Prepare the model trained yourself, export the model to onnx.
Ensure the onnx can be generated to engine file by tensorrt. Only the first dimension dynamic is supported(batch-size dimension)
Build a pipeline with you model, There are two ways to parser the outputs.
1.Use a customParser, you can refer to this library, please follow your model outputs.
Add output_tensor_meta in nvinfer configuration file. then add a probe function at nvinfer src pad to parse the output tensor. You can refer to this FAQ. DeepStream SDK FAQ - #52 by junshengy