Converting DeepLabv3+ made in Tensorflow v2 beta 1

Hi, I’m running this code to try to convert a frozen .pb model I’ve saved to a coreml model:

import tfcoreml as tf_converter
tf_converter.convert(tf_model_path = './archived models/DeepLabv3+_15_epochs/saved_model.pb',
                     mlmodel_path = './archived models/CoreML/DeepLabv3+_15_epochs.mlmodel',
                     output_feature_names = ['Output'],
                     use_coreml_3 = True)

This is my error:

---------------------------------------------------------------------------
DecodeError                               Traceback (most recent call last)
<ipython-input-3-45de1822b420> in <module>()
      3                      mlmodel_path = './archived models/CoreML/DeepLabv3+_15_epochs.mlmodel',
      4                      output_feature_names = ['Output'],
----> 5                      use_coreml_3 = True)

~/anaconda3/envs/Python3.5/lib/python3.5/site-packages/tfcoreml/_tf_coreml_converter.py in convert(tf_model_path, mlmodel_path, output_feature_names, input_name_shape_dict, image_input_names, is_bgr, red_bias, green_bias, blue_bias, gray_bias, image_scale, class_labels, predicted_feature_name, predicted_probabilities_output, add_custom_layers, custom_conversion_functions, use_coreml_3)
    592         tf_model_path,
    593         inputs=input_name_shape_dict,
--> 594         outputs=output_feature_names)
    595     if mlmodel_path is not None:
    596       mlmodel.save(mlmodel_path)

~/anaconda3/envs/Python3.5/lib/python3.5/site-packages/coremltools/converters/tensorflow/_tf_converter.py in convert(filename, inputs, outputs, **kwargs)
      9     try:
     10         from ..nnssa.frontend.tensorflow import load as frontend_load
---> 11         ssa = frontend_load(filename, **kwargs)
     12     except ImportError as err:
     13         raise ImportError("Frontend converter not found! Error message:\n%s" % err)

~/anaconda3/envs/Python3.5/lib/python3.5/site-packages/coremltools/converters/nnssa/frontend/tensorflow/load.py in load(tfgraph, resume_on_errors, **kwargs)
     35         gd = tfgraph
     36 
---> 37     ssa = graphdef_to_ssa(gd)
     38 
     39     placeholder_shape = kwargs.get("placeholder_shape", {})

~/anaconda3/envs/Python3.5/lib/python3.5/site-packages/coremltools/converters/nnssa/frontend/tensorflow/graphdef_to_ssa.py in graphdef_to_ssa(graphdef_or_file, main_method_name)
     36     """
     37     if isinstance(graphdef_or_file, (bytes, str)):
---> 38         gdorig, g = load_tf_graph(graphdef_or_file)
     39     else:
     40         gdorig = graphdef_or_file

~/anaconda3/envs/Python3.5/lib/python3.5/site-packages/coremltools/converters/nnssa/frontend/tensorflow/graphdef_to_ssa.py in load_tf_graph(graph_file)
     21     with tf.gfile.GFile(graph_file, "rb") as f:
     22         graph_def = tf.GraphDef()
---> 23         graph_def.ParseFromString(f.read())
     24 
     25     # Then, we import the graph_def into a new Graph and returns it

DecodeError: Error parsing message

I’ve tried saving the model two different ways, both using Tensorflow 2.0.0b1:

tf.keras.experimental.export_saved_model(model, "./archived models/DeepLabv3+_15_epochs", custom_objects={'relu6': relu6}) tf.keras.experimental.export_saved_model(model, "./archived models/DeepLabv3+_15_epochs", as_text=True, custom_objects={'relu6': relu6})

Note the additional attribute ‘as_text=True’ in the second example. When saving ‘as_text=True’, my conversion code is reading from .../saved_model.pbtxt instead of .../saved_model.pb.

When converting, I’ve been running my conversion code with CoreMLtools 3.0b1, TF-CoreML 0.4.0b1, and I’ve tried reading/converting with both Tensorflow 2.0.0b1 and Tensorflow 1.13.1.

I’ve tried a couple other things the past ~2-3 hours… I’m sort of forgetting all the little paths I’ve gone down.

=====

Most recently, using Tensorflow 1.13.1 and use_coreml_3 = False) in the convert function, I changed code in tf_coreml_converter.py to read the model differently, via the answer on this StackOverflow page by user anujonthemove.

That failed with this error: MetaGraphDef associated with tags could not be found in SavedModel. To inspect available tag-sets in the SavedModel, please use the SavedModel CLI: saved_model_cli``, as I passed in an empty list of tags, and tags=‘None’ didn’t work, either.

=====

Most of the things I’ve tried usually end with the saved model I’ve made not being read correctly and failing somewhere when loading.

Any ideas? Thanks so much.

[Oktai15]

@not-Ian if you convert your model in tf 2.0 to coreml, please, write here.

[Mstronach]

Experimental support for TensorFlow 2.x and tf.keras was added in coremltools 3.1 with some known issues, could you please try again with the latest version of coremltools 3.1 and/or tfcoreml 1.1. Please see the neural network conversion guide and/or Jupyter notebooks for some conversion examples. If you encounter any issues while converting your model, please file a new GitHub issue with steps to reproduce. Thanks!