Summary - Tensor operations

In this group you learned tensor shape, dtype, and reshape operations.

Summary: tensor operations

In this group:

Element-wise operations, broadcasting, @; choosing axes with dim; in-place operations with _.
Tracking shape, unsqueeze/squeeze.
Types: float32 (default), float16/bfloat16 (half), int64 (labels), bool (masks); memory = element size × count.
view/reshape, permute, flatten(1).
Indexing, argmax, gather, computing accuracy.

Exercise Batch processing

Apply a per-row softmax to the (5, 3) logits (torch.softmax(logit, dim=1)) and print the row sums as a list with round(..., 3). Then get each row's most likely class and its probability with max(dim=1); print the classes with tolist().

Show example solution

import torch

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

p = torch.softmax(logit, dim=1)
print([round(v, 3) for v in p.sum(dim=1).tolist()])
olasilik, sinif = p.max(dim=1)
print(sinif.tolist())

Test yourself

Which operation is needed before giving a CNN output (32, 64, 7, 7) to a linear layer?
- unsqueeze(0)
- None
- Flattening to (32, 3136) with flatten(1)
- permute
Show answer

Flattening to (32, 3136) with flatten(1) — A linear layer wants (example, feature); 64 × 7 × 7 = 3136.
Flatten all axes except the batch dimension.
```
d = x.____(1)
```
Show answer
```
d = x.flatten(1)
```
Match the functions.

unsqueeze

Adds an axis
permute

Reorders axes
flatten

Flattens
gather

Picks by index

Next lesson: What is autograd?

We Are Enhancing Your Experience

Cookie Preferences

Strictly Necessary Cookies

Analytical Cookies