Tensor computation for Brats transforms

[Nic-Ma]

Is your feature request related to a problem? Please describe.
Hi @rijobro ,

Thanks for your great work on the Tensor-based transforms.
We are working on optimizing the Brats tutorial performance:
https://github.com/Project-MONAI/tutorials/blob/master/3d_segmentation/brats_segmentation_3d.ipynb
And there are several random transforms in it, it may be helpful to move data to GPU Tensor, so is it possible to add Tensor support for the below 4 transforms first?

RandFlipd(keys=["image", "label"], prob=0.5, spatial_axis=0),
NormalizeIntensityd(keys="image", nonzero=True, channel_wise=True),
RandScaleIntensityd(keys="image", factors=0.1, prob=1.0),
RandShiftIntensityd(keys="image", offsets=0.1, prob=1.0),

I am not sure whether you have implemented them in Tensor locally?

Thanks in advance.

[rijobro]

Hi @Nic-Ma, I've just posted the current backend of the transforms that I'm proposing to update to MONAI. As you can see, RandFlip will use torch.Tensor under the hood, and the other 3 transforms are ambivalent to torch/numpy input.

I will prioritise getting these into MONAI as quickly as possible, once the current PR is merged. If you could help with the code review for that, I'll move onto the next ASAP.

[Nic-Ma]

Cool, sounds great! Thanks very much for your quick help!

[Nic-Ma]

• RandFlipd(keys=["image", "label"], prob=0.5, spatial_axis=0)
• NormalizeIntensityd(keys="image", nonzero=True, channel_wise=True)
• RandScaleIntensityd(keys="image", factors=0.1, prob=1.0)
• RandShiftIntensityd(keys="image", offsets=0.1, prob=1.0)