Model file size on jetson nano with 6.0-full-dims is larger than that on desktop PC with 6.0

I find that the model file size of the detection model (49MB) converted using ONNX-Tensorrt 6.0 on the desktop is slightly larger than the ONNX model (44MB). However, after converting the same ONNX model (44MB) on Jetson Nano using ONNX-Tensorrt 6.0-full-dims, the output Tensorrt model file size increase to 170 MB. I am wondering what is the problem with the significant model file size increase on Jetson Nano?

The command I am using is:

onnx2trt -o detection_model.trt -b 1 -d 16 -l model.onnx

Here, I have set the max batch size to be 1, the model data type is float16.

The onnx-tensorrt code I am using:

The jetson nano is installed with Jetpack 4.3 with tensorrt 6.


The TensorRT version for Jetson and desktop should be slightly different on minor version.
So the output file size will have some different.


1 Like

Thanks for prompt reply. I want to understand a bit more on the representation under the hood of the model on Jetson Nano. Any documentation you could point me looking at? If it is possible, could you provide toolkit suggestion provided by NVIDIA that could improve the model file size and speed?


It’s recommended to use our latest software first.
You should be able to get some improvement when upgrading TensorRT into v7.1.

Most of our TensorRT implementation is not open-sourced.
But you can find some serializer for plugin here: