Diff-SCM aims at developing solutions based diffusion probabilistic modeling (DPM) for causal machine learning. Our algorithm perform counterfactual inference using Pearl's abduction-action-prediction strategy using a deterministic inference based on a deterministic diffusion inference (DDIM). We have also applied to the method for lesion localization in Brain MRI.
The code is this repo implements the following papers:
Sanchez, Pedro, and Sotirios A. Tsaftaris. "Diffusion Causal Models for Counterfactual Estimation." In Conference on Causal Learning and Reasoning (CLeaR). 2022.
Sanchez, Pedro, Kascenas, Antanas, Liu, Xiao, O'Neil, Alison and Sotirios A. Tsaftaris. "What is Healthy? Generative Counterfactual Diffusion for Lesion Localization." In Deep Generative Models workshop at MICCAI 2022 (DGM4Miccai). 2022.
Note. If one is looking for particular bits of code with our contributions, consider checking the diff_scm/sampling/sampling_utils.py file. The code for counterfactual estimation, classfier-free (implicit) inference and dynamic normalisation is there.
The provided .yml file can be use for creating a conda environment with the following command.
conda env create -f environment.yml
A few hyperparameters are particularly important for counterfactual estimation. In particular, the parameters the control the strenght of the intervention. The sampling.classifier_scale, sampling.norm_cond_scale, sampling.sampling_progression_ratio are the some of these and can be found in the diff_scm/configs folder. Also, se this folder for example configuration on MNIST and BRATS dataset.