[sseveran]

I had tried to do the same. However given the CPU/GPU imbalance on AWS GPU instances I have resorted to building a fully "rendered" training set and doing all the transforms in spark.

See: https://github.com/tensorflow/tensorflow/issues/13610