Focal_loss 的实现

最新推荐文章于 2024-08-19 14:20:52 发布

Runner_of_nku

最新推荐文章于 2024-08-19 14:20:52 发布

阅读量685

点赞数 1

分类专栏： python 机器学习 note 文章标签：深度学习

本文链接：https://blog.csdn.net/Runner_of_nku/article/details/110403807

版权

本文探讨了Focal Loss的实现，分别介绍了在TensorFlow 1.13和Pytorch版本下的实现方法。通过设置alpha为0.5，gamma为0，验证了实现与交叉熵损失函数的一致性。

摘要由CSDN通过智能技术生成

TensorFlow 1.13版本

def focal_loss(logits,onehot_labels,  gamma=2.0, alpha=4.0):
    """
    focal loss for multi-classification
    FL(p_t)=-alpha(1-p_t)^{gamma}ln(p_t)
    Notice: logits is probability after softmax
    gradient is d(Fl)/d(p_t) not d(Fl)/d(x) as described in paper
    d(Fl)/d(p_t) * [p_t(1-p_t)] = d(Fl)/d(x)
    Lin, T.-Y., Goyal, P., Girshick, R., He, K., & Dollár, P. (2017).
    Focal Loss for Dense Object Detection, 130(4), 485–491.
    https://doi.org/10.1016/j.ajodo.2005.02.022
    :param labels: ground truth labels, shape of [batch_size]
    :param logits: model's output, shape of [batch_size, num_cls]
    :param gamma:
    :param alpha:正样本*alpha 负样本*（1-alpha） alpha越高 正样本训练的就越好，负样本训练的就越差
    :return: shape of [batch_size]
    """
    epsilon = 1.e-9
    logits = tf.convert_to_tensor(logits, tf.float32)
    model_out = tf.add(logits, epsilo