NeurIPS'18: Masking: A New Perspective of Noisy Supervision
Switch branches/tags
Nothing to show
Clone or download
Latest commit 72411cd Nov 29, 2018


NeurIPS'18: Masking: A New Perspective of Noisy Supervision (Tensorflow implementation).

This is the code for the paper: Masking: A New Perspective of Noisy Supervision
Bo Han*, Jiangchao Yao*, Gang Niu, Mingyuan Zhou, Ivor Tsang, Ya Zhang, Masashi Sugiyama
To be presented at NeurIPS 2018.

If you find this code useful in your research then please cite

  title={Masking: A new perspective of noisy supervision},
  author={Han, Bo and Yao, Jiangchao and Gang, Niu and Zhou, Mingyuan and Tsang, Ivor and Zhang, Ya and Sugiyama, Masashi},
  pages = {5839--5849},

Introduction about the codes


(1) implements the classifier directly trained on the dataset.

(2) implements the loss correction method in

(3) implements the classifier with adaptation of noise transition.

(4) implements our MASKING model.


(1) The CIFAR-10 dataset can be downloaded and placed in the corresponding position by following the introduction in ./data/cifar-10-batches-bin/readme.txt

(2) The noisy datasets is generated by based on the clean CIFAR-10 dataset.

You can switch the noisy dataset for, and by setting the NOISE_TYPE parameter in


(1) Due to the requirements of initialization about the noise transition, some codes must be executed in order. For example, you can execute the codes in the following order,





(2) For evaluation, since the evaluation scripts are separated, you can first launch up the training script and then launch up the evaluation script in another terminal. For example,

python --train_dir events/cifar10_train

python --checkpoint_dir events/cifar10_train --eval_dir events/cifar10_eval

These codes are forked from the Tensorflow official CIFARnet in

Contact: Jiangchao Yao (; Bo Han (