Tensor

A tensor is PyTorch's core data structure, a multi-dimensional array.

The tensor

The tensor is PyTorch's core data structure: an n-dimensional array of numbers of the same type. It is named by its number of dimensions: scalar (0), vector (1), matrix (2), beyond that a tensor. A batch of images is typically 4-dimensional: (example, channel, height, width).

s = torch.tensor(3.0)               # 0 dimensions
v = torch.tensor([1.0, 2.0])        # 1 dimension
m = torch.tensor([[1, 2], [3, 4]])  # 2 dimensions
print(s.dim(), v.dim(), m.dim())    # 0 1 2

Element-wise operations, broadcasting and indexing work as in NumPy.

The bridge to NumPy

torch.from_numpy and .numpy() move data between the two worlds without copying; because they share the same memory, a change in one shows up in the other:

a = np.array([1, 2, 3])
t = torch.from_numpy(a)
a[0] = 100
print(t.tolist())       # [100, 2, 3]  shared memory
k = torch.tensor(a)     # a copy: independent

If an independent copy is wanted, torch.tensor(a) or t.clone() is used.

Exercise Shared memory

From the array a, create t1 with torch.from_numpy and t2 with torch.tensor. Then set a[1] = 50; print t1.tolist() and t2.tolist() on separate lines.

Show example solution

import numpy as np
import torch

a = np.array([1, 2, 3])

t1 = torch.from_numpy(a)
t2 = torch.tensor(a)
a[1] = 50
print(t1.tolist())
print(t2.tolist())

Test yourself

How many dimensions does a batch of 32 colour (3-channel) 64×64 images have as a tensor?
- 2
- 5
- 4
- 3
Show answer

4 — The shape is (32, 3, 64, 64): example, channel, height, width.
Turn the NumPy array into a tensor without copying.
```
t = torch.____(a)
```
Show answer
```
t = torch.from_numpy(a)
```
Match the dimensions.

scalar

0 dimensions
vector

1 dimension
matrix

2 dimensions
image batch

4 dimensions

Next lesson: Creating tensors

We Are Enhancing Your Experience

Cookie Preferences

Strictly Necessary Cookies

Analytical Cookies