Deep learning with PyTorch
Build neural networks that train properly, from tensors and autograd to image classifiers.
Created by Lerato Dlamini, Tshega Mampshika
What you'll learn
- Work with tensors, shapes and broadcasting without guessing
- Explain how autograd computes gradients for a training step
- Build and train feed-forward and convolutional networks
- Diagnose overfitting and apply regularisation that works
- Use transfer learning to train an image classifier on a small dataset
- Export a trained model for inference
Skills you'll gain
- PyTorch
- Neural networks
- Convolutional networks
- Model debugging
- Transfer learning
- Python
Part of these pathways
Course content
4 sections · 12 lectures · 4.1 total hours · ≈ 1.6 GB to download
Tensors and autograd
The foundations that every PyTorch model rests on.
- Tensors, shapes and broadcasting Preview 18 min · 135 MB
- Autograd and gradients explained Locked until you enrol 22 min · 165 MB
- Your first training loop Locked until you enrol 24 min · 180 MB
Neural networks
Building and training networks that generalise.
- Layers, activations and loss functions Locked until you enrol 20 min · 150 MB
- Overfitting, dropout and regularisation Locked until you enrol 22 min · 165 MB
- Optimisers and learning rate schedules Locked until you enrol 18 min · 135 MB
Convolutional networks for images
Image models that recognise patterns in pixels.
- Convolutions and feature maps Locked until you enrol 20 min · 150 MB
- Training an image classifier Locked until you enrol 25 min · 188 MB
- Data augmentation and transfer learning Locked until you enrol 22 min · 165 MB
Sequences and deployment
Models for ordered data and getting a model into use.
- Recurrent networks and sequence data Locked until you enrol 18 min · 135 MB
- Exporting a model for inference Locked until you enrol 14 min · < 1 MB
- Capstone: a defect detector for a production line Locked until you enrol 25 min · < 1 MB
Requirements
- Completed Machine learning with scikit-learn, or equivalent
- Confident Python including classes and functions
- Access to a GPU through Google Colab or a local card
Description
This course takes you from the mathematics of training a model to a working image classifier in PyTorch. You start with tensors and automatic differentiation, then build networks layer by layer and learn how to diagnose the usual failures: overfitting, vanishing gradients and training runs that never converge.
It is an advanced course. Expect to spend more time than the stated hours on the convolutional networks module, and be comfortable with Python and basic machine learning first. The capstone is a defect detector trained on a set of product photos, with a report on its errors and a plan for deployment.
Instructors
Lerato Dlamini
Senior software engineer and AI trainer
- 9 courses
- 193 learners
Lerato leads the AiX software and AI courses. Lerato has built systems for public-sector clients and teaches the testing and code review habits that the courses depend on, with a focus on code that the next person can maintain.
Tshega Mampshika
Co-founder and CTO, AlgoAtWork
- 14 courses
- 257 learners
Tshega leads technology at AlgoAtWork, the company behind AiX Academy. Tshega leads the robotics, electronics and hardware courses and reviews the design of each pathway.
3.9 course rating · 8 ratings
-
Bongani Dube
I would like more on deploying models, but the foundations are strong.
-
Lindiwe Sithole
The image classifier project took most of a weekend on a laptop without a GPU. Be ready to use Colab.
-
Thabo Mokoena
Heavy on maths, but the lessons give the intuition before the equations, which helped.
-
Aisha Hendricks
Good coverage of overfitting and dropout, with real training curves. Clear and no padding.
-
Johan Botha
I stopped after module three. The jump from basic networks to transformers came with little warning.
-
Zanele Nkosi
The GPU setup section did not say which CUDA version to pick, and I lost an evening on it.
-
Nomsa Khumalo
Tensors, autograd and a training loop that actually trains. This is the course I needed after the scikit-learn one.
-
Pieter van der Merwe
Advanced in the true sense. Budget double the stated hours for the convolutional networks module.
Frequently asked questions
Do I need a GPU?
No, but training is faster with one. The course shows how to run every notebook on free Colab hardware.
How much linear algebra is needed?
Vectors, matrices and the idea of a derivative. The first module revisits these with code.
Is the capstone assessed?
Yes. The AiX team reviews your notebook and the error report, and the certificate depends on that review.