Summary - PyTorch

In this group you learned tensor creation, properties, and GPU support.

Summary: PyTorch

In this group:

PyTorch: tensors, autograd and nn; code runs line by line.
Tensor: an n-dimensional array; back to Python with item() and tolist().
from_numpy shares memory, torch.tensor copies.
Creation: zeros, ones, arange, rand, randn; manual_seed.
shape, dtype (input float32, labels int64), device; .to(device) and .cpu().

Exercise Prepare data as tensors

Convert the NumPy arrays X_np (float64) and y_np into tensors: X should be float32 and y int64. Print both dtype values on one line, then X.shape.

Show example solution

import numpy as np
import torch

X_np = np.random.default_rng(0).normal(size=(8, 3))
y_np = np.array([0, 1, 1, 0, 2, 2, 1, 0])

X = torch.from_numpy(X_np).float()
y = torch.from_numpy(y_np).long()
print(X.dtype, y.dtype)
print(X.shape)

Test yourself

What should be done to float64 features from NumPy before giving them to the model?
- Nothing
- Convert to a list
- Convert to float32 with .float()
- Convert to integers with .long()
Show answer

Convert to float32 with .float() — Model weights are float32; mismatched types raise an error.

Make the labels int64.

y = torch.from_numpy(y_np).____()

Show answer

y = torch.from_numpy(y_np).long()

Match the tools.

torch.from_numpy

Shared conversion
torch.manual_seed

Seed
.float()

float32
.cpu().numpy()

Back to NumPy

Next lesson: Tensor operations

We Are Enhancing Your Experience

Cookie Preferences

Strictly Necessary Cookies

Analytical Cookies