CausalVAE: Disentangled Representation Learning via Neural Structural Causal Models

Code for the paper CausalVAE: Disentangled Representation Learning via Neural Structural Causal Models (CVPR, 2021), by Mengyue Yang, Furui Liu, Zhitang Chen, Xinwei Shen, Jianye Hao, and Jun Wang.

If you find it useful, please consider to cite the paper.

This is a causal disentanglement image package.

Requirements

• python 3.10
• torch 2.3.1

Process

Generate synthetic data:

• Generate flow data: python ./causal_data/flow.py
• Generate pendulum data: python ./causal_data/pendulum.py

Train CausalVae:

• Train on flow data: python ./run_flow.py
• Train on pendulum data: python ./run_pendulum.py

Intervention on synthetic data by CausalVae:

• Intervention on flow data: python ./inference_flow.py
• Intervention on pendulum data: python ./inference_pendulum.py