In this section, we explore a selection of operators that can be used to transform tensors. Note that some of these operators work very similar to NumPy array transformations. However, when we are dealing with tensors with ranks higher than 2, we need to be careful in using such transformations, for example, the transpose of a tensor.
First, as in NumPy, we can use the attribute arr.shape to get the shape of a NumPy array. In TensorFlow, we use the tf.get_shape function instead:
>>> import tensorflow as tf
>>> import numpy as np
>>>
>>> g = tf.Graph()
>>> with g.as_default():
... arr = np.array([[1., 2., 3., 3.5],
... [4., 5., 6., 6.5],
... [7., 8., 9., 9.5]])
... T1 = tf.constant(arr, name='T1')
... print(T1)
... s = T1.get_shape()
... print('Shape of T1 is', s)
... T2 = tf.Variable(tf.random_normal(
... shape=s))
... print(T2)
... T3 = tf.Variable(tf.random_normal(
... shape=(s.as_list()[0],)))
... print(T3)The output of the previous code example is as follows:
Tensor("T1:0", shape=(3, 4), dtype=float64)
Shape of T1 is (3, 4)
<tf.Variable 'Variable:0' shape=(3, 4) dtype=float32_ref>
<tf.Variable 'Variable_1:0' shape=(3,) dtype=float32_ref>Notice that we used s to create T2, but we cannot slice or index s for creating T3. Therefore, we converted s into a regular Python list by s.as_list() and then used the usual indexing conventions.
Now, let's see how we can reshape tensors. Recall that in NumPy, we can use np.reshape or arr.reshape for this purpose. In TensorFlow, we use the function tf.reshape to reshape a tensor. As is the case for NumPy, one dimension can be set to -1 so that the size of the new dimension will be inferred based on the total size of the array and the other remaining dimensions that are specified.
In the following code, we reshape the tensor T1 to T4 and T5, both of which have rank 3:
>>> with g.as_default(): ... T4 = tf.reshape(T1, shape=[1, 1, -1], ... name='T4') ... print(T4) ... T5 = tf.reshape(T1, shape=[1, 3, -1], ... name='T5') ... print(T5)
The output is as follows:
Tensor("T4:0", shape=(1, 1, 12), dtype=float64)
Tensor("T5:0", shape=(1, 3, 4), dtype=float64)Next, let's print the elements of T4 and T5:
>>> with tf.Session(graph = g) as sess: ... print(sess.run(T4)) ... print() ... print(sess.run(T5)) [[[ 1. 2. 3. 3.5 4. 5. 6. 6.5 7. 8. 9. 9.5]]] [[[ 1. 2. 3. 3.5] [ 4. 5. 6. 6.5] [ 7. 8. 9. 9.5]]]
As we know, there are three ways to transpose an array in NumPy: arr.T, arr.transpose(), and np.transpose(arr). In TensorFlow, we use the tf.transpose function instead, and in addition to a regular transpose operation, we can change the order of dimensions in any way we want by specifying the order in perm=[...]. Here's an example:
>>> with g.as_default():
... T6 = tf.transpose(T5, perm=[2, 1, 0],
... name='T6')
... print(T6)
... T7 = tf.transpose(T5, perm=[0, 2, 1],
... name='T7')
... print(T7)
Tensor("T6:0", shape=(4, 3, 1), dtype=float64)
Tensor("T7:0", shape=(1, 4, 3), dtype=float64)Next, we can also split a tensor into a list of subtensors using the tf.split function, as follows:
>>> with g.as_default(): ... t5_splt = tf.split(T5, ... num_or_size_splits=2, ... axis=2, name='T8') ... print(t5_splt) [<tf.Tensor 'T8:0' shape=(1, 3, 2) dtype=float64>, <tf.Tensor 'T8:1' shape=(1, 3, 2) dtype=float64>]
Here, it's important to note that the output is not a tensor object anymore; rather, it's a list of tensors. The name of these subtensors are 'T8:0' and 'T8:1'.
Lastly, another useful transformation is the concatenation of multiple tensors. If we have a list of tensors with the same shape and dtype, we can combine them into one big tensor using the tf.concat function. An example is given in the following code:
>>> g = tf.Graph()
>>> with g.as_default():
... t1 = tf.ones(shape=(5, 1),
... dtype=tf.float32, name='t1')
... t2 = tf.zeros(shape=(5, 1),
... dtype=tf.float32, name='t2')
... print(t1)
... print(t2)
>>> with g.as_default():
... t3 = tf.concat([t1, t2], axis=0, name='t3')
... print(t3)
... t4 = tf.concat([t1, t2], axis=1, name='t4')
... print(t4)
Tensor("t1:0", shape=(5, 1), dtype=float32)
Tensor("t2:0", shape=(5, 1), dtype=float32)
Tensor("t3:0", shape=(10, 1), dtype=float32)
Tensor("t4:0", shape=(5, 2), dtype=float32)Let's print the values of these concatenated tensors:
>>> with tf.Session(graph=g) as sess:
... print(t3.eval())
... print()
... print(t4.eval())
[[ 1.]
[ 1.]
[ 1.]
[ 1.]
[ 1.]
[ 0.]
[ 0.]
[ 0.]
[ 0.]
[ 0.]]
[[ 1. 0.]
[ 1. 0.]
[ 1. 0.]
[ 1. 0.]
[ 1. 0.]]