https://github.com/oneflow-inc/oneflow-model-compression

https://github.com/oneflow-inc/oneflow-model-compression

Last synced: 12 months ago
JSON representation

• Host: GitHub
• URL: https://github.com/oneflow-inc/oneflow-model-compression
• Owner: Oneflow-Inc
• License: apache-2.0
• Created: 2021-03-15T05:51:16.000Z (about 5 years ago)
• Default Branch: main
• Last Pushed: 2021-08-16T04:24:17.000Z (almost 5 years ago)
• Last Synced: 2025-03-29T17:12:38.512Z (about 1 year ago)
• Language: Python
• Size: 1.33 MB
• Stars: 7
• Watchers: 50
• Forks: 4
• Open Issues: 0
• Metadata Files:
  • Readme: README.md
  • License: LICENSE

Awesome Lists containing this project

README

          
# Oneflow-Model-Compression

## 概述

炼知技术平台是一个模型压缩平台，包含剪枝、量化、知识蒸馏等一系列模型压缩策略。

提供完整的模型压缩解决方案，可用于各种类型的自然语言和计算机视觉场景，如文本分类、推理，图像分类等。

另外，平台在不断完善各种压缩策略在经典开源任务的Benchmark，以便用户参考。

同时，平台也提供各种压缩策略的功能算子，方便用户使用、复现最新的论文方法，以及利用压缩算子进行二次开发。



    


    

    




## 功能



	

		

			

				功能模块

			

			

				算法

			

			

				相关文档

			

		

		

			

				量化

			

			

				


					


                        deep compression: Han S, Mao H, Dally W J. "Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding" arXiv preprint arXiv:1510.00149 (2017).

					

					


                        NVIDIA TensorRT: a C++ library for high performance inference on NVIDIA GPUs and deep learning accelerators.  

					
					

				

			

			

				


					


						量化API文档

					

                    


						量化快速上手

					

				

			

		

		

			

				剪枝


			

			

				


					


                        bn channel slimming: Zhuang Liu, Jianguo Li, Zhiqiang Shen. "Learning Efficient Convolutional Networks through Network Slimming" arXiv preprint arXiv:1708.06519 (2017).

					

                    


                        conv channel slimming: Hao Li, Asim Kadav, Igor Durdanovic. "Pruning Filters for Efficient ConvNets" arXiv preprint arXiv:1608.08710 (2016).

					

                    


                        conv channel slimming: Hengyuan Hu, Rui Peng, Yu-Wing Tai. "Network Trimming: A Data-Driven Neuron Pruning Approach towards Efficient Deep Architectures" arXiv preprint arXiv:1607.03250 (2016).

					

				

			

			

					


						


							剪枝API文档

						

					

					


						


							剪枝快速上手

						

					

			

		

		

			

				知识蒸馏

			

			

				


					


						Knowledge Distillation: Hinton, Geoffrey, Oriol Vinyals, and Jeff Dean. "Distilling the knowledge in a neural network." arXiv preprint arXiv:1503.02531 (2015).

					

					


						Distilled-BiLSTM:  Tang, Raphael, et al. "Distilling task-specific knowledge from bert into simple neural networks." arXiv preprint arXiv:1903.12136 (2019).

					

					


						BERT-PKD:  Sun, Siqi, et al. "Patient knowledge distillation for bert model compression." arXiv preprint arXiv:1908.09355 (2019).

					

					


						TinyBERT: Jiao, Xiaoqi, et al. "Tinybert: Distilling bert for natural language understanding." arXiv preprint arXiv:1909.10351 (2019).

					

					


						MobileBERT: Sun, Zhiqing, et al. "Mobilebert: a compact task-agnostic bert for resource-limited devices." arXiv preprint arXiv:2004.02984 (2020).

					

					


						BERT-Theseus: Xu, Canwen, et al. "Bert-of-theseus: Compressing bert by progressive module replacing." arXiv preprint arXiv:2002.02925 (2020).

					

					


						改进版的BERT-Theseus: Xu, Canwen, et al. "Bert-of-theseus: Compressing bert by progressive module replacing." arXiv preprint arXiv:2002.02925 (2020).

					

				

			

			

				


					


						知识蒸馏API文档

					

					


					    知识蒸馏快速上手

                   

                   


					    Knowledge Distillation算法文档

                    

                    


					    Distilled-BiLSTM算法文档

                    

                    


					    BERT-PKD算法文档

                    

                    


					    TinyBERT算法文档

                    

                    


					    BERT-Theseus算法文档

					

				

			

		

	

>





## 使用

- Oneflow介绍： 深度学习框架Oneflow[介绍以及环境安装说明](https://github.com/Oneflow-Inc/oneflow)。

- Oneflow快速开始：通过[简单示例](http://docs.oneflow.org/quick_start/quickstart_in_3_min.html)介绍如何快速3分钟上手使用Oneflow。

- 模型压缩API文档：用户接口文档，包含以下功能

    - [量化](./docs/API_quant.md)

    - [剪枝](./docs/API_prune.md)

    - [知识蒸馏](./docs/API_knowledge_distill.md)

- 高阶教程：包括在CV和NLP等应用场景任务的使用示例、算法使用步骤，高级特性的使用教程。

    - 量化功能文档: 介绍量化功能[使用示例](./model_compress/quantization/tutorial.md)，主要包含int8量化。

    - 剪枝功能文档:  介绍通道剪枝实现和[使用示例](./model_compress/ChannelSlimming/readme.md)，只要包括CNN模型、DNN模型的不同剪枝算子。

    - [知识蒸馏功能](./model_compress/distil)文档: 介绍知识蒸馏功能相关论文实现和使用示例，主要包含[KD](./model_compress/distil/examples/knowledge_distillation/README.md), [Distilled-BiLSTM](./model_compress/distil/examples/distilled-bilstm/README.md), [BERT-PKD](./model_compress/distil/examples/bert-pkd/README.md), [TinyBERT](./model_compress/distil/examples/tinybert/README.md), [BERT-Theseus](./model_compress/distil/theseus/README.md)等算法。

    - [TensorRT量化部署](./docs/API_quant.md): 介绍如何使用TensorRT部署量化得到的Oneflow模型。

- [模型库](./docs/model_zoo.md)：各个压缩算法在文本分类、推理，图像分类等数据集上的实验结果，包括模型精度、模型尺寸和推理速度。