Fix KL normalization - Githubissues

szagoruyko / attention-transfer

Improving Convolutional Networks via Attention Transfer (ICLR 2017)

https://arxiv.org/abs/1612.03928

1.43k stars 274 forks source link

Closed szagoruyko closed 6 years ago

szagoruyko commented 6 years ago

F.kl_div normalizes by the number of elements in tensor, fixed to normalize by minibatch size
also removed 2

Checked that this produces the same results as in the paper.

Fixes #18 and #7