├── .gitignore ├── CITATION.cff ├── LICENSE ├── README.md ├── cff-version 1.2.0.cliff.txt ├── data └── alpaca_data.json ├── requirements.txt └── src ├── configs ├── deepspeed_config.json └── hostfile ├── environment_ChatPath.yml ├── generate.py ├── imgs ├── chatpath_logo.png ├── data_process.png ├── pathgpt_instruction.png └── pathgpt_logo.png ├── train.py └── utils.py /.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 | share/python-wheels/ 24 | *.egg-info/ 25 | .installed.cfg 26 | *.egg 27 | MANIFEST 28 | 29 | # PyInstaller 30 | # Usually these files are written by a python script from a template 31 | # before PyInstaller builds the exe, so as to inject date/other infos into it. 32 | *.manifest 33 | *.spec 34 | 35 | # Installer logs 36 | pip-log.txt 37 | pip-delete-this-directory.txt 38 | 39 | # Unit test / coverage reports 40 | htmlcov/ 41 | .tox/ 42 | .nox/ 43 | .coverage 44 | .coverage.* 45 | .cache 46 | nosetests.xml 47 | coverage.xml 48 | *.cover 49 | *.py,cover 50 | .hypothesis/ 51 | .pytest_cache/ 52 | cover/ 53 | 54 | # Translations 55 | *.mo 56 | *.pot 57 | 58 | # Django stuff: 59 | *.log 60 | local_settings.py 61 | db.sqlite3 62 | db.sqlite3-journal 63 | 64 | # Flask stuff: 65 | instance/ 66 | .webassets-cache 67 | 68 | # Scrapy stuff: 69 | .scrapy 70 | 71 | # Sphinx documentation 72 | docs/_build/ 73 | 74 | # PyBuilder 75 | .pybuilder/ 76 | target/ 77 | 78 | # Jupyter Notebook 79 | .ipynb_checkpoints 80 | 81 | # IPython 82 | profile_default/ 83 | ipython_config.py 84 | 85 | # pyenv 86 | # For a library or package, you might want to ignore these files since the code is 87 | # intended to run in multiple environments; otherwise, check them in: 88 | # .python-version 89 | 90 | # pipenv 91 | # According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control. 92 | # However, in case of collaboration, if having platform-specific dependencies or dependencies 93 | # having no cross-platform support, pipenv may install dependencies that don't work, or not 94 | # install all needed dependencies. 95 | #Pipfile.lock 96 | 97 | # poetry 98 | # Similar to Pipfile.lock, it is generally recommended to include poetry.lock in version control. 99 | # This is especially recommended for binary packages to ensure reproducibility, and is more 100 | # commonly ignored for libraries. 101 | # https://python-poetry.org/docs/basic-usage/#commit-your-poetrylock-file-to-version-control 102 | #poetry.lock 103 | 104 | # pdm 105 | # Similar to Pipfile.lock, it is generally recommended to include pdm.lock in version control. 106 | #pdm.lock 107 | # pdm stores project-wide configurations in .pdm.toml, but it is recommended to not include it 108 | # in version control. 109 | # https://pdm.fming.dev/#use-with-ide 110 | .pdm.toml 111 | 112 | # PEP 582; used by e.g. github.com/David-OConnor/pyflow and github.com/pdm-project/pdm 113 | __pypackages__/ 114 | 115 | # Celery stuff 116 | celerybeat-schedule 117 | celerybeat.pid 118 | 119 | # SageMath parsed files 120 | *.sage.py 121 | 122 | # Environments 123 | .env 124 | .venv 125 | env/ 126 | venv/ 127 | ENV/ 128 | env.bak/ 129 | venv.bak/ 130 | 131 | # Spyder project settings 132 | .spyderproject 133 | .spyproject 134 | 135 | # Rope project settings 136 | .ropeproject 137 | 138 | # mkdocs documentation 139 | /site 140 | 141 | # mypy 142 | .mypy_cache/ 143 | .dmypy.json 144 | dmypy.json 145 | 146 | # Pyre type checker 147 | .pyre/ 148 | 149 | # pytype static type analyzer 150 | .pytype/ 151 | 152 | # Cython debug symbols 153 | cython_debug/ 154 | 155 | # PyCharm 156 | # JetBrains specific template is maintained in a separate JetBrains.gitignore that can 157 | # be found at https://github.com/github/gitignore/blob/main/Global/JetBrains.gitignore 158 | # and can be added to the global gitignore or merged into this file. For a more nuclear 159 | # option (not recommended) you can uncomment the following to ignore the entire idea folder. 160 | .idea/ 161 | 162 | transformers/ 163 | .DS_Store -------------------------------------------------------------------------------- /CITATION.cff: -------------------------------------------------------------------------------- 1 | # This CITATION.cff file was generated with cffinit. 2 | # Visit https://bit.ly/cffinit to generate yours today! 3 | 4 | cff-version: 1.2.0 5 | title: 'PathGPT: A Knowledgeable GPT Model for Pathology' 6 | message: >- 7 | If you use this software, please cite it using the 8 | metadata from this file. 9 | type: software 10 | authors: 11 | - given-names: Yuxuan 12 | family-names: Sun 13 | email: sunyuxuan@westlake.edu.cn 14 | name-particle: YUXUAN 15 | repository-code: 'https://github.com/superjamessyx/PathGPT' 16 | abstract: >- 17 | PathGPT is a specialized language model tailored for the 18 | field of pathology. Developed by fine-tuning the Llama-7B 19 | model using a dataset of 13,000 pathology-specific 20 | questions and answers we've collected. 21 | license: Apache-2.0 22 | -------------------------------------------------------------------------------- /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 2023 Rohan Taori, Ishaan Gulrajani, Tianyi Zhang, Yann Dubois, Xuechen Li 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 |

2 | ChatPath 3 |

4 | 5 | 6 | 7 | 8 | 9 | [![Code License](https://img.shields.io/badge/Code%20License-Apache_2.0-green.svg)](https://github.com/tatsu-lab/stanford_alpaca/blob/main/LICENSE) 10 | [![Data License](https://img.shields.io/badge/Data%20License-CC%20By%20NC%204.0-red.svg)](https://github.com/tatsu-lab/stanford_alpaca/blob/main/DATA_LICENSE) 11 | 12 | ## PathGPT: A Knowledgeable GPT Model for Pathology 13 | 14 | Welcome to the PathGPT repository! PathGPT is a specialized language model tailored for the field of pathology. Developed by fine-tuning the Llama-7B model using a dataset of 13,000 pathology-specific questions and answers we've collected. We're excited to announce the release of the PathGPT checkpoint (the weight diff of Llama), with the full 13k dataset to follow shortly. The detailed data processing process is as follows: 15 | 16 |

17 | ChatPath 18 |

19 | 20 | But we're not stopping there! In the future, we plan to expand the dataset to over 100,000 entries, encompassing a diverse range of pathology-related instruction data. We believe that PathGPT will become an valuable tool for pathologists and the entire pathology community. 21 | 22 | ### **Authors** 23 | 24 | This project was completed by **Yuxuan Sun** and **Chenglu Zhu** from the **Artificial Intelligence and Biomedical Image Analysis Lab** of the School of Engineering at Westlake University. We would like to thank **Kai Zhang** (Ohio State University) for participating in the discussion and collaboration, as well as the following individuals who contributed to the annotation process: **Xinheng Lv** and **Ruojia Zhao**. 25 | 26 | 27 | 28 | ## Get your demo experience!! 29 | 30 | We deployed PathGPT on the A100 server and opened it up for user experience. You can follow the instruction illustrated below. The demo webset is: https://f86fb98a1b7bbd2a.gradio.app. Please feel free to point out the problems our model. 31 | 32 |

33 | ChatPath 34 |

35 | 36 | 37 | 38 |

Usage

39 | 40 | - Setup. Install the conda environment: 41 | ```bash 42 | conda create -n pathgpt python=3.10 43 | conda activate pathgpt 44 | git clone https://github.com/superjamessyx/PathGPT.git 45 | cd PathGPT/src 46 | conda install pytorch==1.12.0 torchvision==0.13.0 torchaudio==0.12.0 cudatoolkit=11.3 -c pytorch 47 | git clone https://github.com/huggingface/transformers.git 48 | cd transformers 49 | pip install -e . 50 | cd ../.. 51 | pip install -r requirements.txt 52 | ``` 53 | 54 | - Training data 55 | 56 | Taking into account that our present data collection focuses solely on pathology-related QA pairs, an exclusive emphasis on QA training may hinder PathGPT's ability to effectively carry out other tasks. To address this limitation, we initially fine-tune the model using Stanford [Alpaca](https://crfm.stanford.edu/2023/03/13/alpaca.html) data, followed by further fine-tuning with our pathology-specific data. This approach enables PathGPT to adeptly handle a diverse range of questioning styles. 57 | ```bash 58 | PathGPT/src/data/pathology_alpaca.json 59 | ``` 60 | 61 | - Convert LLaMA checkpoint to HuggingFace format: 62 | ```bash 63 | cd PathGPT/src 64 | python transformers/src/transformers/models/llama/convert_llama_weights_to_hf.py \ 65 | --input_dir /path/to/llama-7B/ \ 66 | --model_size 7B \ 67 | --output_dir /path/to/llama-7B/hf 68 | ``` 69 | 70 | - To train LLaMA-7B with DeepSpeed, you can select either DeepSpeed Zero-2 or Zero-3 by using the following command options: `--deepspeed configs/ds_config_zero2.json` or `configs/ds_config_zero3.json`: 71 | ```bash 72 | deepspeed train.py \ 73 | --model_name_or_path /path/to/llama-7B/hf \ 74 | --data_path /path/to/pathology_alpaca.json \ 75 | --output_dir /path/to/llama-7B/hf/ft \ 76 | --num_train_epochs 3 \ 77 | --model_max_length 512 \ 78 | --per_device_train_batch_size 64 \ 79 | --per_device_eval_batch_size 1 \ 80 | --gradient_accumulation_steps 1 \ 81 | --evaluation_strategy "no" \ 82 | --save_strategy "steps" \ 83 | --save_steps 100 \ 84 | --save_total_limit 2 \ 85 | --learning_rate 2e-5 \ 86 | --warmup_steps 2 \ 87 | --logging_steps 2 \ 88 | --lr_scheduler_type "cosine" \ 89 | --report_to "tensorboard" \ 90 | --gradient_checkpointing True \ 91 | --deepspeed configs/ds_config_zero2.json \ 92 | --fp16 True 93 | ``` 94 | - Train LLaMA-7B on DeepSpeed with Multi-nodes 95 | ```bash 96 | deepspeed --num_gpus num_of_gpus_in_each_node \ 97 | --num_nodes num_of_nodes \ 98 | --master_addr ip_address_of_main_node \ 99 | --master_port 34545 \ 100 | --hostfile configs/hostfile \ 101 | train.py \ 102 | --model_name_or_path /path/to/llama-7B/hf \ 103 | --data_path /path/to/pathology_alpaca.json \ 104 | --output_dir /path/to/llama-7B/hf/ft \ 105 | --num_train_epochs 3 \ 106 | --model_max_length 512 \ 107 | --per_device_train_batch_size 64 \ 108 | --per_device_eval_batch_size 4 \ 109 | --gradient_accumulation_steps 1 \ 110 | --evaluation_strategy "no" \ 111 | --save_strategy "steps" \ 112 | --save_steps 100 \ 113 | --save_total_limit 2 \ 114 | --learning_rate 2e-5 \ 115 | --warmup_steps 2 \ 116 | --logging_steps 2 \ 117 | --lr_scheduler_type "cosine" \ 118 | --report_to "tensorboard" \ 119 | --gradient_checkpointing True \ 120 | --deepspeed configs/ds_config_zero2.json \ 121 | --fp16 True 122 | ``` 123 | 124 | 125 | 126 | - Training Cost 127 | 128 | Training 7B model with `13k examples/ 3 epoch` & `batch_size=32` within `1 hour` on `4 x A100 GPUs`. 129 | 130 | | LLaMA | Batch Size | A100s | Time (h) | 131 | | ----- | ---------- | ----- | -------- | 132 | | 7 B | 32 | 4 | 1 | 133 | 134 | 135 | - Inference 136 | ```bash 137 | # web demo inference 138 | python generate.py 139 | ``` 140 | 141 | ## Download fine-tuned checkpoint 142 | 143 | We have released the weight difference ([huggingface](https://huggingface.co/jamessyx/ChatPath)) between the finetuned weights and the original llama weights. By simply adding this difference to the latter, you can obtain the finetuned weights. 144 | 145 | ## Example outputs 146 | 147 | *How to distinguish the distribution pattern of immune cells in solid tumors of H&E staining images?* 148 | 149 | **Alpaca** 150 | 151 | The distribution pattern of immune cells in solid tumors can be determined by examining H&E stained slides. Immunohistochemical stains such as CD3, CD8, CD4, and CD20 can be used to identify T-cells, B-cells, and antigen-presenting cells, respectively. Additionally, macrophages can be identified with CD68 staining. 152 | 153 | **PathGPT** 154 | 155 | To distinguish the distribution pattern of immune cells in solid tumors of H&E staining images, the following steps can be followed: 156 | 157 | 1. Select the area of interest in the H&E staining image. 158 | 159 | 2. Identify the location of immune cells in the selected area. 160 | 161 | 3. Count the number of immune cells in the selected area. 162 | 163 | 4. Calculate the percentage of immune cells in the selected area. 164 | 165 | 5. Compare the percentage of immune cells in the selected area with the percentage of immune cells in normal tissues. 166 | 167 | 6. Analyze the distribution pattern of immune cells in the selected area. 168 | 169 | 7. Compare the distribution pattern of immune cells in the selected area with the distribution pattern of immune cells in normal tissues. 170 | 171 | 8. Evaluate the significance of the distribution pattern of immune cells in the selected area. 172 | 173 | By following these steps, it is possible to distinguish the distribution pattern of immune cells in solid tumors of H&E staining images. 174 | 175 | 176 | 177 | 178 | 179 | 180 | ## Thanks For 181 | 182 | This project is based on 183 | 184 | [Llama-X](https://github.com/AetherCortex/Llama-X) 185 | 186 | and also relate to multiple open source projects: 187 | 188 | [Meta AI LLaMA](https://arxiv.org/abs/2302.13971v1) 189 | 190 | [Huggingface Transformers Llama](https://github.com/huggingface/transformers/tree/main/src/transformers/models/llama) 191 | 192 | [Alpaca](https://crfm.stanford.edu/2023/03/13/alpaca.html) and [Alpaca-LoRA](https://github.com/tloen/alpaca-lora) 193 | 194 | 195 | ## Disclaimer 196 | 197 | The use of resources(e.g., code, data and model weights) related to this project is limited to academic research and is prohibited for commercial purposes. The content generated by PathGPT is subject to factors such as randomness and uncontrollability, and this project cannot guarantee its accuracy. This project does not assume any legal responsibility for the content of the model output, nor does it assume any responsibility for any losses that may arise from the use of related resources and output results. 198 | 199 | 200 | 201 | -------------------------------------------------------------------------------- /cff-version 1.2.0.cliff.txt: -------------------------------------------------------------------------------- 1 | cff-version: 1.2.0 2 | message: "If you use this software, please cite it as below." 3 | authors: 4 | - family-names: "Yuxuan" 5 | given-names: "Sun" 6 | - family-names: "Chenglu" 7 | given-names: "Zhu" 8 | title: "PathGPT: A Knowledgeable GPT Model for Pathology" 9 | version: 1.0.0 10 | date-released: 2023-4-22 11 | url: "https://github.com/superjamessyx/PathGPT" 12 | 13 | 14 | 15 | 16 | 然而，需要注意的是，PubMed等scientific paper中包含的image-text pair往往较短，yielding a narrow perspective of pathological images instead of a comprehensive analysis of all discernible content. This constraint limits the capability of trained models to provide detailed and accurate descriptions of the images, resulting in typically brief responses，因此，我们需要花费了大量人力去收集来自于书籍、病理图谱或者人为标注来解决。当然，与此同时，研究如何自动化的正确的扩充scientific paper中的caption 详细程度也显得十分重要 17 | 18 | 19 | 20 | However, it's important to note that caption of the image in scientific papers often quite succinct, yielding a narrow description of pathological images instead of a comprehensive analysis of all discernible content. This limitation constrains the capacity of trained models to provide detailed descriptions of the images. Therefore, we have invested considerable effort in collating data from books, pathology atlases, and manual annotations to rectify this issue. Simultaneously, researching how to automatically and correctly augment the details of the captions within scientific papers is of utmost importance. 21 | 22 | 23 | Nevertheless, it's important to note that captions accompanying images in scientific papers are frequently succinct, offering a limited interpretation of pathological images, rather than an exhaustive analysis of all identifiable content. This constrains the ability of models to furnish intricate descriptions of the images. In response to this, we've dedicated a significant effort to collate data from various sources, such as books, pathology atlases, and manual annotations to address this issue. Moreover, it is highly worthwhile to explore appropriate approaches to augment the level of detail in image captions within scientific papers. 24 | 25 | 26 | python -m training.main \ 27 | --save-frequency 1 \ 28 | --zeroshot-frequency 1 \ 29 | --report-to wandb \ 30 | --train-data="/data1/syx/dataset/Final_project_data/train/pathclip.csv" \ 31 | --csv-img-key img \ 32 | --csv-caption-key caption \ 33 | --warmup 10000 \ 34 | --batch-size=128 \ 35 | --pretrained=openai \ 36 | --lr=1e-3 \ 37 | --wd=0.1 \ 38 | --epochs=30 \ 39 | --workers=8 \ 40 | --model openai/clip-vit-base-patch16 -------------------------------------------------------------------------------- /requirements.txt: -------------------------------------------------------------------------------- 1 | numpy 2 | rouge_score 3 | fire 4 | openai 5 | sentencepiece 6 | wandb 7 | gradio==3.9 8 | deepspeed 9 | accelerate 10 | tensorboardX 11 | -------------------------------------------------------------------------------- /src/configs/deepspeed_config.json: -------------------------------------------------------------------------------- 1 | { 2 | "zero_optimization": { 3 | "stage": 3, 4 | "offload_optimizer": { 5 | "device": "cpu", 6 | "pin_memory": true 7 | }, 8 | "offload_param": { 9 | "device": "cpu", 10 | "pin_memory": true 11 | }, 12 | "overlap_comm": true, 13 | "contiguous_gradients": true, 14 | "sub_group_size": 0, 15 | "reduce_bucket_size": "auto", 16 | "stage3_prefetch_bucket_size": "auto", 17 | "stage3_param_persistence_threshold": "auto", 18 | "stage3_max_live_parameters": 0, 19 | "stage3_max_reuse_distance": 0, 20 | "stage3_gather_16bit_weights_on_model_save": true 21 | }, 22 | "fp16": { 23 | "enabled": true, 24 | "auto_cast": false, 25 | "loss_scale": 0, 26 | "initial_scale_power": 32, 27 | "loss_scale_window": 1000, 28 | "hysteresis": 2, 29 | "min_loss_scale": 1 30 | }, 31 | "optimizer": { 32 | "type": "AdamW", 33 | "params": { 34 | "lr": 2e-5, 35 | "betas": [ 36 | 0.9, 37 | 0.999 38 | ], 39 | "eps": 1e-8, 40 | "weight_decay": 0 41 | } 42 | }, 43 | "train_batch_size": "auto", 44 | "train_micro_batch_size_per_gpu": "auto", 45 | "wall_clock_breakdown": false 46 | } 47 | -------------------------------------------------------------------------------- /src/configs/hostfile: -------------------------------------------------------------------------------- 1 | ip_address_of_main_node slots=num_of_gpus_in_each_node 2 | ip_address_of_sub_node1 slots=num_of_gpus_in_each_node -------------------------------------------------------------------------------- /src/environment_ChatPath.yml: -------------------------------------------------------------------------------- 1 | name: chatpath 2 | channels: 3 | - pytorch 4 | - defaults 5 | dependencies: 6 | - _libgcc_mutex=0.1=main 7 | - _openmp_mutex=5.1=1_gnu 8 | - blas=1.0=mkl 9 | - bzip2=1.0.8=h7b6447c_0 10 | - ca-certificates=2023.01.10=h06a4308_0 11 | - charset-normalizer=2.0.4=pyhd3eb1b0_0 12 | - cudatoolkit=11.3.1=h2bc3f7f_2 13 | - ffmpeg=4.3=hf484d3e_0 14 | - freetype=2.12.1=h4a9f257_0 15 | - giflib=5.2.1=h5eee18b_3 16 | - gmp=6.2.1=h295c915_3 17 | - gnutls=3.6.15=he1e5248_0 18 | - intel-openmp=2021.4.0=h06a4308_3561 19 | - jpeg=9e=h5eee18b_1 20 | - lame=3.100=h7b6447c_0 21 | - lcms2=2.12=h3be6417_0 22 | - ld_impl_linux-64=2.38=h1181459_1 23 | - lerc=3.0=h295c915_0 24 | - libdeflate=1.17=h5eee18b_0 25 | - libffi=3.4.2=h6a678d5_6 26 | - libgcc-ng=11.2.0=h1234567_1 27 | - libgomp=11.2.0=h1234567_1 28 | - libiconv=1.16=h7f8727e_2 29 | - libidn2=2.3.2=h7f8727e_0 30 | - libpng=1.6.39=h5eee18b_0 31 | - libstdcxx-ng=11.2.0=h1234567_1 32 | - libtasn1=4.16.0=h27cfd23_0 33 | - libtiff=4.5.0=h6a678d5_2 34 | - libunistring=0.9.10=h27cfd23_0 35 | - libuuid=1.41.5=h5eee18b_0 36 | - libwebp=1.2.4=h11a3e52_1 37 | - libwebp-base=1.2.4=h5eee18b_1 38 | - lz4-c=1.9.4=h6a678d5_0 39 | - mkl=2021.4.0=h06a4308_640 40 | - mkl_fft=1.3.1=py310hd6ae3a3_0 41 | - mkl_random=1.2.2=py310h00e6091_0 42 | - ncurses=6.4=h6a678d5_0 43 | - nettle=3.7.3=hbbd107a_1 44 | - numpy-base=1.23.5=py310h8e6c178_0 45 | - openh264=2.1.1=h4ff587b_0 46 | - openssl=1.1.1t=h7f8727e_0 47 | - pycparser=2.21=pyhd3eb1b0_0 48 | - python=3.10.10=h7a1cb2a_2 49 | - pytorch=1.12.0=py3.10_cuda11.3_cudnn8.3.2_0 50 | - pytorch-mutex=1.0=cuda 51 | - readline=8.2=h5eee18b_0 52 | - six=1.16.0=pyhd3eb1b0_1 53 | - sqlite=3.41.1=h5eee18b_0 54 | - tk=8.6.12=h1ccaba5_0 55 | - typing_extensions=4.4.0=py310h06a4308_0 56 | - tzdata=2022g=h04d1e81_0 57 | - xz=5.2.10=h5eee18b_1 58 | - zlib=1.2.13=h5eee18b_0 59 | - zstd=1.5.2=ha4553b6_0 60 | - pip: 61 | - absl-py==1.4.0 62 | - accelerate==0.18.0 63 | - aiofiles==23.1.0 64 | - aiohttp==3.8.4 65 | - aiosignal==1.3.1 66 | - altair==4.2.2 67 | - anyio==3.6.2 68 | - appdirs==1.4.4 69 | - async-timeout==4.0.2 70 | - attrs==22.2.0 71 | - bcrypt==4.0.1 72 | - beartype==0.12.0 73 | - brotlipy==0.7.0 74 | - cachetools==5.3.0 75 | - certifi==2022.12.7 76 | - cffi==1.15.1 77 | - chatllama-py==0.0.3 78 | - click==8.1.3 79 | - cmake==3.26.1 80 | - contourpy==1.0.7 81 | - cryptography==39.0.1 82 | - cycler==0.11.0 83 | - dataclasses-json==0.5.7 84 | - datasets==2.10.1 85 | - deepspeed==0.8.3 86 | - dill==0.3.6 87 | - docker-pycreds==0.4.0 88 | - einops==0.6.0 89 | - entrypoints==0.4 90 | - fairscale==0.4.13 91 | - fastapi==0.95.0 92 | - ffmpy==0.3.0 93 | - filelock==3.10.5 94 | - fire==0.5.0 95 | - flit-core==3.8.0 96 | - fonttools==4.39.2 97 | - frozenlist==1.3.3 98 | - fsspec==2023.3.0 99 | - gitdb==4.0.10 100 | - gitpython==3.1.31 101 | - google-auth==2.16.3 102 | - google-auth-oauthlib==0.4.6 103 | - gradio==3.9 104 | - greenlet==2.0.2 105 | - grpcio==1.51.3 106 | - h11==0.12.0 107 | - hjson==3.1.0 108 | - httpcore==0.15.0 109 | - httpx==0.23.3 110 | - huggingface-hub==0.13.3 111 | - idna==3.4 112 | - jinja2==3.1.2 113 | - joblib==1.2.0 114 | - jsonschema==4.17.3 115 | - kiwisolver==1.4.4 116 | - langchain==0.0.123 117 | - linkify-it-py==2.0.0 118 | - lit==16.0.0 119 | - markdown==3.4.3 120 | - markdown-it-py==2.2.0 121 | - markupsafe==2.1.2 122 | - marshmallow==3.19.0 123 | - marshmallow-enum==1.5.1 124 | - matplotlib==3.7.1 125 | - mdit-py-plugins==0.3.3 126 | - mdurl==0.1.2 127 | - mkl-fft==1.3.1 128 | - mkl-random==1.2.2 129 | - mkl-service==2.4.0 130 | - multidict==6.0.4 131 | - multiprocess==0.70.14 132 | - mypy-extensions==1.0.0 133 | - ninja==1.11.1 134 | - nltk==3.8.1 135 | - numpy==1.23.5 136 | - oauthlib==3.2.2 137 | - openai==0.27.2 138 | - orjson==3.8.8 139 | - packaging==23.0 140 | - pandas==1.5.3 141 | - paramiko==3.1.0 142 | - pathtools==0.1.2 143 | - pillow==9.4.0 144 | - pip==23.0.1 145 | - plotly==5.13.1 146 | - protobuf==4.22.1 147 | - psutil==5.9.4 148 | - py-cpuinfo==9.0.0 149 | - pyarrow==11.0.0 150 | - pyasn1==0.4.8 151 | - pyasn1-modules==0.2.8 152 | - pycryptodome==3.17 153 | - pydantic==1.10.7 154 | - pydub==0.25.1 155 | - pynacl==1.5.0 156 | - pyopenssl==23.0.0 157 | - pyparsing==3.0.9 158 | - pyrsistent==0.19.3 159 | - pysocks==1.7.1 160 | - python-dateutil==2.8.2 161 | - python-multipart==0.0.6 162 | - pytz==2023.2 163 | - pyyaml==6.0 164 | - regex==2023.3.23 165 | - requests==2.28.1 166 | - requests-oauthlib==1.3.1 167 | - responses==0.18.0 168 | - rfc3986==1.5.0 169 | - rouge-score==0.1.2 170 | - rsa==4.9 171 | - semantic-version==2.10.0 172 | - sentencepiece==0.1.97 173 | - sentry-sdk==1.17.0 174 | - setproctitle==1.3.2 175 | - setuptools==65.6.3 176 | - smmap==5.0.0 177 | - sniffio==1.3.0 178 | - sqlalchemy==1.4.47 179 | - starlette==0.26.1 180 | - tenacity==8.2.2 181 | - tensorboard==2.12.0 182 | - tensorboard-data-server==0.7.0 183 | - tensorboard-plugin-wit==1.8.1 184 | - termcolor==2.2.0 185 | - tokenizers==0.12.1 186 | - toolz==0.12.0 187 | - torch==1.12.0 188 | - torchaudio==0.12.0 189 | - torchvision==0.13.0 190 | - tqdm==4.65.0 191 | - transformers==4.28.0.dev0 192 | - typing-extensions==4.4.0 193 | - typing-inspect==0.8.0 194 | - uc-micro-py==1.0.1 195 | - urllib3==1.26.14 196 | - uvicorn==0.21.1 197 | - wandb==0.14.0 198 | - websockets==10.4 199 | - werkzeug==2.2.3 200 | - wheel==0.38.4 201 | - xxhash==3.2.0 202 | - yarl==1.8.2 203 | prefix: /home/yourname/.conda/envs/chatpath 204 | -------------------------------------------------------------------------------- /src/generate.py: -------------------------------------------------------------------------------- 1 | import sys 2 | 3 | import fire 4 | import torch 5 | # from peft import PeftModel 6 | import transformers 7 | import gradio as gr 8 | 9 | assert ( 10 | "LlamaTokenizer" in transformers._import_structure["models.llama"] 11 | ), "LLaMA is now in HuggingFace's main branch.\nPlease reinstall it: pip uninstall transformers && pip install git+https://github.com/huggingface/transformers.git" 12 | from transformers import LlamaTokenizer, LlamaForCausalLM, GenerationConfig 13 | 14 | if torch.cuda.is_available(): 15 | device = "cuda" 16 | else: 17 | device = "cpu" 18 | 19 | try: 20 | if torch.backends.mps.is_available(): 21 | device = "mps" 22 | except: 23 | pass 24 | 25 | 26 | def main( 27 | load_8bit: bool = False, 28 | base_model: str = "/path/to/llama-7B/hf/ft/checkpoint-300", 29 | # lora_weights: str = "tloen/alpaca-lora-7b", 30 | ): 31 | assert base_model, ( 32 | "Please specify a --base_model, e.g. --base_model='decapoda-research/llama-7b-hf'" 33 | ) 34 | 35 | tokenizer = LlamaTokenizer.from_pretrained(base_model) 36 | if device == "cuda": 37 | model = LlamaForCausalLM.from_pretrained( 38 | base_model, 39 | load_in_8bit=load_8bit, 40 | torch_dtype=torch.float16, 41 | device_map="auto", 42 | ) 43 | elif device == "mps": 44 | model = LlamaForCausalLM.from_pretrained( 45 | base_model, 46 | device_map={"": device}, 47 | torch_dtype=torch.float16, 48 | ) 49 | 50 | # unwind broken decapoda-research config 51 | model.config.pad_token_id = tokenizer.pad_token_id = 0 # unk 52 | model.config.bos_token_id = 1 53 | model.config.eos_token_id = 2 54 | 55 | if not load_8bit: 56 | model.half() # seems to fix bugs for some users. 57 | 58 | model.eval() 59 | if torch.__version__ >= "2" and sys.platform != "win32": 60 | model = torch.compile(model) 61 | 62 | def evaluate( 63 | instruction, 64 | input=None, 65 | temperature=0.6, 66 | top_p=0.9, 67 | top_k=40, 68 | num_beams=4, 69 | max_new_tokens=512, 70 | **kwargs, 71 | ): 72 | prompt = generate_prompt(instruction, input) 73 | inputs = tokenizer(prompt, return_tensors="pt") 74 | input_ids = inputs["input_ids"].to(device) 75 | generation_config = GenerationConfig( 76 | temperature=temperature, 77 | top_p=top_p, 78 | top_k=top_k, 79 | num_beams=num_beams, 80 | **kwargs, 81 | ) 82 | with torch.no_grad(): 83 | generation_output = model.generate( 84 | input_ids=input_ids, 85 | generation_config=generation_config, 86 | return_dict_in_generate=True, 87 | output_scores=True, 88 | max_new_tokens=max_new_tokens, 89 | ) 90 | s = generation_output.sequences[0] 91 | output = tokenizer.decode(s) 92 | return output.split("### Response:")[1].strip() 93 | 94 | gr.Interface( 95 | fn=evaluate, 96 | inputs=[ 97 | gr.components.Textbox( 98 | lines=2, label="Instruction", placeholder="For pathology questions, you can use: In a genuine and professional manner, please assume the role of a pathologist and respond to the provided questions." 99 | ), 100 | gr.components.Textbox(lines=2, label="Input", placeholder="none"), 101 | gr.components.Slider(minimum=0, maximum=1, value=0.1, label="Temperature"), 102 | gr.components.Slider(minimum=0, maximum=1, value=0.75, label="Top p"), 103 | gr.components.Slider( 104 | minimum=0, maximum=100, step=1, value=40, label="Top k" 105 | ), 106 | gr.components.Slider(minimum=1, maximum=4, step=1, value=4, label="Beams"), 107 | gr.components.Slider( 108 | minimum=1, maximum=2000, step=1, value=128, label="Max tokens" 109 | ), 110 | ], 111 | outputs=[ 112 | gr.inputs.Textbox( 113 | lines=5, 114 | label="Output", 115 | ) 116 | ], 117 | title="ChatPath", 118 | description="A knowledgeable Llama-based Chat Model for Pathology.", 119 | ).launch(share=True) 120 | 121 | 122 | def generate_prompt(instruction, input=None): 123 | if input: 124 | return f"""Below is an instruction that describes a task, paired with an input that provides further context. Write a response that appropriately completes the request. 125 | 126 | ### Instruction: 127 | {instruction} 128 | 129 | ### Input: 130 | {input} 131 | 132 | ### Response: 133 | """ 134 | else: 135 | return f"""Below is an instruction that describes a task. Write a response that appropriately completes the request. 136 | 137 | ### Instruction: 138 | {instruction} 139 | 140 | ### Response: 141 | """ 142 | 143 | 144 | if __name__ == "__main__": 145 | fire.Fire(main) -------------------------------------------------------------------------------- /src/imgs/chatpath_logo.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/superjamessyx/PathGPT/dc170d26d1df8d7cac0492de7f02c84cf4520728/src/imgs/chatpath_logo.png -------------------------------------------------------------------------------- /src/imgs/data_process.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/superjamessyx/PathGPT/dc170d26d1df8d7cac0492de7f02c84cf4520728/src/imgs/data_process.png -------------------------------------------------------------------------------- /src/imgs/pathgpt_instruction.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/superjamessyx/PathGPT/dc170d26d1df8d7cac0492de7f02c84cf4520728/src/imgs/pathgpt_instruction.png -------------------------------------------------------------------------------- /src/imgs/pathgpt_logo.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/superjamessyx/PathGPT/dc170d26d1df8d7cac0492de7f02c84cf4520728/src/imgs/pathgpt_logo.png -------------------------------------------------------------------------------- /src/train.py: -------------------------------------------------------------------------------- 1 | # Copyright 2023 Rohan Taori, Ishaan Gulrajani, Tianyi Zhang, Yann Dubois, Xuechen Li 2 | # 3 | # Licensed under the Apache License, Version 2.0 (the "License"); 4 | # you may not use this file except in compliance with the License. 5 | # You may obtain a copy of the License at 6 | # 7 | # http://www.apache.org/licenses/LICENSE-2.0 8 | # 9 | # Unless required by applicable law or agreed to in writing, software 10 | # distributed under the License is distributed on an "AS IS" BASIS, 11 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 | # See the License for the specific language governing permissions and 13 | # limitations under the License. 14 | 15 | import copy 16 | import logging 17 | import random 18 | from dataclasses import dataclass, field 19 | from typing import Optional, Dict, Sequence 20 | 21 | import torch 22 | import torch.distributed 23 | import transformers 24 | from torch.utils.data import Dataset 25 | from transformers import Trainer 26 | from datasets import load_dataset 27 | import utils 28 | 29 | IGNORE_INDEX = -100 30 | DEFAULT_PAD_TOKEN = "[PAD]" 31 | DEFAULT_EOS_TOKEN = "" 32 | DEFAULT_BOS_TOKEN = "" 33 | DEFAULT_UNK_TOKEN = "" 34 | PROMPT_DICT = { 35 | "prompt_input": ( 36 | "Below is an instruction that describes a task, paired with an input that provides further context. " 37 | "Write a response that appropriately completes the request.\n\n" 38 | "### Instruction:\n{instruction}\n\n### Input:\n{input}\n\n### Response:" 39 | ), 40 | "prompt_no_input": ( 41 | "Below is an instruction that describes a task. " 42 | "Write a response that appropriately completes the request.\n\n" 43 | "### Instruction:\n{instruction}\n\n### Response:" 44 | ), 45 | } 46 | 47 | 48 | @dataclass 49 | class ModelArguments: 50 | model_name_or_path: Optional[str] = field(default="facebook/opt-125m") 51 | 52 | 53 | @dataclass 54 | class DataArguments: 55 | data_path: str = field(default=None, metadata={"help": "Path to the training data."}) 56 | 57 | 58 | @dataclass 59 | class TrainingArguments(transformers.TrainingArguments): 60 | cache_dir: Optional[str] = field(default=None) 61 | optim: str = field(default="adamw_torch") 62 | model_max_length: int = field( 63 | default=512, 64 | metadata={"help": "Maximum sequence length. Sequences will be right padded (and possibly truncated)."}, 65 | ) 66 | 67 | 68 | def safe_save_model_for_hf_trainer(trainer: transformers.Trainer, output_dir: str): 69 | """Collects the state dict and dump to disk.""" 70 | state_dict = trainer.model.state_dict() 71 | if trainer.args.should_save: 72 | cpu_state_dict = {key: value.cpu() for key, value in state_dict.items()} 73 | del state_dict 74 | trainer._save(output_dir, state_dict=cpu_state_dict) # noqa 75 | 76 | 77 | def smart_tokenizer_and_embedding_resize( 78 | special_tokens_dict: Dict, 79 | tokenizer: transformers.PreTrainedTokenizer, 80 | model: transformers.PreTrainedModel, 81 | ): 82 | """Resize tokenizer and embedding. 83 | 84 | Note: This is the unoptimized version that may make your embedding size not be divisible by 64. 85 | """ 86 | num_new_tokens = tokenizer.add_special_tokens(special_tokens_dict) 87 | model.resize_token_embeddings(len(tokenizer)) 88 | 89 | if num_new_tokens > 0: 90 | input_embeddings = model.get_input_embeddings().weight.data 91 | output_embeddings = model.get_output_embeddings().weight.data 92 | 93 | input_embeddings_avg = input_embeddings[:-num_new_tokens].mean(dim=0, keepdim=True) 94 | output_embeddings_avg = output_embeddings[:-num_new_tokens].mean(dim=0, keepdim=True) 95 | 96 | input_embeddings[-num_new_tokens:] = input_embeddings_avg 97 | output_embeddings[-num_new_tokens:] = output_embeddings_avg 98 | 99 | 100 | def _tokenize_fn(strings: Sequence[str], tokenizer: transformers.PreTrainedTokenizer) -> Dict: 101 | """Tokenize a list of strings.""" 102 | tokenized_list = [ 103 | tokenizer( 104 | text, 105 | return_tensors="pt", 106 | padding="longest", 107 | max_length=tokenizer.model_max_length, 108 | truncation=True, 109 | ) 110 | for text in strings 111 | ] 112 | input_ids = labels = [tokenized.input_ids[0] for tokenized in tokenized_list] 113 | input_ids_lens = labels_lens = [ 114 | tokenized.input_ids.ne(tokenizer.pad_token_id).sum().item() for tokenized in tokenized_list 115 | ] 116 | return dict( 117 | input_ids=input_ids, 118 | labels=labels, 119 | input_ids_lens=input_ids_lens, 120 | labels_lens=labels_lens, 121 | ) 122 | 123 | 124 | def preprocess( 125 | sources: Sequence[str], 126 | targets: Sequence[str], 127 | tokenizer: transformers.PreTrainedTokenizer, 128 | ) -> Dict: 129 | """Preprocess the data by tokenizing.""" 130 | examples = [s + t for s, t in zip(sources, targets)] 131 | examples_tokenized, sources_tokenized = [_tokenize_fn(strings, tokenizer) for strings in (examples, sources)] 132 | input_ids = examples_tokenized["input_ids"] 133 | labels = copy.deepcopy(input_ids) 134 | for label, source_len in zip(labels, sources_tokenized["input_ids_lens"]): 135 | label[:source_len] = IGNORE_INDEX 136 | return dict(input_ids=input_ids, labels=labels) 137 | 138 | 139 | @dataclass 140 | class DataCollatorForSupervisedDataset(object): 141 | """Collate examples for supervised fine-tuning.""" 142 | 143 | tokenizer: transformers.PreTrainedTokenizer 144 | 145 | def __call__(self, instances: Sequence[Dict]) -> Dict[str, torch.Tensor]: 146 | input_ids, labels = tuple([instance[key] for instance in instances] for key in ("input_ids", "labels")) 147 | input_ids = [torch.tensor(x) for x in input_ids] 148 | input_ids = torch.nn.utils.rnn.pad_sequence( 149 | input_ids, batch_first=True, padding_value=self.tokenizer.pad_token_id 150 | ) 151 | labels = [torch.tensor(x) for x in labels] 152 | labels = torch.nn.utils.rnn.pad_sequence(labels, batch_first=True, padding_value=IGNORE_INDEX) 153 | return dict( 154 | input_ids=input_ids, 155 | labels=labels, 156 | attention_mask=input_ids.ne(self.tokenizer.pad_token_id), 157 | ) 158 | 159 | def train_tokenize_function(examples, tokenizer): 160 | prompt_input, prompt_no_input = PROMPT_DICT["prompt_input"], PROMPT_DICT["prompt_no_input"] 161 | if 'input' in examples: 162 | sources = [ 163 | prompt_input.format_map(dict(instruction=instruction, input=input)) if input != "" \ 164 | else prompt_no_input.format_map(dict(instruction=instruction)) \ 165 | for instruction, input in zip(examples['instruction'], examples['input']) 166 | ] 167 | else: 168 | sources = [ 169 | prompt_no_input.format_map(dict(instruction=instruction)) \ 170 | for instruction in examples['instruction'] 171 | ] 172 | targets = [f"{output}{tokenizer.eos_token}" for output in examples['output']] 173 | data_dict = preprocess(sources, targets, tokenizer) 174 | return data_dict 175 | 176 | def train(): 177 | parser = transformers.HfArgumentParser((ModelArguments, DataArguments, TrainingArguments)) 178 | model_args, data_args, training_args = parser.parse_args_into_dataclasses() 179 | 180 | model = transformers.AutoModelForCausalLM.from_pretrained( 181 | model_args.model_name_or_path, 182 | cache_dir=training_args.cache_dir, 183 | ) 184 | 185 | tokenizer = transformers.AutoTokenizer.from_pretrained( 186 | model_args.model_name_or_path, 187 | cache_dir=training_args.cache_dir, 188 | model_max_length=training_args.model_max_length, 189 | padding_side="right", 190 | use_fast=True, 191 | ) 192 | if tokenizer.pad_token is None: 193 | smart_tokenizer_and_embedding_resize( 194 | special_tokens_dict=dict(pad_token=DEFAULT_PAD_TOKEN), 195 | tokenizer=tokenizer, 196 | model=model, 197 | ) 198 | if "llama" in model_args.model_name_or_path: 199 | tokenizer.add_special_tokens( 200 | { 201 | "eos_token": DEFAULT_EOS_TOKEN, 202 | "bos_token": DEFAULT_BOS_TOKEN, 203 | "unk_token": DEFAULT_UNK_TOKEN, 204 | } 205 | ) 206 | 207 | raw_train_datasets = load_dataset('json', data_files=data_args.data_path, split="train", cache_dir=training_args.cache_dir) 208 | if training_args.local_rank > 0: 209 | torch.distributed.barrier() 210 | 211 | train_dataset = raw_train_datasets.map( 212 | train_tokenize_function, 213 | batched=True, 214 | batch_size=3000, 215 | num_proc=32, 216 | remove_columns=raw_train_datasets.column_names, 217 | load_from_cache_file=True, # not args.overwrite_cache 218 | desc="Running tokenizer on train dataset", 219 | fn_kwargs={"tokenizer": tokenizer} 220 | ) 221 | 222 | if training_args.local_rank == 0: 223 | torch.distributed.barrier() 224 | 225 | if training_args.local_rank == 0: 226 | print(len(train_dataset)) 227 | for index in random.sample(range(len(train_dataset)), 3): 228 | print(f"Sample {index} of the training set: {train_dataset[index]}.") 229 | 230 | data_collator = DataCollatorForSupervisedDataset(tokenizer=tokenizer) 231 | data_module = dict(train_dataset=train_dataset, eval_dataset=None, data_collator=data_collator) 232 | 233 | #Tell Trainer not to attempt DataParallel 234 | model.is_parallelizable = True 235 | model.model_parallel = True 236 | 237 | trainer = Trainer(model=model, tokenizer=tokenizer, args=training_args, **data_module) 238 | model.config.use_cache = False 239 | 240 | trainer.train() 241 | trainer.save_state() 242 | safe_save_model_for_hf_trainer(trainer=trainer, output_dir=training_args.output_dir) 243 | 244 | 245 | if __name__ == "__main__": 246 | train() 247 | -------------------------------------------------------------------------------- /src/utils.py: -------------------------------------------------------------------------------- 1 | import dataclasses 2 | import logging 3 | import math 4 | import os 5 | import io 6 | import sys 7 | import time 8 | import json 9 | from typing import Optional, Sequence, Union 10 | 11 | import openai 12 | import tqdm 13 | from openai import openai_object 14 | import copy 15 | 16 | StrOrOpenAIObject = Union[str, openai_object.OpenAIObject] 17 | 18 | openai_org = os.getenv("OPENAI_ORG") 19 | if openai_org is not None: 20 | openai.organization = openai_org 21 | logging.warning(f"Switching to organization: {openai_org} for OAI API key.") 22 | 23 | 24 | @dataclasses.dataclass 25 | class OpenAIDecodingArguments(object): 26 | max_tokens: int = 1800 27 | temperature: float = 0.2 28 | top_p: float = 1.0 29 | n: int = 1 30 | stream: bool = False 31 | stop: Optional[Sequence[str]] = None 32 | presence_penalty: float = 0.0 33 | frequency_penalty: float = 0.0 34 | suffix: Optional[str] = None 35 | logprobs: Optional[int] = None 36 | echo: bool = False 37 | 38 | 39 | def openai_completion( 40 | prompts: Union[str, Sequence[str], Sequence[dict[str, str]], dict[str, str]], 41 | decoding_args: OpenAIDecodingArguments, 42 | model_name="text-davinci-003", 43 | sleep_time=2, 44 | batch_size=1, 45 | max_instances=sys.maxsize, 46 | max_batches=sys.maxsize, 47 | return_text=False, 48 | **decoding_kwargs, 49 | ) -> Union[Union[StrOrOpenAIObject], Sequence[StrOrOpenAIObject], Sequence[Sequence[StrOrOpenAIObject]],]: 50 | """Decode with OpenAI API. 51 | 52 | Args: 53 | prompts: A string or a list of strings to complete. If it is a chat model the strings should be formatted 54 | as explained here: https://github.com/openai/openai-python/blob/main/chatml.md. If it is a chat model 55 | it can also be a dictionary (or list thereof) as explained here: 56 | https://github.com/openai/openai-cookbook/blob/main/examples/How_to_format_inputs_to_ChatGPT_models.ipynb 57 | decoding_args: Decoding arguments. 58 | model_name: Model name. Can be either in the format of "org/model" or just "model". 59 | sleep_time: Time to sleep once the rate-limit is hit. 60 | batch_size: Number of prompts to send in a single request. Only for non chat model. 61 | max_instances: Maximum number of prompts to decode. 62 | max_batches: Maximum number of batches to decode. This argument will be deprecated in the future. 63 | return_text: If True, return text instead of full completion object (which contains things like logprob). 64 | decoding_kwargs: Additional decoding arguments. Pass in `best_of` and `logit_bias` if you need them. 65 | 66 | Returns: 67 | A completion or a list of completions. 68 | Depending on return_text, return_openai_object, and decoding_args.n, the completion type can be one of 69 | - a string (if return_text is True) 70 | - an openai_object.OpenAIObject object (if return_text is False) 71 | - a list of objects of the above types (if decoding_args.n > 1) 72 | """ 73 | is_single_prompt = isinstance(prompts, (str, dict)) 74 | if is_single_prompt: 75 | prompts = [prompts] 76 | 77 | if max_batches < sys.maxsize: 78 | logging.warning( 79 | "`max_batches` will be deprecated in the future, please use `max_instances` instead." 80 | "Setting `max_instances` to `max_batches * batch_size` for now." 81 | ) 82 | max_instances = max_batches * batch_size 83 | 84 | prompts = prompts[:max_instances] 85 | num_prompts = len(prompts) 86 | prompt_batches = [ 87 | prompts[batch_id * batch_size : (batch_id + 1) * batch_size] 88 | for batch_id in range(int(math.ceil(num_prompts / batch_size))) 89 | ] 90 | 91 | completions = [] 92 | for batch_id, prompt_batch in tqdm.tqdm( 93 | enumerate(prompt_batches), 94 | desc="prompt_batches", 95 | total=len(prompt_batches), 96 | ): 97 | batch_decoding_args = copy.deepcopy(decoding_args) # cloning the decoding_args 98 | 99 | while True: 100 | try: 101 | shared_kwargs = dict( 102 | model=model_name, 103 | **batch_decoding_args.__dict__, 104 | **decoding_kwargs, 105 | ) 106 | completion_batch = openai.Completion.create(prompt=prompt_batch, **shared_kwargs) 107 | choices = completion_batch.choices 108 | 109 | for choice in choices: 110 | choice["total_tokens"] = completion_batch.usage.total_tokens 111 | completions.extend(choices) 112 | break 113 | except openai.error.OpenAIError as e: 114 | logging.warning(f"OpenAIError: {e}.") 115 | if "Please reduce your prompt" in str(e): 116 | batch_decoding_args.max_tokens = int(batch_decoding_args.max_tokens * 0.8) 117 | logging.warning(f"Reducing target length to {batch_decoding_args.max_tokens}, Retrying...") 118 | else: 119 | logging.warning("Hit request rate limit; retrying...") 120 | time.sleep(sleep_time) # Annoying rate limit on requests. 121 | 122 | if return_text: 123 | completions = [completion.text for completion in completions] 124 | if decoding_args.n > 1: 125 | # make completions a nested list, where each entry is a consecutive decoding_args.n of original entries. 126 | completions = [completions[i : i + decoding_args.n] for i in range(0, len(completions), decoding_args.n)] 127 | if is_single_prompt: 128 | # Return non-tuple if only 1 input and 1 generation. 129 | (completions,) = completions 130 | return completions 131 | 132 | 133 | def _make_w_io_base(f, mode: str): 134 | if not isinstance(f, io.IOBase): 135 | f_dirname = os.path.dirname(f) 136 | if f_dirname != "": 137 | os.makedirs(f_dirname, exist_ok=True) 138 | f = open(f, mode=mode) 139 | return f 140 | 141 | 142 | def _make_r_io_base(f, mode: str): 143 | if not isinstance(f, io.IOBase): 144 | f = open(f, mode=mode) 145 | return f 146 | 147 | 148 | def jdump(obj, f, mode="w", indent=4, default=str): 149 | """Dump a str or dictionary to a file in json format. 150 | 151 | Args: 152 | obj: An object to be written. 153 | f: A string path to the location on disk. 154 | mode: Mode for opening the file. 155 | indent: Indent for storing json dictionaries. 156 | default: A function to handle non-serializable entries; defaults to `str`. 157 | """ 158 | f = _make_w_io_base(f, mode) 159 | if isinstance(obj, (dict, list)): 160 | json.dump(obj, f, indent=indent, default=default) 161 | elif isinstance(obj, str): 162 | f.write(obj) 163 | else: 164 | raise ValueError(f"Unexpected type: {type(obj)}") 165 | f.close() 166 | 167 | 168 | def jload(f, mode="r"): 169 | """Load a .json file into a dictionary.""" 170 | f = _make_r_io_base(f, mode) 171 | jdict = json.load(f) 172 | f.close() 173 | return jdict 174 | --------------------------------------------------------------------------------