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

fixed rotationaugmentation for omniglot #2

Closed polo5 closed 6 years ago

polo5 commented 6 years ago

In data.py, the rotation indexes k should be integers in [0,1,2,3] when given to np.rot90.

The current version seems to use k in [0,90,180], which amounts to only 2 rotations (since np.all(np.rot90(x, 90*0) == np.rot90(x, 90*2)))

AntreasAntoniou commented 6 years ago

Good catch! I was using scipy rotate before which took in actual degrees. Thanks for spotting. Going back and forth between packages too much can do this.