Code for "Long-tailed learning with in- and out-of-distribution noisy labels in the open-world"

In this paper, we propose the Imbalanced Noisy Labels Calibration (INLC) approach.

Preparation

Main dependency:

loguru             0.7.3
torch              2.5.1

The datasets used in the experiments:

• ILSVRC2012: Images - Validation images (all tasks). 6.3GB. [link]
• ImageNet32: Download downsampled image data (32x32, 64x64) - Train(32x32), 3 GB [link]
• WebVision1.0: Resized Images (small version) - Google Images Resized (16 GB) [link]

We put the datasets in the ~/data directory, and the directory structure is as follows:

├─cifar-10-batches-py
├─cifar-100-python
├─ILSVRC2012
│  ├─ILSVRC2012_devkit_t12
│  │  ├─data
│  │  └─evaluation
│  └─ILSVRC2012_img_val
├─imagenet32
│  ├─train_data_batch_1
│  ├─...
│  └─val_data
└─webvision1.0
    ├─google
    │  ├─q0001
    │  ├─q0002
    │  ├─...
    |  └─q1632
    ├─info
    └─val_images_256

Usage

Training in CIFAR-10

python main_cifar.py
python main_cifar.py --r_ood 0.2
python main_cifar.py --r_ood 0.2 --r_id 0.2
python main_cifar.py --r_ood 0.2 --r_id 0.2 --asym
python main_cifar.py --r_imb 0.01
python main_cifar.py --r_imb 0.01 --r_ood 0.2
python main_cifar.py --r_imb 0.01 --r_ood 0.2 --r_id 0.2
python main_cifar.py --r_imb 0.01 --r_ood 0.2 --r_id 0.2 --asym

Try strategy 2 or replace loss with focal in CIFAR-10

python main_cifar.py --s2
python main_cifar.py --loss focal

Training in CIFAR-100

python main_cifar.py --dataset 100 --tau 0.6
python main_cifar.py --dataset 100 --tau 0.6 --r_ood 0.2
python main_cifar.py --dataset 100 --tau 0.6 --r_ood 0.2 --r_id 0.2 --warm_epochs 20
python main_cifar.py --dataset 100 --tau 0.6 --r_ood 0.2 --r_id 0.2 --asym --warm_epochs 20
python main_cifar.py --dataset 100 --tau 0.6 --r_imb 0.01
python main_cifar.py --dataset 100 --tau 0.6 --r_imb 0.01 --r_ood 0.2
python main_cifar.py --dataset 100 --tau 0.6 --r_imb 0.01 --r_ood 0.2 --r_id 0.2 --warm_epochs 20
python main_cifar.py --dataset 100 --tau 0.6 --r_imb 0.01 --r_ood 0.2 --r_id 0.2 --asym --warm_epochs 20

Training or evaluating in WebVision 1.0

python main_webvision.py --s2
python main_webvision.py --test

If training collapses, try increasing the value of --warm_epochs.

Reference

• Learning Imbalanced Datasets with Label-Distribution-Aware Margin Loss
• DivideMix: Learning with Noisy Labels as Semi-supervised Learning
• DASO: Distribution-Aware Semantics-Oriented Pseudo-label for Imbalanced Semi-Supervised Learning
• Jo-SRC: A Contrastive Approach for Combating Noisy Labels