Knowledge-Distillation

Pytorch implementation of various Knowledge Distillation (KD) methods.

Lists

Name	Method	Paper Link	Code Link
Baseline	basic model with softmax loss	—	code
ST	soft target	paper	code
AT	attention transfer	paper	code
Fitnet	hints for thin deep nets	paper	code
NST	neural selective transfer	paper	code
FT	factor transfer	paper	code
RKD	relational knowledge distillation	paper	code

Note, there are some differences between this repository and the original papers：
- For AT: I use the sum of absolute values with power p=2 as the attention.
- For Fitnet: The training procedure is one stage without hint layer.
- For NST: I employ polynomial kernel with d=2 and c=0.

Create ./dataset directory and download CIFAR10/CIFAR100 in it.
You can simply specify the hyper-parameters listed in train_xxx.py or manually change them.
- Use train_base.py to train the teacher model in KD and then save the model.
- Before traning, you can choose the method you need in ./kd_losses directory, and run train_kd.py to train the student model.

This repo is partly based on the following repos, thank the authors a lot.

HobbitLong/RepDistiller
bhheo/BSS_distillation
clovaai/overhaul-distillation
passalis/probabilistic_kt
lenscloth/RKD
[AberHu/Knowledge-Distillation-Zoo]（https://github.com/AberHu/Knowledge-Distillation-Zoo）

If you employ the listed KD methods in your research, please cite the corresponding papers.