LoadNetwork crashes on Android while loading sample's face detection xml and bin

[avaish1]

Hi We have recompiled dldt 2019-R1.1 for android P and trying to enable face detection via JNI layer and CMAKE in android app via MKLDNN plugin . The reference for face detection app is as mentioned in https://github.com/opencv/open_model_zoo/tree/master/demos/interactive_face_detection_demo and model tried is face-detection-adas-0001.xml.

As we try to upload face-detection-adas-0001.xml from openvino_models, it results in following error all the time 08-21 16:42:23.133 11235 11253 E libc++abi: terminating with uncaught exception of type InferenceEngine::details::InferenceEngineException: Unsupported primitive of type: PriorBox name: fc7_mbox_priorbox 08-21 16:42:23.133 11235 11253 E libc++abi: vendor/intel/external/project-celadon/nn-hal/intel_nn_hal/dl/../../../dldt/inference-engine/src/mkldnn_plugin/mkldnn_node.cpp:167 08-21 16:42:23.133 11235 11253 E libc++abi: ../../../../src/main/cpp/include/inference_engine\details/ie_exception_conversion.hpp:71

As googled for this issue , compiled libcpu_extension.so on android and included here (attached is the adb output) , but error still remains.

As this interactive_Face_demo works on Linux flawlessly , we are trying to simulate the same on Android .

could someone please help here? Thankyou <img width="288" alt="adb_libs" src="http

s://user-images.githubusercontent.com/25952023/63449476-e08f4880-c45d-11e9-89c2-89eaa2bd77d7.PNG"

Model file is also attached.

[shubha-ramani]

Dear @avaish1 Without stepping through Inference Engine code built as Debug there is no way I can determine what went wrong. JNI definitely adds a complication. I'm glad that you're linked against libcpu_extension.so because that is indeed the first thing I thought about.

My advice to you is first 1) use 2019R2. 2019R1.1 is quite old and so you should always be on the latest release because many bug fixes and improvements have been made. 2) Follow this README to build Inference Engine in Debug and step through your code. If you step through your code you will narrow down the problem. 3) The error seems to say that Inference Engine cannot find this extension right here: https://github.com/opencv/dldt/blob/2019/inference-engine/src/extension/ext_priorbox.cpp. So maybe you need to JNI that part too. But frankly it's confusing because it should be included in your libcpu_extension.so unless somehow it got left out. 4) If all else fails try with a super simple example first, something like hello_classification

Hope it helps,

Thanks,

Shubha

[avaish1]

Yeh Shubha, could load alexnet , would try other steps now.. 2019-08-22 15:23:48.490 4346-4365/? D/FACE_DEMO: layer count of networks are 22 2019-08-22 15:23:48.490 4346-4365/? D/FACE_DEMO: name of networks is AlexNet 2019-08-22 15:23:48.490 4346-4365/? D/FACE_DEMO: batchsize of networks is 1 2019-08-22 15:23:48.490 4346-4365/? D/FACE_DEMO: size of networks is 22

[avaish1]

adding this line helped resolved this unknown primitive issue .. plugin.AddExtension(std::make_shared());

But encountered another one as below .. though xml shows activation as relu, but here it says convolution..

08-22 14:51:31.233 9969 9969 F DEBUG : Abort message: 'terminating with uncaught exception of type InferenceEngine::details::InferenceEngineException: Node conv2_1/dw has unsupported activation primitive: Convolution : Convolution

[shubha-ramani]

Dear @avaish1 As already discussed, please try again on OpenVino 2019R2. Your issue may already be fixed. Your error is strange since Convolution is not an activation primitive within OpenVIno per the Supported Devices Doc

Thanks,

Shubha

[AnastasiaKazantaeva]

It seems that issue nor actual anymore closing it. Feel free to open a new issue in case of some problems/issues