I am currently trying to train sign language and use the jetson nano for inference. I understand the Nano is mainly used for inference and not for training, but from what I’ve read it seems like we can train it to a certain amount.
I was wondering to what extent I can train a pre-trained model to recognize sign language on the Nano. Is it realistic to train up to 10 categories with 1000 images each for at least 10 epochs? How would that affect the power and memory requirements and what can I do to minimize consumption, besides going headless?
To avoid overheating the board, should I train it for 5 epochs, let it cool and then train again for another 5, then 5 more?