DL05 · Transformers
A free, illustrated L3 course of 11 modules in the Deep Learning track. Modules include: Self-attention from scratch; Multi-head attention; Positional information.
DL05 L3 · Advanced 11 modules · about 44 lessons Planned · Ring 2
Part of the Deep Learning track. Ships in Ring 2 (Spine) of the roadmap.
Modules
- Self-attention from scratch
- Multi-head attention
- Positional information
- The transformer block
- Encoder, decoder and encoder-decoder variants
- Tokenization with byte-pair encoding
- Pretraining objectives
- Vision transformers
- Efficient attention and long context
- Mixture of experts and other variants
- Build and train a small GPT
Signature illustrations
The pictures a learner should still see in their head a year later:
Query, key and value matching, attention-matrix heatmaps, tensor shapes flowing through one block, the causal mask, positional-encoding stripes.
Status
This course is planned, not written yet. It ships with Ring 2 (Spine).
Every lesson will follow the same twelve-part blueprint with at least three purpose-made figures. Watch the repository to hear when modules ship, or help write and review it.