mindspore.nn.MultiLabelSoftMarginLoss

class mindspore.nn.MultiLabelSoftMarginLoss(weight=None, reduction='mean')

Calculates the MultiLabelSoftMarginLoss. The multi-label soft margin loss is a commonly used loss function in multi-label classification tasks where an input sample can belong to multiple classes. Given an input $x$ and binary labels $y$ of size $(N,C)$, where $N$ denotes the number of samples and $C$ denotes the number of classes.

$$\mathcal{loss}(\mathcal{x}\mathcal{,}\mathcal{y})=-\frac{1}{N}\frac{1}{C}\sum _{i=1}^N\sum _{j=1}^C(y_{ij}\log \frac{1}{1+e^{-x_{ij}}}+(1-y_{ij})\log \frac{e^{-x_{ij}}}{1+e^{-x_{ij}}})$$

where $x_{ij}$ represents the predicted score of sample $i$ for class $j$. $y_{ij}$ represents the binary label of sample $i$ for class $j$, where sample $i$ belongs to class $j$ if $y_{ij}=1$ , and sample $i$ does not belong to class $j$ if $y_{ij}=0$. For a multi-label classification task, each sample may have multiple labels with a value of 1 in the binary label $y$. weight will multiply to the loss of each class if given.

Parameters

• weight (Union[Tensor, int, float]) – The manual rescaling weight given to each class. Default: None .

• reduction (str, optional) –

  Apply specific reduction method to the output: 'none' , 'mean' , 'sum' . Default: 'mean' .

  • 'none': no reduction will be applied.

  • 'mean': compute and return the weighted mean of elements in the output.

  • 'sum': the output elements will be summed.

Inputs:

• x (Tensor) - A tensor of shape $(N,C)$, where N is batch size and C is number of classes.

• target (Tensor) - The label target Tensor which has the same shape as x.

Outputs:

Tensor, the data type is the same as x, if the reduction is 'none', its shape is (N), otherwise it is zero.

Raises

ValueError – If the rank of x or target is not 2.

Supported Platforms:

Ascend GPU CPU

Examples

>>> import mindspore as ms
>>> import mindspore.nn as nn
>>> x = ms.Tensor([[0.3, 0.6, 0.6], [0.9, 0.4, 0.2]])
>>> target = ms.Tensor([[0.0, 0.0, 1.0], [0.0, 0.0, 1.0]])
>>> loss = nn.MultiLabelSoftMarginLoss(reduction='mean')
>>> out = loss(x, target)
>>> print(out.asnumpy())
0.84693956