Deep learning and PyTorch

What happens inside a network while it trains, with a loop you write yourself first.

TensorsShape, dtype and device, and why shape is the thing you will print most in your life.
Autogradrequires_grad, the computation graph and backward, step by step.
A training loop by handForward, loss, backward, optimiser step, with no Trainer and no shortcut.
Layers and normalisationLinear, activations, dropout and normalisation, and why each one works.
Dataset and DataLoaderBatches, shuffling, parallel loading, and what batch size does to training.
Debugging trainingA loss that will not fall, a loss that goes NaN, a model that memorises: three faults and their cures.
Making it fastMixed precision and torch.compile, and when the complexity is worth it.
Project: a network from zeroA loop you wrote end to end, and three faults we plant in the code for you to find.