GHMC代码中函数含义

代码

import torch
import torch.nn.functional as F
from torch.autograd import Variable

class GHMC_Loss:
    def __init__(self, bins=10, momentum=0.0):
        self.bins = bins
        self.momentum = momentum
        self.edges = [float(x) / bins for x in range(bins+1)]
        self.edges[-1] += 1e-6
        if momentum > 0:
            self.acc_sum = [0.0 for _ in range(bins)]

    def calc(self, input, target, mask):
        """ Args:
        input [batch_num, class_num]:
            The direct prediction of classification fc layer.
        target [batch_num, class_num]:
            Binary target (0 or 1) for each sample each class. The value is -1
            when the sample is ignored.
        """
        edges = self.edges
        mmt = self.momentum
        weights = torch.zeros_like(input)

        # gradient length
        self.g = torch.abs(input.sigmoid().detach() - target)

        valid = mask > 0
        tot = max(valid.float().sum().item(), 1.0)
        n = 0  # n valid bins
        for i in range(self.bins):
            inds = (self.g >= edges[i]) & (self.g < edges[i+1]) & valid
            num_in_bin = inds.sum().item()
            if num_in_bin > 0:
                if mmt > 0:
                    self.acc_sum[i] = mmt * self.acc_sum[i] \
                        + (1 - mmt) * num_in_bin
                    weights[inds] = tot / self.acc_sum[i]
                else:
                    weights[inds] = tot / num_in_bin
                n += 1
        if n > 0:
            weights = weights / n

        loss = F.binary_cross_entropy_with_logits(
            input, target, weights, reduction='sum') / tot
        return loss, self.acc_sum,self.g

---

第一部分

class GHMC_Loss:
def __init__(self, bins=10, momentum=0.0):#构造函数，类实例化是自动执行
‘’’初始化实例属性’’’
        self.bins = bins
        self.momentum = momentum
        self.edges = [float(x) / bins for x in range(bins+1)]
        self.edges[-1] += 1e-6
        if momentum > 0:
            self.acc_sum = [0.0 for _ in range(bins)]

class是关键字（表示要开始创建类了），GHMC_Loss是新建类名

第二部分

    def calc(self, input, target, mask):#创建类中函数，self特殊参数（必填）
        """ Args:
        input [batch_num, class_num]:
            The direct prediction of classification fc layer.
        target [batch_num, class_num]:
            Binary target (0 or 1) for each sample each class. The value is -1
            when the sample is ignored.
        """
        edges = self.edges
        mmt = self.momentum
        weights = torch.zeros_like(input)

zeros_like函数，从函数名不难猜到，这个函数是用于生成和输入tensor大小相同的全零tensor的。

比如：

程序

x=torch.Tensor(5,3)
print(torch.zeros_like(x)) #torch.FloatTensor of size 5x3

结果：

tensor([[0., 0., 0.],

    [0., 0., 0.],

    [0., 0., 0.],

    [0., 0., 0.],

    [0., 0., 0.]])

第三部分

       # gradient length
        self.g = torch.abs(input.sigmoid().detach() - target)

torch.abs(input, out=None):

计算输入张量的每个元素绝对值

input(Tensor) - 输入张量

out(Tensor, optional) - 结果张量

第四部分

        valid = mask > 0
        tot = max(valid.float().sum().item(), 1.0)
        n = 0  # n valid bins
        for i in range(self.bins):
            inds = (self.g >= edges[i]) & (self.g < edges[i+1]) & valid
            num_in_bin = inds.sum().item()
            if num_in_bin > 0:
                if mmt > 0:
                    self.acc_sum[i] = mmt * self.acc_sum[i] \
                        + (1 - mmt) * num_in_bin
                    weights[inds] = tot / self.acc_sum[i]
                else:
                    weights[inds] = tot / num_in_bin
                n += 1
        if n > 0:
            weights = weights / n

        loss = F.binary_cross_entropy_with_logits(
            input, target, weights, reduction='sum') / tot
        return loss, self.acc_sum,self.g

&按位与运算。

max() 方法返回给定参数的最大值，参数可以为序列。