Is this somehow related to keras-contrib? about batchrenormalization HOT 9 CLOSED

Yes, this repository was the base for the Contrib version. As of a few months ago, the Contrib version supercedes this local version.

Though I have not heard of the bug with Tensorflow backend. Could you possibly elaborate?

from batchrenormalization.

Hi @titu1994
thanks for the fast answer. For the bug see here:
https://github.com/keras-team/keras-contrib/blob/master/keras_contrib/layers/normalization.py#L224

from batchrenormalization.

Yes I read that. I was wondering what the actual bug was. Is it an issue in shapes, or wrong updates or something.

I haven't looked into batch renormalization since I implemented Group Normalization.

from batchrenormalization.

I haven't looked into batch renormalization since I implemented Group Normalization.

You talk about this here right? https://github.com/titu1994/Keras-Group-Normalization

from batchrenormalization.

Yes.

from batchrenormalization.

Yes.

Ahh ok nice. Thanks for the reply.
Did you do some experiments with Group N.? How is it compared to BN and Batch Renormalization?

from batchrenormalization.

It's good for batch sizes less than 8 in the few experiments I did. Anything above that and batch norm beats it at test time.

But not by a huge amount or anything. 0.5-1% difference between the two.

Haven't compared against batch renorm, but according to the group norm paper, batch renorm behaves close to batch norm performance for most batch sizes.

from batchrenormalization.

Ok. Thanks! :-)

from batchrenormalization.

Closing this...

from batchrenormalization.