How Does The Machine Learning Library TensorFlow Work?

An awesome deep learning library - TensorFlow

An open-source machine learning framework for everyone. As I believe, when we know the internals of a library, we can get the most out of it. And we feel much more confident while working with that library. Today, TensorFlow is being used by the top companies like Google, Uber, Airbnb, Nvidia, Dropbox, etc.

Believe me, you can achieve anything out of it. It is mainly designed for deep neural network models.

TensorFlow supports the following:

  • TensorFlow lets you deploy computation to one or more CPUs or GPUs in a desktop, server, or mobile device in a very simple way. This way the things can be done very fast.
  • TensorFlow lets you express your computation as a data flow graph.
  • TensorFlow lets you visualize the graph using the in-built tensorboard. You can inspect and debug the graph very easily.
  • TensorFlow gives the best performance with an ability to iterate quickly, train models faster and run more experiments.
  • TensorFlow runs on nearly everything: GPUs and CPUs—including mobile and embedded platforms—and even tensor processing units (TPUs), which are specialized hardware to do the tensor math on.

How does the TensorFlow flexible enough to support all the above features?

The architecture of the TensorFlow helps it supporting all the above and much more.

First of all, you must know that in TensorFlow, you build a graph in which you define the constants, variables, operations and then you execute it. The graph is a data structure which consists of all the constants, variables and operations that you want to do.

The node represents an operation.

The edges are carriers of data structures (tensors), where an output of one operation (from one node) becomes the input for another operation.

You do the things in TensorFlow library like below.

import tensorflow as tf

# Basic constant operations
# The value returned by the constructor represents the output
# of the Constant op.
a = tf.constant(2)
b = tf.constant(3)

# Launch the default graph.
with tf.Session() as sess:
    print "a: %i" %, "b: %i" %
    print "Addition with constants: %i" %
    print "Multiplication with constants: %i" %*b)

# a=2, b=3
# Addition with constants: 5
# Multiplication with constants: 6

# Basic Operations with variable as graph input
# The value returned by the constructor represents the output
# of the Variable op. (define as input when running session)
# tf Graph input
a = tf.placeholder(tf.int16)
b = tf.placeholder(tf.int16)

# Define some operations
add = tf.add(a, b)
mul = tf.multiply(a, b)

# Launch the default graph.
with tf.Session() as sess:
    # Run every operation with variable input
    print "Addition with variables: %i" %, feed_dict={a: 2, b: 3})
    print "Multiplication with variables: %i" %, feed_dict={a: 2, b: 3})

# Addition with variables: 5
# Multiplication with variables: 6

Let’s see it by building a mini TensorFlow. We will try to build a small library just for understanding how the TensorFlow works.

Create a class for the operations called TFOperation.

class TFOperation():

def __init__(self, input_nodes=[]):
    self.input_nodes = input_nodes
    self.output_nodes = []
    for node in input_nodes:
def compute(self):

Then, extend TFOperation class to make add class.

class add(TFOperation, object):
    def __init__(self,x,y):
        super(add, self).__init__([x,y])
    def compute(self, x_var, y_var):
        self.inputs = [x_var, y_var]
        return x_var + y_var

Similarly, we can extend the TFOperation to make multiply class.

class multiply(TFOperation, object):
    def __init__(self,x,y):
        super(multiply, self).__init__([x,y])
    def compute(self, x_var, y_var):
        self.inputs = [x_var, y_var]
        return x_var * y_var

Now, create the Placeholder class for the variables.

class Placeholder():
    def __init__(self):
        self.output_nodes = []

Now, create the Constant class.

class Constant():
    def __init__(self, initial_value=None):
        self.value = initial_value
        self.output_nodes = []

Then, we will create the Graph class.

class Graph():
    def __init__(self):
        self.operations = []
        self.placeholders = []
        self.variables = []
    def set_as_default(self):
        global _default_graph
        _default_graph = self

Now, let's make the utility function traverse_postorder so that the operations happen in the correct order.

def traverse_postorder(operation):
    nodes_postorder = []
    def recurse(node):
        if isinstance(node, TFOperation):
            for input_node in node.input_nodes:
    return nodes_postorder

Finally create the Session class.

import numpy as np

class Session():
    def run(self, operation, feed_dict={}):
        nodes_postorder = traverse_postorder(operation)
        for node in nodes_postorder:
            if isinstance(node, Placeholder):
                node.output = feed_dict[node]
            elif isinstance(node, Constant):
                node.output = node.value
                # OPERATION
                node.inputs = [input_node.output for input_node in node.input_nodes]
                node.output = node.compute(*node.inputs)
            if type(node.output) == list:
                node.output = np.array(node.output)
        return operation.output

So, our mini TensorFlow library is ready.

Let’s do some operations with constants and placeholder(variables).

We will do the following operations with our mini TensorFlow library.

y = mx + c
m = 10
c = 1
y = 10x + 1	
# Create the Graph
g = Graph()
# Set the graph as default
# Create a constant
m = Constant(10)
# Create an another constant
c = Constant(1)
# Create a variable using Placeholder
x = Placeholder()
# Create add and multiply operation
y = add(multiply(m,x), c)
# Create the session
sess = Session()
# Run the session to get the result
result =, feed_dict={x:10})
# Print the result

So, this is how the TensorFlow works. That's it for now.

Also, Let’s become friends on Twitter, Linkedin, Github, Quora, and Facebook.