How to confirm jetson inference train.py is using GPU for processing

Robotics & Edge Computing Jetson & Embedded Systems Jetson Nano
1

I am following the AI Hello World tutorial (being new to the Nvidia Jetson world and AI generally)

(jetson-inference/pytorch-cat-dog.md at master · dusty-nv/jetson-inference · GitHub)

When trying to retrain my first model (cat_dog), everything is working but jtop shows me that the CPU and not GPU appears to be doing all the work.

I am running Jetpack 4.4.1. and using the docker containers from dusty-nv’s github.

Firstly, should the GPU be under load when retraining?

Secondly, if yes, what is the best way to check the GPU is set up correctly?

Thanks for reading.

2

Quick update, it appears that jtop had merely frozen on me and was not displaying the GPU usage and temperature rise whist training.

I have restarted jtop, and running another epoch of training and it is showing that yes indeed the GPU is working hard.

3

OK gotcha, and yes the PyTorch training from Hello AI World will automatically use GPU. For example, here in the code it transfers the model to GPU so that CUDA/cuDNN is used:

The input tensors also get put on the GPU. For these examples, I have the default GPU hardcoded to 0 (since the Jetson has one GPU).