Closed bd3dowling closed 1 month ago
I read around and discovered that there is compute available to the maths department in the form of those described here and here. However, it appears these didn't have GPU capabilities. I contacted Andy Thomas at Imperial who directed me to this page which describes the GPU compute available to maths researchers. These capabilities should suffice (multiple 3090Ti clusters) but unsure. In the event they don't, Imperial's research compute service does have a Pilot system (HX1) with 60 A100-80s, but access is apparently quite restricted. They have other GPU nodes but these aren't as powerful as the maths ones. To get access to this system, Deniz has to register with them and then register me, and we would need to apply to get special access to the HX1 system.
I think, following some discussions I had with Ben B, that focussing on the tractable, low dimension, Gaussian (RF or MM) example cases could be an idea and once we have a method really well fleshed out we can consider any of the above options for testing on higher resource intensive tasks (e.g. image based ones). Working with the lower dimension CIFAR dataset also provides a case for testing locally (i.e. on my 3080) before trying to run on higher dimensional datasets (e.g. FFHQ or ImageNet) on the research clusters (so we can iron out potential image-related idiosyncrasies and avoid naieve experimentation on the clusters).
Continuing discussion in discussions board rather than issue: https://github.com/bd3dowling/diffusion-research/discussions/24
Figure out access to and capabilities of Imperial compute resources.