More on Tensor
Prepared by Sanasam Ranbir Singh
Variable tensor
Unlike constant tensor, you can change the value of a tensor using tf.assign() method. However, variable tensor should be initialized while creating it.
In [1]:
import tensorflow as tf
x = tf.Variable([1,2,3,4]) # initialize with [1,2,3,4]
print(x)<tf.Variable 'Variable:0' shape=(4,) dtype=int32, numpy=array([1, 2, 3, 4])>
In [109]:
x = tf.Variable([[1,2,3,4],[5,6,7,8]])
print(x)
<tf.Variable 'Variable:0' shape=(2, 4) dtype=int32, numpy=
array([[1, 2, 3, 4],
[5, 6, 7, 8]])>
In [111]:
x = tf.Variable([[1,2,3,4],[5,6,7,8]], dtype=tf.float32)
print(x)<tf.Variable 'Variable:0' shape=(2, 4) dtype=float32, numpy=
array([[1., 2., 3., 4.],
[5., 6., 7., 8.]], dtype=float32)>
In [31]:
x = tf.Variable([1,2,3,4])
print(x.name)
print(x.shape)
print(x.dtype)
print(x.numpy())
Variable:0 (4,) <dtype: 'int32'> [1 2 3 4]
Show attributes of a tensor
In [32]:
x = tf.constant([1,2,3,4])
#print(x.name) #possible, when eager execution is disabled
print(x.shape)
print(x.dtype)
print(x.numpy())(4,) <dtype: 'int32'> [1 2 3 4]
In [55]:
x = tf.Variable([[1,2,3,4],[5,6,7,8]])
print(x.name)
print(x.shape)
print(x.dtype)
print(x.numpy())Variable:0 (2, 4) <dtype: 'int32'> [[1 2 3 4] [5 6 7 8]]
Convert a content tensor to a variable tensor and vice versa
In [33]:
x_con = tf.constant([1,2,3,4])
x_var = tf.Variable(t_con)
print(x_var)<tf.Variable 'Variable:0' shape=(4,) dtype=int32, numpy=array([1, 2, 3, 4])>
In [35]:
x_var = tf.Variable([1,2,3,4])
x_con = tf.constant(x_var)
print(x_con)tf.Tensor([1 2 3 4], shape=(4,), dtype=int32)
Reshape a tensor
You can chage the sape of the tensor after creating it. but, the number of element of the source tensor and target sensor should be same.
In [71]:
x = tf.constant([1,2,3,4,5,6,7,8], shape=(2,4))
print(x)tf.Tensor( [[1 2 3 4] [5 6 7 8]], shape=(2, 4), dtype=int32)
In [112]:
x = tf.constant([1,2,3,4,5,6,7,8], shape=(2,4))
tf.reshape(x, (4,2))<tf.Tensor: shape=(4, 2), dtype=int32, numpy=
array([[1, 2],
[3, 4],
[5, 6],
[7, 8]])> In [87]:
x = tf.constant([[1,2,3,4],[5,6,7,8]])
tf.reshape(x, (8))<tf.Tensor: shape=(8,), dtype=int32, numpy=array([1, 2, 3, 4, 5, 6, 7, 8])>
In [5]:
x = tf.constant([[1,2,3,4],[5,6,7,8]], shape=(2,4))
tf.reshape(x, (-1)) # flatten the tensor in 1D<tf.Tensor: shape=(8,), dtype=int32, numpy=array([1, 2, 3, 4, 5, 6, 7, 8])>
Use of -1
The -1 is like a don't care. When you reshape with (x,-1), it generate a new 2D tensor with 4 number of 1D sensors of equal shape. The number of 0D tensor in each 1D tensor depends on the number of elements in the original tensor. Note that the number of elements in the original tensor and new tensor should be same.
In [6]:
tf.reshape(x, (4,-1))<tf.Tensor: shape=(4, 2), dtype=int32, numpy=
array([[1, 2],
[3, 4],
[5, 6],
[7, 8]])> In [115]:
x = tf.constant([1,2,3,4,5,6,7,8])
tf.reshape(x, (2,-1,2))<tf.Tensor: shape=(2, 2, 2), dtype=int32, numpy=
array([[[1, 2],
[3, 4]],
[[5, 6],
[7, 8]]])> In [65]:
x = tf.constant([[1,2,3,4],[5,6,7,8]])
tf.reshape(x, (2,2,-1))<tf.Tensor: shape=(2, 2, 2), dtype=int32, numpy=
array([[[1, 2],
[3, 4]],
[[5, 6],
[7, 8]]])> In [76]:
x = tf.constant([[1,2,3,4],[5,6,7,8]])
tf.reshape(x, (-1,2,2))<tf.Tensor: shape=(2, 2, 2), dtype=int32, numpy=
array([[[1, 2],
[3, 4]],
[[5, 6],
[7, 8]]])> Access the elements of a tensor
argmax()
Return the index of the maximum element in kD tendor and return a (k-1)D index tensor
In [4]:
x = tf.constant([[1,2,3,4], [5,6,7,8]])
tf.argmax(x)<tf.Tensor: shape=(4,), dtype=int64, numpy=array([1, 1, 1, 1], dtype=int64)>
In [10]:
x = tf.constant([[1,2,3,4], [5,6,7,8]])
tf.argmin(x)<tf.Tensor: shape=(4,), dtype=int64, numpy=array([0, 0, 0, 0], dtype=int64)>
In [9]:
x = tf.constant([[9,2,10,4],[5,6,7,8]])
print(tf.math.argmax(x))tf.Tensor([0 1 0 1], shape=(4,), dtype=int64)
In [17]:
x = tf.constant([[2, 20, 30, 3, 6], [3, 11, 16, 1, 8],
[14, 45, 23, 5, 27]])
print(x)
print(tf.math.argmax(x))tf.Tensor( [[ 2 20 30 3 6] [ 3 11 16 1 8] [14 45 23 5 27]], shape=(3, 5), dtype=int32) tf.Tensor([2 2 0 2 2], shape=(5,), dtype=int64)
In [16]:
x = tf.constant([[[2, 20, 30, 3, 6], [3, 11, 16, 1, 8]],[[1,1,1,1,1],
[14, 45, 23, 5, 27]]])
print(x)
print(tf.math.argmax(x))tf.Tensor( [[[ 2 20 30 3 6] [ 3 11 16 1 8]] [[ 1 1 1 1 1] [14 45 23 5 27]]], shape=(2, 2, 5), dtype=int32) tf.Tensor( [[0 0 0 0 0] [1 1 1 1 1]], shape=(2, 5), dtype=int64)
Define the axis of the application.
In [25]:
x = tf.constant([[2, 20, 30, 3, 6], [3, 11, 16, 1, 8],
[14, 45, 23, 5, 27]])
print(x)
print(tf.math.argmax(x,0)) # fine across the 0-axis which is the defaulttf.Tensor( [[ 2 20 30 3 6] [ 3 11 16 1 8] [14 45 23 5 27]], shape=(3, 5), dtype=int32) tf.Tensor([2 2 0 2 2], shape=(5,), dtype=int64)
In [23]:
x = tf.constant([[2, 20, 30, 3, 6], [3, 11, 16, 1, 8],
[14, 45, 23, 5, 27]])
print(x)
print(tf.math.argmax(x,1)) # fine across the 1-axis i.e., across each 1D tensortf.Tensor( [[ 2 20 30 3 6] [ 3 11 16 1 8] [14 45 23 5 27]], shape=(3, 5), dtype=int32) tf.Tensor([2 2 1], shape=(3,), dtype=int64)
In [29]:
x = tf.constant([[[2, 20, 30, 3, 6], [3, 11, 16, 1, 8]],[[1,1,1,1,1],
[14, 45, 23, 5, 27]]])
print(x)
print(tf.math.argmax(x,1)) # across the 2D tensorstf.Tensor( [[[ 2 20 30 3 6] [ 3 11 16 1 8]] [[ 1 1 1 1 1] [14 45 23 5 27]]], shape=(2, 2, 5), dtype=int32) tf.Tensor( [[1 0 0 0 1] [1 1 1 1 1]], shape=(2, 5), dtype=int64)
In [30]:
x = tf.constant([[[2, 20, 30, 3, 6], [3, 11, 16, 1, 8]],[[1,1,1,1,1],
[14, 45, 23, 5, 27]]])
print(x)
print(tf.math.argmax(x,2)) # across the 1D tensorstf.Tensor( [[[ 2 20 30 3 6] [ 3 11 16 1 8]] [[ 1 1 1 1 1] [14 45 23 5 27]]], shape=(2, 2, 5), dtype=int32) tf.Tensor( [[2 2] [0 1]], shape=(2, 2), dtype=int64)
return the maximum element
In [11]:
x = tf.constant([[9,2,10,4],[5,6,7,50]])
print(tf.reduce_max(x))tf.Tensor(50, shape=(), dtype=int32)
In [ ]: