focal loss代码实现

以下是 focal loss 的代码实现（Python）：

import tensorflow as tf

def focal_loss(y_true, y_pred, gamma=2.0, alpha=0.25):
    """
    Focal loss for multi-class classification.
    
    Args:
    y_true (tensor): True labels of shape (batch_size, num_classes).
    y_pred (tensor): Predicted labels of shape (batch_size, num_classes).
    gamma (float): Focusing parameter. Default is 2.0.
    alpha (float): Weighting factor for positive class. Default is 0.25.
    
    Returns:
    Focal loss for the batch.
    """
    # Convert true labels to one-hot encoding
    y_true = tf.one_hot(tf.cast(y_true, tf.int32), depth=y_pred.shape[-1])
    
    # Calculate cross-entropy loss
    pt = tf.where(tf.equal(y_true, 1), y_pred, 1 - y_pred)
    ce_loss = -tf.reduce_sum(alpha * tf.pow(1 - pt, gamma) * tf.log(pt), axis=1)
    
    # Calculate focal loss
    fl_loss = tf.reduce_mean(ce_loss)
    
    return fl_loss

这个实现使用了 TensorFlow 库实现，支持多类别分类问题。gamma 参数控制背景类别相对于前景类别的权重，alpha 参数调节正类别权重。在实现中，先将真实标签转换为 one-hot 编码，然后计算交叉熵损失。最后使用 focal loss 更新损失值。