Train NN with Jetson TX2

Hello all,
I am a beginner in GPU programming and AI. For a project part, I need to train a neural network with NVIDIA Jetson TX2 board. I know there’s plenty of documentation about it but I am running out of time, could someone help me about the following questions ?

The neural network is quite simple :

  • 6 numerical inputs (simply numbers).
  • A large number of possible numerical outputs (let’s say 1000).
  • 1 or 2 hidden layers of approximately 100 neurons each.
  • I have to try the following types of neurons : Perceptron, linear, sigmoid, hyperbolic tangent.
  •   I have the data to train the NN: correct output for given inputs
    

My questions :

  1. What tools should I use to implement that NN as fast as possible ? I think CUDA is used for such implementation but both cuDNN and cuBLAS seem to simplify the task, true ? A template/sample of code implementing a NN similar to the one described above would be perfect !

  2. How to train this implemented NN on the Jetson TX2 ? (connect board and host computer->tools, specific documentation?)

I read the Two Days Demo, but couldn’t understand how to apply it to my own problem, since it is not as complex as Image recognition, segmentation, …

I am sorry, I am asking a lot of simple things and I am sure I would find the solution with a better research but I dont have much time.
Thank you !

Hi,

You can try TensorFlow which is more flexible for customized layer design.
Here is the tutorial for installation:
[url]Installing TensorFlow for Jetson Platform :: NVIDIA Deep Learning Frameworks Documentation

After the installation, the training procedure is same as the standard TensorFlow:
[url]https://www.tensorflow.org/tutorials/[/url]

Thanks.

Hi Xavier,

Like yourself I am new to GPU programming and AI and look for the right course / project. Are you enrolled in a specific course you can share ?