Awesome

Safeguarded-Dynamic-Label-Regression-for-Generalized-Noisy-Supervision

If you use this code in your research, please cite

@inproceedings{aaai19-jiangchao,
  title     = {Safeguarded Dynamic Label Regression for Noisy Supervision},
  author    = {Jiangchao Yao, Hao Wu, Ya Zhang, Ivor W. Tsang and Jun Sun},
  booktitle = {Proceedings of the Association for the Advancement of Artificial Intelligence Conference on
               Artificial Intelligence, {AAAI-19}},
  publisher = {Association for the Advancement of Artificial Intelligence Conference},
  year      = {2019}
}

Form the noisy datasets.

python dataset.py

Train the models.

Train DNNs directly with the cross-entropy loss (CE).

python cifar10_train.py --train_dir results/events_ce/cifar10_train --noise_rate 0.3 # You can train other models like this one

Train Bootstrapping

  python cifar10_train_bootstrapping.py --train_dir results/events_bootstrapping/cifar10_train --noise_rate 0.3 # You can train other models like this one

Train Forward

  python cifar10_train_T.py --init_dir results/events_ce/cifar10_train --train_dir results/events_T/cifar10_train --noise_rate 0.3 # You can train other models like this one

Train S-adaptation

  python cifar10_train_varT.py --init_dir results/events_ce/cifar10_train --train_dir results/events_varT/cifar10_train --noise_rate 0.3 # You can train other models like this one

Train LCCN

  python cifar10_train_varC.py --init_dir results/events_ce/cifar10_train --train_dir results/events_varC/cifar10_train --noise_rate 0.3 # You can train other models like this one

Note that step 3, 4, 5 are based on the classifier pretraining in step 1 and among them, they also have the internal inheritance relationship due to the transition matrix initialization. Please execute them sequentially. However, this does not mean a shortcoming of the model. You can easily modify them into parallel versions by integrating the classifier pretraining and transition initialization into one script.

Test and evaluate

# open a new shell. (Take CE as the example)
python cifar10_eval.py --checkpoint_dir results/events_ce/cifar10_train --eval_dir results/cifar10_eval

Visualization with Tensorboard

# open a new shell. (Take CE as the example)
tensorboard --logdir=results/events_ce --port=8080

For clothing1M and WebVision, please check here

This is the code of the Clothing-1M experiments. Similarly, we have provided multiple scripts to run the baselines and our model. The parameters also have been hard-coded in the script.

In addition, it is almost the same code of the WebVision experiments. You can simply organize the data structure of WebVision like Clothing1M and run the codes.