I think we should port the notebook generated by Ian into the codebase proper.
Personally, I can see a very strong value proposition in having a classifier which can run well on CPU (aka, what Ian started)- especially if we take this onto some kind of website.
Yes, its possible that a CNN approach could do better, but we should be asking ourselves the questions:
by how much?
is the gain even useful given im almost convinced you will need at least a £500 gpu to train it or run inference on it
is it smart to have the input constrained to a certain image dimension? (we currently dont need to worry about that with a pixel-level approach
@murraycutforth, @ipoole any takers on this issue?
I can also do it if not
...................Yes, just for my own interest, I still want us to train a CNN when we have the GT :)
I think we should port the notebook generated by Ian into the codebase proper.
Personally, I can see a very strong value proposition in having a classifier which can run well on CPU (aka, what Ian started)- especially if we take this onto some kind of website.
Yes, its possible that a CNN approach could do better, but we should be asking ourselves the questions:
@murraycutforth, @ipoole any takers on this issue? I can also do it if not
...................Yes, just for my own interest, I still want us to train a CNN when we have the GT :)