[ERROR] Unsupported layer type : FakeQuantize

[lisovoy__andrey]

Hi!

I get error:

[ ERROR ] Check (env.config.ignoreUnknownLayers) failed: Failed to compile layer L0008_ActivationBin-back_bone_seq.conv2_2_sep_relubin_bin_conv_BIN01/Quantize_ : Unsupported layer type : FakeQuantize

when run object_detection_sample_ssd example with face-detection-adas-binary-0001 model on NCS2 in docker:

docker run -it -u 0 --device /dev/dri:/dev/dri --device-cgroup-rule='c 189:* rmw' -v /dev/bus/usb:/dev/bus/usb --rm openvino/ubuntu18_dev:latest /bin/bash

cd /tmp
source /opt/intel/openvino/bin/setupvars.sh
cmake -DCMAKE_BUILD_TYPE=Release /opt/intel/openvino/deployment_tools/inference_engine/samples/cpp
make -j2 object_detection_sample_ssd


curl -O https://download.01.org/opencv/2020/openvinotoolkit/2020.1/open_model_zoo/models_bin/1/face-detection-adas-binary-0001/FP32-INT1/face-detection-adas-binary-0001.bin
curl -O https://download.01.org/opencv/2020/openvinotoolkit/2020.1/open_model_zoo/models_bin/1/face-detection-adas-binary-0001/FP32-INT1/face-detection-adas-binary-0001.xml

intel64/Release/object_detection_sample_ssd -m face-detection-adas-binary-0001.xml -d MYRIAD -i /opt/intel/openvino/deployment_tools/open_model_zoo/models/intel/emotions-recognition-retail-0003/description/emotions-recognition-retail-0003.jpg

But face-detection-adas-0001 on MYRIAD and face-detection-adas-binary-0001 on CPU works.

What could be wrong with that?

[MAURICIOAL_R_Intel]

Thanks for reaching out.

The issue is that you are using a pre-trained model from the open model zoo in FP32-INT1 format. You need to use the FP16 format found here: https://download.01.org/opencv/2020/openvinotoolkit/2020.1/open_model_zoo/models_bin/1/face-detection-adas-0001/FP16/

Regards,

Mauricio R.

[lisovoy__andrey]

It is different models, face-detection-adas-0001 and face-detection-adas-binary-0001, accordingly that page:

• face-detection-adas-0001 - Complexity 2.835 GFLOPs
• face-detection-adas-binary-0001 - Complexity 0.819 GFLOPs

From here:

The network features a pruned MobileNet backbone...

2.8  GFLOPs are a little bit slow for my project. But when i tryed to use more fast model, i had the problem.

[lisovoy__andrey]

Took some tests. Seems to MYRIAD can not handle INT models (FP32-INT1, FP32-INT8). Can somebody explain me why?

[lisovoy__andrey]

Ok, i found answer here:

NCS2  supports FP16 only

[MAURICIOAL_R_Intel]

Thanks for your response.

We are glad that you found the answer to your query.

Regards,

Mauricio R.