class mindspore.ops.BCEWithLogitsLoss(reduction='mean')[source]

Adds sigmoid activation function to input logits, and uses the given logits to compute binary cross entropy between the logits and the label.

Sets input logits as \(X\), input label as \(Y\), input weight as \(W\), output as \(L\). Then,

\[\begin{split}\begin{array}{ll} \\ p_{ij} = sigmoid(X_{ij}) = \frac{1}{1 + e^{-X_{ij}}} \\ L_{ij} = -[Y_{ij} * log(p_{ij}) + (1 - Y_{ij})log(1 - p_{ij})] \end{array}\end{split}\]

\(i\) indicates the \(i^{th}\) sample, \(j\) indicates the category. Then,

\[\begin{split}\ell(x, y) = \begin{cases} L, & \text{if reduction} = \text{'none';}\\ \operatorname{mean}(L), & \text{if reduction} = \text{'mean';}\\ \operatorname{sum}(L), & \text{if reduction} = \text{'sum'.} \end{cases}\end{split}\]

\(\ell\) indicates the method of calculating the loss. There are three methods: the first method is to provide the loss value directly, the second method is to calculate the average value of all losses, and the third method is to calculate the sum of all losses.

This operator will multiply the output by the corresponding weight. The tensor weight assigns different weights to each piece of data in the batch, and the tensor pos_weight adds corresponding weights to the positive examples of each category.

In addition, it can trade off recall and precision by adding weights to positive examples. In the case of multi-label classification the loss can be described as:

\[\begin{split}\begin{array}{ll} \\ p_{ij,c} = sigmoid(X_{ij,c}) = \frac{1}{1 + e^{-X_{ij,c}}} \\ L_{ij,c} = -[P_{c}Y_{ij,c} * log(p_{ij,c}) + (1 - Y_{ij,c})log(1 - p_{ij,c})] \end{array}\end{split}\]

where c is the class number (c>1 for multi-label binary classification, c=1 for single-label binary classification), n is the number of the sample in the batch and \(p_c\) is the weight of the positive answer for the class c. \(p_c>1\) increases the recall, \(p_c<1\) increases the precision.


reduction (str) – Type of reduction to be applied to loss. The optional values are ‘mean’, ‘sum’, and ‘none’, not case sensitive. If ‘none’, do not perform reduction. Default:’mean’.

  • logits (Tensor) - Input logits. Data type must be float16 or float32. Tensor of shape \((N, *)\) where \(*\) means, any number of additional dimensions.

  • label (Tensor) - Ground truth label, has the same shape as logits. Data type must be float16 or float32.

  • weight (Tensor) - A rescaling weight applied to the loss of each batch element. It can be broadcast to a tensor with shape of logits. Data type must be float16 or float32.

  • pos_weight (Tensor) - A weight of positive examples. Must be a vector with length equal to the number of classes. It can be broadcast to a tensor with shape of logits. Data type must be float16 or float32.


Tensor or Scalar, if reduction is ‘none’, it’s a tensor with the same shape and type as input logits. Otherwise, the output is a scalar.

  • TypeError – If data type of any input is neither float16 nor float32.

  • ValueError – If weight or pos_weight can not be broadcast to a tensor with shape of logits.

  • ValueError – If reduction is not one of ‘none’, ‘mean’ or ‘sum’.

Supported Platforms:

Ascend GPU


>>> logits = Tensor(np.array([[-0.8, 1.2, 0.7], [-0.1, -0.4, 0.7]]), mindspore.float32)
>>> label = Tensor(np.array([[0.3, 0.8, 1.2], [-0.6, 0.1, 2.2]]), mindspore.float32)
>>> weight = Tensor(np.array([1.0, 1.0, 1.0]), mindspore.float32)
>>> pos_weight = Tensor(np.array([1.0, 1.0, 1.0]), mindspore.float32)
>>> loss = ops.BCEWithLogitsLoss()
>>> output = loss(logits, label, weight, pos_weight)
>>> print(output)