Hi all, I was wondering if anyone has dealt with what I'm trying to do before. My goal is to take the model I've trained in Keras with the Tensorflow backend and optimize the graph for inference. So far I've been trying this with MNIST hello world example. The problem I'm having is that when I get the graph from the keras backend session, the input node has a -1 where the 'None' was (which should
be there to represent a wildcard value for batch size), but that keeps the graph from being rebuilt properly because the first layer has -1 in its placeholder dimensions! I get the following error.
2017-07-07 17:25:53.977324: W tensorflow/core/framework/op_kernel.cc:1148] Invalid argument: Shape [-1,28,28,1] has negative dimensions
2017-07-07 17:25:53.977393: E tensorflow/core/common_runtime/executor.cc:644] Executor failed to create kernel. Invalid argument: Shape [-1,28,28,1] has negative dimensions
[[Node: input_1 = Placeholder[dtype=DT_FLOAT, shape=[?,28,28,1], _device="/job:localhost/replica:0/task:0/gpu:0"]()]]