Optimizing is a great way to map a new system's territory! Current trying to make a non-interactive machine learning systems interactive: a treasure hunt in a strange land!
2.5s -> 250ms: reimplement on GPU
250ms -> 50ms: fewer stages
50ms -> 25ms: fewer textures
25ms -> 2.5ms: ?
Conversation
Doing this has also made me take the model apart and put it back together in different ways a couple times, which has helped me understand how it works much better than if I'd implemented it very literally from the paper. Active learning is powerful!
And they didn’t really *fail*: most of them were faster than the previous, though still not fast enough…
And actually, the instances which were *slower* (when I expected them to be faster) are even more interesting. One of them I don't understand yet.
5