Wrong classification results for MobilenetV1 TF Model

[Gowtham__Hareesh]

Hello,

I converted the tensorflow pretrained model(frozen pb file) for MobilenetV1 downloaded from below link.

http://download.tensorflow.org/models/mobilenet_v1_2018_08_02/mobilenet_v1_1.0_224_quant.tgz

The command I use to convert is as follows:

python mo_tf.py --input_model C:\OpenVINO\TF_Models\mobilenet_v1_1.0_224_quant\mobilenet_v1_1.0_224_quant_frozen.pb -b 1 --output_dir C:\OpenVINO\TF_Models\mobilenet_v1_1.0_224_quant --reverse_input_channels --log_level=DEBUG

After that, I use classification_sample_async to  run the inference Engine using below command.

classification_sample_async.exe -i C:\Users\user1\Downloads\Banana.bmp -m C:\OpenVINO\TF_Models\mobilenet_v1_1.0_224_quant\mobilenet_v1_1.0_224_quant_frozen.xml

Though the commands are executed successfully without any errors, I results are poor(not even close to correct classes).

I want to know whether I am missing some thing?

Thanks,

Hareesh Gowtham

[Shubha_R_Intel]

Dear Hareesh Gowtham, Please review the Model Optimizer Tensorflow Doc .

In particular note the columns --MEAN_VALUES and --SCALE for Mobilenet V1 models. You need to pass those into the mo_tf.py command. These are pre-processing which the images went through prior to training. Unless you tell Model Optimizer about such pre-processing, there is no way it can ever know about it.

Hope it helps,

Thanks,

Shubha

[Gowtham__Hareesh]

Hello Shubha,

Thank you for your answer.

I checked the document you mentioned.

The MEAN_VALUES and SCALE are for unfrozen model, But I am using  Frozen Quantized model for mobilenetV1.

http://download.tensorflow.org/models/mobilenet_v1_2018_08_02/mobilenet_v1_1.0_224_quant.tgz

When I provide MEAN VALUES and SCALE which is in float [127.5,127.5,127.5] and 127.5 respectively to the model, it gives below error

ValueError: invalid literal for int() with base10: 'input'

Does quantized frozen model also require MEAN_VALUES and SCALE?

Thanks,

Hareesh Gowtham 

[Shubha_R_Intel]

Dear Gowtham, Hareesh

OK. This is a Tensorflow-Lite model, right (since you say "quantized") ?  Or did you pick your Mobilenet V1 from one of the Tensorflow Model Optimizer Supported and Validated List ?

Actually no, --mean_values and --scale are fed along with --input_model (as a frozen pb) to mo_tf.py.  Doing a --h on mo_tf.py in 2019R3 gets this:

--mean_values MEAN_VALUES, -ms MEAN_VALUES
                        Mean values to be used for the input image per
                        channel. Values to be provided in the (R,G,B) or
                        [R,G,B] format. Can be defined for desired input of
                        the model, for example: "--mean_values
                        data[255,255,255],info[255,255,255]". The exact
                        meaning and order of channels depend on how the
                        original model was trained.

--scale SCALE, -s SCALE
                        All input values coming from original network inputs
                        will be divided by this value. When a list of inputs
                        is overridden by the --input parameter, this scale is
                        not applied for any input that does not match with the
                        original input of the model.

Did you input the values correctly ? Kindly give me your full model optimizer command after adding --mean_values and --scale.

Thanks,

Shubha

[Gowtham__Hareesh]

Dear Shubha, 

Thank you for the information.

Yes I am using tensorflow lite quantized frozen model from Supported Frozen Quantized Topologies

I was using the below command:

python mo_tf.py --input_model C:\OpenVINO\TF_Models\mobilenet_v1_1.0_224_quant\mobilenet_v1_1.0_224_quant_frozen.pb -b 1 --output_dir C:\OpenVINO\TF_Models\mobilenet_v1_1.0_224_quant --reverse_input_channels --log_level=DEBUG --mean_values data(127.5,127.5,127.5) --scale 127.5

I figured out problem with --mean_values data(127.5,127.5,127.5) parameter.

I changed the layer name to input i.e.  --mean_values input(127.5,127.5,127.5) and both model conversion and execution of sample works fine.

Again, thank you very much for the support.

Regards,

Hareesh Gowtham 

[Shubha_R_Intel]

Dearest Hareesh Gowtham ,

Fantastic News ! Bravo ! Thanks for sharing your success with the OpenVino community.

Shubha