Indexing

Tensor elements are accessed by index.

Indexing

Indexing and slicing work as in NumPy; in addition there are a few functions often used in batch work:

x = torch.tensor([[1, 2, 3], [4, 5, 6], [7, 8, 9]])
print(x[1].tolist())          # [4, 5, 6]
print(x[:, -1].tolist())      # [3, 6, 9]
print(x[x > 4].tolist())      # [5, 6, 7, 8, 9]  boolean mask
print(x[[0, 2]].tolist())     # [[1, 2, 3], [7, 8, 9]]

In classification the predicted class is taken with cikti.argmax(dim=1) and the top k probabilities with torch.topk.

gather and the correct class's score

Taking the correct class's probability for every example is common in loss computations. gather picks the given column index from each row; fancy indexing does the same:

p = torch.tensor([[0.7, 0.2, 0.1], [0.1, 0.1, 0.8]])
y = torch.tensor([0, 2])
print(p.gather(1, y.unsqueeze(1)).squeeze(1).tolist())   # [0.7..., 0.8...]
print(p[torch.arange(2), y].tolist())                   # the same

Using round when printing floating-point values makes the output readable.

Exercise Computing accuracy

Get the predicted classes from the given logits with argmax(dim=1) and print them with tolist(). Print the accuracy with (tahmin == y).float().mean().item(), then the correct classes' logits with logit[torch.arange(len(y)), y].tolist().

Show example solution

import torch

logit = torch.tensor([[2., 1., 0.], [0., 3., 1.], [1., 0., 4.], [5., 1., 2.]])
y = torch.tensor([0, 1, 1, 0])

tahmin = logit.argmax(dim=1)
print(tahmin.tolist())
print((tahmin == y).float().mean().item())
print(logit[torch.arange(len(y)), y].tolist())

Test yourself

For <code>x</code> above, what does <code>x.argmax(dim=1).tolist()</code> give?
- [0, 1, 2]
- [2, 5, 8]
- 8
- [2, 2, 2]
Show answer

[2, 2, 2] — In every row the largest value is in the last column (index 2).

Get the index of each row's largest value.

tahmin = cikti.____(dim=1)

Show answer

tahmin = cikti.argmax(dim=1)

Match the expressions.

x[:, 0]

First column
x[x > 0]

Selection by mask
argmax(dim=1)

Max per row
gather

Indexed pick per row

Next lesson: Summary - Tensor operations

We Are Enhancing Your Experience

Cookie Preferences

Strictly Necessary Cookies

Analytical Cookies