Many many many thanks -- thank you! The architecture here is certainly specialized, though it should apply quite well to a number of other usecases at similar i

Many many many thanks -- thank you! The architecture here is certainly specialized, though it should apply quite well to a number of other usecases at similar image scales. I believe the only other main changes needed for a larger-image dataset (like ImageNet) would be adding maybe 2-3 blocks, so an additional 6-9 convolutions (making it Resnet15-Resnet18). The hyperparameters are fiddly but the region of good performance is pretty flat, so I'd expect it to do well. I might take it for a spin on something else at some point, let me know if you do! One benefit of having to go extremely fast is that it sort of makes you have to make your architectures as simple and short as possible. This helps with generalization to other datasets, I'd reckon from my experience here and other professional experience. Thanks for your question -- much appreciated -- and there are indeed a lot of potentials for this little research project, hope this helps!