Temperature Scaling

Implement Temperature Scaling from “On Calibration of Modern Neural Networks” (Guo et al., 2017).

Temperature scaling is a simple post-hoc calibration method. It divides logits by a learned temperature T before applying softmax:

$$\hat{q}_i = \text{softmax}(z_i / T)$$

• T > 1: softer (more uniform) distribution — reduces overconfidence
• T < 1: sharper (more peaked) distribution
• T = 1: no change

Input:

• logits: shape (batch, n_classes) — raw model outputs
• temperature: float T > 0

Output: Tensor of shape (batch, n_classes) — calibrated probabilities.