Problem: Cannot pass theano.shared to .itertrain()

[Philipp777]

First of all, thanks a lot for this clean piece of software :+1:

I am using .itertrain() to train a deep autoencoder with a train and a validation dataset (numpy.ndarray) supplied, while doing various monitoring tasks. The problem is the inefficient execution footprint since for every invocation of .itertrain() the datasets are copied from host to GPU. Correct me if I am wrong :) So, in order to overcome this bottleneck, I tried to supply the datasets as theano.shared objects to .itertrain() which doesnt seem to work. After consulting the documentation of .itertrain(), it is not possible to supply a theano.shared object directly, but a downhill.dataset.Dataset object can be supplied, which, in turn, can be populated with a theano.shared object. Unfortunately, this also doesnt work... What am I doing wrong?