PyTorch learning hub

Learn or revisit PyTorch through connected explanations, visual tools, and complete runnable projects. The library is organized by guides, not by progress or course status.

Start PyTorch Fundamentals Open the project gallery

Start here

PyTorch projects become easier to read when every part has one responsibility:

data → batch → model → logits → loss → gradients → parameter update
                            ↓
                   validation and error inspection

Begin with the two long Fundamentals guides:

1. Core workflow explains tensors, Dataset, DataLoader, models, logits, loss, autograd, optimization, and evaluation.
2. Vision and real data explains convolution, CNN shapes, robust image pipelines, generalization, regularization, and saving.

Each guide starts with a map and uses a table of contents for direct return visits. There are no parallel collections containing the same topic.

Choose a question

If you want to understand…	Open this section
What [32, 3, 224, 224] means	Tensors and shapes
How files become shuffled batches	Dataset, transforms, and DataLoader
Why models return logits	Models, activations, and logits
Where weights actually change	Loss, autograd, and optimizers
Why a CNN preserves image structure	Convolution and feature maps
Why training improves while validation worsens	Generalization and regularization
Why a run fails	Reference and troubleshooting

Learn from complete examples

The four projects are maintained programs, not isolated fragments. Each page explains the question, important code, evidence, an interactive check, how to run a small CPU version, the optional longer GPU mode, and the complete source.

Project	What it makes visible
Nonlinear regression	why an activation changes what a network can represent
EMNIST letter classifier	a complete image-classification workflow and CNN shape path
Robust image pipeline	validation, corrupt-file decisions, batches, and diagnostics
Nature CNN	reusable CNN blocks, regularization, curves, and saved artifacts

How this library grows

Every future PyTorch guide will use two substantial pages: one for its core mechanisms and one for applied workflows. A guide appears only when it contains useful material. Shared projects and the reference remain available without duplicating explanations.

For a short review outside the documentation site, open DaZu's PyTorch guides.