\n# Calculate and print the table body\n# TODO", "output": "# This program prints a conversion table from Fahrenheit to Celsius\n\n# Print a heading for the table\nprint(\"Fahrenheit to Celsius Conversion Table\")\n\n# Print the column headings\nprint(\"Fahrenheit \\t Celsius\")\n\n# Calculate and print the table body\nfor fahr in range(-50 , 121):\n celsius = (fahr - 32) * 5 / 9\n print(\"%3d \\t\\t %.2f\" % (fahr, celsius))"} 17 | {"instruction": "Develop a classification algorithm in Python to predict whether a bird is a hawk or a falcon.", "input": "", "output": "import pandas as pd\nimport numpy as np\n\n# Define features\nfeatures = [\"wing_shape\", \"size\", \"tails_length\", \"color_pattern\"] \n\n# Load the data\ndata = pd.read_csv(\"birds.csv\")\n\n# Preprocess the data to generate feature vectors\nX = np.array(data[features]) \n\n# Make a target vector \ny = np.array(data[\"species\"]) \n\n# Divide the data into training and test sets\nfrom sklearn.model_selection import train_test_split\nX_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2, random_state=42)\n\n# Train a classification model\nfrom sklearn.tree import DecisionTreeClassifier\nmodel = DecisionTreeClassifier()\nmodel.fit(X_train, y_train)\n\n# Make predictions\ny_pred = model.predict(X_test)\n\n# Generate evaluation metrics\nfrom sklearn.metrics import accuracy_score\naccuracy = accuracy_score(y_test, y_pred)\n\nprint(\"Model accuracy: %.2f\" % accuracy)"} 18 | {"instruction": "Create a function in Java for computing the factorial of a positive integer number.", "input": "", "output": "public static long factorial(int n) {\n if (n == 0) {\n return 1;\n }\n long result = 1;\n for (int i = 1; i <= n; i++) {\n result *= i;\n }\n return result;\n}"} 19 | {"instruction": "Create a C++ program for randomly shuffling an array of integers.", "input": "", "output": "#include <iostream>\n#include <algorithm>\n#include <random>\n\nvoid shuffleArray(int array[], int size) {\n\tstd::random_device rd;\n\tstd::mt19937 g(rd()); \n\tstd::shuffle(array, array+size, g); \n}\n\nint main() \n{ \n\tint array[] = {1, 2, 3, 4, 5, 6, 7, 8}; \n\tint size = sizeof(array) / sizeof(array[0]); \n\n // Print original array\n std::cout << \"Original array: \";\n for (int i : array)\n std::cout << i << \" \";\n std::cout << \"\\n\";\n\n // Shuffle array\n\tshuffleArray(array, size); \n\n // Print shuffled array\n std::cout << \"Shuffled array: \";\n for (int i : array)\n std::cout << i << \" \";\n std::cout << \"\\n\";\n\n\treturn 0; \n}"} 20 | {"instruction": "Use the function to debug the given program and prevent the segmentation fault.", "input": "#include <stdio.h>\n\nint main() {\n int array[] = {1, 2, 3, 4, 5};\n int n = 5;\n int i;\n // TODO\n for (i = 0; i <= n; i++) {\n printf(\"%d\\n\", array[i]);\n }\n return 0;\n}", "output": "#include <stdio.h>\n\nint main() {\n int array[] = {1, 2, 3, 4, 5};\n int n = 5;\n int i;\n // Limit loop size to the length of the array\n for (i = 0; i < n; i++) {\n printf(\"%d\\n\", array[i]);\n }\n return 0;\n}"} 21 | {"instruction": "Replace the placeholder with a for loop for iterating over the numbers from 1 to n and printing them.", "input": "for i in range(n):\n # TODO", "output": "for i in range(1, n+1):\n print(i)"} 22 | {"instruction": "Use try-except statements to detect and handle exceptions in the code.", "input": "for i in range(n):\n x = 1/i", "output": "try:\n for i in range(n):\n x = 1/i\nexcept ZeroDivisionError:\n print(\"Zero division error encountered!\")"} 23 | {"instruction": "Re-write the code segment as a list comprehension.", "input": "numbers = []\nfor i in range(10):\n if i % 2 == 0:\n numbers.append(i)", "output": "numbers = [i for i in range(10) if i % 2 == 0]"} 24 | {"instruction": "Remove the boilerplate code from the program for printing 10 random numbers.", "input": "import random\n\n# Generate 10 random numbers\nfor i in range(10):\n # TODO", "output": "import random\n\n# Generate 10 random numbers\nfor i in range(10):\n print(random.randrange(1, 101))"} 25 | {"instruction": "Modify the code to perform the mathematical expression “x to the power of y”.", "input": "# Calculate x to the power of y\n# TODO", "output": "# Calculate x to the power of y\nx_pow_y = x**y"} 26 | {"instruction": "Create a dictionary for counting the occurrence of each letter in the given string.", "input": "string = \"hello world\"\n#TODO", "output": "string = \"hello world\"\nletter_counts = {}\n\nfor letter in string:\n if letter in letter_counts:\n letter_counts[letter] += 1\n else:\n letter_counts[letter] = 1\n\nprint(letter_counts)"} 27 | {"instruction": "Follow the comments to complete the code that finds the maximum value in an array.", "input": "# Find the maximum value in an array\narr = [4, 5, 2, 8, 7]\n# TODO", "output": "# Find the maximum value in an array\narr = [4, 5, 2, 8, 7]\n\nmax_value = arr[0]\nfor i in range(1, len(arr)):\n if arr[i] > max_value:\n max_value = arr[i]\n\nprint(max_value)"} 28 | {"instruction": "Write a function in HTML for creating a table of n rows and m columns.", "input": "", "output": "<table>\n <tbody>\n <% for (let i = 0; i < n; i++) { %>\n <tr>\n <% for (let j = 0; j < m; j++) { %>\n <td>Cell Content</td>\n <% } %>\n </tr>\n <% } %>\n </tbody>\n</table>"} 29 | {"instruction": "Create a CSS script that centers the login form on the webpage.", "input": "", "output": ".login-form {\n position: absolute;\n top: 50%;\n left: 50%;\n transform: translate(-50%, -50%);\n}"} 30 | {"instruction": "Create a Java program that prints out the sum of the numbers in the array.", "input": "int[] numbers = {1, 2, 3 , 4, 5};", "output": "public class SumOfArray {\n public static void main(String[] args) {\n int[] numbers = {1, 2, 3 , 4, 5};\n int sum = 0;\n for(int i=0; i < numbers.length;i++) {\n sum = sum + numbers[i];\n }\n System.out.println(\"The sum of numbers in the given array is: \" + sum);\n }\n}"} 31 | {"instruction": "Write a SQL query to calculate the total number of orders in the \"orders\" table.", "input": "", "output": "SELECT COUNT(*) AS \"Total Number of Orders\" \nFROM orders;"} 32 | {"instruction": "Write a Python function to convert a given string to lowercase.", "input": "string = \"FOO\"", "output": "def to_lowercase(string):\n return string.lower()\n\nstring = \"FOO\"\nprint(to_lowercase(string))"} 33 | {"instruction": "Create a C program that prints out the Fibonacci sequence.", "input": "", "output": "#include <stdio.h>\n\nint main()\n{\n int i, num, first=0, second=1, next;\n printf(\"Enter the number of elements: \");\n scanf(\"%d\", &num);\n printf(\"Fibonacci Series: \");\n for (i = 0; i < num; i++)\n {\n if (i <= 1)\n next = i;\n else\n {\n next = first + second;\n first = second;\n second = next;\n }\n printf(\"%d \", next);\n }\n return 0;\n}"} 34 | {"instruction": "Write a function in JavaScript to count the number of vowels in a given string.", "input": "const string = \"abcdefg\"", "output": "function countVowels(string) {\n let vowelCount = 0;\n const vowels = [\"a\",\"e\",\"i\",\"o\",\"u\"];\n \n for (let i = 0; i < string.length; i++) {\n if (vowels.includes(string[i])) {\n vowelCount++;\n }\n }\n \n return vowelCount;\n}\n\nconst string = \"abcdefg\";\nconsole.log(countVowels(string));"} 35 | {"instruction": "Create a JSX code that renders a \"Hello, World!\" message in the browser.", "input": "", "output": "import React from \"react\";\n\nfunction App() {\n return (\n <div>\n <h1>Hello, World!</h1>\n </div>\n );\n}\n\nexport default App;"} 36 | {"instruction": "Write an SQL query to calculate the average order price for each product category in the \"orders\" table.", "input": "", "output": "SELECT category, AVG(price) as \"Average Order Price\"\nFROM orders\nGROUP BY category;"} 37 | {"instruction": "Create a Java program that prints out the contents of a given two-dimensional array.", "input": "int[][] arr = {{1,2},{3,4},{5,6}}", "output": "public class TwoDArray {\n\n public static void main(String[] args) {\n int[][] arr = {{1,2},{3,4},{5,6}};\n\n for(int i=0;i<arr.length;i++){\n for(int j=0;j<arr[i].length;j++){\n System.out.print(arr[i][j]+\" \");\n }\n System.out.println();\n }\n }\n}"} 38 | {"instruction": "Create a C++ program that sorts a given array in descending order.", "input": "int array[] = {6, 3, 8, 12, 9, 4};", "output": "#include<iostream>\nusing namespace std;\n \nint main()\n{\n int array[] = {6, 3, 8, 12, 9, 4};\n int arraySize = sizeof(array)/sizeof(array[0]);\n \n //ascending order sorting\n for(int i = 0; i < arraySize; i++)\n {\n // find the position of smallest element\n int small = i;\n for(int j = i+1; j < arraySize; j++)\n {\n if(array[j] > array[small])\n small = j;\n }\n \n // swap the position of smallest element\n int temp = array[i];\n array[i] = array[small];\n array[small] = temp;\n }\n \n // print the sorted array\n cout << \"Sorted array : \";\n for(int i = 0; i < arraySize; i++)\n cout << array[i] << \" \";\n \n return 0;\n}"} 39 | {"instruction": "Edit the Python code so that it prints out the total number of elements in the list.", "input": "list = [1, 2, 3, 4, 5]", "output": "list = [1, 2, 3, 4, 5]\nprint(\"Total number of elements in the list: \", len(list))"} 40 | {"instruction": "Create a Swift code that show the number of days in the given month.", "input": "int month = 5", "output": "let month = 5\nlet daysEndingInth = [3, 0, 3, 2, 3, 2, 3, 3, 2, 3, 2, 3]\nlet daysInMonth = daysEndingInth[month - 1] \nprint(\"The number of days in the month of \\(month) is \\(daysInMonth).\")"} 41 | {"instruction": "Write a HTML code that create a form for user to fill out their current address.", "input": "", "output": "<form action=\"\">\n <label for=\"street\">Street</label>\n <input type=\"text\" name=\"street\" id=\"street\">\n <label for=\"city\">City</label>\n <input type=\"text\" name=\"city\" id=\"city\">\n <label for=\"state\">State</label>\n <input type=\"text\" name=\"state\" id=\"state\">\n <label for=\"zip\">Zip Code</label>\n <input type=\"text\" name=\"zip\" id=\"zip\"> \n <input type=\"submit\" value=\"Submit\">\n</form>"} 42 | {"instruction": "Write a SQL query to find the total number of orders placed between two given dates.", "input": "start_date = '2020-01-01'\nend_date = '2020-02-01'", "output": "SELECT COUNT(*)\nFROM orders\nWHERE order_date BETWEEN '2020-01-01' AND '2020-02-01';"} 43 | {"instruction": "Create a Ruby code that prints out the median of a given array of numbers.", "input": "nums = [1, 7, 3, 6]", "output": "nums = [1, 7, 3, 6]\nsorted_nums = nums.sort\n\n# Calculation of median \nif sorted_nums.length % 2 == 0\n mid_index = sorted_nums.length/2\n median = (sorted_nums[mid_index] + sorted_nums[mid_index - 1])/2.to_f\nelse\n mid_index = sorted_nums.length/2\n median = sorted_nums[mid_index]\nend\n\nputs median"} 44 | {"instruction": "Create a PHP code to check if an array contains a given item.", "input": "$input_array = array(1,2,3,4,5);\n$item_to_check = 3;", "output": "<?php\n$input_array = array(1,2,3,4,5);\n$item_to_check = 3;\n\nif(in_array($item_to_check, $input_array)) {\n echo \"Array contains the given item.\";\n} else {\n echo \"Array doesn't contain the given item.\";\n}\n?>"} 45 | {"instruction": "Edit the code to create a try-except block to handle the exception.", "input": "a = 5\nb = 0\n\nans = a/b", "output": "try:\n a = 5\n b = 0\n\n ans = a/b\nexcept ZeroDivisionError:\n print(\"Cannot divide by zero.\")"} 46 | {"instruction": "Write a script using Bash that replaces all vowels with a number.", "input": "", "output": "#!/bin/bash\n\nstring=\"Hello World!\"\nvowels=(\"a\" \"e\" \"i\" \"o\" \"u\" \"A\" \"E\" \"I\" \"O\" \"U\")\n\nfor i in \"${vowels[@]}\"\ndo\n string=\"${string//$i/0}\"\ndone\n\necho $string #H0ll0 W0rld!"} 47 | {"instruction": "Change the variable `x` to the data type desired.", "input": "x = 'Hello World!'", "output": "x = \"Hello World!\""} 48 | {"instruction": "Create a function that takes an array as an argument and returns the sum of all the elements in the array.", "input": "arr = [1, 2, 3, 4, 5]", "output": "def arraySum(arr):\n total = 0\n for num in arr:\n total += num\n return total"} 49 | {"instruction": "Sort the array in ascending order.", "input": "arr = [3, 2, 1, 5, 4]", "output": "arr.sort() # [1, 2, 3, 4, 5]"} 50 | {"instruction": "Generate a random integer between 4 and 8 (inclusively).", "input": "", "output": "import random\nx = random.randint(4, 8)"} 51 | -------------------------------------------------------------------------------- /examples/distilgpt2_demo.py: -------------------------------------------------------------------------------- 1 | # -*- coding: utf-8 -*- 2 | """ 3 | @author:XuMing(xuming624@qq.com) 4 | @description: 5 | """ 6 | import sys 7 | 8 | sys.path.append('..') 9 | from codeassist import GPT2Coder 10 | 11 | m = GPT2Coder("shibing624/code-autocomplete-distilgpt2-python") 12 | print(m.generate('def load_csv_file(file_path):')[0]) 13 | -------------------------------------------------------------------------------- /examples/gpt2_demo.py: -------------------------------------------------------------------------------- 1 | # -*- coding: utf-8 -*- 2 | """ 3 | @author:XuMing(xuming624@qq.com) 4 | @description: 5 | """ 6 | import sys 7 | 8 | sys.path.append('..') 9 | from codeassist.gpt2_coder import GPT2Coder 10 | 11 | m = GPT2Coder("shibing624/code-autocomplete-gpt2-base") 12 | print(m.generate('def load_csv_file(file_path):')[0]) 13 | -------------------------------------------------------------------------------- /examples/gradio_demo.py: -------------------------------------------------------------------------------- 1 | """ 2 | @author:XuMing(xuming624@qq.com) 3 | @description: 4 | """ 5 | 6 | import sys 7 | 8 | import gradio as gr 9 | 10 | sys.path.append('..') 11 | from codeassist.gpt2_coder import GPT2Coder 12 | 13 | model = GPT2Coder("shibing624/code-autocomplete-gpt2-base") 14 | 15 | 16 | def ai_text(text): 17 | gen_text = model.generate(text)[0] 18 | print(text, ' => ', gen_text) 19 | return gen_text 20 | 21 | 22 | if __name__ == '__main__': 23 | print(ai_text('import torch.nn as')) 24 | 25 | examples = [ 26 | ['def load_csv_file(file_path):'], 27 | ['import torch.nn as'], 28 | ['parser.add_argument("--num_train_epochs",'], 29 | ['torch.device('], 30 | ['def set_seed('], 31 | ] 32 | 33 | output_text = gr.outputs.Textbox() 34 | gr.Interface(ai_text, "textbox", output_text, 35 | # theme="grass", 36 | title="Code Autocomplete Model shibing624/code-autocomplete-gpt2-base", 37 | description="Copy or input python code here. Submit and the machine will generate code.", 38 | article="Link to <a href='https://github.com/shibing624/code-autocomplete' style='color:blue;' target='_blank\'>Github REPO</a>", 39 | examples=examples).launch() 40 | -------------------------------------------------------------------------------- /examples/inference_demo.py: -------------------------------------------------------------------------------- 1 | # -*- coding: utf-8 -*- 2 | """ 3 | @author:XuMing(xuming624@qq.com) 4 | @description: 5 | """ 6 | import argparse 7 | import sys 8 | 9 | sys.path.append('..') 10 | from codeassist import GPT2Coder, WizardCoder 11 | 12 | if __name__ == '__main__': 13 | parser = argparse.ArgumentParser() 14 | parser.add_argument('--model_type', type=str, default="wizard", help='wizard or gpt2') 15 | parser.add_argument('--model_name', type=str, default="WizardLM/WizardCoder-15B-V1.0", help='model name or path') 16 | parser.add_argument('--max_length', type=int, default=128, help='maximum length for code generation') 17 | parser.add_argument('--temperature', type=float, default=1.0, help='temperature for sampling-based code generation') 18 | args = parser.parse_args() 19 | print(args) 20 | if args.model_type == 'wizard': 21 | model = WizardCoder(args.model_name) 22 | else: 23 | model = GPT2Coder(args.model_name) 24 | 25 | # generate code 26 | while True: 27 | print(f'Enter the context code (exit or python code)') 28 | context = input(">>> ") 29 | if context == "exit": 30 | break 31 | generated_codes = model.generate(context, temperature=args.temperature, max_length=args.max_length) 32 | print("Generated code:") 33 | for i, code in enumerate(generated_codes): 34 | print("{}:\n {}".format(i + 1, code)) 35 | print("=" * 20) 36 | -------------------------------------------------------------------------------- /examples/original_gpt2_demo.py: -------------------------------------------------------------------------------- 1 | # -*- coding: utf-8 -*- 2 | """ 3 | @author:XuMing(xuming624@qq.com) 4 | @description: 5 | """ 6 | 7 | import sys 8 | 9 | sys.path.append('..') 10 | from codeassist.gpt2_coder import GPT2Coder 11 | 12 | if __name__ == '__main__': 13 | prompts = [ 14 | "def load_csv_file(file_path):", 15 | "import numpy as np", 16 | "import torch.nn as", 17 | 'parser.add_argument("--num_train_epochs",', 18 | "def set_seed(", 19 | "def factorial", 20 | ] 21 | 22 | m = GPT2Coder("gpt2") 23 | for i in m.generate('import torch.nn as', num_return_sequences=3): 24 | print(i) 25 | 26 | for prompt in prompts: 27 | decoded = m.generate(prompt, num_return_sequences=1) 28 | print("Input :", prompt) 29 | print("Output:", decoded[0]) 30 | print("=" * 20) 31 | -------------------------------------------------------------------------------- /examples/prepare_code_data.py: -------------------------------------------------------------------------------- 1 | # -*- coding: utf-8 -*- 2 | """ 3 | @author:XuMing(xuming624@qq.com) 4 | @description: 5 | """ 6 | import argparse 7 | import sys 8 | 9 | from sklearn.model_selection import train_test_split 10 | 11 | sys.path.append("..") 12 | from codeassist import create_dataset 13 | 14 | 15 | def main(): 16 | parser = argparse.ArgumentParser() 17 | parser.add_argument("--save_dir", type=str, default="download", help="Save dataset directory") 18 | parser.add_argument("--num_repos", type=int, default=3, help="Number of repos to use") 19 | parser.add_argument("--code", default="python", const='python', nargs='?', 20 | choices=['python', 'java', 'cpp'], help="Download code language source code dataset") 21 | args = parser.parse_args() 22 | print(args) 23 | sources = dict() 24 | try: 25 | sources = create_dataset.get_source_code_by_language(code_languages=args.code, 26 | save_dir=args.save_dir, 27 | each_limit_repos=args.num_repos 28 | ) 29 | except KeyboardInterrupt: 30 | pass 31 | X = sources[f"{args.code}"] 32 | X_train, X_test = train_test_split(X, test_size=0.2, random_state=1) 33 | X_train, X_val = train_test_split(X_train, test_size=0.25, random_state=1) # 0.25 x 0.8 = 0.2 34 | train_file = f'{args.save_dir}/{args.code}/train.txt' 35 | valid_file = f'{args.save_dir}/{args.code}/valid.txt' 36 | test_file = f'{args.save_dir}/{args.code}/test.txt' 37 | create_dataset.merge_and_save(X_train, train_file) 38 | create_dataset.merge_and_save(X_val, valid_file) 39 | create_dataset.merge_and_save(X_test, test_file) 40 | print(f'Save train file: {train_file}, valid file: {valid_file}, test file: {test_file}') 41 | 42 | 43 | if __name__ == '__main__': 44 | main() 45 | -------------------------------------------------------------------------------- /examples/server.py: -------------------------------------------------------------------------------- 1 | """ 2 | @author:XuMing(xuming624@qq.com) 3 | @description: pip install fastapi uvicorn 4 | """ 5 | import argparse 6 | import os 7 | import sys 8 | 9 | import torch 10 | import uvicorn 11 | from fastapi import FastAPI, Query 12 | from loguru import logger 13 | from starlette.middleware.cors import CORSMiddleware 14 | 15 | sys.path.append('..') 16 | from codeassist import GPT2Coder 17 | 18 | pwd_path = os.path.abspath(os.path.dirname(__file__)) 19 | use_cuda = torch.cuda.is_available() 20 | # Use finetuned GPT model 21 | parser = argparse.ArgumentParser() 22 | parser.add_argument("--model_name_or_path", type=str, default="shibing624/code-autocomplete-gpt2-base", 23 | help="Model save dir or model name") 24 | args = parser.parse_args() 25 | model = GPT2Coder(args.model_name_or_path) 26 | 27 | # define the app 28 | app = FastAPI() 29 | app.add_middleware( 30 | CORSMiddleware, 31 | allow_origins=["*"], 32 | allow_credentials=True, 33 | allow_methods=["*"], 34 | allow_headers=["*"]) 35 | 36 | 37 | @app.get('/') 38 | async def index(): 39 | return {"message": "index, docs url: /docs"} 40 | 41 | 42 | @app.get('/autocomplete') 43 | async def autocomplete(q: str = Query(..., min_length=1, max_length=512, title='query')): 44 | try: 45 | # Generate text using the model. Verbose set False to prevent logging generated sequences. 46 | generated = model.generate(q) 47 | result_dict = generated[0] 48 | logger.debug(f"Successfully autocomplete, q:{q}, res:{result_dict}") 49 | return result_dict 50 | except Exception as e: 51 | logger.error(e) 52 | return {'status': False, 'msg': e}, 400 53 | 54 | 55 | if __name__ == '__main__': 56 | uvicorn.run(app=app, host='0.0.0.0', port=8001) 57 | -------------------------------------------------------------------------------- /examples/training_gpt2_mydata.py: -------------------------------------------------------------------------------- 1 | # -*- coding: utf-8 -*- 2 | """ 3 | @author:XuMing(xuming624@qq.com) 4 | @description: 5 | """ 6 | import argparse 7 | import sys 8 | 9 | sys.path.append('..') 10 | from codeassist import GPT2Coder 11 | 12 | if __name__ == '__main__': 13 | parser = argparse.ArgumentParser() 14 | parser.add_argument("--model_name", type=str, default="gpt2", help="Model arch, gpt2, gpt2-medium or distilgpt2") 15 | parser.add_argument("--train_file", type=str, default="data/train_code_5k.txt", help="Train file path") 16 | parser.add_argument("--valid_file", type=str, default="data/train_code_5k.txt", help="Valid file path") 17 | parser.add_argument("--do_train", action="store_true", help="Whether to run training.") 18 | parser.add_argument("--do_predict", action="store_true", help="Whether to run predict.") 19 | parser.add_argument("--output_dir", type=str, default="./outputs-fine-tuned/", help="Output dir") 20 | parser.add_argument("--num_epochs", type=int, default=5, help="Num of training epochs") 21 | parser.add_argument("--batch_size", type=int, default=8, help="Batch size") 22 | args = parser.parse_args() 23 | print(args) 24 | 25 | if args.do_train: 26 | model = GPT2Coder(model_name_or_path=args.model_name) 27 | model.train_model( 28 | args.train_file, 29 | args.output_dir, 30 | eval_file=args.valid_file, 31 | num_epochs=args.num_epochs, 32 | batch_size=args.batch_size 33 | ) 34 | print(f"model saved to {args.output_dir}") 35 | if args.do_predict: 36 | model = GPT2Coder(model_name_or_path=args.output_dir) 37 | prompts = [ 38 | "def load_csv_file(file_path):", 39 | "import numpy as np", 40 | "import torch.nn as", 41 | 'parser.add_argument("--num_train_epochs",', 42 | "def set_seed(", 43 | "def factorial", 44 | ] 45 | for prompt in prompts: 46 | outputs = model.generate(prompt) 47 | print("Input :", prompt) 48 | print("Output:", outputs[0]) 49 | print("=" * 20) 50 | -------------------------------------------------------------------------------- /examples/training_wizardcoder_mydata.py: -------------------------------------------------------------------------------- 1 | # -*- coding: utf-8 -*- 2 | """ 3 | @author:XuMing(xuming624@qq.com) 4 | @description: 5 | """ 6 | import argparse 7 | import sys 8 | 9 | sys.path.append('..') 10 | from codeassist import WizardCoder 11 | 12 | if __name__ == '__main__': 13 | parser = argparse.ArgumentParser() 14 | parser.add_argument("--model_name", type=str, default="WizardLM/WizardCoder-15B-V1.0", 15 | help="Model arch, gpt2, gpt2-medium, distilgpt2 or WizardLM/WizardCoder-15B-V1.0") 16 | parser.add_argument("--train_file", type=str, default="data/code_alpaca_20k_50.jsonl", help="Train file path") 17 | parser.add_argument("--valid_file", type=str, default="data/code_alpaca_20k_50.jsonl", help="Valid file path") 18 | parser.add_argument("--do_train", action="store_true", help="Whether to run training.") 19 | parser.add_argument("--do_predict", action="store_true", help="Whether to run predict.") 20 | parser.add_argument("--output_dir", type=str, default="./outputs-finetuned-wizardcoder/", help="output dir") 21 | parser.add_argument("--num_epochs", type=int, default=5, help="Num of training epochs") 22 | parser.add_argument("--batch_size", type=int, default=2, help="Batch size") 23 | args = parser.parse_args() 24 | print(args) 25 | 26 | if args.do_train: 27 | model = WizardCoder(model_name_or_path=args.model_name) 28 | model.train_model( 29 | args.train_file, 30 | args.output_dir, 31 | eval_file=args.valid_file, 32 | num_epochs=args.num_epochs, 33 | batch_size=args.batch_size 34 | ) 35 | print(f"model saved to {args.output_dir}") 36 | if args.do_predict: 37 | model = WizardCoder(model_name_or_path=args.model_name, peft_name=args.output_dir) 38 | prompts = [ 39 | "def load_csv_file(file_path):", 40 | "write a C++ code to sum 1 to 12.", 41 | "写个python的快排算法", 42 | "生成4到400之间的随机数，用java和python写代码", 43 | ] 44 | for prompt in prompts: 45 | outputs = model.generate(prompt) 46 | print("Input :", prompt) 47 | print("Output:", outputs[0]) 48 | print("=" * 20) 49 | -------------------------------------------------------------------------------- /examples/use_transformers_gpt2.py: -------------------------------------------------------------------------------- 1 | # -*- coding: utf-8 -*- 2 | """ 3 | @author:XuMing(xuming624@qq.com) 4 | @description: 5 | """ 6 | 7 | import os 8 | 9 | import torch 10 | from transformers import GPT2Tokenizer, GPT2LMHeadModel 11 | 12 | os.environ["KMP_DUPLICATE_LIB_OK"] = "TRUE" 13 | device = torch.device("cuda" if torch.cuda.is_available() else "cpu") 14 | 15 | tokenizer = GPT2Tokenizer.from_pretrained("shibing624/code-autocomplete-gpt2-base") 16 | model = GPT2LMHeadModel.from_pretrained("shibing624/code-autocomplete-gpt2-base") 17 | model.to(device) 18 | prompts = [ 19 | "def load_csv_file(file_path):", 20 | "import numpy as np", 21 | "import torch.nn as", 22 | 'parser.add_argument("--num_train_epochs",', 23 | "def set_seed(", 24 | "def factorial", 25 | ] 26 | for prompt in prompts: 27 | input_ids = tokenizer(prompt, return_tensors='pt').to(device).input_ids 28 | outputs = model.generate( 29 | input_ids=input_ids, 30 | max_length=64 + len(input_ids[0]), 31 | temperature=1.0, 32 | top_k=50, 33 | top_p=0.95, 34 | repetition_penalty=1.0, 35 | do_sample=True, 36 | num_return_sequences=1, 37 | length_penalty=2.0, 38 | early_stopping=True, 39 | pad_token_id=tokenizer.eos_token_id, 40 | eos_token_id=tokenizer.eos_token_id, 41 | ) 42 | decoded = tokenizer.decode(outputs[0], skip_special_tokens=True) 43 | print("Input :", prompt) 44 | print("Output:", decoded) 45 | print("=" * 20) 46 | -------------------------------------------------------------------------------- /examples/wizardcoder_demo.py: -------------------------------------------------------------------------------- 1 | # -*- coding: utf-8 -*- 2 | """ 3 | @author:XuMing(xuming624@qq.com) 4 | @description: 5 | """ 6 | import sys 7 | 8 | sys.path.append('..') 9 | from codeassist import WizardCoder 10 | 11 | m = WizardCoder("WizardLM/WizardCoder-15B-V1.0") 12 | print(m.generate('def load_csv_file(file_path):')[0]) 13 | 14 | prompts = [ 15 | "write a function to load csv file and return a dataframe", 16 | "give me odd numbers from 1 to 10", 17 | "write a function to show fibonacci sequence and return a list", 18 | "write a Java code to sum 1 to 10.", 19 | "write a python code to sum 1 to 10.", 20 | "write a C++ code to sum 1 to 12.", 21 | "写个python的快排算法", 22 | "生成4到400之间的随机数，用java和python写代码", 23 | "写java代码，从1累加到100", 24 | "写python代码，从1累加到10", 25 | "写个斐波拉契数列，返回一个列表，python代码", 26 | "给出所有1-20的偶数列表，C++代码", 27 | "给出所有1-20的奇数列表，python代码", 28 | # below is for language model chat 29 | "tell me about beijing", 30 | "give me a plan to NewYork city for three days trip", 31 | "详细介绍下南京", 32 | "列个南京的详细三天旅游计划", 33 | "失眠怎么办", 34 | ] 35 | for prompt in prompts: 36 | print('input :', prompt) 37 | print('output:', m.generate(prompt)[0]) 38 | print() 39 | -------------------------------------------------------------------------------- /requirements.txt: -------------------------------------------------------------------------------- 1 | loguru 2 | transformers 3 | datasets 4 | pandas 5 | tqdm -------------------------------------------------------------------------------- /setup.cfg: -------------------------------------------------------------------------------- 1 | [tool:pytest] 2 | python_functions=test_ 3 | 4 | codestyle_max_line_length = 119 5 | 6 | log_cli = true 7 | log_cli_level = WARNING 8 | 9 | [metadata] 10 | description-file = README.md 11 | license_file = LICENSE 12 | 13 | [pycodestyle] 14 | max-line-length = 119 15 | 16 | [flake8] 17 | max-line-length = 119 18 | ignore = E203 , W503, F401 19 | -------------------------------------------------------------------------------- /setup.py: -------------------------------------------------------------------------------- 1 | # -*- coding: utf-8 -*- 2 | from setuptools import setup, find_packages 3 | 4 | __version__ = "1.0.0" 5 | 6 | with open('README.md', 'r', encoding='utf-8') as f: 7 | readme = f.read() 8 | 9 | setup( 10 | name='codeassist', 11 | version=__version__, 12 | description='Code AutoComplete', 13 | long_description=readme, 14 | long_description_content_type='text/markdown', 15 | author='XuMing', 16 | author_email='xuming624@qq.com', 17 | url='https://github.com/shibing624/autocoder', 18 | license='Apache License 2.0', 19 | zip_safe=False, 20 | python_requires='>=3.5', 21 | classifiers=[ 22 | 'Intended Audience :: Developers', 23 | 'Operating System :: OS Independent', 24 | 'Natural Language :: Chinese (Simplified)', 25 | 'Natural Language :: Chinese (Traditional)', 26 | 'Programming Language :: Python', 27 | 'Programming Language :: Python :: 3', 28 | 'Topic :: Text Processing', 29 | 'Topic :: Text Processing :: Indexing', 30 | 'Topic :: Text Processing :: Linguistic', 31 | ], 32 | keywords='CodeGenie,autocomplete,code-autocomplete', 33 | install_requires=[ 34 | "loguru", 35 | "transformers", 36 | "pandas", 37 | "datasets", 38 | "tqdm", 39 | ], 40 | packages=find_packages(exclude=['tests']), 41 | package_dir={'autocoder': 'autocoder'}, 42 | package_data={'autocoder': ['*.*']} 43 | ) 44 | -------------------------------------------------------------------------------- /tests/test.txt: -------------------------------------------------------------------------------- 1 | # PROJECT: pytorch_audio FILE: torchaudio/functional/functional.py 2 | # -*- coding: utf-8 -*- 3 | 4 | from collections.abc import Sequence 5 | import io 6 | import math 7 | import warnings 8 | from typing import Optional, Tuple 9 | 10 | import torch 11 | from torch import Tensor 12 | from torchaudio._internal import module_utils as _mod_utils 13 | import torchaudio 14 | 15 | __all__ = [ 16 | "spectrogram", 17 | "inverse_spectrogram", 18 | "griffinlim", 19 | "amplitude_to_DB", 20 | "DB_to_amplitude", 21 | "compute_deltas", 22 | "compute_kaldi_pitch", 23 | "melscale_fbanks", 24 | "linear_fbanks", 25 | "create_dct", 26 | "compute_deltas", 27 | "detect_pitch_frequency", 28 | "DB_to_amplitude", 29 | "mu_law_encoding", 30 | "mu_law_decoding", 31 | "phase_vocoder", 32 | 'mask_along_axis', 33 | 'mask_along_axis_iid', 34 | 'sliding_window_cmn', 35 | "spectral_centroid", 36 | "apply_codec", 37 | "resample", 38 | "edit_distance", 39 | "pitch_shift", 40 | "rnnt_loss", 41 | ] 42 | 43 | 44 | def spectrogram( 45 | waveform: Tensor, 46 | pad: int, 47 | window: Tensor, 48 | n_fft: int, 49 | hop_length: int, 50 | win_length: int, 51 | power: Optional[float], 52 | normalized: bool, 53 | center: bool = True, 54 | pad_mode: str = "reflect", 55 | onesided: bool = True, 56 | return_complex: Optional[bool] = None, 57 | ) -> Tensor: 58 | r"""Create a spectrogram or a batch of spectrograms from a raw audio signal. 59 | The spectrogram can be either magnitude-only or complex. 60 | 61 | Args: 62 | waveform (Tensor): Tensor of audio of dimension `(..., time)` 63 | pad (int): Two sided padding of signal 64 | window (Tensor): Window tensor that is applied/multiplied to each frame/window 65 | n_fft (int): Size of FFT 66 | hop_length (int): Length of hop between STFT windows 67 | win_length (int): Window size 68 | power (float or None): Exponent for the magnitude spectrogram, 69 | (must be > 0) e.g., 1 for energy, 2 for power, etc. 70 | If None, then the complex spectrum is returned instead. 71 | normalized (bool): Whether to normalize by magnitude after stft 72 | center (bool, optional): whether to pad :attr:`waveform` on both sides so 73 | that the :math:`t`-th frame is centered at time :math:`t \times \text{hop\_length}`. 74 | Default: ``True`` 75 | pad_mode (string, optional): controls the padding method used when 76 | :attr:`center` is ``True``. Default: ``"reflect"`` 77 | onesided (bool, optional): controls whether to return half of results to 78 | avoid redundancy. Default: ``True`` 79 | return_complex (bool, optional): 80 | Deprecated and not used. 81 | 82 | Returns: 83 | Tensor: Dimension `(..., freq, time)`, freq is 84 | ``n_fft // 2 + 1`` and ``n_fft`` is the number of 85 | Fourier bins, and time is the number of window hops (n_frame). 86 | """ 87 | if return_complex is not None: 88 | warnings.warn( 89 | "`return_complex` argument is now deprecated and is not effective." 90 | "`torchaudio.functional.spectrogram(power=None)` always returns a tensor with " 91 | "complex dtype. Please remove the argument in the function call." 92 | ) 93 | 94 | if pad > 0: 95 | # TODO add "with torch.no_grad():" back when JIT supports it 96 | waveform = torch.nn.functional.pad(waveform, (pad, pad), "constant") 97 | 98 | # pack batch 99 | shape = waveform.size() 100 | waveform = waveform.reshape(-1, shape[-1]) 101 | 102 | # default values are consistent with librosa.core.spectrum._spectrogram 103 | spec_f = torch.stft( 104 | input=waveform, 105 | n_fft=n_fft, 106 | hop_length=hop_length, 107 | win_length=win_length, 108 | window=window, 109 | center=center, 110 | pad_mode=pad_mode, 111 | normalized=False, 112 | onesided=onesided, 113 | return_complex=True, 114 | ) 115 | 116 | # unpack batch 117 | spec_f = spec_f.reshape(shape[:-1] + spec_f.shape[-2:]) 118 | 119 | if normalized: 120 | spec_f /= window.pow(2.).sum().sqrt() 121 | if power is not None: 122 | if power == 1.0: 123 | return spec_f.abs() 124 | return spec_f.abs().pow(power) 125 | return spec_f 126 | 127 | 128 | def inverse_spectrogram( 129 | spectrogram: Tensor, 130 | length: Optional[int], 131 | pad: int, 132 | window: Tensor, 133 | n_fft: int, 134 | hop_length: int, 135 | win_length: int, 136 | normalized: bool, 137 | center: bool = True, 138 | pad_mode: str = "reflect", 139 | onesided: bool = True, 140 | ) -> Tensor: 141 | r"""Create an inverse spectrogram or a batch of inverse spectrograms from the provided 142 | complex-valued spectrogram. 143 | 144 | Args: 145 | spectrogram (Tensor): Complex tensor of audio of dimension (..., freq, time). 146 | length (int or None): The output length of the waveform. 147 | pad (int): Two sided padding of signal. It is only effective when ``length`` is provided. 148 | window (Tensor): Window tensor that is applied/multiplied to each frame/window 149 | n_fft (int): Size of FFT 150 | hop_length (int): Length of hop between STFT windows 151 | win_length (int): Window size 152 | normalized (bool): Whether the stft output was normalized by magnitude 153 | center (bool, optional): whether the waveform was padded on both sides so 154 | that the :math:`t`-th frame is centered at time :math:`t \times \text{hop\_length}`. 155 | Default: ``True`` 156 | pad_mode (string, optional): controls the padding method used when 157 | :attr:`center` is ``True``. This parameter is provided for compatibility with the 158 | spectrogram function and is not used. Default: ``"reflect"`` 159 | onesided (bool, optional): controls whether spectrogram was done in onesided mode. 160 | Default: ``True`` 161 | 162 | Returns: 163 | Tensor: Dimension `(..., time)`. Least squares estimation of the original signal. 164 | """ 165 | 166 | if not spectrogram.is_complex(): 167 | raise ValueError("Expected `spectrogram` to be complex dtype.") 168 | 169 | if normalized: 170 | spectrogram = spectrogram * window.pow(2.).sum().sqrt() 171 | 172 | # pack batch 173 | shape = spectrogram.size() 174 | spectrogram = spectrogram.reshape(-1, shape[-2], shape[-1]) 175 | 176 | # default values are consistent with librosa.core.spectrum._spectrogram 177 | waveform = torch.istft( 178 | input=spectrogram, 179 | n_fft=n_fft, 180 | hop_length=hop_length, 181 | win_length=win_length, 182 | window=window, 183 | center=center, 184 | normalized=False, 185 | onesided=onesided, 186 | length=length + 2 * pad if length is not None else None, 187 | return_complex=False, 188 | ) 189 | 190 | if length is not None and pad > 0: 191 | # remove padding from front and back 192 | waveform = waveform[:, pad:-pad] 193 | 194 | # unpack batch 195 | waveform = waveform.reshape(shape[:-2] + waveform.shape[-1:]) 196 | 197 | return waveform 198 | 199 | 200 | def _get_complex_dtype(real_dtype: torch.dtype): 201 | if real_dtype == torch.double: 202 | return torch.cdouble 203 | if real_dtype == torch.float: 204 | return torch.cfloat 205 | if real_dtype == torch.half: 206 | return torch.complex32 207 | raise ValueError(f'Unexpected dtype {real_dtype}') 208 | 209 | 210 | def griffinlim( 211 | specgram: Tensor, 212 | window: Tensor, 213 | n_fft: int, 214 | hop_length: int, 215 | win_length: int, 216 | power: float, 217 | n_iter: int, 218 | momentum: float, 219 | length: Optional[int], 220 | rand_init: bool 221 | ) -> Tensor: 222 | r"""Compute waveform from a linear scale magnitude spectrogram using the Griffin-Lim transformation. 223 | 224 | Implementation ported from 225 | *librosa* [:footcite:`brian_mcfee-proc-scipy-2015`], *A fast Griffin-Lim algorithm* [:footcite:`6701851`] 226 | and *Signal estimation from modified short-time Fourier transform* [:footcite:`1172092`]. 227 | 228 | Args: 229 | specgram (Tensor): A magnitude-only STFT spectrogram of dimension `(..., freq, frames)` 230 | where freq is ``n_fft // 2 + 1``. 231 | window (Tensor): Window tensor that is applied/multiplied to each frame/window 232 | n_fft (int): Size of FFT, creates ``n_fft // 2 + 1`` bins 233 | hop_length (int): Length of hop between STFT windows. ( 234 | Default: ``win_length // 2``) 235 | win_length (int): Window size. (Default: ``n_fft``) 236 | power (float): Exponent for the magnitude spectrogram, 237 | (must be > 0) e.g., 1 for energy, 2 for power, etc. 238 | n_iter (int): Number of iteration for phase recovery process. 239 | momentum (float): The momentum parameter for fast Griffin-Lim. 240 | Setting this to 0 recovers the original Griffin-Lim method. 241 | Values near 1 can lead to faster convergence, but above 1 may not converge. 242 | length (int or None): Array length of the expected output. 243 | rand_init (bool): Initializes phase randomly if True, to zero otherwise. 244 | 245 | Returns: 246 | Tensor: waveform of `(..., time)`, where time equals the ``length`` parameter if given. 247 | """ 248 | assert momentum < 1, 'momentum={} > 1 can be unstable'.format(momentum) 249 | assert momentum >= 0, 'momentum={} < 0'.format(momentum) 250 | 251 | # pack batch 252 | shape = specgram.size() 253 | specgram = specgram.reshape([-1] + list(shape[-2:])) 254 | 255 | specgram = specgram.pow(1 / power) 256 | 257 | # initialize the phase 258 | if rand_init: 259 | angles = torch.rand( 260 | specgram.size(), 261 | dtype=_get_complex_dtype(specgram.dtype), device=specgram.device) 262 | else: 263 | angles = torch.full( 264 | specgram.size(), 1, 265 | dtype=_get_complex_dtype(specgram.dtype), device=specgram.device) 266 | 267 | # And initialize the previous iterate to 0 268 | tprev = torch.tensor(0., dtype=specgram.dtype, device=specgram.device) 269 | for _ in range(n_iter): 270 | # Invert with our current estimate of the phases 271 | inverse = torch.istft(specgram * angles, 272 | n_fft=n_fft, 273 | hop_length=hop_length, 274 | win_length=win_length, 275 | window=window, 276 | length=length) 277 | 278 | # Rebuild the spectrogram 279 | rebuilt = torch.stft( 280 | input=inverse, 281 | n_fft=n_fft, 282 | hop_length=hop_length, 283 | win_length=win_length, 284 | window=window, 285 | center=True, 286 | pad_mode='reflect', 287 | normalized=False, 288 | onesided=True, 289 | return_complex=True, 290 | ) 291 | 292 | # Update our phase estimates 293 | angles = rebuilt 294 | if momentum: 295 | angles = angles - tprev.mul_(momentum / (1 + momentum)) 296 | angles = angles.div(angles.abs().add(1e-16)) 297 | 298 | # Store the previous iterate 299 | tprev = rebuilt 300 | 301 | # Return the final phase estimates 302 | waveform = torch.istft(specgram * angles, 303 | n_fft=n_fft, 304 | hop_length=hop_length, 305 | win_length=win_length, 306 | window=window, 307 | length=length) 308 | 309 | # unpack batch 310 | waveform = waveform.reshape(shape[:-2] + waveform.shape[-1:]) 311 | 312 | return waveform 313 | 314 | 315 | def amplitude_to_DB( 316 | x: Tensor, 317 | multiplier: float, 318 | amin: float, 319 | db_multiplier: float, 320 | top_db: Optional[float] = None 321 | ) -> Tensor: 322 | r"""Turn a spectrogram from the power/amplitude scale to the decibel scale. 323 | 324 | The output of each tensor in a batch depends on the maximum value of that tensor, 325 | and so may return different values for an audio clip split into snippets vs. a full clip. 326 | 327 | Args: 328 | 329 | x (Tensor): Input spectrogram(s) before being converted to decibel scale. Input should take 330 | the form `(..., freq, time)`. Batched inputs should include a channel dimension and 331 | have the form `(batch, channel, freq, time)`. 332 | multiplier (float): Use 10. for power and 20. for amplitude 333 | amin (float): Number to clamp ``x`` 334 | db_multiplier (float): Log10(max(reference value and amin)) 335 | top_db (float or None, optional): Minimum negative cut-off in decibels. A reasonable number 336 | is 80. (Default: ``None``) 337 | 338 | Returns: 339 | Tensor: Output tensor in decibel scale 340 | """ 341 | x_db = multiplier * torch.log10(torch.clamp(x, min=amin)) 342 | x_db -= multiplier * db_multiplier 343 | 344 | if top_db is not None: 345 | # Expand batch 346 | shape = x_db.size() 347 | packed_channels = shape[-3] if x_db.dim() > 2 else 1 348 | x_db = x_db.reshape(-1, packed_channels, shape[-2], shape[-1]) 349 | 350 | x_db = torch.max(x_db, (x_db.amax(dim=(-3, -2, -1)) - top_db).view(-1, 1, 1, 1)) 351 | 352 | # Repack batch 353 | x_db = x_db.reshape(shape) 354 | 355 | return x_db 356 | 357 | 358 | def DB_to_amplitude( 359 | x: Tensor, 360 | ref: float, 361 | power: float 362 | ) -> Tensor: 363 | r"""Turn a tensor from the decibel scale to the power/amplitude scale. 364 | 365 | Args: 366 | x (Tensor): Input tensor before being converted to power/amplitude scale. 367 | ref (float): Reference which the output will be scaled by. 368 | power (float): If power equals 1, will compute DB to power. If 0.5, will compute DB to amplitude. 369 | 370 | Returns: 371 | Tensor: Output tensor in power/amplitude scale. 372 | """ 373 | return ref * torch.pow(torch.pow(10.0, 0.1 * x), power) 374 | 375 | 376 | def _hz_to_mel(freq: float, mel_scale: str = "htk") -> float: 377 | r"""Convert Hz to Mels. 378 | 379 | Args: 380 | freqs (float): Frequencies in Hz 381 | mel_scale (str, optional): Scale to use: ``htk`` or ``slaney``. (Default: ``htk``) 382 | 383 | Returns: 384 | mels (float): Frequency in Mels 385 | """ 386 | 387 | if mel_scale not in ['slaney', 'htk']: 388 | raise ValueError('mel_scale should be one of "htk" or "slaney".') 389 | 390 | if mel_scale == "htk": 391 | return 2595.0 * math.log10(1.0 + (freq / 700.0)) 392 | 393 | # Fill in the linear part 394 | f_min = 0.0 395 | f_sp = 200.0 / 3 396 | 397 | mels = (freq - f_min) / f_sp 398 | 399 | # Fill in the log-scale part 400 | min_log_hz = 1000.0 401 | min_log_mel = (min_log_hz - f_min) / f_sp 402 | logstep = math.log(6.4) / 27.0 403 | 404 | if freq >= min_log_hz: 405 | mels = min_log_mel + math.log(freq / min_log_hz) / logstep 406 | 407 | return mels 408 | 409 | 410 | def _mel_to_hz(mels: Tensor, mel_scale: str = "htk") -> Tensor: 411 | """Convert mel bin numbers to frequencies. 412 | 413 | Args: 414 | mels (Tensor): Mel frequencies 415 | mel_scale (str, optional): Scale to use: ``htk`` or ``slaney``. (Default: ``htk``) 416 | 417 | Returns: 418 | freqs (Tensor): Mels converted in Hz 419 | """ 420 | 421 | if mel_scale not in ['slaney', 'htk']: 422 | raise ValueError('mel_scale should be one of "htk" or "slaney".') 423 | 424 | if mel_scale == "htk": 425 | return 700.0 * (10.0**(mels / 2595.0) - 1.0) 426 | 427 | # Fill in the linear scale 428 | f_min = 0.0 429 | f_sp = 200.0 / 3 430 | freqs = f_min + f_sp * mels 431 | 432 | # And now the nonlinear scale 433 | min_log_hz = 1000.0 434 | min_log_mel = (min_log_hz - f_min) / f_sp 435 | logstep = math.log(6.4) / 27.0 436 | 437 | log_t = (mels >= min_log_mel) 438 | freqs[log_t] = min_log_hz * torch.exp(logstep * (mels[log_t] - min_log_mel)) 439 | 440 | return freqs 441 | 442 | 443 | def _create_triangular_filterbank( 444 | all_freqs: Tensor, 445 | f_pts: Tensor, 446 | ) -> Tensor: 447 | """Create a triangular filter bank. 448 | 449 | Args: 450 | all_freqs (Tensor): STFT freq points of size (`n_freqs`). 451 | f_pts (Tensor): Filter mid points of size (`n_filter`). 452 | 453 | Returns: 454 | fb (Tensor): The filter bank of size (`n_freqs`, `n_filter`). 455 | """ 456 | # Adopted from Librosa 457 | # calculate the difference between each filter mid point and each stft freq point in hertz 458 | f_diff = f_pts[1:] - f_pts[:-1] # (n_filter + 1) 459 | slopes = f_pts.unsqueeze(0) - all_freqs.unsqueeze(1) # (n_freqs, n_filter + 2) 460 | # create overlapping triangles 461 | zero = torch.zeros(1) 462 | down_slopes = (-1.0 * slopes[:, :-2]) / f_diff[:-1] # (n_freqs, n_filter) 463 | up_slopes = slopes[:, 2:] / f_diff[1:] # (n_freqs, n_filter) 464 | fb = torch.max(zero, torch.min(down_slopes, up_slopes)) 465 | 466 | return fb 467 | 468 | 469 | def melscale_fbanks( 470 | n_freqs: int, 471 | f_min: float, 472 | f_max: float, 473 | n_mels: int, 474 | sample_rate: int, 475 | norm: Optional[str] = None, 476 | mel_scale: str = "htk", 477 | ) -> Tensor: 478 | r"""Create a frequency bin conversion matrix. 479 | 480 | Note: 481 | For the sake of the numerical compatibility with librosa, not all the coefficients 482 | in the resulting filter bank has magnitude of 1. 483 | 484 | .. image:: https://download.pytorch.org/torchaudio/doc-assets/mel_fbanks.png 485 | :alt: Visualization of generated filter bank 486 | 487 | Args: 488 | n_freqs (int): Number of frequencies to highlight/apply 489 | f_min (float): Minimum frequency (Hz) 490 | f_max (float): Maximum frequency (Hz) 491 | n_mels (int): Number of mel filterbanks 492 | sample_rate (int): Sample rate of the audio waveform 493 | norm (str or None, optional): If 'slaney', divide the triangular mel weights by the width of the mel band 494 | (area normalization). (Default: ``None``) 495 | mel_scale (str, optional): Scale to use: ``htk`` or ``slaney``. (Default: ``htk``) 496 | 497 | Returns: 498 | Tensor: Triangular filter banks (fb matrix) of size (``n_freqs``, ``n_mels``) 499 | meaning number of frequencies to highlight/apply to x the number of filterbanks. 500 | Each column is a filterbank so that assuming there is a matrix A of 501 | size (..., ``n_freqs``), the applied result would be 502 | ``A * melscale_fbanks(A.size(-1), ...)``. 503 | 504 | """ 505 | 506 | if norm is not None and norm != "slaney": 507 | raise ValueError("norm must be one of None or 'slaney'") 508 | 509 | # freq bins 510 | all_freqs = torch.linspace(0, sample_rate // 2, n_freqs) 511 | 512 | # calculate mel freq bins 513 | m_min = _hz_to_mel(f_min, mel_scale=mel_scale) 514 | m_max = _hz_to_mel(f_max, mel_scale=mel_scale) 515 | 516 | m_pts = torch.linspace(m_min, m_max, n_mels + 2) 517 | f_pts = _mel_to_hz(m_pts, mel_scale=mel_scale) 518 | 519 | # create filterbank 520 | fb = _create_triangular_filterbank(all_freqs, f_pts) 521 | 522 | if norm is not None and norm == "slaney": 523 | # Slaney-style mel is scaled to be approx constant energy per channel 524 | enorm = 2.0 / (f_pts[2:n_mels + 2] - f_pts[:n_mels]) 525 | fb *= enorm.unsqueeze(0) 526 | 527 | if (fb.max(dim=0).values == 0.).any(): 528 | warnings.warn( 529 | "At least one mel filterbank has all zero values. " 530 | f"The value for `n_mels` ({n_mels}) may be set too high. " 531 | f"Or, the value for `n_freqs` ({n_freqs}) may be set too low." 532 | ) 533 | 534 | return fb 535 | 536 | 537 | def linear_fbanks( 538 | n_freqs: int, 539 | f_min: float, 540 | f_max: float, 541 | n_filter: int, 542 | sample_rate: int, 543 | ) -> Tensor: 544 | r"""Creates a linear triangular filterbank. 545 | 546 | Note: 547 | For the sake of the numerical compatibility with librosa, not all the coefficients 548 | in the resulting filter bank has magnitude of 1. 549 | 550 | .. image:: https://download.pytorch.org/torchaudio/doc-assets/lin_fbanks.png 551 | :alt: Visualization of generated filter bank 552 | 553 | Args: 554 | n_freqs (int): Number of frequencies to highlight/apply 555 | f_min (float): Minimum frequency (Hz) 556 | f_max (float): Maximum frequency (Hz) 557 | n_filter (int): Number of (linear) triangular filter 558 | sample_rate (int): Sample rate of the audio waveform 559 | 560 | Returns: 561 | Tensor: Triangular filter banks (fb matrix) of size (``n_freqs``, ``n_filter``) 562 | meaning number of frequencies to highlight/apply to x the number of filterbanks. 563 | Each column is a filterbank so that assuming there is a matrix A of 564 | size (..., ``n_freqs``), the applied result would be 565 | ``A * linear_fbanks(A.size(-1), ...)``. 566 | """ 567 | # freq bins 568 | all_freqs = torch.linspace(0, sample_rate // 2, n_freqs) 569 | 570 | # filter mid-points 571 | f_pts = torch.linspace(f_min, f_max, n_filter + 2) 572 | 573 | # create filterbank 574 | fb = _create_triangular_filterbank(all_freqs, f_pts) 575 | 576 | return fb 577 | 578 | 579 | def create_dct( 580 | n_mfcc: int, 581 | n_mels: int, 582 | norm: Optional[str] 583 | ) -> Tensor: 584 | r"""Create a DCT transformation matrix with shape (``n_mels``, ``n_mfcc``), 585 | normalized depending on norm. 586 | 587 | Args: 588 | n_mfcc (int): Number of mfc coefficients to retain 589 | n_mels (int): Number of mel filterbanks 590 | norm (str or None): Norm to use (either 'ortho' or None) 591 | 592 | Returns: 593 | Tensor: The transformation matrix, to be right-multiplied to 594 | row-wise data of size (``n_mels``, ``n_mfcc``). 595 | """ 596 | # http://en.wikipedia.org/wiki/Discrete_cosine_transform#DCT-II 597 | n = torch.arange(float(n_mels)) 598 | k = torch.arange(float(n_mfcc)).unsqueeze(1) 599 | dct = torch.cos(math.pi / float(n_mels) * (n + 0.5) * k) # size (n_mfcc, n_mels) 600 | if norm is None: 601 | dct *= 2.0 602 | else: 603 | assert norm == "ortho" 604 | dct[0] *= 1.0 / math.sqrt(2.0) 605 | dct *= math.sqrt(2.0 / float(n_mels)) 606 | return dct.t() 607 | 608 | 609 | def mu_law_encoding( 610 | x: Tensor, 611 | quantization_channels: int 612 | ) -> Tensor: 613 | r"""Encode signal based on mu-law companding. For more info see the 614 | `Wikipedia Entry <https://en.wikipedia.org/wiki/%CE%9C-law_algorithm>`_ 615 | 616 | This algorithm assumes the signal has been scaled to between -1 and 1 and 617 | returns a signal encoded with values from 0 to quantization_channels - 1. 618 | 619 | Args: 620 | x (Tensor): Input tensor 621 | quantization_channels (int): Number of channels 622 | 623 | Returns: 624 | Tensor: Input after mu-law encoding 625 | """ 626 | mu = quantization_channels - 1.0 627 | if not x.is_floating_point(): 628 | x = x.to(torch.float) 629 | mu = torch.tensor(mu, dtype=x.dtype) 630 | x_mu = torch.sign(x) * torch.log1p(mu * torch.abs(x)) / torch.log1p(mu) 631 | x_mu = ((x_mu + 1) / 2 * mu + 0.5).to(torch.int64) 632 | return x_mu 633 | 634 | 635 | def mu_law_decoding( 636 | x_mu: Tensor, 637 | quantization_channels: int 638 | ) -> Tensor: 639 | r"""Decode mu-law encoded signal. For more info see the 640 | `Wikipedia Entry <https://en.wikipedia.org/wiki/%CE%9C-law_algorithm>`_ 641 | 642 | This expects an input with values between 0 and quantization_channels - 1 643 | and returns a signal scaled between -1 and 1. 644 | 645 | Args: 646 | x_mu (Tensor): Input tensor 647 | quantization_channels (int): Number of channels 648 | 649 | Returns: 650 | Tensor: Input after mu-law decoding 651 | """ 652 | mu = quantization_channels - 1.0 653 | if not x_mu.is_floating_point(): 654 | x_mu = x_mu.to(torch.float) 655 | mu = torch.tensor(mu, dtype=x_mu.dtype) 656 | x = ((x_mu) / mu) * 2 - 1.0 657 | x = torch.sign(x) * (torch.exp(torch.abs(x) * torch.log1p(mu)) - 1.0) / mu 658 | return x 659 | 660 | 661 | def phase_vocoder( 662 | complex_specgrams: Tensor, 663 | rate: float, 664 | phase_advance: Tensor 665 | ) -> Tensor: 666 | r"""Given a STFT tensor, speed up in time without modifying pitch by a 667 | factor of ``rate``. 668 | 669 | Args: 670 | complex_specgrams (Tensor): 671 | A tensor of dimension `(..., freq, num_frame)` with complex dtype. 672 | rate (float): Speed-up factor 673 | phase_advance (Tensor): Expected phase advance in each bin. Dimension of `(freq, 1)` 674 | 675 | Returns: 676 | Tensor: 677 | Stretched spectrogram. The resulting tensor is of the same dtype as the input 678 | spectrogram, but the number of frames is changed to ``ceil(num_frame / rate)``. 679 | 680 | Example 681 | >>> freq, hop_length = 1025, 512 682 | >>> # (channel, freq, time) 683 | >>> complex_specgrams = torch.randn(2, freq, 300, dtype=torch.cfloat) 684 | >>> rate = 1.3 # Speed up by 30% 685 | >>> phase_advance = torch.linspace( 686 | >>> 0, math.pi * hop_length, freq)[..., None] 687 | >>> x = phase_vocoder(complex_specgrams, rate, phase_advance) 688 | >>> x.shape # with 231 == ceil(300 / 1.3) 689 | torch.Size([2, 1025, 231]) 690 | """ 691 | if rate == 1.0: 692 | return complex_specgrams 693 | 694 | # pack batch 695 | shape = complex_specgrams.size() 696 | complex_specgrams = complex_specgrams.reshape([-1] + list(shape[-2:])) 697 | 698 | # Figures out the corresponding real dtype, i.e. complex128 -> float64, complex64 -> float32 699 | # Note torch.real is a view so it does not incur any memory copy. 700 | real_dtype = torch.real(complex_specgrams).dtype 701 | time_steps = torch.arange( 702 | 0, 703 | complex_specgrams.size(-1), 704 | rate, 705 | device=complex_specgrams.device, 706 | dtype=real_dtype) 707 | 708 | alphas = time_steps % 1.0 709 | phase_0 = complex_specgrams[..., :1].angle() 710 | 711 | # Time Padding 712 | complex_specgrams = torch.nn.functional.pad(complex_specgrams, [0, 2]) 713 | 714 | # (new_bins, freq, 2) 715 | complex_specgrams_0 = complex_specgrams.index_select(-1, time_steps.long()) 716 | complex_specgrams_1 = complex_specgrams.index_select(-1, (time_steps + 1).long()) 717 | 718 | angle_0 = complex_specgrams_0.angle() 719 | angle_1 = complex_specgrams_1.angle() 720 | 721 | norm_0 = complex_specgrams_0.abs() 722 | norm_1 = complex_specgrams_1.abs() 723 | 724 | phase = angle_1 - angle_0 - phase_advance 725 | phase = phase - 2 * math.pi * torch.round(phase / (2 * math.pi)) 726 | 727 | # Compute Phase Accum 728 | phase = phase + phase_advance 729 | phase = torch.cat([phase_0, phase[..., :-1]], dim=-1) 730 | phase_acc = torch.cumsum(phase, -1) 731 | 732 | mag = alphas * norm_1 + (1 - alphas) * norm_0 733 | 734 | complex_specgrams_stretch = torch.polar(mag, phase_acc) 735 | 736 | # unpack batch 737 | complex_specgrams_stretch = complex_specgrams_stretch.reshape(shape[:-2] + complex_specgrams_stretch.shape[1:]) 738 | return complex_specgrams_stretch 739 | 740 | 741 | def mask_along_axis_iid( 742 | specgrams: Tensor, 743 | mask_param: int, 744 | mask_value: float, 745 | axis: int 746 | ) -> Tensor: 747 | r""" 748 | Apply a mask along ``axis``. Mask will be applied from indices ``[v_0, v_0 + v)``, where 749 | ``v`` is sampled from ``uniform(0, mask_param)``, and ``v_0`` from ``uniform(0, max_v - v)``. 750 | 751 | Args: 752 | specgrams (Tensor): Real spectrograms `(batch, channel, freq, time)` 753 | mask_param (int): Number of columns to be masked will be uniformly sampled from [0, mask_param] 754 | mask_value (float): Value to assign to the masked columns 755 | axis (int): Axis to apply masking on (2 -> frequency, 3 -> time) 756 | 757 | Returns: 758 | Tensor: Masked spectrograms of dimensions `(batch, channel, freq, time)` 759 | """ 760 | 761 | if axis not in [2, 3]: 762 | raise ValueError('Only Frequency and Time masking are supported') 763 | 764 | device = specgrams.device 765 | dtype = specgrams.dtype 766 | 767 | value = torch.rand(specgrams.shape[:2], device=device, dtype=dtype) * mask_param 768 | min_value = torch.rand(specgrams.shape[:2], device=device, dtype=dtype) * (specgrams.size(axis) - value) 769 | 770 | # Create broadcastable mask 771 | mask_start = min_value[..., None, None] 772 | mask_end = (min_value + value)[..., None, None] 773 | mask = torch.arange(0, specgrams.size(axis), device=device, dtype=dtype) 774 | 775 | # Per batch example masking 776 | specgrams = specgrams.transpose(axis, -1) 777 | specgrams = specgrams.masked_fill((mask >= mask_start) & (mask < mask_end), mask_value) 778 | specgrams = specgrams.transpose(axis, -1) 779 | 780 | return specgrams 781 | 782 | 783 | def mask_along_axis( 784 | specgram: Tensor, 785 | mask_param: int, 786 | mask_value: float, 787 | axis: int 788 | ) -> Tensor: 789 | r""" 790 | Apply a mask along ``axis``. Mask will be applied from indices ``[v_0, v_0 + v)``, where 791 | ``v`` is sampled from ``uniform(0, mask_param)``, and ``v_0`` from ``uniform(0, max_v - v)``. 792 | All examples will have the same mask interval. 793 | 794 | Args: 795 | specgram (Tensor): Real spectrogram `(channel, freq, time)` 796 | mask_param (int): Number of columns to be masked will be uniformly sampled from [0, mask_param] 797 | mask_value (float): Value to assign to the masked columns 798 | axis (int): Axis to apply masking on (1 -> frequency, 2 -> time) 799 | 800 | Returns: 801 | Tensor: Masked spectrogram of dimensions `(channel, freq, time)` 802 | """ 803 | if axis not in [1, 2]: 804 | raise ValueError('Only Frequency and Time masking are supported') 805 | 806 | # pack batch 807 | shape = specgram.size() 808 | specgram = specgram.reshape([-1] + list(shape[-2:])) 809 | value = torch.rand(1) * mask_param 810 | min_value = torch.rand(1) * (specgram.size(axis) - value) 811 | 812 | mask_start = (min_value.long()).squeeze() 813 | mask_end = (min_value.long() + value.long()).squeeze() 814 | mask = torch.arange(0, specgram.shape[axis], device=specgram.device, dtype=specgram.dtype) 815 | mask = (mask >= mask_start) & (mask < mask_end) 816 | if axis == 1: 817 | mask = mask.unsqueeze(-1) 818 | 819 | assert mask_end - mask_start < mask_param 820 | 821 | specgram = specgram.masked_fill(mask, mask_value) 822 | 823 | # unpack batch 824 | specgram = specgram.reshape(shape[:-2] + specgram.shape[-2:]) 825 | 826 | return specgram 827 | 828 | 829 | def compute_deltas( 830 | specgram: Tensor, 831 | win_length: int = 5, 832 | mode: str = "replicate" 833 | ) -> Tensor: 834 | r"""Compute delta coefficients of a tensor, usually a spectrogram: 835 | 836 | .. math:: 837 | d_t = \frac{\sum_{n=1}^{\text{N}} n (c_{t+n} - c_{t-n})}{2 \sum_{n=1}^{\text{N}} n^2} 838 | 839 | where :math:`d_t` is the deltas at time :math:`t`, 840 | :math:`c_t` is the spectrogram coeffcients at time :math:`t`, 841 | :math:`N` is ``(win_length-1)//2``. 842 | 843 | Args: 844 | specgram (Tensor): Tensor of audio of dimension `(..., freq, time)` 845 | win_length (int, optional): The window length used for computing delta (Default: ``5``) 846 | mode (str, optional): Mode parameter passed to padding (Default: ``"replicate"``) 847 | 848 | Returns: 849 | Tensor: Tensor of deltas of dimension `(..., freq, time)` 850 | 851 | Example 852 | >>> specgram = torch.randn(1, 40, 1000) 853 | >>> delta = compute_deltas(specgram) 854 | >>> delta2 = compute_deltas(delta) 855 | """ 856 | device = specgram.device 857 | dtype = specgram.dtype 858 | 859 | # pack batch 860 | shape = specgram.size() 861 | specgram = specgram.reshape(1, -1, shape[-1]) 862 | 863 | assert win_length >= 3 864 | 865 | n = (win_length - 1) // 2 866 | 867 | # twice sum of integer squared 868 | denom = n * (n + 1) * (2 * n + 1) / 3 869 | 870 | specgram = torch.nn.functional.pad(specgram, (n, n), mode=mode) 871 | 872 | kernel = torch.arange(-n, n + 1, 1, device=device, dtype=dtype).repeat(specgram.shape[1], 1, 1) 873 | 874 | output = torch.nn.functional.conv1d(specgram, kernel, groups=specgram.shape[1]) / denom 875 | 876 | # unpack batch 877 | output = output.reshape(shape) 878 | 879 | return output 880 | 881 | 882 | def _compute_nccf( 883 | waveform: Tensor, 884 | sample_rate: int, 885 | frame_time: float, 886 | freq_low: int 887 | ) -> Tensor: 888 | r""" 889 | Compute Normalized Cross-Correlation Function (NCCF). 890 | 891 | .. math:: 892 | \phi_i(m) = \frac{\sum_{n=b_i}^{b_i + N-1} w(n) w(m+n)}{\sqrt{E(b_i) E(m+b_i)}}, 893 | 894 | where 895 | :math:`\phi_i(m)` is the NCCF at frame :math:`i` with lag :math:`m`, 896 | :math:`w` is the waveform, 897 | :math:`N` is the length of a frame, 898 | :math:`b_i` is the beginning of frame :math:`i`, 899 | :math:`E(j)` is the energy :math:`\sum_{n=j}^{j+N-1} w^2(n)`. 900 | """ 901 | 902 | EPSILON = 10 ** (-9) 903 | 904 | # Number of lags to check 905 | lags = int(math.ceil(sample_rate / freq_low)) 906 | 907 | frame_size = int(math.ceil(sample_rate * frame_time)) 908 | 909 | waveform_length = waveform.size()[-1] 910 | num_of_frames = int(math.ceil(waveform_length / frame_size)) 911 | 912 | p = lags + num_of_frames * frame_size - waveform_length 913 | waveform = torch.nn.functional.pad(waveform, (0, p)) 914 | 915 | # Compute lags 916 | output_lag = [] 917 | for lag in range(1, lags + 1): 918 | s1 = waveform[..., :-lag].unfold(-1, frame_size, frame_size)[..., :num_of_frames, :] 919 | s2 = waveform[..., lag:].unfold(-1, frame_size, frame_size)[..., :num_of_frames, :] 920 | 921 | output_frames = ( 922 | (s1 * s2).sum(-1) 923 | / (EPSILON + torch.norm(s1, p=2, dim=-1)).pow(2) 924 | / (EPSILON + torch.norm(s2, p=2, dim=-1)).pow(2) 925 | ) 926 | 927 | output_lag.append(output_frames.unsqueeze(-1)) 928 | 929 | nccf = torch.cat(output_lag, -1) 930 | 931 | return nccf 932 | 933 | 934 | def _combine_max( 935 | a: Tuple[Tensor, Tensor], 936 | b: Tuple[Tensor, Tensor], 937 | thresh: float = 0.99 938 | ) -> Tuple[Tensor, Tensor]: 939 | """ 940 | Take value from first if bigger than a multiplicative factor of the second, elementwise. 941 | """ 942 | mask = (a[0] > thresh * b[0]) 943 | values = mask * a[0] + ~mask * b[0] 944 | indices = mask * a[1] + ~mask * b[1] 945 | return values, indices 946 | 947 | 948 | def _find_max_per_frame( 949 | nccf: Tensor, 950 | sample_rate: int, 951 | freq_high: int 952 | ) -> Tensor: 953 | r""" 954 | For each frame, take the highest value of NCCF, 955 | apply centered median smoothing, and convert to frequency. 956 | 957 | Note: If the max among all the lags is very close 958 | to the first half of lags, then the latter is taken. 959 | """ 960 | 961 | lag_min = int(math.ceil(sample_rate / freq_high)) 962 | 963 | # Find near enough max that is smallest 964 | 965 | best = torch.max(nccf[..., lag_min:], -1) 966 | 967 | half_size = nccf.shape[-1] // 2 968 | half = torch.max(nccf[..., lag_min:half_size], -1) 969 | 970 | best = _combine_max(half, best) 971 | indices = best[1] 972 | 973 | # Add back minimal lag 974 | indices += lag_min 975 | # Add 1 empirical calibration offset 976 | indices += 1 977 | 978 | return indices 979 | 980 | 981 | def _median_smoothing( 982 | indices: Tensor, 983 | win_length: int 984 | ) -> Tensor: 985 | r""" 986 | Apply median smoothing to the 1D tensor over the given window. 987 | """ 988 | 989 | # Centered windowed 990 | pad_length = (win_length - 1) // 2 991 | 992 | # "replicate" padding in any dimension 993 | indices = torch.nn.functional.pad( 994 | indices, (pad_length, 0), mode="constant", value=0. 995 | ) 996 | 997 | indices[..., :pad_length] = torch.cat(pad_length * [indices[..., pad_length].unsqueeze(-1)], dim=-1) 998 | roll = indices.unfold(-1, win_length, 1) 999 | 1000 | values, _ = torch.median(roll, -1) 1001 | -------------------------------------------------------------------------------- /tests/test_issue.py: -------------------------------------------------------------------------------- 1 | # -*- coding: utf-8 -*- 2 | """ 3 | @author:XuMing(xuming624@qq.com) 4 | @description: 5 | """ 6 | import sys 7 | import unittest 8 | 9 | import torch 10 | 11 | sys.path.append('..') 12 | from codeassist.gpt2_coder import GPT2Coder 13 | 14 | 15 | class IssueTestCase(unittest.TestCase): 16 | 17 | def test_code_predict(self): 18 | prompts = [ 19 | """from torch import nn 20 | class LSTM(Module): 21 | def __init__(self, *, 22 | n_tokens: int, 23 | embedding_size: int, 24 | hidden_size: int, 25 | n_layers: int):""", 26 | """import numpy as np 27 | import torch 28 | import torch.nn as""", 29 | "import java.util.ArrayList;", 30 | ] 31 | infer = GPT2Coder("shibing624/code-autocomplete-gpt2-base") 32 | results = [] 33 | for prompt in prompts: 34 | res = infer.generate(prompt) 35 | print("Query:", prompt) 36 | print("Result:", res[0]) 37 | print("=" * 20) 38 | results.append(res[0]) 39 | self.assertEqual(len(results), 3) 40 | 41 | 42 | if __name__ == '__main__': 43 | unittest.main() 44 | -------------------------------------------------------------------------------- /tests/test_qps.py: -------------------------------------------------------------------------------- 1 | # -*- coding: utf-8 -*- 2 | """ 3 | @author:XuMing(xuming624@qq.com) 4 | @description: 5 | """ 6 | import os 7 | import sys 8 | import unittest 9 | from time import time 10 | 11 | sys.path.append('..') 12 | from codeassist.gpt2_coder import GPT2Coder 13 | 14 | pwd_path = os.path.abspath(os.path.dirname(__file__)) 15 | 16 | test_file = os.path.join(pwd_path, 'test.txt') 17 | model = GPT2Coder("shibing624/code-autocomplete-gpt2-base") 18 | 19 | 20 | def load_data(file_path): 21 | res = [] 22 | with open(file_path, 'r', encoding='utf-8') as f: 23 | for line in f: 24 | line = line.strip() 25 | res.append(line) 26 | if len(res) >= 200: 27 | break 28 | return res 29 | 30 | 31 | class QPSTestCase(unittest.TestCase): 32 | def test_code_infer_speed(self): 33 | """Test code_infer_speed""" 34 | codes = load_data(test_file) 35 | codes = codes[100:110] 36 | t1 = time() 37 | for prompt in codes: 38 | res = model.generate(prompt) 39 | print("Query:", prompt) 40 | print("Result:", res[0]) 41 | print("=" * 20) 42 | spend_time = time() - t1 43 | print('spend time:', spend_time, ' seconds') 44 | print('size:', len(codes), ' qps:', len(codes) / spend_time) 45 | 46 | 47 | if __name__ == '__main__': 48 | unittest.main() 49 | --------------------------------------------------------------------------------