I have used the tutorial as specified to download about 3k images but needed to gather more images for my project in order to generalize to the problem I’m dealing with.
My question is a general one: I have now about 4.6k images which the optimal learning rate was suggested to be ~ 1e-3. Using this learning rate, batch=16, I’ve seen a consistent increase in f1 and slow (but very spiky) loss decrease.
I’ve now reached the point where i’m still seeing small improvements at 200 epochs which seems quit high. Is this reasonable?
Obviously I could increase batch size, decrease learning rate. But the answer to many ML questions is: “It depends”. Curious to hear other thoughts on other routes to go.