PyTorch

What is it?

PyTorch is an open-source deep learning framework for Python. Its core unit is the tensor: a multi-dimensional array much like NumPy's, but with two superpowers — it runs on the GPU and ships with autograd (automatic differentiation). On top of that it offers ready-made building blocks (torch.nn, optimizers, loss functions) to construct and train neural networks.

What is it for?

• Computer vision. Image classification, object detection, segmentation — with convolutional networks and pretrained models.
• NLP and LLMs. It is the reference framework for language processing and for training or fine-tuning large language models; most of the ecosystem (e.g. Hugging Face) is built on top of it.
• Custom models. Any architecture of your own that needs gradients: autograd handles the derivatives, you define the model and the loss.
• Research and production. Its dynamic API makes it comfortable to experiment with, while tools like TorchScript and torch.compile take it to production.

When to use it / when not to?

Use it for deep learning on unstructured data — images, text, audio, complex sequences — where neural networks add real value, and when you want to leverage the GPU to train at scale.

Think twice for classic ML on tabular data: there scikit-learn is simpler, faster to get going, and almost always enough. If a linear model or a decision tree solves your problem, don't take it to PyTorch — you'd add complexity without gaining accuracy. The honest boundary is this: scikit-learn for tabular and baselines; PyTorch once you step into deep learning.

Start in 1 minute

pip install torch

import torch

# A tensor that tracks its gradient
x = torch.tensor([2.0], requires_grad=True)

# Any operation: y = x^2 + 3x
y = x ** 2 + 3 * x

# autograd computes dy/dx automatically
y.backward()

print(x.grad)  # dy/dx = 2x + 3 = 7.0 at x=2

x.grad shows the gradient PyTorch worked out on its own: that same mechanism, applied to millions of parameters, is what trains a neural network. The natural next step is to build a model with torch.nn.Module and train it with an optimizer from torch.optim. The official documentation includes tutorials for every case.