Oh, and I forgot to mention. Some of the easily doable callbacks include: 1.
Verbose Logs (maybe progress bars ? Saw an issue earlier)
2.
Model Checkpoints
3.
Early Stopping
4.
Learning Rate annealing
As a second alternative, we can boil everything down and simply define learning
rate strategies, like linear, polynomial or exponential decreasing after fixed
amount of epochs.For a very naive alternative, we can even end up allowing the
user to provide a list having length equal to max_iter if it is specified. But
this doesn't sound too appetizing to me.
Karan.
_______________________________________________
scikit-learn mailing list
[email protected]
https://mail.python.org/mailman/listinfo/scikit-learn
