Home

Awesome

Robust Inference via Generative Classifiers for Handling Noisy Labels

This project is for the paper "Robust Inference via Generative Classifiers for Handling Noisy Labels ". Codes will be updated.

Preliminaries

It is tested under Ubuntu Linux 16.04.1 and Python 3.6 environment, and requries Pytorch package to be installed:

Datasets

Training networks with noisy labels

1. Generate noisy labels:

# dataset: CIFAR-10, noise type: uniform, noise fraction: 60%
python generate_labels.py --dataset cifar10 --noise_type uniform --noise_fraction 60

2. Train networks

# model: DenseNet, dataset: CIFAR-10, noise type: uniform, noise fraction: 60%, gpu 0
python train.py --net_type densenet --dataset cifar10 --noise_type uniform --noise_fraction 60 --gpu 0

Performance evaluation

# model: DenseNet, dataset: CIFAR-10, noise type: uniform, noise fraction: 60%
python inference.py --net_type densenet --dataset cifar10 --noise_type uniform --noise_fraction 60 --gpu 0