NetGraft

Abstract:

Knowledge distillation has demonstrated encouraging performances in deep model compression. Most existing approaches, however, require massive labeled data to accomplish the knowledge transfer, making the model compression a cumbersome and costly process. In this paper, we investigate the practical few-shot knowledge distillation scenario, where we assume only a few samples without human annotations are available for each category. To this end, we introduce a principled dual-stage distillation scheme tailored for few-shot data. In the first step, we graft the student blocks one by one onto the teacher, and learn the parameters of the grafted block intertwined with those of the other teacher blocks. In the second step, the trained student blocks are progressively connected and then together grafted onto the teacher network, allowing the learned student blocks to adapt themselves to each other and eventually replace the teacher network. Experiments demonstrate that our approach, with only a few unlabeled samples, achieves gratifying results on CIFAR10, CIFAR100, and ILSVRC-2012. On CIFAR10 and CIFAR100, our performances are even on par with those of knowledge distillation schemes that utilize the full datasets.

@inproceedings{shen2021progressive,
  author    = {Shen, Chengchao and Wang, Xinchao and Yin, Youtan and Song, Jie and Luo, Sihui and Song, Mingli},
  title     = {Progressive Network Grafting for Few-Shot Knowledge Distillation},
  booktitle = {AAAI Conference on Artificial Intelligence (AAAI)},
  year      = {2021}
}

Requirements

To install requirements:

conda create -n netgraft python=3.7
pip install -r requirements.txt

Pre-trained Teacher Models

You can download pretrained teacher models here (Github Releases):

Teacher model trained on full CIFAR10.
Teacher model trained on full CIFAR100.

Note: put the pre-trained teacher models in the directory: ckpt/teacher/

Dataset Preparation

To download and build datasets (CIFAR10 and CIFAR100) for few-shot distillation, run this command:

python build_dataset.py

Training

To train the model(s) in the paper, run this command:

# ----------- Run on CIFAR10 -----------
python train.py --dataset CIFAR10 # Training [1~10, 20, 50]-Shot Distillation 

# ----------- Run on CIFAR100 -----------
python train.py --dataset CIFAR100 # Training [1~10, 20, 50]-Shot Distillation

Evaluation

To evaluate my model on CIFAR10 and CIFAR100, run:

# ----------- Run on CIFAR10 -----------
python evaluate.py --dataset='CIFAR10' --nshot=1 # 1-Shot Distillation
python evaluate.py --dataset='CIFAR10' --nshot=5 # 5-Shot Distillation
python evaluate.py --dataset='CIFAR10' --nshot=10 # 10-Shot Distillation

# ----------- Run on CIFAR100 -----------
python evaluate.py --dataset='CIFAR100' --nshot=1 # 1-Shot Distillation
python evaluate.py --dataset='CIFAR100' --nshot=5 # 5-Shot Distillation
python evaluate.py --dataset='CIFAR100' --nshot=10 # 10-Shot Distillation

Experimental Results

Few-Shot Distillation on CIFAR10 and CIFAR100

N-Shot	Accuracy on CIFAR10 (%)	Accuracy on CIFAR100 (%)
1	90.74 ± 0.49	64.22 ± 0.17
2	92.60 ± 0.06	66.51 ± 0.11
3	92.70 ± 0.07	67.35 ± 0.10
4	92.77 ± 0.04	67.69 ± 0.03
5	92.88 ± 0.07	68.16 ± 0.20
6	92.84 ± 0.08	68.38 ± 0.11
7	92.77 ± 0.05	68.46 ± 0.10
8	92.83 ± 0.06	68.78 ± 0.22
9	92.88 ± 0.05	68.77 ± 0.10
10	92.89 ± 0.06	68.86 ± 0.03
20	92.78 ± 0.09	69.04 ± 0.08
50	92.76 ± 0.09	69.06 ± 0.10

zju-vipa / NetGraft