BinaryCrossEntropy — Binary Classification Loss¶
Compute binary cross-entropy for probability predictions.
Overview¶
Appropriate when model outputs Bernoulli probabilities, usually after sigmoid. Penalizes overconfident incorrect predictions heavily while rewarding accurate, calibrated probabilities. Includes numerical stability floor (eps=1e-12).
Standard binary classification objective using -mean(y_true · log(y_pred) + (1 - y_true) · log(1 - y_pred)). No gnostic scale S or θ coordinate applies.
Value range: [0, ∞).
Key Features¶
- Standard binary classification objective
- Numerical stability with
epsclipping - MAGNET
Tensorcompatible
Parameters¶
| Parameter | Type | Default | Description |
|---|---|---|---|
y_pred |
Tensor / array |
— | Predicted probabilities in [0, 1] |
y_true |
Tensor / array |
— | Binary targets (0 / 1) |
eps |
float |
1e-12 |
Stability floor before log |
Methods¶
forward(y_pred, y_true, eps=1e-12)¶
Return mean binary cross-entropy.
- Returns: Scalar
Tensor
Example Usage¶
import numpy as np
from machinegnostics.magnet import BinaryCrossEntropy
loss = BinaryCrossEntropy()
value = loss(np.array([[0.9]]), np.array([[1.0]]))
Author: Nirmal Parmar