Our work suggests that: when a neural network is too powerful to find stronger teacher models, or computation resource is limited to train teacher models, "self-training" or "manually-designed regularization" can be applied.

For example, ResNeXt101-32x8d is a powerful model with 88.79M parameters and 16.51G FLOPs on ImageNet, and it is hard or computation expensive to train a stronger teacher model for this student. Our strategy can further improve this powerful student model by 0.48% without extra computation on ImageNet. Similarly, when taking a powerful single model ResNeXt29-8x64d with 34.53M parameters as a student model, our self-training implementation achieves more than 1.0% improvement on CIFAR100 (from 81.03% to 82.08%).

Clone this repository:

git clone https://github.com/yuanli2333/Teacher-free-Knowledge-Distillation.git

Build a new environment and install:

pip install -r requirements.txt

Better use: NVIDIA GPU + CUDA9.0 + Pytorch 1.2.0

Please do not use other versions of pytorch, otherwise, some experiment results may not be reproduced because some slight difference would make the hyper-parameters different.

CIFAR100, CIFAR10 and Tiny_ImageNet; For CIFAR100 and CIFAR10, our codes will download the datasets automatically. For Tiny-ImageNet, you should download and put in the dir: "data/". The follow instruction and commands are for CIFAR100.

You can skip this step by using our pre-trained models in here. Download and unzip to: experiments/pretrained_teacher_models/

Use ''--model_dir'' to specify the directory of "parameters", model saving and log saving.

For example, normally train ResNet18 to obtain the pre-trained teacher:

CUDA_VISIBLE_DEVICES=0 python main.py --model_dir experiments/base_experiments/base_resnet18/

We ignore the command ''CUDA_VISIBLE_DEVICES=gpu_id'' in the following commands

Normally train MobileNetV2 to obtain the baseline model and baseline accuracy:

python main.py --model_dir experiments/base_experiments/base_mobilenetv2/

Normally train ResNeXt29 to obtain the baseline model and baseline accuracy:

python main.py --model_dir experiments/base_experiments/base_resnext29/

The baseline accuracy (in %) on CIFAR100 is:

Normal KD: ResNet18 teach MobileNetV2

python main.py --model_dir experiments/kd_experiments/mobilenet_v2_distill/resnet18_teacher/

Re-KD: MobileNetV2 teach ResNet18

python main.py --model_dir experiments/kd_experiments/resnet18_distill/mobilenet_v2_teacher/

Re-KD: ShuffleNetV2 teach ResNet18

python main.py --model_dir experiments/kd_experiments/resnet18_distill/shufflenet_v2_teacher/

Re-KD experiment results on CIFAR100:

Use the arguments "--pt_teacher" to switch to Defective KD experiment.

For expample, use a pooly-trained Teacher ResNet18 with 15.48% accuracy (just trained one epoch) to teach MobileNetV2:

python main.py --model_dir experiments/kd_experiments/mobilenet_v2_distill/resnet18_teacher/ --pt_teacher

Use one-epoch trained teacher ResNet18 to teach ShuffleNetV2:

python main.py --model_dir experiments/kd_experiments/shufflenet_v2_distill/resnet18_teacher/ --pt_teacher

Use 50-epoch-trained teacher ResNet50 (acc:45.82%) to teach ShuffleNetV2:

python main.py --model_dir experiments/kd_experiments/shufflenet_v2_distill/resnet50_teacher/ --pt_teacher

Use 50-epoch-trained teacher ResNet50 (acc:45.82%) to teach ResNet18:

python main.py --model_dir experiments/kd_experiments/resnet18_distill/resnet50_teacher/ --pt_teacher

Use 50-epoch-trained teacher ResNeXt29 (acc:51.94%) to teach MobileNetV2:

python main.py --model_dir experiments/kd_experiments/mobilenet_v2_distill/resnext29_teacher/ --pt_teacher

De-KD experiment results on CIFAR100:

We have two implementations to achieve Tf-KD, the first one is self-training, the second one is manually-designed teacher(regularization)

Use the arguments ''--self_training'' to control this training. The --model_dir should be experiment dirctory, should be "experiments/kd_experiments/student/student_self_teacher"

MobileNetV2 self training:

python main.py --model_dir experiments/kd_experiments/mobilenet_v2_distill/mobilenet_self_teacher/ --self_training

ShuffleNetV2 self training:

python main.py --model_dir experiments/kd_experiments/shufflenet_v2_distill/shufflenet_self_teacher/ --self_training

ResNet18 self training:

python main.py --model_dir experiments/kd_experiments/resnet18_distill/resnet18_self_teacher/ --self_training

Our method achieve more than 1.0% improvement for a big single model ResNeXt29, run self-training for ResNeXt29:

python main.py --model_dir experiments/kd_experiments/resnext29_distill/resnext29_self_teacher/ --self_training

Tf-KD self-training experiment results on CIFAR100:

MobileNetV2 taught by mannually-designed regularization:

python main.py --model_dir experiments/base_experiments/base_mobilenetv2/  --regularization

ShuffleNetV2 taught by mannually-designed regularization:

python main.py --model_dir experiments/base_experiments/base_shufflenetv2/ --regularization

ResNet18 taught by mannually-designed regularization:

python main.py --model_dir experiments/base_experiments/base_resnet18/  --regularization

GoogLeNet taught by mannually-designed regularization:

python main.py --model_dir experiments/base_experiments/base_googlenet/ --regularization

MobileNetV2 Lable Smoothing:

python main.py --model_dir experiments/base_experiments/base_mobilenetv2/  --label_smoothing

ShuffleNetV2 Lable Smoothing:

python main.py --model_dir experiments/base_experiments/base_shufflenetv2/ --label_smoothing

Tf-KD regularization and LSR experiment results on CIFAR100:

If you find this repo useful, please consider citing:

@inproceedings{yuan2020revisiting,
  title={Revisiting Knowledge Distillation via Label Smoothing Regularization},
  author={Yuan, Li and Tay, Francis EH and Li, Guilin and Wang, Tao and Feng, Jiashi},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={3903--3911},
  year={2020}
}