`decollate_batch` return an empty list

[yiheng-wang-nv]

decollate_batch will return an empty list, if input is a dict, and one of its value is an empty list.
For example:

data = {
    "class": ['ToTensord', 'ToTensord'],
    "id": torch.tensor([1, 2]),
    "orig_size": [],
}

decollate_batch(data) will return []

This case comes from a classification task, where I used SupervisedTrainer, and labels are achieved from ToNumpyd. The data in the above example belongs to label_transforms which is in engine.state.batch

[yiheng-wang-nv]

For classification tasks, labels may only have one dimension, thus in this line, info[TraceKeys.ORIG_SIZE] will be (), which can lead to the issue above mentioned.

For example:

a = np.random.rand(2)
print(a.shape[1:])

will get ()

[wyli]

#3373 is one way to address this:

>>>decollate_batch({"a": [1, 2], "b": []}, pad_zip=True, fillvalue=None)
[{'a': 1, 'b': None}, {'a': 2, 'b': None}]
>>>decollate_batch({"a": [1, 2], "b": []}, pad_zip=True, fillvalue=0)
[{'a': 1, 'b': 0}, {'a': 2, 'b': 0}]

is this good enough?

[yiheng-wang-nv]

Thanks @wyli , I think it solves the issue.

In addition, should this place be modified too? For input that only has one dimension.

MONAI/monai/transforms/inverse.py

Line 64 in 9048166

info[TraceKeys.ORIG_SIZE] = data[key].shape[1:]

[wyli]

In addition, should this place be modified too? For input that only has one dimension.

MONAI/monai/transforms/inverse.py

Line 64 in 9048166

info[TraceKeys.ORIG_SIZE] = data[key].shape[1:]

I feel it makes sense to have [] spatial shape for non-spatial data. or we can use TraceKey.NONE but I don't see any use cases that require this value.