Representing Features in Tensorflow Neural Network

1
votes

I've been playing around with the tensorflow playground: http://playground.tensorflow.org/ and have designed some interesting networks that I'd like to deploy in an actual tensorflow job.

I've noticed on the playground they represent feature sets as functions x, x^2, sine(x) and so on.

Feature Sets on Tensorflow Playground

My question: what's the best way to represent these types of transformation functions in a tensorflow job? How do people usually handle this? In transformations before the input layer OR within the network itself?

I've had some experience writing basic regression in tensorflow and playing with custom datasets but am a relative noob to deep learning.

1
neural-networktensorflowdeep-learning
In TensorFlow, those features will look like this: x1, x2, tf.square(x1), x1*x2, tf.sin(x1) – Yaroslav Bulatov

1 Answers

0
votes

Basically you need to extract some features from the dataset and they should describe your problem the best way. So, try to generate data from your dataset using different algorithms\functions first, reshape it to fit your model inputs (or vise versa) and then feed this stuff to Neural Net.