02. TensorFlow

Basic of tensorflow

TOC

• TensorFlow
• Data Flow Graph
• Hello TensorFlow!
• Computation Graph
• TensorFlow Mechanics
• Placeholder
• Everything is Tensor

TensorFlow

• An open source software library for numerical computation using data flow graphs.

Data Flow Graph

• Nodes in the graph represent mathematical operations.
• Egges represent the multidimensional data arrays(tensors) communicated between them.

Image 1. Data Flow Graph (src: https://www.tensorflow.org/)

Hello TensorFlow!

import tensorflow as tf

# Creates a constant node
hello = tf.constant("Hello, TensorFlow!")

# Start a TensorFlow session
sess = tf.Session()

# Run the operation and print result
print(sess.run(hello))

b'Hello, TensorFlow!'

Computation Graph

import tensorflow as tf

# Defines constant and operation nodes
node1 = tf.constant(3.0, tf.float32)
node2 = tf.constant(4.0) # tf.float32 can be implicitly given.
node3 = tf.add(node1, node2)

# If node is printed, only type information will be displayed
print("node1:", node1)
print("node2:", node2)
print("node3:", node3)

# To get outputs of each nodes, session should run the data flow graph.
sess = tf.Session()

# Session can run single node or list of nodes
print("sess.run(node1, node2): ", sess.run([node1, node2]))
print("sess.run(node3): ", sess.run(node3))

node1: Tensor("Const_3:0", shape=(), dtype=float32)
node2: Tensor("Const_4:0", shape=(), dtype=float32)
node3: Tensor("Add:0", shape=(), dtype=float32)
sess.run(node1, node2):  [3.0, 4.0]
sess.run(node3):  7.0

TensorFlow Mechanics

1. Build graph using TensoFlow operations
   • define nodes
2. Feed data and run grap (operation)
   • sess.run(node)
3. Update variables in the graph and return values
   • print(output of sess.run(node))

Image 2. TensorFlow Mechanics (src: www.mathwarehouse.com)

Placeholder

• For the second step of TensorFlow Mechanics, data should be fed to the node in the data graph.
• Placeholder is a special tensorflow's node, which can be fed.

import tensorflow as tf

# Defines placeholders
node1 = tf.placeholder(tf.float32)
node2 = tf.placeholder(tf.float32)

# Defines an operation
node3 = node1 + node2

# Defines a session
sess = tf.Session()

# Run session with fed data
# feed_dict is a dictionary data type,
#  and single value and list can be injected.
print(sess.run(node3, feed_dict={node1: 3, node2: 4}))

print(sess.run(node3, feed_dict={node1: [3, 5], node2: [4, 6]}))

7.0
[  7.  11.]

Everything is Tensor

Tensor rank

Rank	Math entity	Python example
0	Scalar (magnitude only)	s = 100
1	Vector (magnitude and direction)	v = [1,2,3]
2	Matrix (table of numbers)	m = [[1,2,3], [1,2,3]]
3	3-Tensor (cube of numbers)	t = [[[2], [4], [6]], [[8], [9], [10]]]
4	n-Tensor (you get the idea)	...

Table 1. Rank

Tensor shape

Rank	Shape	Dimension number	Example
0	[]	0-D	A 0-D tensor. A Scalar
1	[D0]	1-D	A 1-D tensor with shape [5]
2	[D0, D1]	2-D	A 2-D tensor with shape [3, 4]
3	[D0, D1, D2]	3-D	A 3-D tensor with shape [1,4,3]
n	[D0, D1,..., Dn-1]	n-D	A tensor with shape [D0, D1, ..., Dn-1]

Table 2. Shape

Data type

Data type	Python type	Description
DT_FLOAT	tf.float32	32 bits floating point
DT_DOUBLE	tf.float64	64 bits floating point
DT_INT8	tf.int8	8 bits signed integer
DT_INT16	tf.int16	16 bits signed integer
DT_INT32	tf.int32	32 bits sigend integer
DT_INT64	tf.int64	64 bits sigend integer

Table 3. Type