The best way to make the modules stick is to build something. These five projects take you from a neural network and backpropagation written from scratch to a CNN image classifier, transfer learning, a character-level language model, and a transformer built from scratch.
Each one is a code-along video, so you can pause, build it yourself and adapt it to your own data. Pick the project that matches the module you've just finished and come back for the rest.
Andrej Karpathy builds micrograd, a tiny automatic-differentiation engine, from an empty file, then uses it to train a small neural network and see backpropagation work step by step.
Patrick Loeber builds and trains a convolutional neural network in PyTorch, then evaluates it on images it has not seen.
Patrick Loeber starts from a pretrained image model and adapts it to a new, small dataset, showing why transfer learning saves so much training time.
Andrej Karpathy builds a language model that learns to generate new names one character at a time, starting from simple counts and moving to a neural network.
Umar Jamil codes a full transformer in PyTorch, including attention, training and inference, with each part explained as it is written.