GPU acceleration

[camsooper]

• Which cloud platform will support this?
• Can it be used for featurisation
• Can it be used for encoding
• (Or for UMAP calculation if we decide to add that?

[rmdocherty]

1. There are some providers (AWS, GCP, Oblivius (?)) who provision VMs with GPUS but that will incur much higher costs than a server less consumption model

2. Depending on the features used (Blurs, Sobel definitely - others unclear) yes. If I can find a way to map a function across a footprint kernel (I.e min/max/median) then could implement most.
   I believe Ilastix uses a reduced feature set (Gaussian, Sobel and Hessian) compared to Weka and these might all be relatively straightforward (except hessian) on GPU

3. Yes - it’s a major draw of the GPU model. The SAM embedding computation would be effectively instant on a good GPU w/ sufficient VRAM

4. Depends. GPUMAP (fork of UMAP with GPU support) hasn’t been updated to latest CuPy, but I think it’s literally just one function/line that needs fixing. I would need to go in depth into UMAP and CuPy to fix though.

[rmdocherty]

This looks very promising for Python GPU featurisations:
https://github.com/clEsperanto/pyclesperanto_prototype/

[rmdocherty]

Plan:

1. Add option to GPU accelerate SAM encoding
2. Make new branch with option to a) use pyclesperanto featurisation (need to write this) and b) GPU XGB as ensemble classifier
3. Consider alternate web hosting that makes use of 1. and 2.