RandGaussianNoise and SavitzkyGolaySmooth Torch/NumpyTransforms

[rijobro]

Related #2231, #2278.

Description

This is the first of many PRs to add support for Torch and Numpy input in our transforms, minimising conversion between the two and allowing for GPU data.

In this PR I've added the basic components:

• One example of a transform that is ambivalent to Torch/Numpy input (RandGaussianSmooth)
• One example of a transform that potentially requires conversion (SavitzkyGolaySmooth)
• Functionality for switching between types
• Functionality for harmonising dtypes
• Tests use numpy, torch and torch.gpu input data
• Convenience enums (DataObjects)

Status

Ready/Work in progress/Hold

Types of changes

• Non-breaking change (fix or new feature that would not break existing functionality).
• New tests added to cover the changes.
• 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.

[wyli]

thanks for the PR, I'm looking into the details. @Nic-Ma could you please help review it as well?

[rijobro]

@wyli The test seems to be failing on GaussianFilter, but that hasn't been modified. Any ideas? I tried re-running the tests.

[Nic-Ma]

Hi @rijobro @wyli ,

Thanks for the quick PR, I planned to review it today, but I am blocked by other internal task now.
Will try to review it ASAP.

Thanks.

[wyli]

looks good to me, I mainly have some comments on the styles and type hints for discussions @Project-MONAI/core-reviewers (and testing the github team feature...)

[wyli]

@wyli The test seems to be failing on GaussianFilter, but that hasn't been modified. Any ideas? I tried re-running the tests.

looks like it's been addressed by rerunning, probably a GPU testing job scheduling issue, we'll migrate to blossom soon

[madil90]

/build

[Nic-Ma]

Sorry for the delay response, very happy to see the great progress of this task.
The overall design looks good to me, put some minor comments inline.
And I want to double confirm the direction: do we want to change transforms to accept both numpy and tensor data and use PyTorch tensor for the real computation logic to support differentiable transforms later?

Thanks.

[rijobro]

@Nic-Ma

do we want to change transforms to accept both numpy and tensor data and use PyTorch tensor for the real computation logic to support differentiable transforms later?

Correct, the majority of transforms won't need numpy<->torch conversion, so adding in differentiable transforms should be possible.

[rijobro]

@wyli some flake8 errors have appeared since switching to Ndarraytensor: https://github.com/Project-MONAI/MONAI/pull/2740/checks?check_run_id=3322770593#step:7:376.

In this snippet:

   def __call__(self, img: NdarrayTensor) -> NdarrayTensor:
        """
        Apply the transform to `img`.
        """
        self.randomize(img.shape)
        if self._noise is None:
            raise AssertionError
        if not self._do_transform:
            return img
        noise, *_ = convert_data_type(
            self._noise, type(img), dtype=img.dtype, device=img.device if isinstance(img, torch.Tensor) else None
        )
        return img + noise

It's complaining about img.device for np.ndarray. As in, it thinks img is only of type np.ndarray, even though it's of type NdarrayTensor, which should be both np.ndarray and torch.Tensor.

[Nic-Ma]

Thanks for the quick update.
Looks good to me now.

Thanks.

[rijobro]

@wyli I'm still struggling with the flake8 problem, do you think you could take a look? Cheers

[wyli]

Sure I can look into this later today