Background:
I've been looking at how to create a recurrent seq-to-seq model, that's not transformers. The ideas I implement do not work. It seems like off the well trodden path, there are traps everywhere - how should I tune parameters, add biases, normalize, is this dataset impossible, gradient explosion and vanishing, etc.
From a "research = gradient descent" point of view, I'm stuck at a point with no gradient - I have no idea what I'm doing wrong, or what to will get a better result. Am I missing a workflow. intuition, or tools, or other things?
