get output of inference model.

[imohammadhossein]

hey, thanks for the nice work. I wonder if there is any way that we can access to the outputs of the Nvinfer directly ? I have trained a custom model in python and it does have a specific custom bbox parse function in Python. Moreover, when I run it in process_mode=1 (primary mode), I get the following error, how should I fix it using Python ?

nvinfer gstnvinfer.cpp:632:gst_nvinfer_logger:<primary-inference> NvDsInferContext[UID 2]: Error in NvDsInferContextImpl::fillDetectionOutput() <nvdsinfer_context_impl_output_parsing.cpp:735> [UID = 2]: Failed to parse bboxes

While using process_mode=2, has no error with no prediction.

[marcoslucianops]

Are you using an YOLO model?

[imohammadhossein]

Yes for now; but I'm looking for a general solution for any custom model.

[marcoslucianops]

I can help you with the things related to this repo and models supported.

For general models:

• To access the output tensors from the model directly in Python, you need to use the tensor meta.
• Custom models have different layers and different outputs, you need to create the kernel to "decode" the outputs to the format you want to receive to parse.

[imohammadhossein]

Thankyou! but the tensor meta is empty. is that necessary to pass .so file or similar config file ?

[marcoslucianops]

You need to use the same config_infer file. The kernel for the yoloLayer is compiled in the so lib.

[imohammadhossein]

Ok. Is there any powerful and complete tutorial ?

[marcoslucianops]

As far as I know, there isn't. There's no good documentation/turoriais for the DeepStream.