I am trying to create a simple application on Jetson Nano 2 Gb with DeepStream SDK 5.1.
I have a PyTorch classification model converted to ONNX. This is ResNet18 based custom trained binary classification that I would like to implement to the whole input image. No object detection is required.
I was able to convert it to TRT engine but cannot figure out how to connect it to NVINFER layer.
The overall pipeline should look like: video source → convert to 224x224 with required by ResNet color conversion → inference with custom model → OSD with class lable (0 or 1) → sink
Can you please provide any relevant example? All examples that I can find deal with primary detection and secondary classification. I can substitute the model in config.txt file with my engine file but I do not understand what parameters should be changed to tell that I need to classify the whole frame and there will be two classes.
With my model it still expects to get information on bounding boxes.
ERROR nvinfer gstnvinfer.cpp:613:gst_nvinfer_logger: NvDsInferContext[UID 1]: Error in NvDsInferContextImpl::parseBoundingBox() <nvdsinfer_context_impl_output_parsing.cpp:59> [UID = 1]: Could not find output coverage layer for parsing objects