Improving Contrastive Learning by Visualizing Feature Transformation

This project hosts the codes, models and visualization tools for the paper:

Improving Contrastive Learning by Visualizing Feature Transformation,
Rui Zhu, Bingchen Zhao, Jingen Liu, Zhenglong Sun, Chang Wen Chen
Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021, Oral
arXiv preprint (arXiv 2108.02982)

@inproceedings{zhu2021improving,
  title={Improving Contrastive Learning by Visualizing Feature Transformation},
  author={Zhu, Rui and Zhao, Bingchen and Liu, Jingen and Sun, Zhenglong and Chen, Chang Wen},
  booktitle =  {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
  year={2021}
}

Highlights

• Visualization Tools: We provide a visualization tool for pos/neg score distribution, which enables us to analyze, interpret and understand the contrastive learning process.
• Feature Transformation: Inspired by the visualization, we propose a simple yet effective feature transformation (FT), which creates both hard positives and diversified negatives to enhance the training. FT enables to learn more view-invariant and discriminative representations.
• Less Task-biased: FT makes the model less “task-bias”, which means we can achievesignificant performance improvement on various downstream tasks (object detection, instance segmentation, and long-tailed classification).

Updates

• Code, pre-trained models and visualization tools are released. (07/08/2021)

Installation

This project is mainly based on the open-source code PyContrast.

Please refer to the INSTALL.md and RUN.md for installation and dataset preparation.

Models

For your convenience, we provide the following pre-trained models on ImageNet-1K and ImageNet-100.

Note:

• See our paper for more results on different benchmarks.

Usage

Training on IN-1K

python main_contrast.py --method MoCov2 --data_folder your/path/to/imagenet-1K/dataset  --dataset imagenet  --epochs 200 --input_res 224 --cosine --batch_size 256 --learning_rate 0.03   --mixnorm --mixnorm_target posneg --sep_alpha --pos_alpha 2.0 --neg_alpha 1.6 --mask_distribution beta --expolation_mask --alpha 0.999 --multiprocessing-distributed --world-size 1 --rank 0 --save_score --num_workers 8

Linear Evaluation on IN-1K

python main_linear.py --method MoCov2 --data_folder your/path/to/imagenet-1K/dataset --ckpt your/path/to/pretrain_model   --n_class 1000 --multiprocessing-distributed --world-size 1 --rank 0 --epochs 100 --lr_decay_epochs 60,80  --num_workers 8

Training on IN-100

python main_contrast.py --method MoCov2 --data_folder your/path/to/imagenet-1K/dataset  --dataset imagenet100  --imagenet100path your/path/to/imagenet100.class  --epochs 200 --input_res 224 --cosine --batch_size 256 --learning_rate 0.03   --mixnorm --mixnorm_target posneg --sep_alpha --pos_alpha 2.0 --neg_alpha 1.6 --mask_distribution beta --expolation_mask --alpha 0.99 --multiprocessing-distributed --world-size 1 --rank 0 --save_score  --num_workers 8

Linear Evaluation on IN-100

python main_linear.py --method MoCov2 --data_folder your/path/to/imagenet-1K/dataset  --dataset imagenet100  --imagenet100path your/path/to/imagenet100.class  --n_class 100  --ckpt your/path/to/pretrain_model  --multiprocessing-distributed --world-size 1 --rank 0  --num_workers 8

Transferring to Object Detection

Please refer to DenseCL and MoCo for transferring to object detection.

Visualization Tools

• Our visualization is offline, which almost does not affect the training speed. Instead of storing K (65536) pair scores, we save their statistical mean and variance to represent the scores’ distribution. You can refer to the original paper for the details.

• Visualization code is line 69-74 to store the scores. And then we further process the scores in the IpythonNotebook for drawing.

Citations

Please consider citing our paper in your publications if the project helps your research. BibTeX reference is as follow.

@inproceedings{zhu2021improving,
  title={Improving Contrastive Learning by Visualizing Feature Transformation},
  author={Zhu, Rui and Zhao, Bingchen and Liu, Jingen and Sun, Zhenglong and Chen, Chang Wen},
  booktitle =  {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
  year={2021}
}