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
@inproceedings{han2018masking,
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},
booktitle={NeurIPS},
pages = {5839--5849},
year={2018}
}
Introduction about the codes
Models:
(1) cifar10_train.py implements the classifier directly trained on the dataset.
(2) cifar10_train_T.py implements the loss correction method in https://github.com/giorgiop/loss-correction
(3) cifar10_train_varT.py implements the classifier with adaptation of noise transition.
(4) cifar10_train_GANT.py implements our MASKING model.
Datasets:
(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 noise.py based on the clean CIFAR-10 dataset.
You can switch the noisy dataset for cifar10_train_T.py, cifar10_train_varT.py and cifar10_train_GANT.py by setting the NOISE_TYPE parameter in cifar10_input.py
Example:
(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,
python cifar10_train.py
python cifar10_train_T.py
python cifar10_train_varT.py
python cifar10_train_GANT.py
(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 cifar10_train.py --train_dir events/cifar10_train
python cifar10_eval.py --checkpoint_dir events/cifar10_train --eval_dir events/cifar10_eval
These codes are forked from the Tensorflow official CIFARnet in https://github.com/tensorflow/models/tree/master/tutorials/image/cifar10.
Contact: Jiangchao Yao (sunarker@sjtu.edu.cn); Bo Han (bo.han@riken.jp).