Training a model on AGX Xavier

Hi,

I noticed TensorRT is used for prediction only, but is there a way to train a model efficiently on AGX Xavier?

Use case: suppose you have a network of robots. Each robot has two major states - work and stand by. In the work mode - the robot harvests (predicts and collects) data, and in the stand by mode - the robot doesn’t use any resources, which offers a great opportunity to analyze the collected data locally. Obviously, shipping the data centrally is not very efficient especially when it comes to huge amounts of data and large network of robots, but transferring knowledge is. Is there a way to utilize the GPU on AGX Xavier efficiently in order to optimize a model parameters?

Hi,

Based on your use case, you can train the model with PyTorch.
We have provide some GPU enabled prebuilt package in the below topic:

Or you can run l4t-pytorch or l4t-ml docker container to get the environment directly.

Thanks.

Hi AastaLLL,

Thanks for your reply!

I need a C++ solution. Do I get it correctly - the strategy is to use Python for training, and TensorRT C++ for prediction?

Hi,

You can train the model with PyTorch directly.
Then deploy it with trtexec after converting the model into an ONNX format.

$ /usr/src/tensorrt/bin/trtexec --onnx=[your/model]

The source code can be found in the /usr/src/tensorrt/samples/trtexec.

Thanks.