Course
Deep Learning with PyTorch
From a single neuron to transformers: build, train and debug neural networks in PyTorch, each lesson with code that runs on a laptop.
- Lessons
- 14
- Reading time
- 3 h 25 min
- Level
- Intermediate
- Taught in
- English
- Format
- Articles with runnable code
Course
What you will learn
Fourteen article lessons in four parts. Foundations: tensors and autograd, artificial neural networks, and training them well with optimisers, dropout and batch normalisation. Vision: convolutional networks, classic architectures and transfer learning. Sequences and language: recurrent networks, LSTM and GRU, word embeddings, attention and the transformer. Generative models: autoencoders, GANs and neural style transfer. Every example is small enough to train on a CPU.
- Foundations
- Computer vision
- Sequences and language
- Generative models
Before you start
A computer that runs Python. Each lesson says what it builds on.
Syllabus
Syllabus
Every lesson explains one idea, then runs it in code and shows the output. The first lesson is open to everyone.
01 · Foundations
- 01Tensors and autogradPyTorch tensors, shapes and devices, and automatic differentiation: how gradients are computed and a model learns by gradient descent.Free preview15 min
- 02Artificial neural networksFrom a neuron to a multilayer perceptron: layers, activations, the forward pass, backpropagation and the training loop.14 min
- 03Training neural networks wellLosses, SGD, momentum and Adam, learning-rate schedules, initialisation, dropout, batch normalisation, weight decay and early stopping.15 min
02 · Computer vision
- 04Convolutional neural networksConvolutions, filters, padding, stride and pooling: why CNNs see images better than dense layers, built from scratch on small images.13 min
- 05CNN architecturesLeNet, AlexNet, VGG, Inception and ResNet: the ideas each one added, residual connections, and data augmentation.14 min
- 06Transfer learningReuse a network trained on millions of images: feature extraction, fine-tuning, freezing layers and adapting to a small dataset.15 min
03 · Sequences and language
- 07Recurrent neural networksProcess sequences step by step with a hidden state, unroll through time, and see why plain RNNs forget.15 min
- 08LSTM and GRUGates that decide what to keep and forget: LSTM and GRU cells, stacking and bidirectional layers, and forecasting a time series.15 min
- 09Word embeddingsTurn words into vectors: tokenisation, one-hot versus dense embeddings, word2vec, GloVe, similarity and analogies.15 min
- 10Sequence-to-sequence and attentionEncoder–decoder models for translation, the bottleneck they hit, and attention that lets the decoder look back at every input.14 min
- 11The transformerSelf-attention, multi-head attention, positional encoding and the transformer block, built in PyTorch, then BERT and GPT in context.15 min
04 · Generative models
- 12AutoencodersCompress and reconstruct: undercomplete, denoising and variational autoencoders, the latent space, and anomaly detection.15 min
- 13Generative adversarial networksA generator and a discriminator in a contest: the GAN objective, DCGAN, training instability and mode collapse.15 min
- 14Neural style transferPaint one image in the style of another: content and style losses from a pretrained CNN, Gram matrices, and optimising the pixels.15 min
Feedback
Help us improve this course
Suggest a lesson, point out something unclear, or report a mistake in the code. A person reads every message and replies by email.