keras support

[twiecki]

Are there plans to add keras support to gelato?

[ferrine]

I think not in nearest future. I've just tried to check if it is supported out of the box but keras seems to force .eval() every symbolic weight.

CC @fchollet

%env KERAS_BACKEND=theano
import keras
from keras.layers import Dense, Activation
from keras.models import Sequential
import pymc3 as pm
from gelato.specs import NormalSpec, LognormalSpec

# create prior spec
hyper = LognormalSpec(sd=1).with_shape(())
init = NormalSpec(sd=hyper)
# create model
model = Sequential()
# try to init keras layer with gelato spec -> fail
with pm.Model():
    model.add(Dense(kernel_initializer=init, bias_initializer=init, units=64, input_dim=100))

----------
/usr/local/lib/python3.4/dist-packages/keras/backend/theano_backend.py in variable(value, dtype, name)
    141                               theano.tensor.TensorConstant)):
    142             # Support for RandomStreams().normal(), .uniform().
--> 143             value = value.eval()
MissingInputError: Undeclared input

[ferrine]

I've opened an issue there, hope it can be supported by Keras