Add signature customisation to SavedModel conversion functions

[patlevin]

Description

• The TF SavedModel format allows for multiple model signatures (i.e. specification of inputs and outputs) per model.
• Right now the converter will try and figure out all possible inputs and outputs and add them to a single SavedModel signature.
• It would be nice to have the ability to select custom SavedModel signatures from available inputs and outputs.

Scope

• Extend the API so that graph-model-to-saved-model and graph-models-to-saved-model allow specifying outputs and signature keys

• Add a command line option to the CLI utility to select the requested outputs for adding to the generated signature

Details

API

• The new arguments for the API functions should be optional.
• The functions should verify that the requested outputs exist in the model and are indeed output nodes.

• The optional parameter should be named signature_def_map and contain a mapping from a signature key (of type string) to dict containing the keys inputs and outputs, which map to an enumerable of strings (output names) and an optional method_name:

  signature_map_def = {
  'my_signature': {outputs: ['Identity1', 'Identity2']}
  }
  
  signature_map_def2 = {
  'my_signature': {outputs: ['classes', 'scores'], method_name: 'my_model/classify'}
  }

• An empty key (None or '') should map to the default signature key.
• Duplicate keys are considered to be an invalid argument and raise a ValueError.

CLI

• The signature name can be selected using the command line option --signature_key, which is a singular non-empty string argument. If omitted, the default signature key is applied.
• The outputs for the singular signature can be selected using the command line option --outputs, which is a comma-separated list of output names. The option can be specified multiple times. Duplicate names are only considered once.
• The method name can be set using the command line option --method_name
• --signature_key, --outputs, and --method_name are only valid in conjunction with the output format tf_saved_model
• if any of the three new options is specified along with an output format other than tf_saved_model, a warning message is printed