Delving into Noisy Label Detection with Clean Data

This repository provides codes for the manuscript 'Delving into Noisy Label Detection with Clean Data'.

Preferred Prerequisites

• Python (3.8)
• CUDA
• numpy~=1.21.2
• torch~=1.9.1
• matplotlib~=3.4.3
• Pillow~=8.3.2
• torchvision~=0.10.1
• pandas~=1.3.4
• statsmodels~=0.13.2

You can download the python packages manually or run pip install -r requirements.txt.

Code Overview

.
├── data  
    ├── __init__.py
    ├── dataloader_clothing1M_BH.py # Code for load Clothing1M
    ├── utils.py        # Code for noisify clean dataset
├── models                # Code for defining network architectures
    ├── __init__.py
    ├── imagenet_resnet.py
    ├── nine_layer_cnn.py
    ├── resnet.py  
├── pvalue
    ├── __init__.py
    ├── methods.py # Code for calculate p-value
├── BHN.py # Code for method BHN
├── utility.py # Code for customized utilities (setting gpu, setting outdir, etc.)
└── utils_experiments.py # Code for evaluation

Training the neural network with a part of clean data and detecting corrupted examples with the score function using BHN

Here is an example.

• Train ResNet-18 model on 20% of the overall training set of CIFAR-10.

python BHN.py --dataset cifar10 --noise_rate 0.4 --noise_type instance --leave_ratio 0.4 --net resnet18 --n_epochs 200 --l_train_second_ratio 1.0 --score_evaluation

• Detect corrupted labels on the synthetic CIFAR-10 with the Inst. 0.4 noise using final_model.tar in $INPUT_DIR$, where $INPUT_DIR$ is the directory of final_model.tar.

python BHN.py --dataset cifar10 --net resnet18 --noise_type instance --leave_ratio 0.4 --input_dir $INPUT_DIR$ --model_file final_model.tar --load_model_direct_score --score_evaluation --alpha 0.1

Training on the selected data by BHN

Here is an example.

• Train ResNet-18 model on the selected CIFAR-10 examples in sel_clean.txt

python BHN.py --dataset cifar10 --noise_rate 0.4 --noise_type instance --leave_ratio 0.4 --net resnet18 --n_epochs 200 --l_train_second_ratio 1.0 --lr 0.1 --batch_size 128 --weight_decay 5e-4 --momentum 0.9 --input_dir $INPUT_DIR$ --labeledsetpth sel_clean.txt --retrain_selectedset --lr 0.1 --batch_size 128 --weight_decay 5e-4 --momentum 0.9 

, where $INPUT_ DIR$ is the directory where the file that stores the selected examples lies in.