segment_volume fails with past models containing NumpyToTensor key

[mariehbourget]

Issue description

The function segment_volume fails after the merge of #818 when segmenting with a model containing the "NumpyToTensor": {} field in transformation parameters. (tagging @dyt811 and @cakester so they are aware)

I found this #758 (comment) on the opened PR #758 with what seems to be a fix for this. If that is the case, I think the fix should be done in a dedicated PR as soon as possible.

Current behavior

Config file of models trained prior to the merge of #818 contains the "NumpyToTensor": {} field. This config file is used by the function segment_volume for segmentation.

The segmentation fails with the following error when undoing transforms:

Traceback (most recent call last):
  File "/home/mhbourget/venv-ivadomed-38c/bin/ivadomed", line 11, in <module>
    load_entry_point('ivadomed', 'console_scripts', 'ivadomed')()
  File "/home/mhbourget/ivadomed/ivadomed/main.py", line 560, in run_main
    run_command(context=context,
  File "/home/mhbourget/ivadomed/ivadomed/main.py", line 340, in run_command
    run_segment_command(context, model_params)
  File "/home/mhbourget/ivadomed/ivadomed/main.py", line 274, in run_segment_command
    pred_list, target_list = imed_inference.segment_volume(path_model,
  File "/home/mhbourget/ivadomed/ivadomed/inference.py", line 407, in segment_volume
    pred_list, target_list, last_sample_bool, weight_matrix, volume, image = reconstruct_3d_object(
  File "/home/mhbourget/ivadomed/ivadomed/inference.py", line 483, in reconstruct_3d_object
    image_reconstruction(batch, preds, undo_transforms, i_slice, image, weight_matrix)
  File "/home/mhbourget/ivadomed/ivadomed/inference.py", line 609, in image_reconstruction
    pred_undo, metadata = undo_transforms(image, batch['gt_metadata'][smp_idx], data_type='gt')
  File "/home/mhbourget/ivadomed/ivadomed/transforms.py", line 184, in __call__
    sample, metadata = tr.undo_transform(sample, metadata)
  File "/home/mhbourget/ivadomed/ivadomed/transforms.py", line 210, in undo_transform
    return list(sample.numpy()), metadata
AttributeError: 'list' object has no attribute 'numpy'

Note: a newly trained model without the NumpyToTensor key does not produced this error.

Expected behavior

• "NumpyToTensor": {} field from old config files should be ignored somehow to ensure retrocompatibility and avoid this error.
• functional_tests/test_segment_volume.py should test for this. This is not the case right now and tests are passing because they don't use old config files.