Add categorical focal cross-entropy loss algorithm (#11248)

This commit is contained in:
AtomicVar 2024-01-18 20:41:29 +08:00 committed by GitHub
parent 05a5cdacc3
commit 3952ba703a
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194

View File

@ -148,6 +148,108 @@ def categorical_cross_entropy(
return -np.sum(y_true * np.log(y_pred))
def categorical_focal_cross_entropy(
y_true: np.ndarray,
y_pred: np.ndarray,
alpha: np.ndarray = None,
gamma: float = 2.0,
epsilon: float = 1e-15,
) -> float:
"""
Calculate the mean categorical focal cross-entropy (CFCE) loss between true
labels and predicted probabilities for multi-class classification.
CFCE loss is a generalization of binary focal cross-entropy for multi-class
classification. It addresses class imbalance by focusing on hard examples.
CFCE = -Σ alpha * (1 - y_pred)**gamma * y_true * log(y_pred)
Reference: [Lin et al., 2018](https://arxiv.org/pdf/1708.02002.pdf)
Parameters:
- y_true: True labels in one-hot encoded form.
- y_pred: Predicted probabilities for each class.
- alpha: Array of weighting factors for each class.
- gamma: Focusing parameter for modulating the loss (default: 2.0).
- epsilon: Small constant to avoid numerical instability.
Returns:
- The mean categorical focal cross-entropy loss.
>>> true_labels = np.array([[1, 0, 0], [0, 1, 0], [0, 0, 1]])
>>> pred_probs = np.array([[0.9, 0.1, 0.0], [0.2, 0.7, 0.1], [0.0, 0.1, 0.9]])
>>> alpha = np.array([0.6, 0.2, 0.7])
>>> categorical_focal_cross_entropy(true_labels, pred_probs, alpha)
0.0025966118981496423
>>> true_labels = np.array([[0, 1, 0], [0, 0, 1]])
>>> pred_probs = np.array([[0.05, 0.95, 0], [0.1, 0.8, 0.1]])
>>> alpha = np.array([0.25, 0.25, 0.25])
>>> categorical_focal_cross_entropy(true_labels, pred_probs, alpha)
0.23315276982014324
>>> true_labels = np.array([[1, 0], [0, 1]])
>>> pred_probs = np.array([[0.9, 0.1, 0.0], [0.2, 0.7, 0.1]])
>>> categorical_cross_entropy(true_labels, pred_probs)
Traceback (most recent call last):
...
ValueError: Input arrays must have the same shape.
>>> true_labels = np.array([[2, 0, 1], [1, 0, 0]])
>>> pred_probs = np.array([[0.9, 0.1, 0.0], [0.2, 0.7, 0.1]])
>>> categorical_focal_cross_entropy(true_labels, pred_probs)
Traceback (most recent call last):
...
ValueError: y_true must be one-hot encoded.
>>> true_labels = np.array([[1, 0, 1], [1, 0, 0]])
>>> pred_probs = np.array([[0.9, 0.1, 0.0], [0.2, 0.7, 0.1]])
>>> categorical_focal_cross_entropy(true_labels, pred_probs)
Traceback (most recent call last):
...
ValueError: y_true must be one-hot encoded.
>>> true_labels = np.array([[1, 0, 0], [0, 1, 0]])
>>> pred_probs = np.array([[0.9, 0.1, 0.1], [0.2, 0.7, 0.1]])
>>> categorical_focal_cross_entropy(true_labels, pred_probs)
Traceback (most recent call last):
...
ValueError: Predicted probabilities must sum to approximately 1.
>>> true_labels = np.array([[1, 0, 0], [0, 1, 0], [0, 0, 1]])
>>> pred_probs = np.array([[0.9, 0.1, 0.0], [0.2, 0.7, 0.1], [0.0, 0.1, 0.9]])
>>> alpha = np.array([0.6, 0.2])
>>> categorical_focal_cross_entropy(true_labels, pred_probs, alpha)
Traceback (most recent call last):
...
ValueError: Length of alpha must match the number of classes.
"""
if y_true.shape != y_pred.shape:
raise ValueError("Shape of y_true and y_pred must be the same.")
if alpha is None:
alpha = np.ones(y_true.shape[1])
if np.any((y_true != 0) & (y_true != 1)) or np.any(y_true.sum(axis=1) != 1):
raise ValueError("y_true must be one-hot encoded.")
if len(alpha) != y_true.shape[1]:
raise ValueError("Length of alpha must match the number of classes.")
if not np.all(np.isclose(np.sum(y_pred, axis=1), 1, rtol=epsilon, atol=epsilon)):
raise ValueError("Predicted probabilities must sum to approximately 1.")
# Clip predicted probabilities to avoid log(0)
y_pred = np.clip(y_pred, epsilon, 1 - epsilon)
# Calculate loss for each class and sum across classes
cfce_loss = -np.sum(
alpha * np.power(1 - y_pred, gamma) * y_true * np.log(y_pred), axis=1
)
return np.mean(cfce_loss)
def hinge_loss(y_true: np.ndarray, y_pred: np.ndarray) -> float:
"""
Calculate the mean hinge loss for between true labels and predicted probabilities