/L_DMI

Code for NeurIPS 2019 Paper, "L_DMI: An Information-theoretic Noise-robust Loss Function"

Primary LanguagePython

Introduction

Implementation for NeurIPS 2019 paper :

paper link: https://arxiv.org/abs/1909.03388

Fashion MNIST dataset

  • To run experiments of Fashion MNIST dataset in fashion directory:
python3 fashion.py --r noisy_amount --s seed --c case_num --device device_num

noise_amount: the amount of noise amount r of label flipping. (0 <= r <= 1)
seed: random seed
case_num :  1: class-independent; 2: class-dependent (a); 3: class-dependent (b)
device_num: GPU number

CIFAR-10 dataset:

  • To run experiments of CIFAR10 dataset in CIFAR-10 directory, pleases run all the baseline in the following order:

    python3 CE.py --r noise_amount --s seed --device device_num --root path
    python3 FW.py --r noise_amount --s seed --device device_num --root path
    python3 GCE.py --r noise_amount --s seed --device device_num --root path
    python3 LCCN.py --r noise_amount --s seed --device device_num --root path
    python3 DMI.py --r noise_amount --s seed --device device_num --root path
    
    noise_amount: the amount of noise amount r of label flipping. (0 <= r <= 1)
    seed: random seed
    device_num: GPU number
    root: path to the CIFAR-10 dataset

Dog & Cat datasete:

  • To run experiments of Dog vs. Cats dataset in dogcat directory, pleases run all the baseline in the following order:

    python3 CE.py --r noise_amount --s seed --device device_num
    python3 FW.py --r noise_amount --s seed --device device_num
    python3 GCE.py --r noise_amount --s seed --device device_num
    python3 LCCN.py --r noise_amount --s seed --device device_num
    python3 DMI.py --r noise_amount --s seed --device device_num
    
    noise_amount: the amount of noise amount r of label flipping. (0 <= r <= 1)
    seed: random seed
    device_num: GPU number

MR dataset

  • To download the dataset:

    python3 download_dataset.py MR
  • To run experiments of MR dataset in MR directory:

    python3 main.py --r noise_amount --s seed --device device_num WordCNN
    
    noise_amount: the amount of noise amount r of label flipping. (0 <= r <= 1)
    seed: random seed
    device_num: GPU number

Clothing1M dataset

  • To run experiments of Clothing1M dataset in clothing directory, pleases run all the baseline in the following order:

    python3 main.py --device device_num
    
    device_num: GPU number

TODO:

Combine all the APIs to dataset into one file.