Excluding voxels with no labels

[shmann]

For transfer learning, jaccard loss decreases while dice scores remain relatively constant, and the resulting model spits out empty predictions. Blocks without labels may be interfering with training--per the AMS paper: "The cubes of the MRI scans that did not contain any meningioma voxels were excluded from the training set."

I would like to modify nobrainer.dataset.get_dataset to include a flag to discard blocks without associated labels. Does this seem reasonable? For example:

if(exclude_no_labels): dataset=dataset.filter(lambda x, y: tf.reduce_max(y) != 0)

[satra]

@shmann - i think that would be a good addition. it would have to be limited to dataset generators that map to tensors that have segmentation information.