Unable to desearilize object Validation error

[prateethvnayak]

I am trying to do something new here. So I have the yolov2 model in frozen_pb format from tensorflow. I have successfully converted it to mlmodel and got it working using this repo.

Now, I have a another model where I have quantized the weights in the frozen_pb model from float32 to 8-bits (but the numbers are still in float32 format just the different levels of unique values are now just 8-bits i.e 255 unique float values only). This kinda compresses the model.

I was able to successfully convert the model using the tf-coreml repo. Same as the float32 model.

Now on adding this model to the Xcode.proj , it gives this error.
HERE

There was a problem decoding this CoreML document
validation error : Unable to deserialize object 

The model is still float32 data-type (and same size) . Any ideas where it might be going wrong ? Is there a way that the mlmodel protobuf stores differently for lesser unique values ?

[hollance]

That means the Core ML file is not a correct protobuf file, or that the contents in the file are not correct. Most likely something is wrong with the weights for the layers (since that's what you changed) but I can't see that from here. ;-)

[prateethvnayak]

I believe the weights of the layers are correct. After quantizing, I do get the expected predictions on the tensorflow platform. Do think the frozen_pb itself is not created correct after the quantization ?

            input_node_names = ["Input/Placeholder"] 
            output_node_names = ["Convolutional_9/add"]
            print(input_node_names,"\n",output_node_names)

            gdef = tf.graph_util.convert_variables_to_constants(
                    sess,
                    sess.graph_def,
                    output_node_names)
            gdef = strip_unused_lib.strip_unused(
                    input_graph_def = gdef,
                    input_node_names = input_node_names,
                    output_node_names = output_node_names,
                    placeholder_type_enum = dtypes.float32.as_datatype_enum)
            # Save it to an output file
            frozen_model_file = self.base_dir_out+'/frozen_model.pb'
            with gfile.GFile(frozen_model_file, "wb") as f:
                f.write(gdef.SerializeToString())

This is the same snippet I use to create the model before and after quantizing. Do you think the model frozen_pb from the TF itself is wrong before I bring it to mlmodel ?