"Conversion from FP16 to U8 is unsupported" on MyriadX when all I am using is FP16 precision

[berrylizzard]

Using OpenVINO 2022.2 with OpenCV 4.7.0 (both built from source) and Python 3.9.16 on Ubuntu 22.04.

I have a Movidius MyriadX VPU and an IP camera and want to run the inference on a live stream using MyriadX for object detection. I have downloaded and converted the model ssd_mobilenet_v1_coco with

omz_downloader --name ssd_mobilenet_v1_coco --precisions FP16 --output_dir /home/user0/Downloads/ssd_mobilenet_v1_coco_downloaded

and

omz_converter --download_dir /home/user0/Downloads/ssd_mobilenet_v1_coco_downloaded --output_dir /home/user0/Downloads/ssd_mobilenet_v1_coco_converted --name ssd_mobilenet_v1_coco --precisions FP16

Then trying to run code attached openvino_error.zip (zip includes model files too; in code some info is redacted, like the IP camera stream URL). It includes the following relevant lines:

net = cv2.dnn.readNetFromModelOptimizer(bin = model_bin, xml = config_xml)
net.setPreferableBackend(cv2.dnn.DNN_BACKEND_INFERENCE_ENGINE)
net.setPreferableTarget(cv2.dnn.DNN_TARGET_MYRIAD)
...
video_stream = cv2.VideoCapture(filename = stream_url)
...
ret, frame = video_stream.read()
...
frame = cv2.resize(frame, (300,300), interpolation = cv2.INTER_AREA)
params = cv2.dnn.Image2BlobParams(scalefactor = 1.0/127.5, size = (300, 300), mean = 127.5, swapRB = True, ddepth = cv2.CV_32F, datalayout = cv2.dnn.DNN_LAYOUT_NHWC) 
blob = cv2.dnn.blobFromImageWithParams(frame, params)
net.setInput(blob)
detections = net.forward() # ERRORS HERE
...

Calling the program as

/home/user0/build-opencv/setup_vars.sh python3 opencv_app.py

but facing the error:

...
[ERROR]: OpenCV(4.7.0-dev) /home/user0/opencv/modules/dnn/src/ie_ngraph.cpp:865: error: (-2:Unspecified error) in function 'initPlugin'
> Failed to initialize Inference Engine backend (device = MYRIAD): Convert_5316 of type Convert: [ GENERAL_ERROR ]
> /home/jenkins/agent/workspace/private-ci/ie/build-linux-ubuntu20/b/repos/openvino/src/plugins/intel_myriad/graph_transformer/src/stages/convert.cpp:67 [Internal Error]: Final check for stage Convert_5316 with type Convert has failed: Conversion from FP16 to U8 is unsupported

Here is a reference on backend/target compatibility: https://docs.opencv.org/4.x/db/d30/classcv_1_1dnn_1_1Net.html#a9dddbefbc7f3defbe3eeb5dc3d3483f4
As I understand, VPU only supports FP16 precision hence when converting I specified that precision and am not sure what exactly is trying to convert FP16 to U8 and why it is being done.

Any pointers appreciated!

[berrylizzard]

I have converted the same model to FP32 (just switched --precisions FP32 in the omz_converter command) but the error persists

[berrylizzard]

Using a different model (ssdlite_mobilenet_v2) in FP16 did not solve the error either.

I will ask the same question to OpenCV community too, see if anyone there is familiar.

[berrylizzard]

FYI Link to the same question on the OpenCV Q&A forum.

[berrylizzard]

Myriad X VPU is recognized:

$ lsusb
...
Bus 003 Device 002: ID 03e7:2485 Intel Movidius MyriadX
...

and when running

from openvino import runtime
ie = runtime.Core()
for device in ie.available_devices:
    device_name = ie.get_property(device, "FULL_DEVICE_NAME")
    print(f"{device}: {device_name}")

output is

CPU:    Intel Atom(R) x6425RE Processor @ 1.90GHz
GNA: GNA_SW
MYRIAD: Intel Movidius Myriad X VPU

Additionally, the error is sometimes referring to Convert_5316, and sometimes to Convert_1132.

[berrylizzard]

Re-wrote my code following this reference to create a model object using the openvino.runtime.Core()'s read_model function instead of cv2.dnn’s readNetFromModelOptimizer and now I am not getting any errors!

New code here: opencv_error_20230628.zip

Still don't know what was wrong with cv2.dnn.readNetFromModelOptimizer, but I'll accept this as an answer for now.