Implement rescaling inputs for domain adaptation

[stmartineau99]

motivation

• a source model was trained on data with vsize 10A, so the adapted model should also be trained on 10A data.

two ideas to solve this problem:

1. rescaling after patch extraction
2. rescale full volume inputs then write to temp files

I decided to try option 2. because option 1. requires rescaling then resizing to get back to the original patch size. We are already writing to temp files so I thought rescaling could be done at this step, because just rescaling (one interpolation) would probably result in less information lost than rescale + resize (two interpolations).

changes to semisupervised_training.py get_unsupervised_loader

• rescaling inputs to target_vsize, applied differently to float data and int data (mask)
• rescaling factor is automatically calculated by reading the voxel size from an .mrc file
• rescaled volumes are written to temporary files, which are given to RawDataset()
• rewrite get_stacked_paths() to be more memory efficient, write to $TMPDIR instead of /tmp
• tested case 2 (raw + sample mask) - works
• tested case 4 (raw only) - works

changes to domain_adaptation.py mean_teacher_adaptation

• add target_vsize argument
• moved data loaders and check_loader, to before torch_em.load_model
• before it was loading the model from the checkpoint which would spawn a bunch of workers resulting in freezing; this can be avoided when checking the loader

[constantinpape]

It looks like this was added by accident.