Advanced level · Module 4: Deploying Deep Learning Models · Lesson 2 of 4
Compressing Models
Make models smaller and faster with quantisation, pruning and distillation while keeping accuracy.
In this lesson you will
- Apply post-training quantisation
- Describe pruning and knowledge distillation
Key concepts
quantisation (8-bit)accuracy droppruningknowledge distillation
The lesson
Quantisation stores weights and calculations in 8-bit integers instead of 32-bit floats, making models about four times smaller and often faster, with a small accuracy drop.
Pruning removes weights or channels that contribute little. Knowledge distillation trains a small 'student' model to imitate a large 'teacher', keeping much of its accuracy.
Measure size, speed and accuracy after each step; stop when the targets are met or accuracy drops too far.
Step by step
- Quantise the exported model.
- Measure size, speed and accuracy.
- Compare with the original.
- Decide whether to try distillation.
Practice activity
Compress the project model to meet the size and speed targets.
Expected result: A compressed model with measured trade-offs.
Check your understanding
-
How much smaller does 8-bit quantisation make a 32-bit model, roughly?
Show answer
About four times smaller.
-
What is knowledge distillation?
Show answer
Training a small model to imitate a larger one.
Taught live by an instructor — online or in person. Free course lessons are for self-study and do not lead to a certificate — see registration options.