LaplaceConfidence

This is the code for our submisssion "LaplaceConfidence: a Graph-based Approach for Learning with Noisy Labels"

Motivation

Method

Usage

PyTorch CIFAR Training

optional arguments:
  -h, --help            show this help message and exit
  --batch_size BATCH_SIZE
                        train batchsize
  --warm_up WARM_UP     warm epochs
  --lr LR, --learning_rate LR
                        learning rate
  --num_epochs NUM_EPOCHS
                        number of trainig epochs
  --gpuid GPUID
  --seed SEED
  --save_name SAVE_NAME
  --data_path DATA_PATH
                        path to dataset
  --dataset DATASET
  --resume RESUME
  --T T                 temperature for sharping pseudo-labels
  --knn KNN             knn number for constructing graph
  --pca PCA             PCA dimension
  --r R                 noise ratio
  --noise_mode NOISE_MODE choose symmetrical or asymmetrical noise

To obtain the results on CIFAR-10 or CIFAR-100

You first need to download the public dataset CIFAR in here, then run (the label noise will be generated automaticlly):

python3 GLR_ce.py --data_path ** --dataset ** --num_class ** --r ** --knn **

To obtain the results on WebVision

You first need to download the public dataset WebVision in here Facing a real-world noisy dataset, we don't need to preprocess the label information. Just run

python3 LC_cifar.py --data_path ** --save_name LC_loss

Requirement

python==3.6.8
scikit-learn==0.23.2
torch==1.7.0+cu101
scipy==1.6.2
Pillow==8.2.0
pandas==1.2.4
numpy==1.22.4

chenmc1996 / LaplaceConfidence