ts-model-refining/OneFlow-KAE

• 模型炼知（Model Knowledge Amalgamation, KA）
• • Introduction
  • Table of contents
  • Algorithms
  • • • 0. 数据集
      • 1. 依赖包
      • 2. 算法
      • 3. 权重
  • Authors

模型炼知（Model Knowledge Amalgamation, KA）

Introduction

模型炼知是用于知识融合和模型迁移性度量,而建立的轻量级算法包,详情请见 KAE。

Table of contents

• Introduction
• Algorithms
• Authors

Algorithms

0. 数据集

• data/StanfordCars: https://ai.stanford.edu/~jkrause/cars/car_dataset.html
• data/StanfordDogs: http://vision.stanford.edu/aditya86/ImageNetDogs/
• data/FGVCAircraft: http://www.robots.ox.ac.uk/~vgg/data/fgvc-aircraft/
• data/Flowers102: http://www.robots.ox.ac.uk/~vgg/data/fgvc-aircraft/

1. 依赖包

• requirements.txt

2. 算法

• TW_Layerwise

   python3 TTL.py --car_ckpt ./ckpt/car_res50_model stanford_dogs  --dog_ckpt ./ckpt/dog_res50_model

• TH_Layerwise

   python3 THL.py --car_ckpt ./ckpt/car_res50_model stanford_dogs  --dog_ckpt ./ckpt/dog_res50_model --aircraft_ckpt ./ckpt/aircraft_res50_model

• TF_Layerwise

  python3 TFL.py --car_ckpt ./ckpt/car_res50_model stanford_dogs  --dog_ckpt ./ckpt/dog_res50_model --aircraft_ckpt ./ckpt/aircraft_res50_model --flower_ckpt ./ckpt/flower_res50_model

• flow

   pytorch weights convert oneflow

3. 权重

 文件夹MODEL_ckpt保存的是三个模型重组/炼知算法的权重,oneflow格式,供复现时参考。ckpt文件夹为pytorch权重转为oneflow权重,用于教师模型加载权重，内存占用较大，改为放在下方链接。

• 教师模型权重 [链接]：https://pan.baidu.com/s/19pRLZKQvjgEQ7CuD_RiPPw 提取码：FLOW

Authors

This project is simplified by KAE, which is developed by VIPA Lab from Zhejiang University and Zhejiang Lab