Awesome

SSKT(Accepted WACV2022)

Concept map

Dataset

• Image dataset
  • CIFAR10 (torchvision)
  • CIFAR100 (torchvision)
  • STL10 (torchvision)
  • Pascal VOC (torchvision)
  • ImageNet(I) (torchvision)
  • Places365(P)
• Video dataset
  • UCF101
  • HMDB51

Pre-trained models

• Imagenet
  • we used the pre-trained model in torchvision.
  • using resnet18, 50
• Places365
  • Resnet50

Option

• isSource
  • Single Source Transfer Module
  • Transfer Module X, Only using auxiliary layer
• transfer_module
  • Single Source Transfer Module
• multi_source
  • multiple task transfer learning

Training

• 2D PreLeKT

 python main.py --model resnet20  --source_arch resnet50 --sourceKind places365 --result /raid/video_data/output/PreLeKT --dataset stl10 --lr 0.1 --wd 5e-4 --epochs 200 --classifier_loss_method ce --auxiliary_loss_method kd --isSource --multi_source --transfer_module

• 3D PreLeKT

 python main.py --root_path /raid/video_data/ucf101/ --video_path frames --annotation_path ucf101_01.json  --result_path /raid/video_data/output/PreLeKT --n_classes 400 --n_finetune_classes 101 --model resnet --model_depth 18 --resnet_shortcut A --batch_size 128 --n_threads 4 --pretrain_path /nvadmin/Pretrained_model/resnet-18-kinetics.pth --ft_begin_index 4 --dataset ucf101 --isSource --transfer_module --multi_source

Experiment

Comparison with other knowledge transfer methods.

• For a further analysis of SSKT, we compared its performance with those of typical knowledge transfer methods, namely KD[1] and DML[3]
• For KD, the details for learning were set the same as in [1], and for DML, training was performed in the same way as in [3].
• In the case of 3D-CNN-based action classification[2], both learning from scratch and fine tuning results were included

The performance comparison with MAXL[4], another auxiliary learning-based transfer learning method

• The difference between the learning scheduler in MAXL and in our experiment is whether cosine annealing scheduler and focal loss are used or not.
• In VGG16, SSKT showed better performance in all settings. In ResNet20, we also showed better performance in our settings than MAXL in all settings.

Citation

If you use SSKD in your research, please consider citing:

@InProceedings{SSKD_2022_WACV,
author = {Seungbum Hong, Jihun Yoon, and Min-Kook Choi},
title = {Self-Supervised Knowledge Transfer via Loosely Supervised Auxiliary Tasks},
booktitle = {In The IEEE Winter Conference on Applications of Computer Vision (WACV)},
month = {January},
year = {2022}
}

References

• [1] Hinton et al. - "Distilling the knowledge in a neural network" (NIPSW, 2014)
• [2] Hara et al. - "Can Spatiotemporal 3D CNNs Retrace the History of 2D CNNs and ImageNet?" (CVPR, 2018)
• [3] Zhang et al. - "Deep mutual learning" (CVPR, 2018)
• [4] Davison and Johns - "Self-Supervised Generalisation with Meta Auxiliary Learning" (NeurIPS, 2019)