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

Fix dictionary order dependent behavior #12

Closed namsan96 closed 5 years ago

namsan96 commented 5 years ago

Previous code was using dictionary to store current parameters (names_weights). However, apply_inner_loop_update highly depends on the order of this dictionary key or value. It would cause issue #9 or #11 because of the order mismatch between training time and test time. So the dictionary is simply replaced with OrderedDict in order to preserve its key/value order.

issue: #9, #11

AntreasAntoniou commented 5 years ago

In my local version, the inner loop optimizer actually uses keys for robustness against this problem. I'll have a look at your proposal and potentially integrate it with my own.