Mixnet conversion

Hi!

I’m trying to convert the frozen model of mixnet-s. But when I feed the .pb file of this model to the converter, I get the following error:

OutOfRangeError: Node 'mixnet-s/mixnet_model/stem/batch_normalization/FusedBatchNormV3' (type: 'Add', num of outputs: 1) does not have output 5

I’m relatively new to TF, so I’m still figuring out my ways. Please let me know how to solve this issue.

Thanks in advance!

[dlawrences]

Hi

Just ran into the same issue.

Environment config:

• tensorflow: 1.15.0
• coremltools: 3.1
• tfcoreml: 1.1

The model that I am trying to convert is based on the TensorFlow ssdlite_mobilenet_v2_coco, retrained on 4 classes on an input of 768x768.

I have exported the graph for inference using export_inference_graph.py.

I have optimized the saved model after loading the graph using strip unused:

def optimize_graph(graph):
    """Strips unused subgraphs and save it as another frozen TF model."""
    gdef = strip_unused_lib.strip_unused(
        input_graph_def=graph.as_graph_def(),
        input_node_names=[input_node],
        output_node_names=[bbox_output_node, class_output_node],
        placeholder_type_enum=dtypes.float32.as_datatype_enum)

    with gfile.GFile(frozen_model_file, "wb") as f:
        f.write(gdef.SerializeToString())

I have attempted to convert the model using the following:

# Convert to Core ML model.
ssd_model = tfcoreml.convert(
    tf_model_path=frozen_model_file,
    mlmodel_path=coreml_model_path,
    input_name_shape_dict={input_tensor: [1, input_height, input_width, 3]},
    image_input_names=input_tensor,
    output_feature_names=[bbox_output_tensor, class_output_tensor],
    is_bgr=False,
    red_bias=-1.0,
    green_bias=-1.0,
    blue_bias=-1.0,
    image_scale=2. / 255)

In my case:

• input_node = “Preprocessor/sub”
• bbox_output_node = “concat”
• class_output_node = “Postprocessor/convert_scores”

Any thoughts? Thanks

[vade]

Hi.

Im having a similar issue.

Im using Google Colab, and they recently changed some back end config where Tensorflow 1.14-GPU no longer finds some libraries needed to load the GPU (so training takes forever).

In an attempt to use the default Tensorflow - 1.15, and finding am having the same exact error when converting Mobilenet 224 1.0 to CoreML.

With Tensorflow 1.14, tf-coreml was able to load frozen PBs and export them to CoreML sans issue.

With 1.15, and tf-coreml 1.1, and coremltools 3.1, I receive the same error:

OutOfRangeError: Node 'mixnet-s/mixnet_model/stem/batch_normalization/FusedBatchNormV3' type: 'Add', num of outputs: 1) does not have output 5

Models with TF 1.15:

My Tensorflow PB models are segmented graphs, with a feature extractor multiple classifiers turned into a pipeline. Here’s the PB

Feature Extractor.pb

Classifier.pb