将 Keras 模型导出到具有 (None, 2) 输出形状的 protobuf

Question

I have a Keras model I'm trying to export to ProtoBuf

The final couple of layers look like this:

features (Dense)                (None, 128)          49280       concatenate_1[0][0]              
__________________________________________________________________________________________________
gaze_target (Dense)             (None, 2)            258         features[0][0]      

I try exporting it like this:

sess = K.get_session()

constant_graph = graph_util.convert_variables_to_constants(sess, sess.graph.as_graph_def(), 'gaze_target')
graph_io.write_graph(constant_graph, 'export', 'output.pb', as_text=False)

This errors with this:

~/anaconda3/envs/tensorflow_p36/lib/python3.6/site-packages/tensorflow_core/python/framework/graph_util_impl.py in extract_sub_graph(graph_def, dest_nodes)
    191 
    192   if isinstance(dest_nodes, six.string_types):
--> 193     raise TypeError("dest_nodes must be a list.")
    194 
    195   name_to_input_name, name_to_node, name_to_seq_num = _extract_graph_summary(

TypeError: dest_nodes must be a list.

How do I export this model to ProtoBuf? (Ultimately for use on SageMaker)

Answer 1

Thanks to a colleague at work we worked this out. The parameter to the graph_util.convert_variables_to_constants method isn't the layer name, but instead is the operation name ( op.name ).

The correct code is:

sess = K.get_session()

outputs = [out.op.name for out in model.outputs] # Note this new line

constant_graph = graph_util.convert_variables_to_constants(sess, 
                                             sess.graph.as_graph_def(), 
                                             outputs)

graph_io.write_graph(constant_graph, 'export', 'output.pb', as_text=False)