Often you have too much data and it can all fit on the GPU's RAM. If it is your case, you can now provide keep_on_cpu=True as argument to the Dataset constructor. Theano will transfer data as it needs (one batch at the time) to the GPU. See this for more infos https://groups.google.com/forum/#!topic/theano-users/ZagTpkB4hGg
Often you have too much data and it can all fit on the GPU's RAM. If it is your case, you can now provide
keep_on_cpu=Trueas argument to the Dataset constructor. Theano will transfer data as it needs (one batch at the time) to the GPU. See this for more infos https://groups.google.com/forum/#!topic/theano-users/ZagTpkB4hGg