Creating tensors

A tensor is created from a list with torch.tensor().

Creating tensors

Most NumPy functions have a PyTorch counterpart:

torch.zeros(2, 3)              # zeros
torch.ones(4)                  # ones
torch.full((2, 2), 7.0)        # a constant value
torch.arange(0, 10, 2)         # tensor([0, 2, 4, 6, 8])
torch.linspace(0, 1, 5)        # 5 evenly spaced values
torch.eye(3)                   # identity matrix

Sizes can be given as separate arguments (zeros(2, 3)) or as a tuple (zeros((2, 3))).

Random tensors and the seed

Random tensors are used for weight initialisation and shuffling data. torch.manual_seed makes the results reproducible:

torch.manual_seed(0)
a = torch.rand(3)        # uniform on [0, 1)
b = torch.randn(2, 2)    # standard normal
torch.manual_seed(0)
print(torch.equal(a, torch.rand(3)))   # True: same seed, same numbers

The _like versions copy another tensor's shape and type: torch.zeros_like(x).

Exercise Tensor factory

Turn torch.arange(6) into a 2×3 matrix and print it with tolist(). Print the sum of torch.full((2, 2), 5) with item(). Then print with torch.equal that two torch.rand(4) tensors produced with seed 0 are equal.

Show example solution

import torch

print(torch.arange(6).reshape(2, 3).tolist())
print(torch.full((2, 2), 5).sum().item())
torch.manual_seed(0)
a = torch.rand(4)
torch.manual_seed(0)
print(torch.equal(a, torch.rand(4)))

Test yourself

What does <code>torch.arange(1, 8, 3).tolist()</code> give?
- [3, 6]
- [1, 4, 7]
- [1, 4]
- [1, 3, 5, 7]
Show answer

[1, 4, 7] — The end (8) is excluded; it starts at 1 and steps by three.
Produce a zero tensor with the same shape as another tensor.
```
z = torch.____(x)
```
Show answer
```
z = torch.zeros_like(x)
```
Match the functions.

torch.rand

Uniform [0, 1)
torch.randn

Standard normal
torch.eye

Identity matrix
torch.manual_seed

Reproducibility

Next lesson: Tensor properties

We Are Enhancing Your Experience

Cookie Preferences

Strictly Necessary Cookies

Analytical Cookies