skip RandAffine if possible

[rijobro]

Fixes #2124.

Description

Skip RandAffine if possible.

Status

Ready

Types of changes

• Non-breaking change (fix or new feature that would not break existing functionality).
• Integration tests passed locally by running ./runtests.sh -f -u --net --coverage.
• Quick tests passed locally by running ./runtests.sh --quick --unittests.
• In-line docstrings updated.
• Documentation updated, tested make html command in the docs/ folder.

[rijobro]

When debugging, I put a print statement to verify that the first unit test skips the resampling, as expected.

[wyli]

thanks, need something like

        if not self.rand_affine.do_transform and self.rand_affine.spatial_size is None:
            for idx, key in enumerate(self.keys):
                if self.rand_affine.resampler.as_tensor_output:
                    d[key] = torch.tensor(key)
                elif torch.is_tensor(d[key]):                    
                    d[key] = d[key].cpu().numpy()
            return d

for RandAffined

need to double-check the output data type, make sure that the PR doesn't change the dtype for integer image

[rijobro]

also need to think about the inverse for RandAffined...

[rijobro]

@wyli why would the dtype change if the image isn't modified? Because of the torch/numpy conversion?

[wyli]

the root cause is that the current resampling requires floats...

MONAI/monai/transforms/spatial/array.py

Lines 1240 to 1242 in 41e62e5

	out = torch.nn.functional.grid_sample(
	img.unsqueeze(0).float(),
	grid.unsqueeze(0).float(),

if we bypass the resample step, I feel the datatype might be different, but I haven't checked all the details...

[wyli]

thanks!

[wyli]

the dev branch had its 1024th merged pull request 🎉

[rijobro]

that was quick, it's only a few weeks old! 😅