Dependencies

• Python 3.6.9
• Pytorch 1.1.0
• torchvision 0.3.0
• numpy 1.17
• statsmodels 0.10.1
• matplotlib 3.1.1
• Pillow 6.1
• scipy 1.3.0

The dataset will be downloaded automatically for CIFAR10. For CelebA, please refer to the official website.

For toy Data, 'python train_toy.py ./config/Mog_pid.yaml'. For CIFAR10 and celebA, the baseline can be reproduced by the following:

'python train.py ./config/cifar_pid.yaml'

'python train.py ./config/celeba_pid.yaml'

For our proposed method, use the following commands:

'python train_pid.py ./config/cifar_pid.yaml'

'python train_pid.py ./config/celeba_pid.yaml'

The hyperparameters can be adjusted in the corresponding config files. Specifically, the 'iv' denotes the coefficient for CLC.