Is it possible to train CNN on jetson tx2? Or jetson is used only for a pre-trained neural network?

Hello. Is it possible to train CNN on jetson tx2? Or jetson is used only for a pre-trained neural network?

Hi marcel.nikmon, it is possible to train on Jetson TX2, it’s 8GB memory makes it feasible. However you normally see training done on a PC, workstation, or in the cloud with larger GPU, because of the size of the datasets involved and training recognition/detection networks like GoogLeNet and ResNet consume lots of resources. Training from scratch on Jetson could consume quite some time, but transfer learning less so.

You can install frameworks like Caffe, TensorFlow, and pyTorch to the TX2 and perform training onboard, for large CNN it will take longer. But in reinforcement learning and online learning applications like unsupervised autoencoders, the training performance may be sufficient for realtime use.

See here for TX2 install recipes from the wiki: https://elinux.org/Jetson_TX2#Deep_Learning

Hi Dusty,

Thanks for your help. Can you advise on the exact type of datasets that can smoothly run directly on TX2 for training ?

Technically if you have attached an SSD, then you can use large datasets too, they will just take longer to train.

See here for some example datasets that you can train onboard Jetson in a few hours:
https://github.com/dusty-nv/jetson-inference/blob/master/docs/pytorch-transfer-learning.md