This is a Keras implementation for the paper Understanding and Utilizing Deep Neural Networks Trained with Noisy Labels (Proceedings of ICML, 2019).

@inproceedings{chen2019understanding,
  title={Understanding and Utilizing Deep Neural Networks Trained with Noisy Labels},
  author={Chen, Pengfei and Liao, Ben Ben and Chen, Guangyong and Zhang, Shengyu},
  booktitle={International Conference on Machine Learning},
  pages={1062--1070},
  year={2019}
}

Python 3.6.4, Keras 2.1.6, Tensorflow 1.7.0, numpy, sklearn.

Please be aware of the bug caused by different versions of Keras/tf. For example, in the callback functions in model.fit_generator, new keras versions use "val_accuracy" instead of "val_acc", for which you may not directly get an error but may fail to save the model. Please check the Documentation of Keras carefully if you use a different version.

To set up experiments, we need to download the CIFAR-10 data and extract it to:

data/cifar-10-batches-py

The code will automatically add noise to CIFAR-10 by randomly flipping original labels.

Note

To quantitatively characterize the generalization performance of deep neural networks normally trained with noisy labels, we split the noisy dataset into two halves and perform cross-validation: training on a subset and testing on the other.

We firstly theoretically characterize on the test set the confusion matrix (w.r.t. ground-truth labels) and test accuracy (w.r.t. noisy labels).

We then propose to select a testing sample as a clean one, if the trained model predict the same label with its observed label. The performance is evaluated by label precision and label recall, which can be theoretically estimated using the noise ratio according to our paper.

Train

Experimrental resluts justify our theoretical analysis. To reproduce the experimental results, we can run Verify_Theory.py and specify the noise pattern and noise ratio, e.g.,

• Symmetric noise with ratio 0.5:

  python Verify_Theory.py --noise_pattern sym --noise_ratio 0.5

• Asymmetric noise with ratio 0.4:

  python Verify_Theory.py --noise_pattern asym --noise_ratio 0.4

Results

Test accuracy, label precision and label recall w.r.t noise ratio on manually corrupted CIFAR-10.

Confusion matrix M approximates noise transistion matrix T.

Train

If you only want to use INCV to clean a noisy dataset, you can run INCV.py only, e.g., on CIFAR-10 with

• 50% symmetric noise:

  python INCV.py --noise_pattern sym --noise_ratio 0.5 --dataset cifar10

• 40% asymmetric noise:

  python INCV.py --noise_pattern asym --noise_ratio 0.4 --dataset cifar10

The results will be saved in 'results/(dataset)/(noise_pattern)/(noise_ratio)/(XXX.csv)' with columns ('y', 'y_noisy', 'select', 'candidate', 'eval_ratio').

Results

label precision and label recall on the manually corrupted CIFAR-10.

Our INCV accurately identifies most clean samples. For example, under symmetric noise of ratio 0.5, it selects about 90% (=LR) of the clean samples, and the noise ratio of the selected set is reduced to around 10% (=1−LP).

Note

We present the Iterative Noisy Cross-Validation (INCV) to select a subset of clean samples, then modify the Co-teaching strategy to train noise-robust deep neural networks.

Train

E.g., use our method to train on CIFAR-10 with

• 50% symmetric noise:

  python INCV_main.py --noise_pattern sym --noise_ratio 0.5 --dataset cifar10

• 40% asymmetric noise:

  python INCV_main.py --noise_pattern asym --noise_ratio 0.4 --dataset cifar10

Results

Average test accuracy (%, 5 runs) with standard deviation:

Average test accuracy (%, 5 runs) during training:

Please cite our paper if you use this code in your research work.

Please submit a Github issue or contact [email protected] if you have any questions or find any bugs.