raise error if signature is missing

[guschmue]

Signed-off-by: Guenther Schmuelling guschmue@microsoft.com

[guschmue]

I think its hard for the in-memory model to select the signature by number.
Much easier to specify the signature.
If none is given raise error.

[guschmue]

#1377

[HalfVoxel]

Nice. However I think the function signature should probably be changed so that the parameter doesn't have a default value. Right now it says input_signature=None which kinda implies to the user that leaving it out is acceptable.

[felker]

I am a bit confused what input_signature type should be passed under this new API. If I pass tf.TensorSpec(batch_input_shape) as the README indicates:

Traceback (most recent call last):
  File "mpi_learn.py", line 133, in <module>
    mpi_train(conf, shot_list_train, shot_list_validate, loader,
  File "/Users/felker/Desktop/plasma-python/plasma/models/mpi_runner.py", line 970, in mpi_train
    specific_builder.save_model_weights(train_model, int(round(e)))
  File "/Users/felker/Desktop/plasma-python/plasma/models/builder.py", line 407, in save_model_weights
    model_proto, external_tensor_storage = tf2onnx.convert.from_keras(
  File "/usr/local/Caskroom/miniconda/base/envs/frnn-tf2/lib/python3.8/site-packages/tf2onnx/convert.py", line 315, in from_keras
    function = _saving_utils.trace_model_call(model, input_signature)
  File "/usr/local/Caskroom/miniconda/base/envs/frnn-tf2/lib/python3.8/site-packages/tensorflow/python/keras/saving/saving_utils.py", line 127, in trace_model_call
    def _wrapped_model(*args):
  File "/usr/local/Caskroom/miniconda/base/envs/frnn-tf2/lib/python3.8/site-packages/tensorflow/python/eager/def_function.py", line 1598, in decorated
    decorator_func=Function(
  File "/usr/local/Caskroom/miniconda/base/envs/frnn-tf2/lib/python3.8/site-packages/tensorflow/python/eager/def_function.py", line 536, in __init__
    self._function_spec = function_lib.FunctionSpec.from_function_and_signature(
  File "/usr/local/Caskroom/miniconda/base/envs/frnn-tf2/lib/python3.8/site-packages/tensorflow/python/eager/function.py", line 2410, in from_function_and_signature
    return FunctionSpec(
  File "/usr/local/Caskroom/miniconda/base/envs/frnn-tf2/lib/python3.8/site-packages/tensorflow/python/eager/function.py", line 2477, in __init__
    raise TypeError("input_signature must be either a tuple or a "
TypeError: input_signature must be either a tuple or a list, received <class 'tensorflow.python.framework.tensor_spec.TensorSpec'>

If I pass a tuple batch_input_shape:

Traceback (most recent call last):
  File "mpi_learn.py", line 133, in <module>
    mpi_train(conf, shot_list_train, shot_list_validate, loader,
  File "/Users/felker/Desktop/plasma-python/plasma/models/mpi_runner.py", line 970, in mpi_train
    specific_builder.save_model_weights(train_model, int(round(e)))
  File "/Users/felker/Desktop/plasma-python/plasma/models/builder.py", line 408, in save_model_weights
    model_proto, external_tensor_storage = tf2onnx.convert.from_keras(
  File "/usr/local/Caskroom/miniconda/base/envs/frnn-tf2/lib/python3.8/site-packages/tf2onnx/convert.py", line 315, in from_keras
    function = _saving_utils.trace_model_call(model, input_signature)
  File "/usr/local/Caskroom/miniconda/base/envs/frnn-tf2/lib/python3.8/site-packages/tensorflow/python/keras/saving/saving_utils.py", line 126, in trace_model_call
    @def_function.function(input_signature=input_signature, autograph=False)
  File "/usr/local/Caskroom/miniconda/base/envs/frnn-tf2/lib/python3.8/site-packages/tensorflow/python/eager/def_function.py", line 1586, in function
    function_lib.validate_signature(input_signature)
  File "/usr/local/Caskroom/miniconda/base/envs/frnn-tf2/lib/python3.8/site-packages/tensorflow/python/eager/function.py", line 3397, in validate_signature
    raise TypeError("Invalid input_signature {}; input_signature must be "
TypeError: Invalid input_signature (128, 128, 14); input_signature must be a possibly nested sequence of TensorSpec objects.

[TomWildenhain-Microsoft]

It should be a list of TensorSpec objects. Try [tf.TensorSpec(batch_input_shape)]

BTW If you use the latest tf2onnx from github an input signature is actually optional.

[guschmue]

I guess we could improve the docs a little. As Tom said - the latest version will not require it but that could fail.
If you need a spec, for keras-resnet50 it might look like this:
(tf.TensorSpec((None, 224, 224, 3), tf.float32, name="input"),)
basically a tuple of tf.TensorSpec with shape, type and name of each input.