'saved_model_cli.py' bug fix!

[huaxz1986]

In file python/tools/saved_model_cli.py at function def _print_tensor_info(tensor_info):
The first line should be:

print(' dtype: ' + {value:key for (key,value) in types_pb2.DataType.items()}[tensor_info.dtype])

Not be : print(' dtype: ' + types_pb2.DataType.keyss()[tensor_info.dtype])

because tensor_info.dtype is an Integer which is the value of types(not the index of type values).

[qmick]

It's not necessary to do that, types_pb2.DataType is not python dict. The order of EnumTypeWrapper.keys() is not arbitrary.

See the definition of EnumTypeWrapper.keys():

  def keys(self):
    """Return a list of the string names in the enum.
    These are returned in the order they were defined in the .proto file.
    """

    return [value_descriptor.name
            for value_descriptor in self._enum_type.values]

and self._enum_type.values is something like:

  values=[
    _descriptor.EnumValueDescriptor(
      name='DT_INVALID', index=0, number=0,
      options=None,
      type=None),
    _descriptor.EnumValueDescriptor(
      name='DT_FLOAT', index=1, number=1,
      options=None,
      type=None),
    _descriptor.EnumValueDescriptor(
      name='DT_DOUBLE', index=2, number=2,
      options=None,
      type=None),
      ...
  ]

[huaxz1986]

When I run tfdbg：

graph = tf.get_default_graph()
with  graph.as_default():
    v1 = tf.get_variable('v1', initializer=[1, 2, 3])
    v2 = tf.square(v1, name='v2')
    sess = tf.Session()
    sess = tf_dbg.LocalCLIDebugWrapperSession(sess)
    with sess.as_default():
      sess.run(tf.global_variables_initializer())
      _v2 = sess.run(v2)
      print(_v2)

The type of v1 is DT_INT_REF，and saved_model_cli.py throws exception.

[qmick]

Yes, my mistake, for REF type, the value is not the same as its index. Can you add a minimum repuducible code and its error log to better describe this problem?

[huaxz1986]

This is a simple example to show the usage of saved_model_cli：

export_dir = 'out'
builder = tf.saved_model.builder.SavedModelBuilder(export_dir)
graph = tf.get_default_graph()
with graph.as_default():
  v1 = tf.get_variable('v1',shape=[3],
                  initializer=tf.zeros_initializer)
  v2 = tf.square(v1,name='v2')
  with tf.Session(graph=graph) as sess:
    sess.run(v1.initializer)
    builder.add_meta_graph_and_variables(sess,['xxx'])
    predict_signature_def = tf.saved_model.\
    	signature_def_utils.build_signature_def(\
             inputs={
        		 'input_x': \
                  tf.saved_model.utils.build_tensor_info(v1),
             }, outputs={'pred_y':\
				  tf.saved_model.utils.build_tensor_info(v2)},\
             method_name=tf.saved_model.\
 			 signature_constants.PREDICT_OUTPUTS)
    builder.add_meta_graph([tf.saved_model.\
             tag_constants.TRAINING,
             tf.saved_model.tag_constants.SERVING],\
             signature_def_map={'predict_graph': predict_signature_def})
    builder.save(as_text=True)

Then run saved_model_cli in shell：

saved_model_cli show --dir out/ --tag_set serve,train --signature_def predict_graph

The result is :

The given SavedModel SignatureDef contains the following input(s):
inputs['input_x'] tensor_info:
tensor type: 101
    dtype: DT_FLOAT_REF
    shape: (3)
    name: v1:0
The given SavedModel SignatureDef contains the following output(s):
outputs['pred_y'] tensor_info:
tensor type: 1
    dtype: DT_FLOAT
    shape: (3)
    name: v2:0
Method name is: outputs

I change tensorflow/python/tools/saved_model_cli._print_tensor_info() like this:

def _print_tensor_info(tensor_info):
  """Prints details of the given tensor_info.

  Args:
    tensor_info: TensorInfo object to be printed.
  """
#####  this is what I do ####
  print('tensor type:',tensor_info.dtype)
  print('    dtype: ' + {value:key for (key,value) in types_pb2.DataType.items()}[tensor_info.dtype])
##########
  # Display shape as tuple.
  if tensor_info.tensor_shape.unknown_rank:
    shape = 'unknown_rank'
  else:
    dims = [str(dim.size) for dim in tensor_info.tensor_shape.dim]
    shape = ', '.join(dims)
    shape = '(' + shape + ')'
  print('    shape: ' + shape)
  print('    name: ' + tensor_info.name)

By the way, I use python3.6+ubuntu 16.04+ tensorflow 1.4

[qmick]

Thanks for the supplement. I am not sure who is responsible for this module, friendly ping @yifeif @drpngx

[drpngx]

/CC @sukritiramesh

[yifeif]

Thanks @huaxz1986! Will fix this internally!