Because models might take in very different data, we write a data pipeline for each model (if needed).
Each model should also have a predict function that outputs results in (ground truth, prediction) tuples into a output file for downstream evaluation and visualization.
The output results are stored in the output folder in the format of <dataset>-<model>_out.pkl.
A bunch of UQ algorithms. For post-hoc methods, it takes in a output pickle file.