ImPULSeS

Implementation of ImPULSeS: Robust and Efficient Imbalanced Positive-Unlabeled Learning with Self-supervision

Create conda environment

conda env create -f environment.yml
conda activate impulses

Do pretraining (options can be modified in config/temp.yaml)

python main_pretrain.py -c temp

Afterwards: finetuning of classification head (options can be modified in config/temp.yaml)

python main_finetune.py -c temp

Default setting of config/temp.yaml: Pre-training with debiased contrastive loss and fine-tuning with imbalanced nnPU loss both on imbalanced CIFAR10 data.

Parts of Code used from:

1. https://github.com/spijkervet/SimCLR
2. https://github.com/chingyaoc/DCL
3. https://github.com/guangxinsuuu/Positive-and-Unlabeled-Learning-from-Imbalanced-Data