Support for GPU-accelerated affine transformations

[lorenzoh]

Affine transformations on CPU are already pretty fast, but GPU-accelerated transformations are useful in CPU-constrained environments like Google Colab where a GPU is available but only 2 mediocre CPUs which can lead to a bottleneck in the data pipeline.

Some open questions:

• Would the existing transforms work out-of-the box on CuArrays (i.e. does ImageTransformations.warp[!] work on CuArrays) ? What about for masks, i.e. integer arrays? Otherwise, what would be necessary to implement that?
• Performance-wise, do the transforms need to be applied to a whole batch at once or is it as fast to apply them to samples individually? Former would require first resizing images to the same size and wrapping in a Batch wrapper item.

Can someone who has experience with image transformations on GPU chime in?

@jsamaroo

[ToucheSir]

Think you meant to tag @jpsamaroo? Also, https://github.com/JuliaImages/ImageTransformations.jl/pull/156 looks promising here!

[lorenzoh]

Think you meant to tag @jpsamaroo?

Yup 😅

[jpsamaroo]

Writing image transformations with KernelAbstractions.jl is probably the best approach, since you get CPU+CUDA+AMDGPU from just a single kernel (although you might want to write the CPU kernels differently for better performance). That sort of code should probably go in ImageTransformations.jl in some shape or form.