PyTorch Introduction & Tensors

What is PyTorch?#

PyTorch is the most widely used framework for deep learning research and a leading choice in industry. It gives you:

• Tensors — fast multi-dimensional arrays that run on CPUs and GPUs.
• Autograd — automatic computation of gradients for training.
• torch.nn — ready-made layers, loss functions and model building blocks.
• A huge ecosystem — Hugging Face Transformers, torchvision, PyTorch Lightning and most open-source LLM tooling are built on it.

PyTorch code is plain Python: models are classes, training is a for loop, and you can debug with print() — which is why students find it easier to learn than older graph-based frameworks.

Installing PyTorch#

pip install torch

For NVIDIA GPU support, use the install command for your CUDA version from the official PyTorch website. Check your setup:

import torch

print(torch.__version__)
print("GPU available:", torch.cuda.is_available())

No GPU? Everything in this tutorial runs on a CPU; free cloud notebooks provide a GPU when you need one.

Tensors: the core data structure#

A tensor is an n-dimensional array, like a NumPy array that can also live on a GPU and track gradients.

import torch

x = torch.tensor([[1.0, 2.0], [3.0, 4.0]])
print(x.shape, x.dtype)

torch.Size([2, 2]) torch.float32

Creating tensors#

torch.zeros(2, 3)            # 2×3 of zeros
torch.ones(2, 3)             # 2×3 of ones
torch.rand(2, 3)             # uniform random values in [0, 1)
torch.randn(2, 3)            # standard normal random values
torch.arange(0, 10, 2)       # tensor([0, 2, 4, 6, 8])
torch.linspace(0, 1, 5)      # 5 evenly spaced values from 0 to 1

Tensor operations#

x = torch.tensor([[1.0, 2.0], [3.0, 4.0]])

print(x @ x)          # matrix multiplication
print(x * 10)         # element-wise multiplication
print(x.reshape(4))   # change shape
print(x.T)            # transpose
print(x.mean().item())

tensor([[ 7., 10.],
        [15., 22.]])
tensor([[10., 20.],
        [30., 40.]])
tensor([1., 2., 3., 4.])
tensor([[1., 3.],
        [2., 4.]])
2.5

.item() converts a one-element tensor into a plain Python number.

Common shape tools: reshape/view (change shape), unsqueeze(dim) (add a dimension of size 1), squeeze() (remove size-1 dimensions), permute(...) (reorder dimensions) and torch.cat / torch.stack (combine tensors).

Moving tensors to the GPU#

device = "cuda" if torch.cuda.is_available() else "cpu"
x = x.to(device)
print(x.device)

Rule: all tensors in one operation must be on the same device. Mixing a CPU tensor with a GPU tensor raises an error — one of the most common beginner bugs.

NumPy interoperability#

import numpy as np

a = np.array([1, 2, 3])
t = torch.from_numpy(a)     # shares memory with the NumPy array
back = t.numpy()            # back to NumPy (CPU tensors only)

Use this to load data with NumPy or Pandas and train with PyTorch.

PyTorch vs TensorFlow#

Both are capable; for learning modern AI and working with open-source LLMs, PyTorch is the most practical first choice.

Common mistakes#

• Mixing devices — move the model and every batch to the same device.
• Integer tensors for training data — models expect float32; torch.tensor([1, 2]) creates an integer tensor.
• Shape mismatches — print .shape often; most errors are a missing batch dimension.
• Calling .numpy() on a GPU tensor — move it with .cpu() first.

Key takeaways#

• PyTorch = tensors + autograd + torch.nn, all in plain Python.
• Tensors are GPU-capable n-dimensional arrays with a shape, dtype and device.
• Keep data and models on the same device, and check shapes constantly.