TestTimeAugmentation assumes label is present

[rijobro]

Describe the bug
During training/testing/validation, ground truth labels are often present. However, at deployment only the image will be present. The way TTA is currently set up assumes that both image and label are present, which needs to be modified.

cc @diazandr3s

[wyli]

do you think this could be addressed by InvertD @Nic-Ma ?

MONAI/monai/transforms/post/dictionary.py

Line 392 in a566dc6

class Invertd(MapTransform):

(looks like this num_workers is not used

MONAI/monai/transforms/post/dictionary.py

Line 432 in a566dc6

num_workers: Optional[int] = 0,

)

[Nic-Ma]

Hi @wyli ,

You are right, let me try to update the Invertd ASAP, and I also double confirm the TTA logic in the mean time.

Thanks.

[Nic-Ma]

Hi @rijobro ,

I checked the TTA code and this part confused me:

# do model forward pass
batch_output = self.inferrer_fn(batch_images)
if isinstance(batch_output, torch.Tensor):
    batch_output = batch_output.detach().cpu()
if isinstance(batch_output, np.ndarray):
    batch_output = torch.Tensor(batch_output)

If the output is Tensor, it will be changed to numpy array and then change back to Tensor?

Thanks.