focal loss的tf代码

以下是使用TensorFlow实现Focal Loss的代码：

import tensorflow as tf

def focal_loss(y_true, y_pred, alpha=0.25, gamma=2.0):
    """
    Focal Loss for multi-class classification
    :param y_true: true labels, shape of [batch_size, num_classes]
    :param y_pred: predicted labels, shape of [batch_size, num_classes]
    :param alpha: balance parameter for adjusting the impact of different classes, default to 0.25
    :param gamma: focusing parameter for emphasizing hard examples, default to 2.0
    :return: focal loss
    """
    # convert y_true to one-hot encoding
    y_true = tf.one_hot(tf.cast(y_true, tf.int32), depth=y_pred.shape[-1])
    # compute cross entropy
    ce = tf.losses.categorical_crossentropy(y_true, y_pred, from_logits=True)
    # compute softmax probabilities
    probs = tf.nn.softmax(y_pred, axis=-1)
    # get the probabilities of true labels
    p_t = tf.reduce_sum(y_true * probs, axis=-1)
    # compute the weight for each sample
    weights = alpha * y_true + (1 - alpha) * (1 - y_true)
    # compute the focal loss for each sample
    fl = -weights * tf.pow(1 - p_t, gamma) * ce
    # compute the mean focal loss over all samples
    return tf.reduce_mean(fl)

该函数接受两个参数，`y_true`和`y_pred`，分别代表真实标签和预测标签。此外，还可以调整两个超参数`alpha`和`gamma`，分别代表类别权重平衡和难样本的加权系数。函数内部首先将`y_true`转换为one-hot编码，然后计算交叉熵，并使用softmax函数将预测标签转换为概率值。接着计算每个样本的权重，然后根据公式计算每个样本的focal loss，并最后求出平均focal loss。