1198 efficient differentiable bilateral filter

[charliebudd]

Fixes #1198 .

Description

Adds two bilateral filter algorithms to the c++ extention module. Each have both a cpu and a cuda implementation.
One works by a bruteforce kernel, the other using a more refined approximate solution presented here...https://graphics.stanford.edu/papers/permutohedral/
The desired implementation is chosen via a boolean flag at the python interface level.

There are improvements which could be made but this is long overdue and already a very large PR.
There are also two issues I'm aware of but both are imperceptible at normal scales...

• The approximate cuda implementation shows some small variablity in its output.
• Both the approximate cpu and approximate cuda implementations show slight asymetry in the filtered output (This may be an artifact of the approximation method)

Status

Ready

Types of changes

• Non-breaking change (fix or new feature that would not break existing functionality).
• Breaking change (fix or new feature that would cause existing functionality to change).
• New tests added to cover the changes.
• Integration tests passed locally by running ./runtests.sh --codeformat --coverage.
• Quick tests passed locally by running ./runtests.sh --quick.
• In-line docstrings updated.
• Documentation updated, tested make html command in the docs/ folder.

[wyli]

/black

[wyli]

thanks @charliebudd this is very nice, would be great to have a short notebook demo

[faebstn96]

Hi @charliebudd ,
Thanks for your implementation. Unfortunately, I am failing to get the layer running. When trying to call the forward pass of the bilateral filter with
filter_layer = monai.networks.layers.BilateralFilter(spatial_sigma=5, color_sigma=0.5, fast_approx=True)
filter_layer.apply(tensor_in) # tensor_in: arbitrary 3,4 or 5 dim torch tensor

I get problems with the save_for_backward function, as it does not accept the filter parameters (I think save_for_backward is actually made for storing the input tensor and not for the parameters):
TypeError: save_for_backward can only save variables, but argument 0 is of type int

Wouldn't it be easier to implement the bilateral filter class as a nn.Module instead of torch.autograd.Function (like, e.g., the Gaussian filter)? Then, one could store the parameters in an instance variable.

Thanks for your help!

[faebstn96]

I just recognized, you can just assign the parameters directly to ctx instead of using save_for_backward. E.g., like
ctx.spatial_sigma = spatial_sigma

Then the layer at least does not throw an error.

[charliebudd]

Indeed, thanks for pointing out this out. I have made the changes and have drafted a PR #1888. Sorry for the error.

[faebstn96]

Thank you so much!