App下載

BCEWithLogitsLoss樣本不均衡的處理方案

流年絮語 2021-08-17 11:18:12 瀏覽數(shù) (5792)
反饋

在做deepfake檢測(cè)任務(wù)(可以將其視為二分類問題,label為1和0)的時(shí)候,可能會(huì)遇到正負(fù)樣本不均衡的問題,正樣本數(shù)目是負(fù)樣本的5倍,這樣會(huì)導(dǎo)致FP率較高。那么怎么解決這樣的問題呢?來看看小編的解決方案。

嘗試將正樣本的loss權(quán)重增高,看BCEWithLogitsLoss的源碼

Examples::
 
    >>> target = torch.ones([10, 64], dtype=torch.float32)  # 64 classes, batch size = 10
    >>> output = torch.full([10, 64], 0.999)  # A prediction (logit)
    >>> pos_weight = torch.ones([64])  # All weights are equal to 1
    >>> criterion = torch.nn.BCEWithLogitsLoss(pos_weight=pos_weight)
    >>> criterion(output, target)  # -log(sigmoid(0.999))
    tensor(0.3135)
 
Args:
    weight (Tensor, optional): a manual rescaling weight given to the loss
        of each batch element. If given, has to be a Tensor of size `nbatch`.
    size_average (bool, optional): Deprecated (see :attr:`reduction`). By default,
        the losses are averaged over each loss element in the batch. Note that for
        some losses, there are multiple elements per sample. If the field :attr:`size_average`
        is set to ``False``, the losses are instead summed for each minibatch. Ignored
        when reduce is ``False``. Default: ``True``
    reduce (bool, optional): Deprecated (see :attr:`reduction`). By default, the
        losses are averaged or summed over observations for each minibatch depending
        on :attr:`size_average`. When :attr:`reduce` is ``False``, returns a loss per
        batch element instead and ignores :attr:`size_average`. Default: ``True``
    reduction (string, optional): Specifies the reduction to apply to the output:
        ``'none'`` | ``'mean'`` | ``'sum'``. ``'none'``: no reduction will be applied,
        ``'mean'``: the sum of the output will be divided by the number of
        elements in the output, ``'sum'``: the output will be summed. Note: :attr:`size_average`
        and :attr:`reduce` are in the process of being deprecated, and in the meantime,
        specifying either of those two args will override :attr:`reduction`. Default: ``'mean'``
    pos_weight (Tensor, optional): a weight of positive examples.
            Must be a vector with length equal to the number of classes.

對(duì)其中的參數(shù)pos_weight的使用存在疑惑,BCEloss里的例子pos_weight = torch.ones([64]) # All weights are equal to 1,不懂為什么會(huì)有64個(gè)class,因?yàn)锽CEloss是針對(duì)二分類問題的loss,后經(jīng)過檢索,得知還有多標(biāo)簽分類,

分類問題

多標(biāo)簽分類就是多個(gè)標(biāo)簽,每個(gè)標(biāo)簽有兩個(gè)label(0和1),這類任務(wù)同樣可以使用BCEloss。

BCEloss

現(xiàn)在講一下BCEWithLogitsLoss里的pos_weight使用方法

比如我們有正負(fù)兩類樣本,正樣本數(shù)量為100個(gè),負(fù)樣本為400個(gè),我們想要對(duì)正負(fù)樣本的loss進(jìn)行加權(quán)處理,將正樣本的loss權(quán)重放大4倍,通過這樣的方式緩解樣本不均衡問題。

criterion = nn.BCEWithLogitsLoss(pos_weight=torch.tensor([4]))
 
# pos_weight (Tensor, optional): a weight of positive examples.
#            Must be a vector with length equal to the number of classes.

pos_weight里是一個(gè)tensor列表,需要和標(biāo)簽個(gè)數(shù)相同,比如我們現(xiàn)在是二分類,只需要將正樣本loss的權(quán)重寫上即可。

如果是多標(biāo)簽分類,有64個(gè)標(biāo)簽,則

Examples::
 
    >>> target = torch.ones([10, 64], dtype=torch.float32)  # 64 classes, batch size = 10
    >>> output = torch.full([10, 64], 0.999)  # A prediction (logit)
    >>> pos_weight = torch.ones([64])  # All weights are equal to 1
    >>> criterion = torch.nn.BCEWithLogitsLoss(pos_weight=pos_weight)
    >>> criterion(output, target)  # -log(sigmoid(0.999))
    tensor(0.3135)

補(bǔ)充:Pytorch —— BCEWithLogitsLoss()的一些問題

一、等價(jià)表達(dá)

1、pytorch:

torch.sigmoid() + torch.nn.BCELoss()

2、自己編寫

def ce_loss(y_pred, y_train, alpha=1):
    
    p = torch.sigmoid(y_pred)
    # p = torch.clamp(p, min=1e-9, max=0.99)  
    loss = torch.sum(- alpha * torch.log(p) * y_train 
           - torch.log(1 - p) * (1 - y_train))/len(y_train)
    return loss~

3、驗(yàn)證

import torch
import torch.nn as nn
torch.cuda.manual_seed(300)       # 為當(dāng)前GPU設(shè)置隨機(jī)種子
torch.manual_seed(300)            # 為CPU設(shè)置隨機(jī)種子
def ce_loss(y_pred, y_train, alpha=1):
   # 計(jì)算loss
   p = torch.sigmoid(y_pred)
   # p = torch.clamp(p, min=1e-9, max=0.99)
   loss = torch.sum(- alpha * torch.log(p) * y_train 
          - torch.log(1 - p) * (1 - y_train))/len(y_train)
   return loss
py_lossFun = nn.BCEWithLogitsLoss()
input = torch.randn((10000,1), requires_grad=True)
target = torch.ones((10000,1))
target.requires_grad_(True)
py_loss = py_lossFun(input, target)
py_loss.backward()
print("*********BCEWithLogitsLoss***********")
print("loss: ")
print(py_loss.item())
print("梯度: ")
print(input.grad)
input = input.detach()
input.requires_grad_(True)
self_loss = ce_loss(input, target)
self_loss.backward()
print("*********SelfCELoss***********")
print("loss: ")
print(self_loss.item())
print("梯度: ")
print(input.grad)

測(cè)試結(jié)果:

測(cè)試結(jié)果

– 由上結(jié)果可知,我編寫的loss和pytorch中提供的j基本一致。

– 但是僅僅這樣就可以了嗎?NO! 下面介紹BCEWithLogitsLoss()的強(qiáng)大之處:

– BCEWithLogitsLoss()具有很好的對(duì)nan的處理能力,對(duì)于我寫的代碼(四層神經(jīng)網(wǎng)絡(luò),層之間的激活函數(shù)采用的是ReLU,輸出層激活函數(shù)采用sigmoid(),由于數(shù)據(jù)處理的問題,所以會(huì)導(dǎo)致我們編寫的CE的loss出現(xiàn)nan:原因如下:

–首先神經(jīng)網(wǎng)絡(luò)輸出的pre_target較大,就會(huì)導(dǎo)致sigmoid之后的p為1,則torch.log(1 - p)為nan;

– 使用clamp(函數(shù)雖然會(huì)解除這個(gè)nan,但是由于在迭代過程中,網(wǎng)絡(luò)輸出可能越來越大(層之間使用的是ReLU),則導(dǎo)致我們寫的loss陷入到某一個(gè)數(shù)值而無法進(jìn)行優(yōu)化。但是BCEWithLogitsLoss()對(duì)這種情況下出現(xiàn)的nan有很好的處理,從而得到更好的結(jié)果。

– 我此實(shí)驗(yàn)的目的是為了比較CE和FL的區(qū)別,自己編寫FL,則必須也要自己編寫CE,不能使用BCEWithLogitsLoss()。

二、使用場(chǎng)景

二分類 + sigmoid()

使用sigmoid作為輸出層非線性表達(dá)的分類問題(雖然可以處理多分類問題,但是一般用于二分類,并且最后一層只放一個(gè)節(jié)點(diǎn))

三、注意事項(xiàng)

輸入格式

要求輸入的input和target均為float類型

以上就是BCEWithLogitsLoss樣本不均衡的處理方案,希望能給大家一個(gè)參考,也希望大家多多支持W3Cschool


0 人點(diǎn)贊