Hi guys, I think I fixed the bug in keras. Here is the corresponding pull request : https://github.com/fchollet/keras/pull/3968
Best, Aloïs Le mercredi 5 octobre 2016 23:47:00 UTC+2, Pascal Lamblin a écrit : > > On Wed, Oct 05, 2016, Daπid wrote: > > On 5 October 2016 at 21:01, Pascal Lamblin <[email protected] > <javascript:>> wrote: > > > Just a hunch: is it possible that the "axis=2" parameter of > > > BatchNormalization has to be changed between TF an Theano, since they > > > may not use the same memory layout for convolutions? > > > > I don't think so, the layout is fixed between backends, and I am sure > > it is correct because the number of parameters is what I would expect. > > Using axis=1 throws an error in Keras (before ever dispatching the > > backend), since its dimension is None. > > Then, I don't know. It may be an issue in the gradient of some operation > in Theano. Can you try with test values and pdb, to try to pinpoint > which gradient operation inserts the tensor with a wrong size? > > > > > -- > > > > --- > > You received this message because you are subscribed to the Google > Groups "theano-users" group. > > To unsubscribe from this group and stop receiving emails from it, send > an email to [email protected] <javascript:>. > > For more options, visit https://groups.google.com/d/optout. > > -- > Pascal > -- --- You received this message because you are subscribed to the Google Groups "theano-users" group. To unsubscribe from this group and stop receiving emails from it, send an email to [email protected]. For more options, visit https://groups.google.com/d/optout.
