Contrastive Self-Supervised Learning on CIFAR-10

Paper

"Towards the Generalization of Contrastive Self-Supervised Learning", Weiran Huang, Mingyang Yi and Xuyang Zhao, arXiv:2111.00743, 2021.

The most critical argument we made in our paper is that the quality of data augmentation exhibits great impact on the quality of contrastive-learned encoder. The data augmentation with sharper intra-class concentration enables the model to have better generalization on downstream tasks. We verify it through a variety of experiments in this repository.

Supported methods

• SimCLR
• Barlow Twins
• MoCo
• SimSiam

Installation

pip install -r requirement.txt

Dependencies

• torch==1.4.0
• torchvision==0.5.0
• pytorch-lightning==1.3.8
• lightly==1.0.8 (important!)

Evaluation

KNN evaluation protocol. Code from here.

Results

ResNet-18 trained by SimCLR and Barlow Twins over various data augmentation combinations.

Example: python main.py --model=twins --epoch=800 --batch=512 --round=3 --augs=abcde

Augmentation operations include:

(a) random cropping with a scaling factor chosen in [0.08, 1.0];

(b) random Gaussian blur with a probability 0.5;

(c) color dropping (i.e., randomly convert images to grayscale with 0.2 probability for each image);

(d) color distortion with a probability of 0.8 and with strength of [0.4, 0.4, 0.4, 0.1];

(e) random horizontal flipping with a probability of 0.5.

ResNet18 trained by SimCLR and Barlow Twins over various color distortion strengths.

Example: python main.py --model=simclr --epoch=800 --batch=512 --round=3 --augs=color --strength=1

Acknowledgement

This code is based on:

• IgorSusmelj/barlowtwins
• lightly/imagenette_benchmark.py

Citation

If you find our work useful in your research, please consider citing:

@misc{huang2021generalization,
      title={Towards the Generalization of Contrastive Self-Supervised Learning}, 
      author={Weiran Huang and Mingyang Yi and Xuyang Zhao},
      year={2021},
      eprint={2111.00743},
      archivePrefix={arXiv}
}