mask_rcnn_resnet50_atrous_coco in HETERO mode fails

[SDabh1]

Hello All,

I am trying to run one of the mask rcnn python demo (mask_rcnn_resnet50_atrous_coco) in an HETERO mode. But it runs into a following issue:

[ INFO ] Preparing output blobs
[ INFO ] Loading model to the device
[ INFO ] Create infer request
Emulator Context callback: Requested memory object size exceeds device limits
ERROR: CL_INVALID_BUFFER_SIZE
Location: /teamcity/work/scoring_engine_build/releases_2019_R3/thirdparty/dla/runtime/core/src/device_buffer.cpp:128
Failed clCreateBuffer
[ ERROR ]

Initially, I was trying it to run with input size of (800x800) but later I tried the smaller input size model (1,1365) but still having the same issue. I am wondering if this issue is because of input size? Or there is something else that I am missing on?

FPGA board I am using is "Mustang sg1"

Any help would be appreciated. Thank you.

[JAIVIN_J_Intel]

Hi Sunil,

Please confirm that you have followed the steps to Configure and Set Up the Intel® Vision Accelerator Design with an Intel® Arria® 10 FPGA.

Have you programmed the bitstream corresponding to the topology of the model?

Also, please mention the command you used to optimize the model to IR.

Regards,

Jaivin

[SDabh1]

Hi Jaivin,

1. Yes, I have followed the steps as mentioned in the document and I am able to run all other models in HETERO as well as in only FPGA mode. So FPGA is configured correctly.

2. I have tried all the available bitstreams but still no luck. Also, I am running it in HETERO mode so it should work irrespective of what bitstream I am using. Because unsupported layers will fall back to CPU. Please note I am able to run it in pure CPU mode. Openvino version I am using is 2019.3.376.

3. Command in use:

/home/demo/omz_demos_build/intel64/Release/mask_rcnn_demo -d HETERO:FPGA,CPU -i /opt/intel/openvino_fpga_2019.3.376/deployment_tools/demo/car_1.bmp -m /opt/intel/openvino_2019.3.376/deployment_tools/open_model_zoo/tools/downloader/public/mask_rcnn_resnet50_atrous_coco/FP16/mask_rcnn_resnet50_atrous_coco.xml

Please let me know if you need any other information from my side.

Thank you

-Sunil