pytorch实现focal loss的两种方式小结

我就废话不多说了，直接上代码吧！

import torch
import torch.nn.functional as F
import numpy as np
from torch.autograd import Variable
'''
pytorch实现focal loss的两种方式(现在讨论的是基于分割任务)
在计算损失函数的过程中考虑到类别不平衡的问题，假设加上背景类别共有6个类别
'''
def compute_class_weights(histogram):
 classWeights = np.ones(6, dtype=np.float32)
 normHist = histogram / np.sum(histogram)
 for i in range(6):
   classWeights[i] = 1 / (np.log(1.10 + normHist[i]))
 return classWeights
def focal_loss_my(input,target):
 '''
 :param input: shape [batch_size,num_classes,H,W] 仅仅经过卷积操作后的输出，并没有经过任何激活函数的作用
 :param target: shape [batch_size,H,W]
 :return:
 '''
 n, c, h, w = input.size()

target = target.long()
 input = input.transpose(1, 2).transpose(2, 3).contiguous().view(-1, c)
 target = target.contiguous().view(-1)

number_0 = torch.sum(target == 0).item()
 number_1 = torch.sum(target == 1).item()
 number_2 = torch.sum(target == 2).item()
 number_3 = torch.sum(target == 3).item()
 number_4 = torch.sum(target == 4).item()
 number_5 = torch.sum(target == 5).item()

frequency = torch.tensor((number_0, number_1, number_2, number_3, number_4, number_5), dtype=torch.float32)
 frequency = frequency.numpy()
 classWeights = compute_class_weights(frequency)
 '''
 根据当前给出的ground truth label计算出每个类别所占据的权重
 '''

# weights=torch.from_numpy(classWeights).float().cuda()
 weights = torch.from_numpy(classWeights).float()
 focal_frequency = F.nll_loss(F.softmax(input, dim=1), target, reduction='none')
 '''
 上面一篇博文讲过
 F.nll_loss(torch.log(F.softmax(inputs, dim=1)，target)的函数功能与F.cross_entropy相同
 可见F.nll_loss中实现了对于target的one-hot encoding编码功能，将其编码成与input shape相同的tensor
 然后与前面那一项（即F.nll_loss输入的第一项）进行 element-wise production
 相当于取出了 log(p_gt)即当前样本点被分类为正确类别的概率
 现在去掉取log的操作，相当于 focal_frequency shape [num_samples]
 即取出ground truth类别的概率数值，并取了负号
 '''

focal_frequency += 1.0#shape [num_samples] 1-P（gt_classes）

focal_frequency = torch.pow(focal_frequency, 2) # torch.Size([75])
 focal_frequency = focal_frequency.repeat(c, 1)
 '''
 进行repeat操作后，focal_frequency shape [num_classes,num_samples]
 '''
 focal_frequency = focal_frequency.transpose(1, 0)
 loss = F.nll_loss(focal_frequency * (torch.log(F.softmax(input, dim=1))), target, weight=None,
          reduction='elementwise_mean')
 return loss

def focal_loss_zhihu(input, target):
 '''
 :param input: 使用知乎上面大神给出的方案 https://zhuanlan.zhihu.com/p/28527749
 :param target:
 :return:
 '''
 n, c, h, w = input.size()

target = target.long()
 inputs = input.transpose(1, 2).transpose(2, 3).contiguous().view(-1, c)
 target = target.contiguous().view(-1)

N = inputs.size(0)
 C = inputs.size(1)

number_0 = torch.sum(target == 0).item()
 number_1 = torch.sum(target == 1).item()
 number_2 = torch.sum(target == 2).item()
 number_3 = torch.sum(target == 3).item()
 number_4 = torch.sum(target == 4).item()
 number_5 = torch.sum(target == 5).item()

frequency = torch.tensor((number_0, number_1, number_2, number_3, number_4, number_5), dtype=torch.float32)
 frequency = frequency.numpy()
 classWeights = compute_class_weights(frequency)

weights = torch.from_numpy(classWeights).float()
 weights=weights[target.view(-1)]#这行代码非常重要

gamma = 2

P = F.softmax(inputs, dim=1)#shape [num_samples,num_classes]

class_mask = inputs.data.new(N, C).fill_(0)
 class_mask = Variable(class_mask)
 ids = target.view(-1, 1)
 class_mask.scatter_(1, ids.data, 1.)#shape [num_samples,num_classes] one-hot encoding

probs = (P * class_mask).sum(1).view(-1, 1)#shape [num_samples,]
 log_p = probs.log()

print('in calculating batch_loss',weights.shape,probs.shape,log_p.shape)

# batch_loss = -weights * (torch.pow((1 - probs), gamma)) * log_p
 batch_loss = -(torch.pow((1 - probs), gamma)) * log_p

print(batch_loss.shape)

loss = batch_loss.mean()
 return loss

if __name__=='__main__':
 pred=torch.rand((2,6,5,5))
 y=torch.from_numpy(np.random.randint(0,6,(2,5,5)))
 loss1=focal_loss_my(pred,y)
 loss2=focal_loss_zhihu(pred,y)

print('loss1',loss1)
 print('loss2', loss2)
'''
in calculating batch_loss torch.Size([50]) torch.Size([50, 1]) torch.Size([50, 1])
torch.Size([50, 1])
loss1 tensor(1.3166)
loss2 tensor(1.3166)
'''

来源：https://blog.csdn.net/WYXHAHAHA123/article/details/88343945