文章目录
- 损失函数
- 1、损失函数概念
- 2、交叉熵损失函数
- 3、NLL/BCE/BCEWithLogits Loss
- 4、其余14种损失函数介绍
- (1)`nn.L1Loss`
- (2)`nn.MSELoss`
- (3)`SmoothL1Loss`
- (4)`PoissonNLLLoss`
- (5)`nn.KLDivLoss`
- (6)`nn.MarginRankingLoss`
- (7)`nn.MultiLabelMarginLoss`
- (8)`nn.SoftMarginLoss`
- (9)`nn.MultiLabelSoftMarginLoss`
- (10)`nn.MultiMarginLoss`
- (11)`nn.TripletMarginLoss`
- (12)`nn.HingeEmbeddingLoss`
- (13)`nn.CosineEmbeddingLoss`
- (14)`nn.CTCLoss`
- 5、损失函数总结
损失函数
1、损失函数概念
(1)概述
损失函数用于衡量模型输出与真实标签的差异,需要注意区分损失函数,代价函数,目标函数三者的区别:
- 损失函数: L o s s = f ( y ^ , y ) Loss = f(\hat{y},y) Loss=f(y^,y)
- 代价函数: C o s t = 1 N ∑ i = 1 N f ( y ^ i , y i ) Cost = \frac{1}{N}\sum_{i=1}^{N}f(\hat{y}_i,y_i) Cost=N1∑i=1Nf(y^i,yi)
- 目标函数: O b j = C o s t + R e g u l a t i o n Obj = Cost + Regulation Obj=Cost+Regulation
(2)PyTorch
中的Loss
class _Loss(Module):
def __init__(self, size_average=None, reduce=None,reduction='mean'):
super(_Loss, self).__init__()
if size_average is not None or reduce is not None:
self.reduction = _Reduction.legacy_get_string(size_average, reduce)
else:
self.reduction = reduction
从上面的代码中可以看出,PyTorch
的_Loss
类是继承的Module
,因此可以把Loss
视为一个网络层,这里需要注意的是size_average
和reduce
这两个参数之后会把舍弃,其功能完全由reduction
参数替代
2、交叉熵损失函数
交叉熵等于信息熵加上相对熵: H ( P , Q ) = − ∑ i = 1 N P ( x i ) log Q ( x i ) H(P,Q) = -\sum_{i=1}^{N}P(x_i)\log Q(x_i) H(P,Q)=−∑i=1NP(xi)logQ(xi)
PyTorch
实现:nn.CrossEntropyLoss
:
nn.CrossEntropyLoss(weight=None,
size_average=None,
ignore_index=-100,
reduce=None,
reduction='mean')
- 功能:
nn.LogSoftmax()
与nn.NLLLoss()
结合,进行交叉熵计算 - 主要参数:
weight
:各类别的loss设置权值ignore_index
:忽略某个类别reduction
:计算模式,可为none/sum/mean
,"none"
表示逐个元素计算,"sum"
表示所有元素求和,返回标量,"mean"
表示加权平均,返回标量- 计算公式:
loss ( x , class ) = − log ( exp ( x [ class ] ) ∑ j exp ( x [ j ] ) ) = − x [ class ] + log ( ∑ j exp ( x [ j ] ) ) \operatorname{loss}(x, \text { class })=-\log \left(\frac{\exp (x[\text { class }])}{\sum_{j} \exp (x[j])}\right)=-x[\text { class }]+\log \left(\sum_{j} \exp (x[j])\right) loss(x, class )=−log(∑jexp(x[j])exp(x[ class ]))=−x[ class ]+log(j∑exp(x[j]))
loss ( x , class ) = weight [ class ] ( − x [ class ] + log ( ∑ j exp ( x [ j ] ) ) ) \operatorname{loss}(x, \text { class })=\operatorname{weight}[\text { class }]\left(-x[\text { class }]+\log \left(\sum_{j} \exp (x[\mathrm{j}])\right)\right) loss(x, class )=weight[ class ](−x[ class ]+log(j∑exp(x[j])))
- 实例:
# def loss function
loss_f_none = nn.CrossEntropyLoss(weight=None, reduction='none')
loss_f_sum = nn.CrossEntropyLoss(weight=None, reduction='sum')
loss_f_mean = nn.CrossEntropyLoss(weight=None, reduction='mean')
# forward
loss_none = loss_f_none(inputs, target)
loss_sum = loss_f_sum(inputs, target)
loss_mean = loss_f_mean(inputs, target)
# view
print("Cross Entropy Loss:\n ", loss_none, loss_sum, loss_mean)
# Cross Entropy Loss:
# tensor([1.3133, 0.1269, 0.1269]) tensor(1.5671) tensor(0.5224)
手动计算验证:
idx = 0
input_1 = inputs.detach().numpy()[idx] # [1, 2]
target_1 = target.numpy()[idx] # [0]
# 第一项
x_class = input_1[target_1]
# 第二项
sigma_exp_x = np.sum(list(map(np.exp, input_1)))
log_sigma_exp_x = np.log(sigma_exp_x)
# 输出loss
loss_1 = -x_class + log_sigma_exp_x
print("第一个样本loss为: ", loss_1)
# 第一个样本loss为: 1.3132617
如果带weight
参数:
# def loss function
weights = torch.tensor([1, 2], dtype=torch.float)
# weights = torch.tensor([0.7, 0.3], dtype=torch.float)
loss_f_none_w = nn.CrossEntropyLoss(weight=weights, reduction='none')
loss_f_sum = nn.CrossEntropyLoss(weight=weights, reduction='sum')
loss_f_mean = nn.CrossEntropyLoss(weight=weights, reduction='mean')
# forward
loss_none_w = loss_f_none_w(inputs, target)
loss_sum = loss_f_sum(inputs, target)
loss_mean = loss_f_mean(inputs, target)
# view
print("\nweights: ", weights)
print(loss_none_w, loss_sum, loss_mean)
# weights: tensor([1., 2.])
# tensor([1.3133, 0.2539, 0.2539]) tensor(1.8210) tensor(0.3642)
这里注意:mean模式下是除以总权值份数,即1+2+2=5而非样本个数3
- 小结:交叉熵损失函数多用于分类问题,共有三种模式可以选择
3、NLL/BCE/BCEWithLogits Loss
(1)nn.NLLLoss
nn.NLLLoss(weight=None,
size_average=None,
ignore_index=-100,
reduce=None,
reduction='mean')
- 功能:实现负对数似然函数中的负号功能(仅仅是取负号,不要被名称迷惑)
- 计算公式:
ℓ ( x , y ) = L = { l 1 , … , l N } ′ , l n = − w y n x n , y n \ell(x, y)=L=\left\{l_{1}, \dots, l_{N}\right\}^{\prime}, \quad l_{n}=-w_{y_{n}} x_{n, y_{n}} ℓ(x,y)=L={l1,…,lN}′,ln=−wynxn,yn - 主要参数:
weight
:各类别的loss设置权值ignore_index
:忽略某个类别reduction
:计算模式,可为none/sum/mean
,"none"
表示逐个元素计算,"sum"
表示所有元素求和,返回标量,"mean"
表示加权平均,返回标量- 实例:
weights = torch.tensor([1, 1], dtype=torch.float)
loss_f_none_w = nn.NLLLoss(weight=weights, reduction='none')
loss_f_sum = nn.NLLLoss(weight=weights, reduction='sum')
loss_f_mean = nn.NLLLoss(weight=weights, reduction='mean')
# forward
loss_none_w = loss_f_none_w(inputs, target)
loss_sum = loss_f_sum(inputs, target)
loss_mean = loss_f_mean(inputs, target)
# view
print("\nweights: ", weights)
print("NLL Loss", loss_none_w, loss_sum, loss_mean)
# weights: tensor([1., 1.])
# NLL Loss tensor([-1., -3., -3.]) tensor(-7.) tensor(-2.3333)
这里的结果[-1,-3,-3]
是这样计算的,由于第一个样本类别为0,因此[1,2]
中取第一个元素并取负,因此为-1,第二个样本和第三个样本类别为1,因此[1,3]
中取第二个元素并取负,因此为-3
(2)nn.BCELoss
nn.BCELoss(weight=None,
size_average=None,
reduce=None,
reduction='mean')
- 功能:二分类交叉熵
- 计算公式:(注意输入值取值在
[
0
,
1
]
[0,1]
[0,1])
l n = − w n [ y n ⋅ log x n + ( 1 − y n ) ⋅ log ( 1 − x n ) ] l_{n}=-w_{n}\left[y_{n} \cdot \log x_{n}+\left(1-y_{n}\right) \cdot \log \left(1-x_{n}\right)\right] ln=−wn[yn⋅logxn+(1−yn)⋅log(1−xn)] - 主要参数:
weight
:各类别的loss设置权值ignore_index
:忽略某个类别reduction
:计算模式,可为none/sum/mean
,"none"
表示逐个元素计算,"sum"
表示所有元素求和,返回标量,"mean"
表示加权平均,返回标量- 实例:
inputs = torch.tensor([[1, 2], [2, 2], [3, 4], [4, 5]], dtype=torch.float)
target = torch.tensor([[1, 0], [1, 0], [0, 1], [0, 1]], dtype=torch.float)
target_bce = target
# itarget
inputs = torch.sigmoid(inputs)
weights = torch.tensor([1, 1], dtype=torch.float)
loss_f_none_w = nn.BCELoss(weight=weights, reduction='none')
loss_f_sum = nn.BCELoss(weight=weights, reduction='sum')
loss_f_mean = nn.BCELoss(weight=weights, reduction='mean')
# forward
loss_none_w = loss_f_none_w(inputs, target_bce)
loss_sum = loss_f_sum(inputs, target_bce)
loss_mean = loss_f_mean(inputs, target_bce)
# view
print("\nweights: ", weights)
print("BCE Loss", loss_none_w, loss_sum, loss_mean)
# weights: tensor([1., 1.])
# BCE Loss tensor([[0.3133, 2.1269],
# [0.1269, 2.1269],
# [3.0486, 0.0181],
# [4.0181, 0.0067]]) tensor(11.7856) tensor(1.4732)
从上面的代码可以看出,一共得到了8个值,充分说明了逐个计算的含义,下面手动计算验证:
idx = 0
x_i = inputs.detach().numpy()[idx, idx]
y_i = target.numpy()[idx, idx] #
# loss
# l_i = -[ y_i * np.log(x_i) + (1-y_i) * np.log(1-y_i) ] # np.log(0) = nan
l_i = -y_i * np.log(x_i) if y_i else -(1-y_i) * np.log(1-x_i)
# 输出loss
print("BCE inputs: ", inputs)
print("第一个loss为: ", l_i)
# 第一个loss为: 0.31326166
(3)nn.BCEWithLogitsLoss
当不希望神经网络中出现
S
i
g
m
o
i
d
Sigmoid
Sigmoid函数,而计算损失函数时需要
S
i
g
m
o
i
d
Sigmoid
Sigmoid函数的时候,就可以使用nn.BCEWithLogitsLoss
nn.BCEWithLogitsLoss(weight=None,
size_average=None,
reduce=None,
reduction='mean',
pos_weight=None)
- 功能:结合
Sigmoid
与二分类交叉熵 - 计算公式:
l n = − w n [ y n ⋅ log σ ( x n ) + ( 1 − y n ) ⋅ log ( 1 − σ ( x n ) ) ] σ 就是Sigmoid函数 l_{n}=-w_{n}\left[y_{n} \cdot \log \sigma\left(x_{n}\right)+\left(1-y_{n}\right) \cdot \log \left(1-\sigma\left(x_{n}\right)\right)\right] \qquad \sigma\text{就是Sigmoid函数} ln=−wn[yn⋅logσ(xn)+(1−yn)⋅log(1−σ(xn))]σ就是Sigmoid函数 - 主要参数:
pos_weight
:正样本的权值,这个参数用于平衡样本权值,例如有100个正样本和300个负样本,则可把这个参数设置为3,这样的话正负样本就相互平衡了weight
:各类别的loss设置权值ignore_index
:忽略某个类别reduction
:计算模式,可为none/sum/mean
,"none"
表示逐个元素计算,"sum"
表示所有元素求和,返回标量,"mean"
表示加权平均,返回标量- 注意事项:由于这里已经有
S
i
g
m
o
i
d
Sigmoid
Sigmoid函数,所以网络最后不能再加
Sigmoid
函数 - 实例:
inputs = torch.tensor([[1, 2], [2, 2], [3, 4], [4, 5]], dtype=torch.float)
target = torch.tensor([[1, 0], [1, 0], [0, 1], [0, 1]], dtype=torch.float)
target_bce = target
# inputs = torch.sigmoid(inputs)
weights = torch.tensor([1, 1], dtype=torch.float)
loss_f_none_w = nn.BCEWithLogitsLoss(weight=weights, reduction='none')
loss_f_sum = nn.BCEWithLogitsLoss(weight=weights, reduction='sum')
loss_f_mean = nn.BCEWithLogitsLoss(weight=weights, reduction='mean')
# forward
loss_none_w = loss_f_none_w(inputs, target_bce)
loss_sum = loss_f_sum(inputs, target_bce)
loss_mean = loss_f_mean(inputs, target_bce)
# view
print("\nweights: ", weights)
print(loss_none_w, loss_sum, loss_mean)
# weights: tensor([1., 1.])
# tensor([[0.3133, 2.1269],
# [0.1269, 2.1269],
# [3.0486, 0.0181],
# [4.0181, 0.0067]]) tensor(11.7856) tensor(1.4732)
pos_weight
参数:
inputs = torch.tensor([[1, 2], [2, 2], [3, 4], [4, 5]], dtype=torch.float)
target = torch.tensor([[1, 0], [1, 0], [0, 1], [0, 1]], dtype=torch.float)
target_bce = target
# itarget
# inputs = torch.sigmoid(inputs)
weights = torch.tensor([1], dtype=torch.float)
pos_w = torch.tensor([3], dtype=torch.float) # 3
loss_f_none_w = nn.BCEWithLogitsLoss(weight=weights, reduction='none', pos_weight=pos_w)
loss_f_sum = nn.BCEWithLogitsLoss(weight=weights, reduction='sum', pos_weight=pos_w)
loss_f_mean = nn.BCEWithLogitsLoss(weight=weights, reduction='mean', pos_weight=pos_w)
# forward
loss_none_w = loss_f_none_w(inputs, target_bce)
loss_sum = loss_f_sum(inputs, target_bce)
loss_mean = loss_f_mean(inputs, target_bce)
# view
print("\npos_weights: ", pos_w)
print(loss_none_w, loss_sum, loss_mean)
# pos_weights: tensor([3.])
# tensor([[0.9398, 2.1269],
# [0.3808, 2.1269],
# [3.0486, 0.0544],
# [4.0181, 0.0201]]) tensor(12.7158) tensor(1.5895)
可以看出,第二段代码加上pos_weight=3
后,输出结果中,为1的位置均乘上了3
4、其余14种损失函数介绍
(1)nn.L1Loss
nn.L1Loss(size_average=None,
reduce=None,
reduction='mean’)
- 功能:计算input和target之差的绝对值,可选返回同维度的张量或者是一个标量
- 计算公式:
ℓ ( x , y ) = L = { l 1 , … , l N } ⊤ , l n = ∣ x n − y n ∣ \ell(x, y)=L=\left\{l_{1}, \ldots, l_{N}\right\}^{\top}, \quad l_{n}=\left|x_{n}-y_{n}\right| ℓ(x,y)=L={l1,…,lN}⊤,ln=∣xn−yn∣ - 参数:
reduction
:计算模式,可为none/sum/mean
,"none"
表示逐个元素计算,"sum"
表示所有元素求和,返回标量,"mean"
表示加权平均,返回标量- 实例:
inputs = torch.ones((2, 2))
target = torch.ones((2, 2)) * 3
loss_f = nn.L1Loss(reduction='none')
loss = loss_f(inputs, target)
print("input:{}\ntarget:{}\nL1 loss:{}".format(inputs, target, loss))
# input:tensor([[1., 1.],
# [1., 1.]])
# target:tensor([[3., 3.],
# [3., 3.]])
# L1 loss:tensor([[2., 2.],
# [2., 2.]])
(2)nn.MSELoss
nn.MSELoss(size_average=None,
reduce=None,
reduction='mean’)
- 功能:计算input和target之差的平方,可选返回同维度的张量或者是一个标量
- 计算公式:
ℓ ( x , y ) = L = { l 1 , … , l N } ⊤ , l n = ( x n − y n ) 2 \ell(x, y)=L=\left\{l_{1}, \ldots, l_{N}\right\}^{\top}, \quad l_{n}=\left(x_{n}-y_{n}\right)^2 ℓ(x,y)=L={l1,…,lN}⊤,ln=(xn−yn)2 - 参数:
reduction
:计算模式,可为none/sum/mean
,"none"
表示逐个元素计算,"sum"
表示所有元素求和,返回标量,"mean"
表示加权平均,返回标量- 实例:
inputs = torch.ones((2, 2))
target = torch.ones((2, 2)) * 3
loss_f = nn.MSELoss(reduction='none')
loss = loss_f(inputs, target)
print("input:{}\ntarget:{}\nMSE loss:{}".format(inputs, target, loss))
# input:tensor([[1., 1.],
# [1., 1.]])
# target:tensor([[3., 3.],
# [3., 3.]])
# MSE loss:tensor([[4., 4.],
# [4., 4.]])
(3)SmoothL1Loss
nn.SmoothL1Loss(size_average=None,
reduce=None,
reduction='mean’)
- 功能:计算平滑 L 1 L1 L1损失,属于Huber Loss中的一种(固定参数 δ \delta δ固定为1)
- 计算公式与补充说明:
计算平滑 L 1 L1 L1损失,属于Huber Loss中的一种,Huber Loss 常用于回归问题,其最大的特点是对离群点(outliers)、噪声不敏感,具有较强的鲁棒性,其计算公式为:
L δ ( y , f ( x ) ) = { 1 2 ( y − f ( x ) ) 2 for ∣ y − f ( x ) ∣ ≤ δ δ ∣ y − f ( x ) ∣ − 1 2 δ 2 otherwise L_{\delta}(y, f(x))=\left\{\begin{array}{ll} \frac{1}{2}(y-f(x))^{2} & \text { for }|y-f(x)| \leq \delta \\ \delta|y-f(x)|-\frac{1}{2} \delta^{2} & \text { otherwise } \end{array}\right. Lδ(y,f(x))={21(y−f(x))2δ∣y−f(x)∣−21δ2 for ∣y−f(x)∣≤δ otherwise
这里的 δ \delta δ理解为误差控制率,即当误差绝对值小于 δ \delta δ,采用 L 2 L2 L2损失;若大于 δ \delta δ,采用 L 1 L1 L1损失。如果固定 δ = 1 \delta=1 δ=1,就是这里的SmoothL1Loss
,其计算公式就变为:
loss ( x , y ) = 1 n ∑ i z i \operatorname{loss}(x, y)=\frac{1}{n} \sum_{i} z_{i} loss(x,y)=n1i∑zi
where z i z_{i} zi is given by:
z i = { 0.5 ( x i − y i ) 2 , if ∣ x i − y i ∣ < 1 ∣ x i − y i ∣ − 0.5 , otherwise z_{i}=\left\{\begin{array}{ll} 0.5\left(x_{i}-y_{i}\right)^{2}, & \text { if }\left|x_{i}-y_{i}\right|<1 \\ \left|x_{i}-y_{i}\right|-0.5, & \text { otherwise } \end{array}\right. zi={0.5(xi−yi)2,∣xi−yi∣−0.5, if ∣xi−yi∣<1 otherwise
SmoothL1Loss
与L1Loss
的区别如下图所示:
- 参数:
reduction
:计算模式,可为none/sum/mean
,"none"
表示逐个元素计算,"sum"
表示所有元素求和,返回标量,"mean"
表示加权平均,返回标量- 实例:
inputs = torch.ones((2, 2))
target = torch.ones((2, 2)) * 3
loss_f = nn.SmoothL1Loss(reduction='none')
loss = loss_f(inputs, target)
print("input:{}\ntarget:{}\nSmooth L1 loss:{}".format(inputs, target, loss))
# input:tensor([[1., 1.],
# [1., 1.]])
# target:tensor([[3., 3.],
# [3., 3.]])
# Smooth L1 loss:tensor([[1.5000, 1.5000],
# [1.5000, 1.5000]])
(4)PoissonNLLLoss
nn.PoissonNLLLoss(log_input=True,
full=False,
size_average=None,
eps=1e-08,
reduce=None,
reduction='mean')
- 功能:计算泊松分布的负对数似然损失函数,用于target服从泊松分布的分类任务
- 计算公式:
{ loss = exp ( i n p u t ) − t a r g e t × i n p u t l o g _ i n p u t = T r u e loss = exp ( i n p u t ) − t a r g e t × log ( i n p u t + e p s ) l o g _ i n p u t = F a l s e \left\{ \begin{array}{lc} \operatorname{loss} = \exp(input) - target \times input & \qquad log\_input = True \\ \operatorname{loss} = \exp(input) - target \times \log (input+eps) & \qquad log\_input = False \\ \end{array}\right. {loss=exp(input)−target×inputloss=exp(input)−target×log(input+eps)log_input=Truelog_input=False
-
主要参数:
-
log_input
:输入是否为对数形式,决定计算公式 -
full
:计算所有loss,默认为False
-
eps
:修正项,避免log(input)
为nan
-
实例:
inputs = torch.randn((2, 2))
target = torch.randn((2, 2))
loss_f = nn.PoissonNLLLoss(log_input=True, full=False, reduction='none')
loss = loss_f(inputs, target)
print("input:{}\ntarget:{}\nPoisson NLL loss:{}".format(inputs, target, loss))
# --------------------------------- compute by hand
idx = 0
loss_1 = torch.exp(inputs[idx, idx]) - target[idx, idx]*inputs[idx, idx]
print("第一个元素loss:", loss_1)
# input:tensor([[ 0.7218, 0.0206],
# [-0.2426, 0.7005]])
# target:tensor([[ 8.7604e-01, 1.3784e+00],
# [-9.6065e-01, 2.9672e-04]])
# Poisson NLL loss:tensor([[1.4258, 0.9924],
# [0.5514, 2.0146]])
# 第一个元素loss: tensor(1.4258)
(5)nn.KLDivLoss
nn.KLDivLoss(size_average=None,
reduce=None,
reduction='mean')
- 功能:计算input和target之间的 K L KL KL散度(Kullback–Leibler divergence)
- 计算公式:
l ( x , y ) = L : = { l 1 , … , l N } , l n = y n ⋅ ( log y n − x n ) l(x, y)=L:=\left\{l_{1}, \ldots, l_{N}\right\}, \quad l_{n}=y_{n} \cdot\left(\log y_{n}-x_{n}\right) l(x,y)=L:={l1,…,lN},ln=yn⋅(logyn−xn) - 注意事项:
- 函数要求输入服从概率分布,即在0到1之间,因此需提前将输入计算log-probabilities,可以通过
nn.logsoftmax()
实现 - 从上面的公式可以看出,并没有对 x n x_n xn求 log \log log,这是与交叉熵不同的地方
- 主要参数:
reduction
:计算模式,可为none/sum/mean/batchsize
,"none"
表示逐个元素计算,"sum"
表示所有元素求和,返回标量,"mean"
表示加权平均,"batchsize"
返回标量batchsize维度求平均值- 实例:
inputs = torch.tensor([[0.5, 0.3, 0.2], [0.2, 0.3, 0.5]])
inputs_log = torch.log(inputs)
target = torch.tensor([[0.9, 0.05, 0.05], [0.1, 0.7, 0.2]], dtype=torch.float)
loss_f_none = nn.KLDivLoss(reduction='none')
loss_f_mean = nn.KLDivLoss(reduction='mean')
loss_f_bs_mean = nn.KLDivLoss(reduction='batchmean')
loss_none = loss_f_none(inputs, target)
loss_mean = loss_f_mean(inputs, target)
loss_bs_mean = loss_f_bs_mean(inputs, target)
print("loss_none:\n{}\nloss_mean:\n{}\nloss_bs_mean:\n{}".format(loss_none, loss_mean, loss_bs_mean))
# --------------------------------- compute by hand
idx = 0
loss_1 = target[idx, idx] * (torch.log(target[idx, idx]) - inputs[idx, idx])
print("第一个元素loss:", loss_1)
# loss_none:
# tensor([[-0.5448, -0.1648, -0.1598],
# [-0.2503, -0.4597, -0.4219]])
# loss_mean:
# -0.3335360586643219
# loss_bs_mean:
# -1.000608205795288
# 第一个元素loss: tensor(-0.5448)
说明:从上面的代码可以看出mean
和batchmean
的区别,上面的代码中如果是mean
,是除以6,而如果是batchmean
则是除以2
(6)nn.MarginRankingLoss
nn.MarginRankingLoss(margin=0.0,
size_average=None,
reduce=None,
reduction='mean')
- 功能:计算两个向量之间的相似度,当两个向量之间的距离大于
margin
,则 loss 为正,小于 margin,loss 为 0;用于排序任务该方法,计算两组数据之间的差异,返回一个$n\times n $的loss矩阵 - 计算公式与说明:
loss ( x , y ) = max ( 0 , − y ∗ ( x 1 − x 2 ) + margin ) \operatorname{loss}(x, y)=\max (0,-y *(x 1-x 2)+\operatorname{margin}) loss(x,y)=max(0,−y∗(x1−x2)+margin)
说明: y y y的取值为+1或-1,从上面的公式可以看出,当 y = 1 y = 1 y=1时,希望 x 1 x1 x1比 x 2 x2 x2大,当 x 1 > x 2 x1>x2 x1>x2时,不产生loss;而当 y = − 1 y = -1 y=−1时,希望 x 2 x2 x2比 x 1 x1 x1大,当 x 2 > x 1 x2>x1 x2>x1时,不产生loss - 主要参数:
margin
:边界值, x 1 x1 x1与 x 2 x2 x2之间的差异值- 实例
x1 = torch.tensor([[1], [2], [3]], dtype=torch.float)
x2 = torch.tensor([[2], [2], [2]], dtype=torch.float)
target = torch.tensor([1, 1, -1], dtype=torch.float)
loss_f_none = nn.MarginRankingLoss(margin=0, reduction='none')
loss = loss_f_none(x1, x2, target)
print(loss)
# tensor([[1., 1., 0.],
# [0., 0., 0.],
# [0., 0., 1.]])
(7)nn.MultiLabelMarginLoss
nn.MultiLabelMarginLoss(size_average=None,
reduce=None,
reduction='mean')
- 功能:用于一个样本属于多个类别时的分类任务。例如一个四分类任务,样本 x x x属于第 0 类,第 1 类,不属于第 2 类,第 3 类。
- 计算公式:
loss ( x , y ) = ∑ i j max ( 0 , 1 − ( x [ y [ j ] ] − x [ i ] ) ) x ⋅ size ( 0 ) \operatorname{loss}(x, y)=\sum_{i j} \frac{\max (0,1-(x[y[j]]-x[i]))}{x \cdot \operatorname{size}(0)} loss(x,y)=ij∑x⋅size(0)max(0,1−(x[y[j]]−x[i]))
x [ y [ j ] ] x[y[j]] x[y[j]]表示样本 x x x所属类的输出值, x [ i ] x[i] x[i]表示不等于该类的输出值。 - 参数:
reduction
:计算模式,可为none/sum/mean
,"none"
表示逐个元素计算,"sum"
表示所有元素求和,返回标量,"mean"
表示加权平均,返回标量- 实例:
x = torch.tensor([[0.1, 0.2, 0.4, 0.8]])
y = torch.tensor([[0, 3, -1, -1]], dtype=torch.long)
loss_f = nn.MultiLabelMarginLoss(reduction='none')
loss = loss_f(x, y)
print(loss)
# tensor([0.8500])
# --------------------------------- compute by hand
# flag = 0
x = x[0]
item_1 = (1-(x[0] - x[1])) + (1 - (x[0] - x[2])) # [0]
item_2 = (1-(x[3] - x[1])) + (1 - (x[3] - x[2])) # [3]
loss_h = (item_1 + item_2) / x.shape[0]
print(loss_h)
# tensor(0.8500)
(8)nn.SoftMarginLoss
nn.SoftMarginLoss(size_average=None,
reduce=None,
reduction='mean')
- 功能:计算二分类的logistic损失
- 计算公式:
loss ( x , y ) = ∑ i log ( 1 + exp ( − y [ i ] ∗ x [ i ] ) ) x.nelement() \operatorname{loss}(x, y)=\sum_{i} \frac{\log (1+\exp (-y[i] * x[i]))}{\text { x.nelement()}} loss(x,y)=i∑ x.nelement()log(1+exp(−y[i]∗x[i])) - 参数:
reduction
:计算模式,可为none/sum/mean
,"none"
表示逐个元素计算,"sum"
表示所有元素求和,返回标量,"mean"
表示加权平均,返回标量- 实例:
inputs = torch.tensor([[0.3, 0.7], [0.5, 0.5]])
target = torch.tensor([[-1, 1], [1, -1]], dtype=torch.float)
loss_f = nn.SoftMarginLoss(reduction='none')
loss = loss_f(inputs, target)
print("SoftMargin: ", loss)
# SoftMargin: tensor([[0.8544, 0.4032],
# [0.4741, 0.9741]])
# --------------------------------- compute by hand
idx = 0
inputs_i = inputs[idx, idx]
target_i = target[idx, idx]
loss_h = np.log(1 + np.exp(-target_i * inputs_i))
print(loss_h)
# tensor(0.8544)
(9)nn.MultiLabelSoftMarginLoss
nn.MultiLabelSoftMarginLoss(weight=None,
size_average=None,
reduce=None,
reduction='mean')
- 功能:
SoftMarginLoss
的多标签版本 - 计算公式:
loss ( x , y ) = − 1 C ∑ i [ A + B ] 其中: A = y [ i ] ⋅ log ( ( 1 + exp ( − x [ i ] ) ) − 1 ) B = ( 1 − y [ i ] ) ⋅ log ( exp ( − x [ i ] ) ( 1 + exp ( − x [ i ] ) ) ) C 为类别总数 \begin{aligned} & \qquad \qquad \operatorname{loss}(x, y)= - \frac{1}{C} \sum_i [A +B] \\ & \text{其中:} \\ & \qquad \qquad A = y[i] \cdot \log \left((1+\exp (-x[i]))^{-1}\right) \\ & \qquad \qquad B = (1-y[i]) \cdot\log \left(\frac{\exp (-x[i])}{(1+\exp (-x[i]))}\right) \\ & \qquad \qquad C \text{为类别总数} \end{aligned} loss(x,y)=−C1i∑[A+B]其中:A=y[i]⋅log((1+exp(−x[i]))−1)B=(1−y[i])⋅log((1+exp(−x[i]))exp(−x[i]))C为类别总数 - 主要参数:
weight
:为每个类别的loss设置权值。weight
必须是float
类型的tensor
,其长度要 于类别 C C C一致,即每一个类别都要设置有weight
。- 实例:
inputs = torch.tensor([[0.3, 0.7, 0.8]])
target = torch.tensor([[0, 1, 1]], dtype=torch.float)
loss_f = nn.MultiLabelSoftMarginLoss(reduction='none')
loss = loss_f(inputs, target)
print("MultiLabel SoftMargin: ", loss)
# MultiLabel SoftMargin: tensor([0.5429])
# --------------------------------- compute by hand
i_0 = torch.log(torch.exp(-inputs[0, 0]) / (1 + torch.exp(-inputs[0, 0])))
i_1 = torch.log(1 / (1 + torch.exp(-inputs[0, 1])))
i_2 = torch.log(1 / (1 + torch.exp(-inputs[0, 2])))
loss_h = (i_0 + i_1 + i_2) / -3
print(loss_h)
# tensor(0.5429)
(10)nn.MultiMarginLoss
nn.MultiMarginLoss(p=1,
margin=1.0,
weight=None,
size_average=None,
reduce=None,
reduction='mean')
- 功能:计算多分类的折页损失
- 计算公式:
loss ( x , y ) = ∑ i max ( 0 , w [ y ] ∗ ( margin − x [ y ] + x [ i ] ) ) p ) x . size ( 0 ) \operatorname{loss}(x, y)=\frac{\left.\sum_{i} \max (0, w[y] *(\operatorname{margin}-x[y]+x[i]))^{p}\right)}{\mathrm{x}.\operatorname{size}(0)} loss(x,y)=x.size(0)∑imax(0,w[y]∗(margin−x[y]+x[i]))p)
注意:$x \in{0, \cdots, \text { x.size }(0)-1}\ 且 且 且 \ y \in{0, \cdots, \text { y.size }(0)-1}$;同时 0 ≤ y [ j ] ≤ x . size ( 0 ) − 1 0 \leq y[j] \leq \mathrm{x} . \operatorname{size}(0)-1 0≤y[j]≤x.size(0)−1且 i ≠ y [ j ] i \neq y[j] i=y[j] for all i i i and j j j - 主要参数:
p(int)
:默认值为 1,仅可选 1 或者 2margin(float)
:默认值为 1weight(Tensor)
:为每个类别的loss设置权值。weight
必须是float
类型的tensor
,其长度要 于类别 C C C一致,即每一个类别都要设置有weight
- 实例:
x = torch.tensor([[0.1, 0.2, 0.7], [0.2, 0.5, 0.3]])
y = torch.tensor([1, 2], dtype=torch.long)
loss_f = nn.MultiMarginLoss(reduction='none')
loss = loss_f(x, y)
print("Multi Margin Loss: ", loss)
# Multi Margin Loss: tensor([0.8000, 0.7000])
# --------------------------------- compute by hand
x = x[0]
margin = 1
i_0 = margin - (x[1] - x[0])
# i_1 = margin - (x[1] - x[1])
i_2 = margin - (x[1] - x[2])
loss_h = (i_0 + i_2) / x.shape[0] # x.shape[0]=3
print(loss_h)
# tensor(0.8000)
(11)nn.TripletMarginLoss
nn.TripletMarginLoss(margin=1.0,
p=2.0,
eps=1e-06,
swap=False,
size_average=None,
reduce=None,
reduction='mean')
-
功能:计算三元组损失,人脸验证中常用。如下图 Anchor、Negative、Positive,目标是让 Positive 元和 Anchor 元之间的距离尽可能的小,Positive 元和 Negative 元之间的距离尽可能的大
-
计算公式:
L ( a , p , n ) = max { d ( a i , p i ) − d ( a i , n i ) + margin, 0 } 其中 d ( x i , y i ) = ∥ x i − y i ∥ p p \begin{aligned} & \qquad \qquad L(a, p, n)=\max \left\{d\left(a_{i}, p_{i}\right)-d\left(a_{i}, n_{i}\right)+\text { margin, } 0\right\} \\ & \text{其中} \\ & \qquad \qquad d\left(x_{i}, y_{i}\right)=\left\|\mathbf{x}_{i}-\mathbf{y}_{i}\right\|_{p_{p}} \end{aligned} L(a,p,n)=max{d(ai,pi)−d(ai,ni)+ margin, 0}其中d(xi,yi)=∥xi−yi∥pp -
主要参数:
-
margin(float)
:边界值,默认值为 1 -
p(int)
:范数的阶,默认为2 -
实例:
anchor = torch.tensor([[1.]])
pos = torch.tensor([[2.]])
neg = torch.tensor([[0.5]])
loss_f = nn.TripletMarginLoss(margin=1.0, p=1)
loss = loss_f(anchor, pos, neg)
print("Triplet Margin Loss", loss)
# Triplet Margin Loss tensor(1.5000)
# --------------------------------- compute by hand
margin = 1
a, p, n = anchor[0], pos[0], neg[0]
d_ap = torch.abs(a-p)
d_an = torch.abs(a-n)
loss = d_ap - d_an + margin
print(loss)
# tensor([1.5000])
(12)nn.HingeEmbeddingLoss
nn.HingeEmbeddingLoss(margin=1.0,
size_average=None,
reduce=None,
reduction='mean’)
- 功能:计算两个输入的相似性,常用于非线性embedding和半监督学习
- 注意:输入 x x x应为两个输入之差的绝对值
- 计算公式:
l n = { x n , if y n = 1 max { 0 , Δ − x n } , if y n = − 1 l_{n}=\left\{\begin{array}{ll} x_{n}, & \text { if } y_{n}=1 \\ \max \left\{0, \Delta-x_{n}\right\}, & \text { if } y_{n}=-1 \end{array}\right. ln={xn,max{0,Δ−xn}, if yn=1 if yn=−1 - 主要参数:
margin
:边界值,默认值为1- 实例:
inputs = torch.tensor([[1., 0.8, 0.5]])
target = torch.tensor([[1, 1, -1]])
loss_f = nn.HingeEmbeddingLoss(margin=1, reduction='none')
loss = loss_f(inputs, target)
print("Hinge Embedding Loss", loss)
# Hinge Embedding Loss tensor([[1.0000, 0.8000, 0.5000]])
# --------------------------------- compute by hand
margin = 1.
loss = max(0, margin - inputs.numpy()[0, 2])
print(loss)
# 0.5
(13)nn.CosineEmbeddingLoss
nn.CosineEmbeddingLoss(margin=0.0,
size_average=None,
reduce=None,
reduction='mean')
- 功能:采用余弦相似度计算两个输入的相似性(使用余弦相似度说明不关注大小的相似性而是关注方向的相似性)
- 计算公式:
loss ( x , y ) = { 1 − cos ( x 1 , x 2 ) , if y = 1 max ( 0 , cos ( x 1 , x 2 ) − margin ) , if y = − 1 \operatorname{loss}(x, y)=\left\{\begin{array}{ll} 1-\cos \left(x_{1}, x_{2}\right), & \text { if } y=1 \\ \max \left(0, \cos \left(x_{1}, x_{2}\right)-\operatorname{margin}\right), & \text { if } y=-1 \end{array}\right. loss(x,y)={1−cos(x1,x2),max(0,cos(x1,x2)−margin), if y=1 if y=−1 - 主要参数:
margin
:可取值 [ − 1 , 1 ] [-1, 1] [−1,1],推荐为 [ 0 , 0.5 ] [0, 0.5] [0,0.5]- 实例:
x1 = torch.tensor([[0.3, 0.5, 0.7], [0.3, 0.5, 0.7]])
x2 = torch.tensor([[0.1, 0.3, 0.5], [0.1, 0.3, 0.5]])
target = torch.tensor([[1, -1]], dtype=torch.float)
loss_f = nn.CosineEmbeddingLoss(margin=0., reduction='none')
loss = loss_f(x1, x2, target)
print("Cosine Embedding Loss", loss)
# Cosine Embedding Loss tensor([[0.0167, 0.9833]])
# --------------------------------- compute by hand
margin = 0.
def cosine(a, b):
numerator = torch.dot(a, b)
denominator = torch.norm(a, 2) * torch.norm(b, 2)
return float(numerator/denominator)
l_1 = 1 - (cosine(x1[0], x2[0]))
l_2 = max(0, cosine(x1[0], x2[0]))
print(l_1, l_2)
# 0.016662120819091797 0.9833378791809082
(14)nn.CTCLoss
nn.CTCLoss(blank=0,
reduction='mean',
zero_infinity=False)
- 功能:计算 C T C CTC CTC损失,解决时序类数据的分类
- 主要参数:
blank
:blank labelzero_infinity
:无穷大的值或梯度置0- 实例:
T = 50 # Input sequence length
C = 20 # Number of classes (including blank)
N = 16 # Batch size
S = 30 # Target sequence length of longest target in batch
S_min = 10 # Minimum target length, for demonstration purposes
# Initialize random batch of input vectors, for *size = (T,N,C)
inputs = torch.randn(T, N, C).log_softmax(2).detach().requires_grad_()
# Initialize random batch of targets (0 = blank, 1:C = classes)
target = torch.randint(low=1, high=C, size=(N, S), dtype=torch.long)
input_lengths = torch.full(size=(N,), fill_value=T, dtype=torch.long)
target_lengths = torch.randint(low=S_min, high=S, size=(N,), dtype=torch.long)
ctc_loss = nn.CTCLoss()
loss = ctc_loss(inputs, target, input_lengths, target_lengths)
print("CTC loss: ", loss)
# CTC loss: tensor(6.5103, grad_fn=<MeanBackward0>)
5、损失函数总结
PyTorch
提供了18种损失函数,即:
- nn.CrossEntropyLoss
- nn.NLLLoss
- nn.BCELoss
- nn.BCEWithLogitsLoss
- nn.L1Loss
- nn.MSELoss
- nn.SmoothL1Loss
- nn.PoissonNLLLoss
- nn.KLDivLoss
- nn.MarginRankingLoss
- nn.MultiLabelMarginLoss
- nn.SoftMarginLoss
- nn.MultiLabelSoftMarginLoss
- nn.MultiMarginLoss
- nn.TripletMarginLoss
- nn.HingeEmbeddingLoss
- nn.CosineEmbeddingLoss
- nn.CTCLoss