search

EderSantana / seya

Bringing up some extra Cosmo to Keras.
Other
377 stars 103 forks source link

About batch_size of ConvGRU #17

Closed t13m closed 8 years ago

t13m commented 8 years ago

Hi, I'm using ConvGRU of seya doing some experiments, and here is my question, why there must be a batch_size in ConvGRU.init()? Is there any way to create an instance of ConvGRU without knowing batch_size?

EderSantana commented 8 years ago

Its mainly to support both Theano and TF backend. TensorFlow doesn't know how to reshape things without having explicit batch sizes.

But, if that is crucial for you work, you can try to modify the code to use Theano.repeat and Theano.reshape using symbolic values. When TF learns how to handle symbolic resize, I'll update the code.

I made it this way so that ConvGRU could be pushed to Keras main repo at one point.