Jump-teaching: Combating Sample Selection Bias via Temporal Disagreement

Official implementation of Jump-teaching, accepted at AAAI 2026.

Repository Structure

Jump-teaching/
├── algorithms/          # Algorithm implementations
│   └── jumpteaching.py # Main Jump-teaching implementation
├── configs/            # Configuration files for different datasets
│   ├── jumpteaching_cifar.py
│   ├── jumpteaching_clothing1M.py
│   ├── jumpteaching_food101n.py
│   └── jumpteaching_webvision.py
├── datasets/           # Dataset loaders and preprocessing
├── models/             # Neural network architectures 
├── losses/             # Loss function implementations
├── utils/              # Utility functions
├── labels/             # Pre-computed Hadamard hash codes
├── scripts/            # Shell scripts for experiments
├── main.py             # Main training script
└── generate_labelcodes.py  # Generate hash codes

Installation

Requirements

• Python 3.8+
• PyTorch 1.8+
• CUDA 10.2+ (for GPU support)

Setup

git clone https://github.com/ky-ji/Jump-teaching.git
cd Jump-teaching
pip install -e .

Quick Start

1. Download Datasets

Download CIFAR-10 and CIFAR-100:

python download_cifar.py

2. Run Jump-teaching on CIFAR-10

# Using the script
bash scripts/jumpteaching_cifar10.sh

# Or run directly
python main.py -c=./configs/jumpteaching_cifar.py \
    --gpu 0 \
    --seed 1 \
    --noise_type sym \
    --percent 0.5

3. Run Jump-teaching on CIFAR-100

# Using the provided script
bash scripts/jumpteaching_cifar100.sh

# Or run directly
python main.py -c=./configs/jumpteaching_cifar.py \
    --gpu 0 \
    --seed 1 \
    --noise_type sym \
    --percent 0.5 \
    --dataset cifar-100 \
    --num_classes 100

Label Hash Codes

Jump-teaching uses Hadamard-based hash codes to represent class labels efficiently. Pre-computed codes are stored in ./labels/. To generate hash codes for new datasets:

python generate_labelcodes.py \
    --hashbits <BITS> \
    --dataset <DATASET_NAME> \
    --num_classes <NUM_CLASSES> \
    [--output_dir ./labels] \
    [--verify]

Common setups:

# CIFAR-10 (10 classes, 32-bit codes)
python generate_labelcodes.py --hashbits 32 --dataset cifar-10 --num_classes 10

# CIFAR-100 (100 classes, 64-bit codes)
python generate_labelcodes.py --hashbits 64 --dataset cifar-100 --num_classes 100

# Clothing1M (14 classes, 32-bit codes)
python generate_labelcodes.py --hashbits 32 --dataset clothing1M --num_classes 14

The script saves tensors as <hashbits>_<dataset>_<num_classes>_class.pkl. Ensure your config points to the directory containing those files:

config['labelhashcodes_path'] = './labels/'
# Automatically loads: f"{hashbits}_{dataset}_{num_classes}_class.pkl"

Use --verify to confirm the Hadamard properties (balanced, orthogonal, ±1 values).

Configuration

All hyperparameters are defined in config files (configs/*.py). Key parameters:

Dataset Configuration

• dataset: Dataset name (cifar-10, cifar-100, clothing1M, food101N, webvision)
• num_classes: Number of classes

Model Configuration

• model1_type: Backbone architecture (PreResNet18SH, resnet50SH, InceptionResNetV2SH)
• hashbits: Length of hash codes (32 for CIFAR-10, 64 for CIFAR-100/WebVision, 128 for Food-101N)
• tau: Threshold for sample selection (default: 0.001)
• T: Temperature parameter for softmax (default: 2)
• Step: Update frequency parameter (default: 2)

Command Line Overrides

You can override config values via command line:

• --gpu: GPU ID
• --seed: Random seed
• --noise_type: Noise type (sym, asym, ins)
• --percent: Noise rate (0.2, 0.5, 0.8)
• --batch_size: Override batch size

Experiments

We provide shell scripts for reproducing paper results:

CIFAR Experiments

# CIFAR-10 
bash scripts/jumpteaching_cifar10.sh

# CIFAR-100 
bash scripts/jumpteaching_cifar100.sh

# CIFAR 
bash scripts/jumpteaching_cifar10_pairflip.sh
bash scripts/jumpteaching_cifar100_pairflip.sh

Real-world Noisy Datasets

# Clothing1M 
bash scripts/jumpteaching_clothing1M.sh

# Food-101N
bash scripts/jumpteaching_food101N.sh

# WebVision
bash scripts/jumpteaching_webvision.sh

Citation

If you find this code useful for your research, please cite our paper:

@article{ji2024jump,
  title={Jump-teaching: Ultra Efficient and Robust Learning with Noisy Label},
  author={Ji, Kangye and Cheng, Fei and Wang, Zeqing and Huang, Bohu},
  journal={arXiv preprint arXiv:2405.17137},
  year={2024}
}

You can also use the provided CITATION.bib.

License

This project is licensed under the MIT License - see the LICENSE file for details.

Acknowledgements

We thank the authors of the repositories Co-teaching, DivideMix and DISC for their excellent work!

Contact

For questions or issues, please:

• Open an issue in this repository
• Contact: [email protected]