Weakly Supervised Disentangled Generative Causal Representation Learning

This repository contains the code for the paper Weakly Supervised Disentangled Generative Causal Representation Learning published at JMLR.

Model

Install prerequisites

pip install -r requirements.txt

Datasets

• CelebA
• Pendulum: the dataset is generated based on the source code from this repository.

Run

• Run DEAR on CelebA:

sh run_celeba_smile.sh

• Run DEAR on Pendulum:

sh run_pendulum.sh

Output

This will create a directory ./results/<dataset>/<save_name> which will contain:

• model.sav: a Python distionary containing the generator, encoder, and discriminator.
• gen.png: generated images.
• recon.png: real images (odd columns) along with the reconstructions (even columns).
• trav.png: traversal images.
• log.txt: All losses computed during training.
• config.txt: training configurations.

Help

Important arguments:

Generative model:
  --latent_dim          	dimension of the latent variable
  --prior {linscm, nlrscm, gaussian, uniform}
                        	prior distribution p_z (linear SCM, Nonlinear SCM, or independent ones)
  --labels {smile, age, pend}
                        	name of the underlying structure
                     
Supervised regularizer:
  --sup_coef          		coefficient of the supervised regularizer
  --sup_prop          		proportion of supervised labels
  --sup_type {ce, l2}		type of the supervised loss

Dataset:
  --dataset          		name of the data
  --data_dir          		directory of the dataset

Results

CelebA:

Pendulum:

Contact information

If you meet any problems with the code, please submit an issue or contact Xinwei Shen (xinwei.shen@connect.ust.hk).

Citation

If you would refer to or extend our work, please cite the following paper:

@article{JMLR:v23:21-0080,
  author  = {Xinwei Shen and Furui Liu and Hanze Dong and Qing Lian and Zhitang Chen and Tong Zhang},
  title   = {Weakly Supervised Disentangled Generative Causal Representation Learning},
  journal = {Journal of Machine Learning Research},
  year    = {2022},
  volume  = {23},
  number  = {241},
  pages   = {1--55},
  url     = {http://jmlr.org/papers/v23/21-0080.html}
}