Tensor properties

A tensor has shape, dtype, and device properties.

shape, dtype, device

Every tensor has three basic properties:

x = torch.zeros(2, 3)
print(x.shape)     # torch.Size([2, 3])
print(x.dtype)     # torch.float32
print(x.device)    # cpu
print(x.numel())   # 6 elements

In deep learning the default type is float32; a tensor created from a list of integers is int64. Class labels should be int64 (torch.long), inputs float32.

Type conversion

To change the type, .to(dtype) or its shortcuts (.float(), .long()) are used. Operating on different types sometimes gives an error, sometimes a silent type promotion:

a = torch.tensor([1, 2, 3])
b = a.float()                 # float32
c = a.to(torch.float64)
print((a / 2).dtype)          # torch.float32: division produces decimals
print((a // 2).tolist())      # [0, 1, 1]

Decimal arrays coming from NumPy are float64; before giving them to a model they must be converted to float32 with .float().

Exercise Types

For a = torch.tensor([1, 2, 3]), print a.dtype, (a / 2).dtype and a.float().dtype on separate lines. Then print the shape and numel() of torch.zeros(3, 4, 5) on one line.

Show example solution

import torch

a = torch.tensor([1, 2, 3])
print(a.dtype)
print((a / 2).dtype)
print(a.float().dtype)
z = torch.zeros(3, 4, 5)
print(z.shape, z.numel())

Test yourself

What is <code>torch.tensor([1, 2, 3]).dtype</code>?
- torch.int64
- torch.float32
- torch.int32
- torch.float64
Show answer

torch.int64 — Integer lists produce an int64 tensor.
Convert the tensor to float32.
```
x = a.____()
```
Show answer
```
x = a.float()
```
Match the properties.

shape

Dimensions
dtype

Element type
device

cpu or cuda
numel()

Element count

Next lesson: GPU support

We Are Enhancing Your Experience

Cookie Preferences

Strictly Necessary Cookies

Analytical Cookies