Contrast to Divide: self-supervised pre-training for learning with noisy labels

This is an official implementation of "Contrast to Divide: self-supervised pre-training for learning with noisy labels". The code is based on DivideMix implementation.

Results

Following tables summarize main resutls of the paper:

CIFAR-10:

CIFAR-100:

Clothing1M:

mini-WebVision:

Running the code

First you need to install dependencies by running pip install -r requirements.txt.

You can download pretrained self-supervised models from Google Drive. Alternatively, you can train them by yourself, using SimCLR implementation. Put them into ./pretrained folder.

Then you can run the code for CIFAR

python3 main_cifar.py --r 0.8 --lambda_u 500 --dataset cifar100 --p_threshold 0.03 --data_path ./cifar-100 --experiment-name simclr_resnet18 --method selfsup --net resnet50

for Clothing1M

python3 main_clothing1M.py --data_path /path/to/clothing1m --experiment-name selfsup --method selfsup --p_threshold 0.7 --warmup 5 --num_epochs 120

or for mini-WebVision

python3 Train_webvision.py --p_threshold 0.03 --num_class 50 --data_path /path/to/webvision --imagenet_data_path /path/to/imagenet --method selfsup```

To run C2D with ELR+ just use the self-suprevised pretrained models with the original code.

License

This project is licensed under the terms of the MIT license.