[data.augmentations] Move transforms to augmentations; Implement auto-parellisation for ElasticDeformation (partial affects #5); Reflect changes in all relevant submodules

[data.datamodule] Fixes cached dataset model transform application; Improves logging and documentation; Estimate loaded sample sizes and RAM occupance

[data.datamodule] Double-check model_transforms are set before datasets are instantiated; Remove model_transforms from constructor (non-sensical); Improve documentation on model-transforms

[data.montgomery/shenzhen.loader] Simplify loaders, adjust some variable names, comments and add some (commented out) test code to visualize generated images

Maxime DELITROZ authored 1 year ago and André Anjos committed 1 year ago

We decided to handle the resizing of the input at the model level and not at the dataset level anymore. Loaders, docstrings and tests of Shenzhen and Montgomery were updated accordingly to only load data, remove black borders and crop to square shape. [image_utils.RemoveBlackBorders] had to be modified to allow functional utilization

[data.montgomery.loader] Use load_pil() instead of load_pil_baw() since Montgomery images are already in grayscale PNG (not black-and-white!)

[data.transforms] Properly implement ElasticDeformation so it works with images or batches; Update test case for transforms; Create new test case for image_utils; Update models to reflect changes (partially addresses #5)

[ptbench.data.datamodule] Implemented typing, added more logging, implemented sampler-balancing depending on sample class prevalence, parallelized dataset caching

Daniel CARRON authored 1 year ago and André Anjos committed 1 year ago

Implemented BCEWithLogitsLoss reweighting function
Removed save_hyperparameters from model
Apply augmentation transforms on singular images
Fixed model summary

[datamodule] Slightly streamlines the datamodule approach; adds documentation; adds type annotations; adds TODOs