Course book
Deep Learning with PyTorch
Seven illustrated chapters, one for each video. They explain the ideas in original words and work through shapes, gradients and losses by hand.
- Chapter 1The PyTorch WorkflowTensors and matrix multiplication rules, the five-step training loop with a known line worked by hand, non-linear classification, convolutional networks and custom image datasets.
- Chapter 2Tensors, Gradients and Linear RegressionNotebooks and tensors, NumPy interoperability, autograd checked against calculus, a linear regression gradient step worked by hand, and the same model with PyTorch built-ins.
- Chapter 3Classifying Digits with Logistic RegressionLoading MNIST with transforms, training and validation splits, a linear layer with softmax, cross-entropy worked by hand, why accuracy is not a loss, testing and saving.
- Chapter 4Feedforward Networks on GPUsHidden layers and ReLU, an exclusive-or network solved by hand, the universal approximation theorem with ReLU hinges, why GPUs help, and moving models and data to a GPU.
- Chapter 5Convolutional Networks for Colour ImagesCIFAR-10 and channel layout, convolution worked by hand, the output-size formula, a three-block network with every parameter counted, and reading curves for under- and overfitting.
- Chapter 6Training a State-of-the-Art Image ClassifierChannel normalisation and augmentation, residual blocks and batch normalisation worked by hand, the one-cycle schedule, weight decay and gradient clipping combined.
- Chapter 7Generative Adversarial NetworksGenerative modelling, the generator and discriminator, adversarial losses worked by hand, network shapes and transposed convolutions, the two-phase training loop and evaluation.