MNIST-based causal effect dataset simulation + Neural Network trainers

Code of the paper: Learning high-dimensional causal effect.

In this repository, we propose simulation of causal effect and study of deep learning models on the simulated dataset.

To run the code:

python code/main.py --encoder '<encoder-name>' --treatment 'odd-even'

For the argument encoder, the following are options:

encoder-name: 'resnet', 'vit', or 'fc'

resnet is for ResNet50 as representation learner (encoder model) vit is for Vision Transformer model as representation learner (encoder model) fc is for Dragonnet based model where feed-forward layers are representation learner (encoder model).

This work is inspired by Claudia Shi et. al. https://arxiv.org/abs/1906.02120; https://github.com/claudiashi57/dragonnet. We adapted a few parts of their code.

For the Vision Transformer class, we refer code presented here: https://keras.io/examples/vision/image_classification_with_vision_transformer/