Introducing tensorflow¶

Why tensorflow?¶

• It's not specifically for neural networks - it's more generally an architecture for executing a graph of numberical operations
• Tensorflow can optimize the processing of that graph, and distribute it's processing across a network
  • Sounds a lot like Apache Spark
• It can also distribute work accross GPUs
  • Can handle massive scale - it was made by Google
• Runs on about anything
• Highly efficient C++ code with easy to use Python APIs

Tensorflow basics¶

• Install conda with conda install tensorflow or conda install tensorflow-gpu
• A tensor is just a fancy name for an array of matrix of values
• To use tensorflow you must
  • Construct a graph to compute your tensors
  • Initialize your variables
  • Execute that graph - nothing actually heppens until then

import tensorflow as tf
a = tf.Variable(1, name="a")
b = tf.Variable(2, name"b")
f = a + b
tf.print(f)

Creating a neural network with tensorflow¶

• Mathematical insights
  • All those interconeected arrows multiplying weights can be thought as a big matrix multiplication
  • The bias term can just be added onto the result of that matrix multiplication
• So in tensorflow, we can define layer of a neural network as

output = tf.matmul(previous_layer, layer_weights) + layer_biases

• By using tensorflow directly we're kinda doing this the hard way.

• Load up our training and testing data
• Construct a graph describing our neural network
• Associate an optimizer (ie gradient descent) to the network
• Run the optimizer with your training data
• Evaluate your trained network with your testing data.

Make sure your features are normalized¶

• Neural networks usually work best if your input data is normalized
  • That is, 0 mean and unit variance
  • The real goal is that every input feature is comparable in terms of magnitude
• ssikit_learn's StandardScaler can do this for you.
• Many data sets are normalized to begin with - such as the one we're about to use.