Correct Twice at Once: Learning to Correct Noisy Labels for Robust Deep Learning

Published at conference ACM MM 2022.

Introduction

This is a PyTorch implementation of ["Correct Twice at Once: Learning to Correct Noisy Labels for Robust Deep Learning"].

Requirements:

• Python 3.7
• PyTorch 1.8.0
• torchvision 0.9.0

Train:

• The code can be run on cifar10,cifar100, and Clothing1M datasets, where the datasets can be downloaded automatically.

  sh run.sh

Log:

• We provided a training log of the dataset Clothing1M which could be used to visualize the training process through tensorboard for reference. The log file can be found at: https://mega.nz/folder/58dFFahZ#cCR-HsLBlzbHQ6L7ztXCXQ.

• Place the log file in logs/, and then execute the command.

  tensorboard --logdir=/logs/ --host= `host address`

Note: Our code will be further improved to make it cleaner.