Troubleshooting

Q: Why is the performance sometimes on CPU better than on GPU, when running pipeline benchmarking like stream density?

A: The performance of pipeline benchmarking strongly depends on the models. Specifically for yolov5s object detection, it is recommended to use the model precision FP32 when it is running on device GPU. If supported, then you can change the model precision by going to the folder configs/opencv-ovms/models/2022 from the root of project folder and editing the base_path for that particular model in the config_template.json file. For example, you can change the base_path of FP16 to FP32 assuming the precision FP32 of the model yolov5s is available:

    ...
    "config": {
    "name": "yolov5s",
    "base_path": "/models/yolov5s/FP32-INT8",
    "layout": "NHWC:NCHW",
    ...
    }

Q: What happens if the system keeps crashing when building the dlstreamer-realsense image?

A: Some systems may run into issues with memory when building the dlstreamer-realsense image. In the Dockerfile.dlstreamer file, change the make command to not use the -j threading option.

- make -j"$(cat < /proc/cpuinfo |grep -c proc)" &&
+ make &&