search

AntreasAntoniou / HowToTrainYourMAMLPytorch

The original code for the paper "How to train your MAML" along with a replication of the original "Model Agnostic Meta Learning" (MAML) paper in Pytorch.
https://arxiv.org/abs/1810.09502
Other
773 stars 137 forks source link

How often are you calling the cosine annealer? (scheduler) #43

Closed brando90 closed 2 years ago

brando90 commented 2 years ago

Are you calling the cosine annealing every epoch? Are you restarting it? Where are those details?

AntreasAntoniou commented 2 years ago

The cosine annealer is updated once every epoch, for a total of 200-250 epochs and anneals all the way down to 0.00001 from 0.001. By updated here I mean that I simply take the next step in the cosine cycle that I am using. There are no warm restarts.

brando90 commented 2 years ago

Thank you!