Depth Multiplier becomes 0, on conversion to TFLITE

[SanthoshRajendiran]

MODEL: DeeplabV3+ MobilenetV2 Pretrained with Pascal VOC 2012
Tensorflow Version: 1.12.0
TFLITE Version: 1.12.0
Bazel Version: 0.17.2
OS version: Ubuntu 16.04
GPU Version: GTX 1080 Ti
Command to reproduce:

bazel run //tensorflow/lite/toco:toco --
--input_file=frozen_inference_graph.pb
--output_file=converted.tflite
--input_format=TENSORFLOW_GRAPHDEF
--output_format=TFLITE
--input_arrays=ImageTensor
--output_arrays=SemanticPredictions
--input_shapes=1,513,513,3
--allow_custom_ops
--inference_type=FLOAT
--inference_input_type=QUANTIZED_UINT8

(or)

CUDA_VISIBLE_DEVICES="0"
tflite_convert
--output_file=converted.tflite
--graph_def_file=frozen_inference_graph.pb
--inference_type=FLOAT
--inference_input_type=QUANTIZED_UINT8
--input_arrays=ImageTensor
--output_arrays=SemanticPredictions
--input_shapes=1,513,513,3
--mean_values=513
--std_dev_values=513

Benchmarking Command:
bazel run -c opt tensorflow/lite/tools/benchmark:benchmark_model -- --graph='realpath converted.tflite'

Conversion of pre-trained models provided on the model zoo to TFLITE, on benchmarking and running on Android provided issues on depth multiplier value getting converted to 0.

Stack Overflow Issue: https://stackoverflow.com/questions/53228969/unable-to-test-and-deploy-a-deeplabv3-mobilenetv2-tensorflow-lite-segmentation-m

[tensorflowbutler]

Thank you for your post. We noticed you have not filled out the following field in the issue template. Could you update them if they are relevant in your case, or leave them as N/A? Thanks.
What is the top-level directory of the model you are using
Have I written custom code
OS Platform and Distribution
TensorFlow installed from
CUDA/cuDNN version
GPU model and memory
Exact command to reproduce

[SanthoshRajendiran]

Custom Code: No
OS PLATFORM: Ubuntu 16.04
Tensorflow Installed from Source (tensorflow site)
CUDA Version: 9.0
GPU Model: GTX 1080 Ti
Exact Command to Reproduce the issue is given above.

Depth multiplier turning out to be zero, was visualized using Netron which is used to view Tensorflow graph and Tensorflow Lite models.
Confirming the issue, is the stack overflow links provided in the previous comment.

[jdduke]

Hmm, good catch, I think there might be a build issue with that specific example test op. Let me take a look.

[jdduke]

I think if you change rate= to dilations= @ https://github.com/tensorflow/tensorflow/blob/master/tensorflow/lite/testing/op_tests/depthwiseconv.py#L99, it should compile if you run:

bazel test //tensorflow/lite/testing:zip_test_depthwiseconv

Looks like there have been some recent changes that have change the arguments for this op.

[haozha111]

Could you download the most recent tf-nightly and try convert this model using the new MLIR converter? It seems that this is a bug associated with the old TOCO converter.

https://www.tensorflow.org/api_docs/python/tf/lite/TFLiteConverter?version=stable