Stretch: Class Imbalance Handling

The problem, from first principles

When one label is scarce, an unweighted average can let a classifier look good by prioritizing the frequent class. Per-class weights make mistakes on selected labels count more without changing which prediction is correct.

From theory to code

Implement weighted_bce_loss(p, y, class_weights): use the same clipped BCE terms as 02-bce-loss, then multiply each sample by the weight selected from its actual class.

Constraints

• p and binary y have matching sample shapes.
• class_weights provides keys 0 and 1.
• Clip probabilities to [1e-12, 1 - 1e-12].
• Select class-1 weight where y == 1, otherwise class-0 weight.
• Return the negative mean as a Python float without loops.

Hints

Open one at a time. Each gives away a little more than the last.