Stopping early, termination criterion

Tutorials
1

Hi

The examples in the Pyro docs usually run a given number of iterations. Is there any convenient convergence/stopping criterion code lurking in Pyro, e.g. tolerance of norm of gradient of the ELBO (or whatever Stan is using)?

Thanks in advance,

2

depending on what you want to do you can look at the grad of the loss or use a learning rate scheduler though i think you’d have to use the dev version of pyro since it fixes a surfaced bug.

3

Thanks! But, a look at the lr scheduler code in PyTorch suggests that they don’t have any sort of stopping criterion. I’ll take a look at the grad of the loss.

4

did you mean the differentiable_loss method? it seems to be dev only, and the loss & loss_and_gradients methods return floats.

5

did you mean the differentiable_loss method

yes and yep you’ll have to use dev until the next release.