calibration
To begin with
To lower a model’s precision without damaging it too much, you need to know which values it really handles. Calibration means watching it work before choosing.
The facts
Some quantization methods run a little representative data through the model and observe the resulting values, such as the range of activations. They use them to choose scales or rounding. The model is not trained for a new task: it is observed so it can be approximated better. Quality is then checked on the target hardware.
Key takeaway
Calibration observes; it teaches nothing. It is not fine-tuning.
The stories about it
The notebook opens soon.The first stories are on their way.