This repo is the implementation of paper Knowledge Distillation Meets Self-Supervision (ECCV 2020).
This repo is tested with Ubuntu 16.04.5, Python 3.7, PyTorch 1.5.0, CUDA 10.2. Make sure to install pytorch, torchvision, tensorboardX, numpy before using this repo.
An example of teacher training is:
python teacher.py --arch wrn_40_2 --lr 0.05 --gpu-id 0
where you can specify the architecture via flag --arch
You can also download all the pre-trained teacher models here.
If you want to run student.py
directly, you have to re-organise the directory. For instance, when you download vgg13.pth, you have to make a directory for it, say teacher_vgg13, and then make a new directory ckpt inside teacher_vgg13. Move the vgg13.pth into teacher_vgg13/ckpt and rename it as best.pth. If you want a simpler way to use pre-trained model, you can edit the code in student.py
(line 90).
An example of student training is:
python student.py --t-path ./experiments/teacher_wrn_40_2_seed0/ --s-arch wrn_16_2 --lr 0.05 --gpu-id 0
The meanings of flags are:
--t-path
: teacher's checkpoint path. Automatically search the checkpoint containing 'best' keyword in its name.
--s-arch
: student's architecture.
All the commands can be found in command.sh
Teacher Student |
wrn40-2 wrn16-2 |
wrn40-2 wrn40-1 |
resnet56 resnet20 |
resnet32x4 resnet8x4 |
vgg13 vgg8 |
---|---|---|---|---|---|
Teacher Student |
76.46 73.64 |
76.46 72.24 |
73.44 69.63 |
79.63 72.51 |
75.38 70.68 |
KD | 74.92 | 73.54 | 70.66 | 73.33 | 72.98 |
FitNet | 75.75 | 74.12 | 71.60 | 74.31 | 73.54 |
AT | 75.28 | 74.45 | 71.78 | 74.26 | 73.62 |
SP | 75.34 | 73.15 | 71.48 | 74.74 | 73.44 |
VID | 74.79 | 74.20 | 71.71 | 74.82 | 73.96 |
RKD | 75.40 | 73.87 | 71.48 | 74.47 | 73.72 |
PKT | 76.01 | 74.40 | 71.44 | 74.17 | 73.37 |
AB | 68.89 | 75.06 | 71.49 | 74.45 | 74.27 |
FT | 75.15 | 74.37 | 71.52 | 75.02 | 73.42 |
CRD | 76.04 | 75.52 | 71.68 | 75.90 | 74.06 |
SSKD | 76.04 | 76.13 | 71.49 | 76.20 | 75.33 |
Teacher Student |
vgg13 MobieleNetV2 |
ResNet50 MobileNetV2 |
ResNet50 vgg8 |
resnet32x4 ShuffleV1 |
resnet32x4 ShuffleV2 |
wrn40-2 ShuffleV1 |
---|---|---|---|---|---|---|
Teacher Student |
75.38 65.79 |
79.10 65.79 |
79.10 70.68 |
79.63 70.77 |
79.63 73.12 |
76.46 70.77 |
KD | 67.37 | 67.35 | 73.81 | 74.07 | 74.45 | 74.83 |
FitNet | 68.58 | 68.54 | 73.84 | 74.82 | 75.11 | 75.55 |
AT | 69.34 | 69.28 | 73.45 | 74.76 | 75.30 | 75.61 |
SP | 66.89 | 68.99 | 73.86 | 73.80 | 75.15 | 75.56 |
VID | 66.91 | 68.88 | 73.75 | 74.28 | 75.78 | 75.36 |
RKD | 68.50 | 68.46 | 73.73 | 74.20 | 75.74 | 75.45 |
PKT | 67.89 | 68.44 | 73.53 | 74.06 | 75.18 | 75.51 |
AB | 68.86 | 69.32 | 74.20 | 76.24 | 75.66 | 76.58 |
FT | 69.19 | 69.01 | 73.58 | 74.31 | 74.95 | 75.18 |
CRD | 68.49 | 70.32 | 74.42 | 75.46 | 75.72 | 75.96 |
SSKD | 71.53 | 72.57 | 75.76 | 78.44 | 78.61 | 77.40 |
If you find this repo useful for your research, please consider citing the paper
@inproceedings{xu2020knowledge,
title={Knowledge Distillation Meets Self-Supervision},
author={Xu, Guodong and Liu, Ziwei and Li, Xiaoxiao and Loy, Chen Change},
booktitle={European Conference on Computer Vision (ECCV)},
year={2020},
}
The implementation of models
is borrowed from CRD