Jupyter Notebook

Tensors Description

We often use $\mathbf{X}$ to denote the feature tensor of the samples. The values in a tensor object is not mutable. The following table shows some conventional ways to structure different feature [1]. In this note, we specifically focus on $2$-dimensional tensor for simplicity.

Conventional Structure of Features

Tensors Declaration

Identity matrix

tf.eye(dim) API

Create an identity matrix

> tf.eye(4)
<tf.Tensor: shape=(4, 4), dtype=float32, numpy=
array([[1., 0., 0., 0.],
       [0., 1., 0., 0.],
       [0., 0., 1., 0.],
       [0., 0., 0., 1.]], dtype=float32)>

Tensor of ones

tf.ones(shape) API

Create a tensor full of ones with the given shape

> tf.ones((4, 5))
<tf.Tensor: shape=(4, 5), dtype=float32, numpy=
array([[1., 1., 1., 1., 1.],
       [1., 1., 1., 1., 1.],
       [1., 1., 1., 1., 1.],
       [1., 1., 1., 1., 1.]], dtype=float32)>

Tensor of zeros

tf.zeros(shape) API

Create a tensor full of zeros with the given shape

> tf.zeros((4, 5))
<tf.Tensor: shape=(4, 5), dtype=float32, numpy=
array([[0., 0., 0., 0., 0.],
       [0., 0., 0., 0., 0.],
       [0., 0., 0., 0., 0.],
       [0., 0., 0., 0., 0.]], dtype=float32)>

Tensor of ones

tf.ones_like(tensor) API

Create a tensor full of ones with same shape as the given tensor