## Paper title: Mitigating Bias Using Per-sample Gradients From a Biased Classifier

## Packages:
## pip packages : ./packages/requirements.txt
## conda packages : ./packages/package_list.txt



## How to run
# Variables
# RUN: Run title (e.g., RUN0, DEBUG, ...)
# GPU: GPU index in CUDA (e.g., 0,1,...)
# JOB: Training and testing job (e.g., train_noise_bias_final for end to end training, test)
# DATA: Benchmark (e.g., CM, WM)
# ALG: Algorithm (e.g., vanilla, rubi, mixin, aflite, repair, rebias, lff, ours)
# NOISE_RATIO: Noise ratio (e.g., 1.00: w/o noise, 0.00: all data samples have flipped label)
# BIAS_RATIO: Majority ratio (e.g., 1.00: All data samples are aligned, 0.00: all data samples are not aligned)
# DENOISE: Applying denoise module (e.g., True, False)

## Example: 

# Train
python main.py --run=test --gpu=0 --job=train_noise_bias_final --data=CM --alg=ours --noise=0.99 --bias=0.99 --denoise=True

# Evaluate
python main.py --run=test --gpu=0 --job=test --data=CM --alg=ours --noise=0.99 --bias=0.99




Due to the file size issue, we upload CM, WM cases, which is officially downloadable from the torchvision.