/black

/black

Hi @rijobro ,

I slightly changed the _BatchInverseDataset to avoid circle-import, I didn't figure out why the TTA unit tests failed...
If adding "self.XXX = transform" in the init of _BatchInverseDataset, the unit tests can pass (XXX can be any word).
Could you please help take a look? Maybe this is some potential logic in your TTA module that I ignored?

Thanks in advance.

Paste the error log:

======================================================================
ERROR: test_single_transform (tests.test_testtimeaugmentation.TestTestTimeAugmentation)
----------------------------------------------------------------------
Traceback (most recent call last):
  File "/home/runner/work/MONAI/MONAI/monai/data/utils.py", line 264, in list_data_collate
    ret[k] = default_collate([d[k] for d in data])
  File "/opt/hostedtoolcache/Python/3.8.10/x64/lib/python3.8/site-packages/torch/utils/data/_utils/collate.py", line 55, in default_collate
    return torch.stack(batch, 0, out=out)
TypeError: expected Tensor as element 1 in argument 0, but got numpy.ndarray

During handling of the above exception, another exception occurred:

Traceback (most recent call last):
  File "/home/runner/work/MONAI/MONAI/tests/test_testtimeaugmentation.py", line 147, in test_single_transform
    tta(self.get_data(1, (20, 20)))
  File "/home/runner/work/MONAI/MONAI/monai/data/test_time_augmentation.py", line 183, in __call__
    inv_batch = inverter(inferred_dict)
  File "/home/runner/work/MONAI/MONAI/monai/transforms/inverse_batch_transform.py", line 92, in __call__
    return first(inv_loader)
  File "/home/runner/work/MONAI/MONAI/monai/utils/misc.py", line 71, in first
    for i in iterable:
  File "/opt/hostedtoolcache/Python/3.8.10/x64/lib/python3.8/site-packages/torch/utils/data/dataloader.py", line 517, in __next__
    data = self._next_data()
  File "/opt/hostedtoolcache/Python/3.8.10/x64/lib/python3.8/site-packages/torch/utils/data/dataloader.py", line 557, in _next_data
    data = self._dataset_fetcher.fetch(index)  # may raise StopIteration
  File "/opt/hostedtoolcache/Python/3.8.10/x64/lib/python3.8/site-packages/torch/utils/data/_utils/fetch.py", line 47, in fetch
    return self.collate_fn(data)
  File "/home/runner/work/MONAI/MONAI/monai/data/utils.py", line 288, in list_data_collate
    raise TypeError(re_str)
TypeError: expected Tensor as element 1 in argument 0, but got numpy.ndarray
Collate error on the key 'label' of dictionary data.

MONAI hint: if your transforms intentionally create mixtures of torch Tensor and numpy ndarray, creating your `DataLoader` with `collate_fn=pad_list_data_collate` might solve this problem (check its documentation).

I just tried the current state of your PR and test_testtimeaugmentation.py passes. Have you solved the problem? Or do I need to change something to be able to reproduce the problem?

Hi @rijobro ,

I think the root cause is that: the 2 failed unit tests of TTA missed ToTensorD() transform, so it means that we invert a Tensor data in a random transform, if do_transform, it will convert to numpy array, if not, still be Tensor, so sometimes can't collate a batch.
About my previous finding: as it's random behavior, adding self.transform = transform may affect the random result and if I change to prob=0.2, it also failed.
Now I just changed prob=1.0 in the TTA unit tests.
In the long term, I think maybe we can change the inverse logic to convert Tensor -> numpy no matter do_transform is True or not, or we change the TTA logic to convert Tensor -> numpy before inverse, now it's numpy -> Tensor.
what do you think?

Thanks.

I see. Perhaps best to implement both, since the inverse of ToTensor is ToNumpy, so that makes sense. Also inverse transforms will always need to be preceded by a ToTensor logic.

Hi @rijobro ,

Thanks for your help, yes, ToTensorD is supposed to be used in all MONAI transform chains, so the test failure is not a regular use case.
But any way, we can enhance the numpy and Tensor logic to make it more clear later.
This PR mainly decoupled invert transform with ignite, users can use it as post transform in regular program.
Could you please help review it?

Thanks in advance.

Hi @vfdev-5 ,

In this PR, I temporarily combined state.batch and state.output to construct a unique dict for the dict-based transform which is decoupled from ignite, just as we discussed in the ignite ticket before.
Could you please double confirm whether this PR has any potential issues?

Thanks.

@Nic-Ma I left a small comment in the code. Otherwise, I do not see any other issues with this approach.

Thanks for pinging!