How to port a pytorch model to jetson nano ?

Autonomous Machines Jetson & Embedded Systems Jetson Nano
1

I have a faster-rcnn.pytorch model. The repository address for this project is: https://github.com/jwyang/faster-rcnn.pytorch. I want to port this model to jetson nano. Is there a tutorial for reference? .

The jetson I bought has already installed python 3.6 and JetPack 4.2, so I didn’t build any wheels according to “Build Instructions”.
I downloaded the swap partition just by following “Build Instructions”.
I downloaded the “torch-1.1.0a0+b457266-cp36-cp36m-linux_aarch64.whl” offline, and then used “pip3 install .whl” to install pytorch1.1. I also used the following command to the torchvision:
$ git clone GitHub - pytorch/vision: Datasets, Transforms and Models specific to Computer Vision

$ cd vision
$ sudo python setup.py install

But when I install scipy with “pip3 install scipy” it always fails.

Any help will be greatly appreciated.

2

Hi,

The simplest way is to install pytorch on the Jetson Nano and run it directly:
https://devtalk.nvidia.com/default/topic/1049071/pytorch-for-jetson-nano/

Another approach is to convert the model into TensorRT.
It will increase the inference speed and also save the memory.
Here is a tutorial for doing so: [url]https://docs.nvidia.com/deeplearning/sdk/tensorrt-sample-support-guide/index.html#network_api_pytorch_mnist[/url]

Thanks.

1 Like
3

The provided sample: Sample Support Guide :: NVIDIA Deep Learning TensorRT Documentation

does not use GPU for training ! (my Platform is Jetson Nano)

Could you please provide a fix or solution for that, as Training on CPUs is really time consuming !

4

Here is a tutorial that uses PyTorch GPU for training: https://github.com/dusty-nv/jetson-inference/blob/master/docs/pytorch-transfer-learning.md

If you prefer to modify the existing example above to use GPU for training, you need to call .cuda() on the PyTorch model and tensors.

5

Thanks ! It was resolved.

Training time dropped drastically and GPU utilization began.