Class-specific Augmentation based Disentanglement (CAD) for ID-PLL

[CVPR 2026] This is the implementation of the paper: Mitigating Instance Entanglement in Instance-Dependent Partial Label Learning.

Requirements

Python 3.8.13
numpy 1.22.3
torch 1.10.0
torchvision 0.11.0
diffusers 0.28.2

Training

data preparing

To synthesize candidate labels, the annotation model weights should be downloaded from this link and place them into the ./partial_models/weights/ directory.

demo

First, generate class-specific augmentations:

python -u csaugmentation.py --dataset cifar10
python -u csaugmentation.py --dataset cifar100
python -u csaugmentation.py --dataset pet37
python -u csaugmentation.py --dataset flower102
python -u csaugmentation.py --dataset fmnist

Then, train the model:

python -u main.py --dataset cifar10
python -u main.py --dataset cifar100
python -u main.py --dataset pet37
python -u main.py --dataset flower102
python -u main.py --dataset fmnist

Reference

https://github.com/wu-dd/DIRK