class balanced loss pytorch 实现

cb loss pytorch 实现，可直接调用
参考：https://github.com/vandit15/Class-balanced-loss-pytorch/blob/master/class_balanced_loss.py

import numpy as np
import torch
import torch.nn.functional as Fdef focal_loss(logits, labels, alpha, gamma):"""Compute the focal loss between `logits` and the ground truth `labels`.Focal loss = -alpha_t * (1-pt)^gamma * log(pt)where pt is the probability of being classified to the true class.pt = p (if true class), otherwise pt = 1 - p. p = sigmoid(logit).Args:logits: A float tensor of size [batch, num_classes].labels: A float tensor of size [batch, num_classes].alpha: A float tensor of size [batch_size]specifying per-example weight for balanced cross entropy.gamma: A float scalar modulating loss from hard and easy examples.Returns:focal_loss: A float32 scalar representing normalized total loss."""bce_loss = F.binary_cross_entropy_with_logits(input=logits, target=labels, reduction="none")if gamma == 0.0:modulator = 1.0else:modulator = torch.exp(-gamma * labels * logits - gamma * torch.log(1 + torch.exp(-1.0 * logits)))loss = modulator * bce_lossweighted_loss = alpha * lossloss = torch.sum(weighted_loss)loss /= torch.sum(labels)return lossclass ClassBalancedLoss(torch.nn.Module):def __init__(self, samples_per_class=None, beta=0.9999, gamma=0.5, loss_type="focal"):super(ClassBalancedLoss, self).__init__()if loss_type not in ["focal", "sigmoid", "softmax"]:loss_type = "focal"if samples_per_class is None:num_classes = 5000samples_per_class = [1] * num_classeseffective_num = 1.0 - np.power(beta, samples_per_class)weights = (1.0 - beta) / np.array(effective_num)self.constant_sum = len(samples_per_class)weights = (weights / np.sum(weights) * self.constant_sum).astype(np.float32)self.class_weights = weightsself.beta = betaself.gamma = gammaself.loss_type = loss_typedef update(self, samples_per_class):if samples_per_class is None:returneffective_num = 1.0 - np.power(self.beta, samples_per_class)weights = (1.0 - self.beta) / np.array(effective_num)self.constant_sum = len(samples_per_class)weights = (weights / np.sum(weights) * self.constant_sum).astype(np.float32)self.class_weights = weightsdef forward(self, x, y):_, num_classes = x.shapelabels_one_hot = F.one_hot(y, num_classes).float()weights = torch.tensor(self.class_weights, device=x.device).index_select(0, y)weights = weights.unsqueeze(1)if self.loss_type == "focal":cb_loss = focal_loss(x, labels_one_hot, weights, self.gamma)elif self.loss_type == "sigmoid":cb_loss = F.binary_cross_entropy_with_logits(x, labels_one_hot, weights)else:  # softmaxpred = x.softmax(dim=1)cb_loss = F.binary_cross_entropy(pred, labels_one_hot, weights)return cb_lossdef test():torch.manual_seed(123)batch_size = 10num_classes = 5x = torch.rand(batch_size, num_classes)y = torch.randint(0, 5, size=(batch_size,))samples_per_class = [1, 2, 3, 4, 5]loss_type = "focal"loss_fn = ClassBalancedLoss(samples_per_class, loss_type=loss_type)loss = loss_fn(x, y)print(loss)if __name__ == '__main__':test()

class balanced loss pytorch 实现相关推荐

[论文解读] A Ranking-based, Balanced Loss Function Unifying Classification and Localisation in Object De
文章内容相关研究现状 1. 定位任务和分类任务的平衡/耦合 2. 基于排名的目标检测算法本文工作基于排序损失的误差驱动优化方法推广定理1:基于概率分布的损失函数重定义定理2:正负样本梯度总和 ...
pytorch训练Class-Balanced Loss
1. 提示:文章写完后,目录可以自动生成,如何生成可参考右边的帮助文档文章目录 1. pytorch版Class-Balanced Loss训练模型一.数据准备二.模型训练三.模型预测总结 ...
pytorch版Class-Balanced Loss训练模型
pytorch版Class-Balanced Loss训练模型 1.论文参考原文 https://arxiv.org/pdf/1901.05555.pdf 2.数据准备将自己的数据集按照一下格式进行 ...
SRGAN loss部分的pytorch代码实现
转载地址:https://bbs.huaweicloud.com/forum/thread-137101-1-1.html 作者: 雨丝儿最近在参加华为与高校合做开发mindspore模型的活动,使 ...
深度学习100+经典模型TensorFlow与Pytorch代码实现大合集
关注上方"深度学习技术前沿",选择"星标公众号", 资源干货,第一时间送达! [导读]深度学习在过去十年获得了极大进展,出现很多新的模型,并且伴随TensorF ...
BCE loss和 CE理解
1. BCE loss:Binary Cross Entropy Loss BCE loss pytorch官网链接 1.1 解释 pytorch中调用如下.设置weight,使得不同类别的损失权值不 ...
【Pytorch】常见的人脸身份识别损失函数
[Pytorch]常见的人脸身份识别损失函数实验环境准备:人脸多角度多光照的图像数据集MUCT(276个受试者)+ MobileNetV3 说明:对于人脸身份数据集MUCT,是少样本数据集,应该使用 ...
Style Transfer（PyTorch）
Style Transfer-PyTorch Content Loss content loss用来计算原图片和生成的图片之间像素的差距,这里用的是卷积层获取的 feature map 之间的差距通 ...
Pytorch以及tensorflow中KLdivergence的计算
1. KL divergence是什么 KL 散度是一个距离衡量指标,衡量的是两个概率分布之间的差异. y p r e d y_{pred} ypred指的是模型的输出的预测概率,形如[0.35,0 ...

class balanced loss pytorch 实现

class balanced loss pytorch 实现相关推荐

最新文章

热门文章