[ICML2020] Xuxi Chen*, Wuyang Chen*, Tianlong Chen, Ye Yuan, Chen Gong, Kewei Chen, Zhangyang Wang

We proposed Self-PU Framework that introduces self-paced, self-calibrated and self-supervised learning to the PU field.

• Self-paced learning: We gradually selected confident samples from the unlabeled set and assign labels to them.
• Self-calibrated learning: Find optimal weights for unlabeled samples in order to obtain more source of supervision.
• Self-supervised learning: Fully exploit the learning ability of models by teacher-student structure.

conda install pytorch==0.4.1 cuda92 torchvision -c pytorch
conda install matplotlib scikit-learn tqdm
pip install opencv-python

Download cifar-10 and extract it into cifar/.

MNIST: Google Drive, Accuracy: 94.45%

CIFAR-10: Google Drive, Accuracy: 90.05%

MNIST:

python evaluation.py --model mnist.pth.tar 

CIFAR-10:

python evaluation.py --model cifar.pth.tar --datapath cifar --dataset cifar

python train.py --self-paced False --mean-teacher False 

python train.py --self-paced False --mean-teacher False --dataset cifar --datapath cifar

Training with self-calibation would be expensive. A cheap alternative:

python train_2s2t.py --soft-label

python train_2s2t.py --dataset cifar --datapath cifar --soft-label

python train_2s2t_mix.py --soft-label

python train_2s2t_mix.py --dataset cifar --datapath cifar --soft-label