How to evaluate metrics every batch instead of epoch?

Part 1 (2020)
1

I’d like to evaluate metrics for training every batch instead of every epoch. And I want to apply SaveModelCallback based on the best result of the batch in an epoch. How to do that? Thanks.

2

I’d like to evaluate metrics for training every batch instead of every epoch

Look into the callback system. To do something every batch, you’ll likely use the after_batch event and then do whatever evaluation you want there.

I want to apply SaveModelCallback based on the best result of the batch in an epoch

Again you’ll likely want something in the after_batch event. You can probably copy SaveModelCallback and instead of having it compare models in after_epoch and then save in after_fit, you’ll do the comparisons in after_batch and save whenever you want.

1 Like