Tensorflow GPU #145
Description
I've been running a modified version of the ImagenetTF app on the Places dataset, and whenever I check the worker container logs it looks like all of the TF graph nodes are being put on CPU. I've looked at the model defs (.pb and py files) and they are correctly requesting all nodes to be placed on GPU, but the soft-placement constraint is allowing them to failover to CPU I guess.
Does anyone know what could be causing this? Could it be the JavaCPP bindings? I'm using the 03/05 prebuilt GPU jars.