最近在做deepfake檢測任務(可以將其視為二分類問題,label為1和0),遇到了正負樣本不均衡的問題,正樣本數(shù)目是負樣本的5倍,這樣會導致FP率較高。
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,不懂為什么會有64個class,因為BCEloss是針對二分類問題的loss,后經(jīng)過檢索,得知還有多標簽分類,
多標簽分類就是多個標簽,每個標簽有兩個label(0和1),這類任務同樣可以使用BCEloss。
比如我們有正負兩類樣本,正樣本數(shù)量為100個,負樣本為400個,我們想要對正負樣本的loss進行加權處理,將正樣本的loss權重放大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tensor列表,需要和標簽個數(shù)相同,比如我們現(xiàn)在是二分類,只需要將正樣本loss的權重寫上即可。
如果是多標簽分類,有64個標簽,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)
補充:Pytorch —— BCEWithLogitsLoss()的一些問題
torch.sigmoid() + torch.nn.BCELoss()
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~
import torch import torch.nn as nn torch.cuda.manual_seed(300) # 為當前GPU設置隨機種子 torch.manual_seed(300) # 為CPU設置隨機種子 def ce_loss(y_pred, y_train, alpha=1): # 計算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)
測試結果:
– 由上結果可知,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()的強大之處:
– BCEWithLogitsLoss()具有很好的對nan的處理能力,對于我寫的代碼(四層神經(jīng)網(wǎng)絡,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,所以會導致我們編寫的CE的loss出現(xiàn)nan:原因如下:
–首先神經(jīng)網(wǎng)絡輸出的pre_target較大,就會導致sigmoid之后的p為1,則torch.log(1 - p)為nan;
– 使用clamp(函數(shù)雖然會解除這個nan,但是由于在迭代過程中,網(wǎng)絡輸出可能越來越大(層之間使用的是ReLU),則導致我們寫的loss陷入到某一個數(shù)值而無法進行優(yōu)化。但是BCEWithLogitsLoss()對這種情況下出現(xiàn)的nan有很好的處理,從而得到更好的結果。
– 我此實驗的目的是為了比較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()。
二分類 + sigmoid()
使用sigmoid作為輸出層非線性表達的分類問題(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節(jié)點)
輸入格式
要求輸入的input和target均為float類型
以上為個人經(jīng)驗,希望能給大家一個參考,也希望大家多多支持腳本之家。