quantization - AshokBhat/ml GitHub Wiki

• Reduce the precision of the numbers used to represent a model's parameters - FP32 by default

• Pros - Smaller model size and faster computation.
• Cons - Not trivial

• Post-training FP16 quantization
• Post-training dynamic range quantization
• Post-training integer quantization
• Quantization-aware training

• What is quantization?
• What are the downsides?
• When is it used?
• What support do various frameworks have for quantization?

• Training