Guidance for custom datasets

[nudro]

Thanks for releasing deepscm. I really enjoyed reading the paper, and appreciate the code. I see under deepscam/datasets there are a few scripts for the mnist and medical datasets. Do you have any guidance on the best way to prepare a custom dataset to train?

[pawni]

Hi there and thanks for your interest in the paper and code.

Sorry for only responding now -- somehow I didn't get notified about this issue. Could you share some more information about what exactly your looking for?

In general, we have done little data preprocessing and the main pointers that I can think of straight away is to add transformations to the code that do the normalisation for you (see https://github.com/biomedia-mira/deepscm/blob/master/deepscm/experiments/morphomnist/base_experiment.py#L233-L237). Apart from this you will need some idea of how to model the causal graph.

You can find an example of integration of a custom dataset here as well: https://github.com/jcreinhold/counterfactualms