The base of TensorFlow is the computational graph, which we discussed earlier in this chapter, and tensors. A tensor is an n-dimensional vector. Thus, a scalar and a matrix variable is also a tensor. Here, we will try some of the basic computations to start with TensorFlow. Please try to implement this section in a python IDE such as Jupyter Notebook.
For the TensorFlow installation and dependencies please refer to the following link:
https://www.tensorflow.org/install/
Import tensorflow by the following command:
import tensorflow as tf
tf.zeros() and tf.ones() are some of the functions that instantiate basic tensors. The tf.zeros() takes a tensor shape (that is, a tuple) and returns a tensor of that shape with all the values being zero. Similarly, tf.ones() takes a tensor shape but returns a tensor of that shape containing only ones. Try the following commands in python shell to create a tensor:
>>> tf.zeros(3)
<tf.Tensor 'zeros:0' shape=(3,) dtype=float32>
>>>tf.ones(3)
<tf.Tensor 'ones:0' shape=(3,) dtype=float32>
As you can see, TensorFlow returns a reference to the tensor and not the value of the tensor. In order to get the value, we can use eval() or run(), a function of tensor objects by running a session as follows:
>>> a = tf.zeros(3)
>>> with tf.Session() as sess:
sess.run(a)
a.eval()
array([0., 0.,0.], dtype=float32)
array([0., 0.,0.], dtype=float32)
Next come the tf.fill() and tf.constant() methods to create a tensor of a certain shape and value:
>>> a = tf.fill((2,2),value=4.)
>>> b = tf.constant(4.,shape=(2,2))
>>> with tf.Session() as sess:
sess.run(a)
sess.run(b)
array([[ 4., 4.],
[ 4., 4.]], dtype=float32)
array([[ 4., 4.],
[ 4., 4.]], dtype=float32)
Next, we have functions that can randomly initialize a tensor. Among them, the most frequently used ones are:
- tf.random_normal: Samples random values from the Normal distribution of specified mean and standard deviation
- tf.random_uniform(): Samples random values from the Uniform distribution of a specified range
>>> a = tf.random_normal((2,2),mean=0,stddev=1)
>>> b = tf.random_uniform((2,2),minval=-3,maxval=3)
>>> with tf.Session() as sess:
sess.run(a)
sess.run(b)
array([[-0.31790468, 1.30740941],
[-0.52323157, -0.2980336 ]], dtype=float32)
array([[ 1.38419437, -2.91128755],
[-0.80171156, -0.84285879]], dtype=float32)
Variables in TensorFlow are holders for tensors and are defined by the function tf.Variable():
>>> a = tf.Variable(tf.ones((2,2)))
>>> a
<tf.Variable 'Variable:0' shape=(2, 2) dtype=float32_ref>
The evaluation fails in case of variables because they have to be explicitly initialized by using tf.global_variables_initializer within a session:
>>> a = tf.Variable(tf.ones((2,2)))
>>> with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
a.eval()
array([[ 1., 1.],
[ 1., 1.]], dtype=float32)
Next in the queue, we have matrices. Identity matrices are square matrices with ones in the diagonal and zeros elsewhere. This can be done with the function tf.eye():
>>> id = tf.eye(4) #size of the square matrix = 4
>>> with tf.Session() as sess:
sess.run(id)
array([[ 1., 0., 0., 0.],
[ 0., 1., 0., 0.],
[ 0., 0., 1., 0.],
[ 0., 0., 0., 1.]], dtype=float32)
Similarly, there are diagonal matrices, which have values in the diagonal and zeros elsewhere, as shown here:
>>> a = tf.range(1,5,1)
>>> md = tf.diag(a)
>>> mdn = tf.diag([1,2,5,3,2])
>>> with tf.Session() as sess:
sess.run(md)
sess.run(mdn)
array([[1, 0, 0, 0],
[0, 2, 0, 0],
[0, 0, 3, 0],
[0, 0, 0, 4]], dtype=int32)
array([[1, 0, 0, 0, 0],
[0, 2, 0, 0, 0],
[0, 0, 5, 0, 0],
[0, 0, 0, 3, 0],
[0, 0, 0, 0, 2]], dtype=int32)
We use the tf.matrix_transpose() function to transpose the given matrix, as shown here:
>>> a = tf.ones((2,3))
>>> b = tf.transpose(a)
>>> with tf.Session() as sess:
sess.run(a)
sess.run(b)
array([[ 1., 1., 1.],
[ 1., 1., 1.]], dtype=float32)
array([[ 1., 1.],
[ 1., 1.],
[ 1., 1.]], dtype=float32)
The next matrix operation is the matrix multiplication function as shown here. This is done by the function tf.matmul():
>>> a = tf.ones((3,2))
>>> b = tf.ones((2,4))
>>> c = tf.matmul(a,b)
>>> with tf.Session() as sess:
sess.run(a)
sess.run(b)
sess.run(c)
array([[ 1., 1.],
[ 1., 1.],
[ 1., 1.]], dtype=float32)
array([[ 1., 1., 1., 1.],
[ 1., 1., 1., 1.]], dtype=float32)
array([[ 2., 2., 2., 2.],
[ 2., 2., 2., 2.],
[ 2., 2., 2., 2.]], dtype=float32)
Reshaping of tensors from one to another is done by using the tf.reshape() function, as shown here:
>>> a = tf.ones((2,4)) #initial shape is (2,4)
>>> b = tf.reshape(a,(8,)) # reshaping it to a vector of size 8. Thus shape is (8,)
>>> c = tf.reshape(a,(2,2,2)) #reshaping tensor a to shape (2,2,2)
>>> d = tf.reshape(b,(2,2,2)) #reshaping tensor b to shape (2,2,2)
#####Thus, tensor 'c' and 'd' will be similar
>>> with tf.Session() as sess:
sess.run(a)
sess.run(b)
sess.run(c)
sess.run(d)
array([[ 1., 1., 1., 1.],
[ 1., 1., 1., 1.]], dtype=float32)
array([ 1., 1., 1., 1., 1., 1., 1., 1.], dtype=float32)
array([[[ 1., 1.],
[ 1., 1.]],
[[ 1., 1.],
[ 1., 1.]]], dtype=float32)
>
array([[[ 1., 1.],
[ 1., 1.]],
[[ 1., 1.],
[ 1., 1.]]], dtype=float32)
The flow of computation in TensorFlow is represented as a computational graph, which is as instance of tf.Graph. The graph contains tensors and operation objects, and keeps track of a series of operations and tensors involved. The default instance of the graph can be fetched by tf.get_default_graph():
>>> tf.get_default_graph()
<tensorflow.python.framework.ops.Graph object at 0x7fa3e139b550>
We will explore complex operations, the creation of neural networks, and much more in TensorFlow in the coming chapters.
