calibration

class fgvc.special.calibration.ModelWithTemperature(model: Module)

A model wrapper that adds classifier calibration - temperature scaling method.

Paper: https://arxiv.org/abs/1706.04599 GitHub repository: https://github.com/gpleiss/temperature_scaling

forward(x: Tensor, *args, **kwargs) → Tensor

Apply forward pass on the model and scale output logits.

fgvc.special.calibration.get_temperature(logits: ndarray, targs: ndarray, device: device | None = None) → float

Tune the temperature parameter after training the model.

Paper: https://arxiv.org/abs/1706.04599 GitHub repository: https://github.com/gpleiss/temperature_scaling

The authors tune the temperature using the validation set. Tuning temperature is done using Cross Entropy loss (Softmax + NLL).

Parameters:

• logits – Numpy array with classifier raw predictions (before softmax).

• targs – Numpy array with ground-truth targets.

• device – Device to use (cpu,0,1,2,…).

Returns:

A scalar value for scaling model logits.

Return type:

temperature