Tensorflow -2

So this is my attempt to learn the very basics of tensorflow. I’m currently following a book Fundatmentals of Deep Learning by Nikhil Buduma. It is available in the public domain and provides an excellent intuition of neural networks in TensorFlow. I’ve been trying to master TensorFlow for previous couple of weeks and i decided to start with a book rather than learning online.

Thankfully I’m an expert in installing tensorflow in Anaconda i’ve done it multiple times. lets start with the next basic step.

Import tensorflow

import tensorflow as tf # This import tensorflow with tf as an alias.

Hello World

let’s print ‘hello World’ in tensorflow

import tensorflow as tf
helloWorld= tf.constant('Hello World')
sess = tf.session()
session.run(helloWorld)
----> 'Hello World'

bit length when compared to python. Interestingly as per latest tensor flow version, we have something called eager execution where in we can do away with tf.Session()

Constants

import tensorflow as tf
tf.enable_eager_execution()
const=tf.constant('Hello World')
print(const)
---->tf.Tensor(b'Hello World', shape=(), dtype=string)

I’m still learning the nuances of tensorflow enable_eager execution. I shall write more on this when i have a clear picture about it. lets go back to the graph and sessions paradigm once again.

a = tf.constant(2)
b = tf.constant(3)
mul = tf.multiply(a,b)
session.run(mul)
----> 6

Variables

  • Variables must be explicitly initialized before a graph is used for the first time
  • Gradient Method could be used to modify the variables after each iteration
  • We can save values stored in variables to disk and restore them for later use.
weights = tf.Variable(tf.random_normal([300,200],stddev=0.5), name='weights', trainable=False)

  • tf.random_normal is an operation that produce a tensor initialised using a normal distribution with standard deviation of 0.5, we could use other initializers as well they are

    tf.zeros(shape, dtype=tf.float32)
tf.ones(shape, dtype=tf.float32)
tf.random_normal(shape, dtype=tf.float32,mean=0, stddev=1,seed=0)
tf.truncated_normal(shape, dtype=tf.float32,mean=0, stddev=1,seed=0)
tf.random_uniform(shape, dtype=tf.float32,mean=0, stddev=1,seed=0)

  • tensor shape is 300x200, implying weights connect a layer with 300 neurons to a layer with 200 neurons

  • we have also passed a name to the variable ‘weights’

  • trainable = True /False lets the tensorflow know that if the weights are meant to be trainable.