I train SSD-MobileNet v2 via jetson-inference, export it as onnx model and use it in Jetson nano. I have a question here.
jetson-inference uses the TensorRT format from the github description to export the final model to onnx. Can I call this TensorRT format(.trt)?
I want to compare jetson-inference’s SSD-MobileNet v2 and Darknet’s YOLOv4-tiny.
But, YOLO’s format is .trt.
(Train with Darknet and convert weight->onnx->trt via tenssort_demos.)
Is it appropriate compare the two model?
In other words, does jetson-inference’s onnx model include TensorRT?
Hi @3629701, yes, jetson-inference uses TensorRT to run the ONNX models. The first time it loads a new model, it optimizes it with TensorRT and saves the TensorRT engine to disk (in a .engine file, probably the same as what you mean with .trt file)
The pre/post-processing for detection models in jetson-inference isn’t setup for supporting YOLO, but if you have other code that runs your YOLO model, yes you can compare the inferencing performance with SSD-Mobilenet. It sounds like both will have been run with TensorRT.