I have a Jetson Nano development kit. I tried the Hello AI World and now am attempting to figure out how to train a new neural net.
Do I really have to have a PC with a good GPU video card or go to the cloud? Or, can I actually use the Jetson Nano itself?
All I want to do at this point is to retrain the neural net to identify about a dozen insects. Not sure which way to go as there is no ‘complete’ tutorial. Lots for the other Jetson boards, just not the Nano.
Where should I look for the info I need?
The Jetson Nano is like one-half of the Jetson TX1, but running the latest software, so you should be able to adapt any of those tutorials.
The re-training “should work” but will run … not fast.
Like, REALLY not fast.
Thank you.
If I understand you correctly, I can download and install DIGITS on the Jetson Nano. Just that it will run really slow. I do realize that I have to go through the download/compile/install cycle to start with.
This is for the “2 days to demo” tutorial.
My concern is that I kept coming up to parts in the tutorial where it said I need to install on a PC or in the cloud.
As mentioned elsewhere, this is a daunting task for someone like me (hobbiest) with no AI education and attempting to understand.
I am using a 5V/5A power supply on the barrel jack. Put the jumper on J48 (by memory?) and it does seem faster than running off the 2.5A USB power supply from my RPi.
I realize that this is early for the Jetson Nano and hopefully there will be more (easier) tutorials coming.
Found one referenced in another post referring to “Getting started with the Jetson Nano - part 3” on RAREblog. It is one of the more clear tutorials I have seen, but uses Jupyter. The other tutorial for beginners uses DIGITS. I am sure you can understand my confusion. :)
Would be nice if I had a degree in AI before I started. It would likely be much farther ahead at the moment. But am entirely new to this field. It is not like the inventory control/POS/accounting and such I have done over the years.
Again, thanks for your response. I will keep working on it over the weekend. Worse case, use the cloud. :)
DIGITS isn’t supported on Jetson, but some community members previously reported getting it to work. If you try it with Nano, you will want to detach the display to conserve memory and also because the display output can become unstable when you are running a long-running training job that fully utilizes the GPU.
DIGITS and Caffe really aren’t ideal for training on embedded platform, and the datasets from that tutorial are pretty big.
You will also probably want to attach an SSD for those with a USB3-to-SATA dongle. Also you will need to mount SWAP file, preferably on the SSD.
I have done transfer learning (re-training) with PyTorch on Nano, it seems to work better and be more memory efficient. This is what I used: examples/imagenet at main · pytorch/examples · GitHub
Hopefully soon I will be able to move the training aspect of the tutorial over to PyTorch. I am waiting for a couple things with ONNX for object detection, so the models can be exported from PyTorch and loaded by TensorRT.
Hi,
I am new to Jetson Nano and IoT boards in general. Can you pls tell me how did you manage to find the memory usage for Pytorch and Tensorflow?
(“I have done transfer learning (re-training) with PyTorch on Nano, it seems to work better and be more memory efficient.”)
Or point me to the appropriate web site that explain how to measure memory usage in general.
Your response is appreciated!
Many thanks!
Please use Tegrastats, you can refer to L4T doc to know how to use it, see https://docs.nvidia.com/jetson/l4t/index.html#page/Tegra%2520Linux%2520Driver%2520Package%2520Development%2520Guide%2FAppendixTegraStats.html%23
Many thanks for your response!