Failed to run with python benchmark.py --cfg ./config/face_detection_yunet.yaml

[ddwolf]

Here is the output of the command

[xxx@fedora-riscv benchmark]$ python benchmark.py --cfg ./config/face_detection_yunet.yaml
Benchmarking ...
mean       median     min        input size   model
12.16      12.25      11.91      [160, 120]   YuNet with ['face_detection_yunet_2023mar.onnx']
11.73      11.83      11.22      [160, 120]   YuNet with ['face_detection_yunet_2023mar_int8.onnx']
[ERROR:[email protected]] global onnx_importer.cpp:1035 handleNode DNN/ONNX: ERROR during processing node with 3 inputs and 1 outputs: [DequantizeLinear]:(onnx_node!bbox_head.multi_level_kps.2.conv2.weight_quantized_node) from domain='ai.onnx'
Traceback (most recent call last):
  File "/home/xxx/repos/opencv_zoo/benchmark/benchmark.py", line 226, in <module>
    model = model_handler(*model_path, **model_config)
            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/xxx/repos/opencv_zoo/models/face_detection_yunet/yunet.py", line 22, in __init__
    self._model = cv.FaceDetectorYN.create(
                  ^^^^^^^^^^^^^^^^^^^^^^^^^
cv2.error: OpenCV(4.10.0) /tmp/pip-install-uvko15ge/opencv-python_5d2ac5030ab94a15b5dcfd252013fa45/opencv/modules/dnn/src/onnx/onnx_importer.cpp:1057: error: (-2:Unspecified error) in function 'handleNode'
> Node [[email protected]]:(onnx_node!bbox_head.multi_level_kps.2.conv2.weight_quantized_node) parse error: OpenCV(4.10.0) /tmp/pip-install-uvko15ge/opencv-python_5d2ac5030ab94a15b5dcfd252013fa45/opencv/modules/dnn/src/int8layers/quantization_utils.cpp:18: error: (-215:Assertion failed) axis >= 0 && targetShape.size() > axis && data.total() == targetShape[axis] in function 'broadcast1D2TargetMat'
>

[fengyuentau]

cv2.error: OpenCV(4.10.0)

Hello @ddwolf , OpenCV 4.x does not support block quantized models, which are models with suffix int8bq. If you stick with 4.x, you could use filters run others without block quantized models, e.g. python3 benchmark.py --fp32 --int8.

[ddwolf]

cv2.error: OpenCV(4.10.0)

Hello @ddwolf , OpenCV 4.x does not support block quantized models, which are models with suffix int8bq. If you stick with 4.x, you could use filters run others without block quantized models, e.g. python3 benchmark.py --fp32 --int8.

Thanks for your reply

Using your comand python3 benchmark.py --fp32 --int8 failed, suggested me to use --all

Finally I use this command, seems right.
python3 benchmark.py --model_exclude int8bq --all

After dozens of cases, it complains module cv2 has no attribute wechat_qrcode_WeChatQRCode

[ddwolf]

By the way, could you please tell me, among the three numbers: mean, median, min, what number do you use to fill into this image?

[fengyuentau]

After dozens of cases, it complains module cv2 has no attribute wechat_qrcode_WeChatQRCode

It is in opencv-contrib-python, you can either exclude it or install the package along with opencv-python.

[fengyuentau]

By the way, could you please tell me, among the three numbers: mean, median, min, what number do you use to fill into this image?

We have https://github.com/fengyuentau/opencv_zoo/blob/main/benchmark/generate_table.py to parse data in benchmark/README.md. So just put data following the same pattern in readme, make items in benchmark/table_config.yaml with the same pattern, and call generate_table.py.

[ddwolf]

Thank you so much, @fengyuentau, issue solved.