Deep Learning
Tensors to transformers, and making one cheap enough to ship.
Read Week One Free Download the 499-page workbookWhy Deep Learning?
Neural networks are the default answer now. The valuable skill is knowing when they are the right one.
Almost every deep learning tutorial ends where the interesting part begins: the model trains, the loss goes down, and nothing is said about whether it beat a simpler method, whether the improvement survives a change of random seed, or what it costs to run. This course builds attention, transformers, GANs and diffusion from the components up, and then measures all of it, including the times the measurement was unflattering.
What Makes This Course Different?
This course publishes the results that did not go the way the textbook says. Every number below came from code run on a real machine, and is printed on the lesson page next to the technique it undercuts.
What This Programme Delivers
Skills You Will Build
Technical Skills
- Tensors, shapes, autograd and the five-step training loop
- nn.Module architectures that are not a straight line
- Momentum, Adam, schedules, clipping and batch size
- Convolutional networks, ResNets and transfer learning
- RNN, LSTM and GRU with padding handled properly
- Scaled dot-product attention built by hand, then multi-head
- Transformer encoders and decoders, and a character-level language model
- Frozen features, fine-tuning and LoRA
- Autoencoders, VAEs, GANs and DDPM diffusion
- Contrastive self-supervision and linear probing
- Quantisation, pruning, distillation, ONNX export and serving
Judgement
- Recognising the shape of a failing training run from its curve
- Telling a real improvement from seed variance
- Knowing when transfer beats training from scratch
- Choosing between frozen features, fine-tuning and LoRA on cost
- Measuring generative models rather than eyeballing samples
- Deciding when a smaller model is the correct answer
PyTorch First, Keras Alongside
Week 15 writes the same models twice and measures both, rather than arguing about which framework is better.
| Aspect | PyTorch (main framework) | Keras (written alongside) |
|---|---|---|
| Training loop | Written by hand, every step visible | fit() handles it |
| Best for | Understanding what the library does | Getting a standard model running fast |
| Debugging | Step into any tensor at any point | Callbacks and inspection hooks |
| Custom layers | Ordinary Python classes | Subclassing with more ceremony |
| Research code | What most papers ship | Less common |
| Course role | Weeks 1 to 14 and 16 to 18 | Week 15, then used where it wins |
| Export | ONNX, measured at four times faster | SavedModel and TFLite |
18-Week Learning Roadmap
View Detailed Week-by-Week Curriculum →
Foundations
Weeks 1 to 4. Tensors and automatic differentiation, nn.Module, optimisation and the failure shapes worth recognising, then regularisation.
Vision
Weeks 5 and 6. Convolution built up from one kernel to a full CIFAR-10 model, why depth stopped working, and the residual connection that fixed it.
Sequences
Week 7. Carrying state across a sequence with RNN, LSTM and GRU, padding done properly, and next-element prediction.
Transformers
Weeks 8 to 11. Attention from scratch, the full transformer, a character-level language model trained end to end, then fine-tuning and LoRA.
Generative and Self-Supervised
Weeks 12 to 14. Autoencoders and VAEs, GANs and diffusion measured against each other, and training an encoder with no labels at all.
Production and Capstone
Weeks 15 to 18. Keras compared directly, making a model cheap enough to ship, deployment and monitoring, then a capstone on untouched data.
What You Build Along The Way
CIFAR-10 From One Kernel Up
Convolution implemented by hand, then built into a network that actually classifies, with each architectural change measured separately.
PyTorchA Transformer That Generates Text
Attention written from scratch, stacked into a decoder, and trained into a character-level language model with the baselines it has to beat.
From scratchA Diffusion Sampler
A DDPM built end to end, then compared with a GAN on mode coverage. The diffusion model scored slightly lower, and the lesson keeps that result.
DDPMAn Exported Model, Timed
Quantisation, pruning and distillation each measured, then ONNX export, which is the one that actually ran four times faster.
ONNXWhat The Capstone Asks Of You
Framing
Data the course has not touched
No CIFAR-10, no course dataset. Something you found, with the messiness that comes with it.
A stated non-neural baseline
Name what you have to beat before you start, and measure it.
Measurement
Seed variance before improvements
The capstone measures run-to-run spread first, at 0.0147, which turned out wider than five of six proposed improvements.
Cost as a reported number
Parameters, latency and memory alongside accuracy, because a model nobody can afford to run has not shipped.
Handover
A repository that reproduces
Installs, runs and reproduces your numbers on a machine that is not yours.
An honest write-up
What worked, what did not survive a second seed, and what you would try next.
Finish With
Where This Material Leads
The roles this course is written for:
Written for people who will be asked why the model costs what it costs, not only whether it works.
Tools & Technologies
Core Tools
Also Covered
Every tool used is free and open source. Nothing on this course needs a paid licence.
Who This Is For
ML Practitioners
Comfortable with scikit-learn, moving to neural networks
Python Developers
Confident with classes and functions
Anyone Tired Of Black Boxes
Attention gets built by hand before it gets imported
Assumes Python and some machine learning. Start with the Machine Learning course if that is not you yet.
Certificate of Completion
Certificate of Completion in Deep Learning
Issued by EDUSHARK TRAINING once you finish the eighteen weeks and submit your capstone. It records what you built, not a grade.
Start With Tensors
Eighteen weeks from a single tensor to a model small enough to deploy.