This loss combines a Sigmoid layer and the BCELoss in one single
class. This version is more numerically stable than using a plain Sigmoid
followed by a BCELoss as, by combining the operations into one layer,
we take advantage of the log-sum-exp trick for numerical stability.
Arguments
- 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.- 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.- pos_weight
(Tensor, optional): a weight of positive examples. Must be a vector with length equal to the number of classes.
Details
The unreduced (i.e. with reduction set to 'none') loss can be described as:
$$ \ell(x, y) = L = \{l_1,\dots,l_N\}^\top, \quad l_n = - w_n \left[ y_n \cdot \log \sigma(x_n) + (1 - y_n) \cdot \log (1 - \sigma(x_n)) \right], $$
where \(N\) is the batch size. If reduction is not 'none'
(default 'mean'), then
$$ \ell(x, y) = \begin{array}{ll} \mbox{mean}(L), & \mbox{if reduction} = \mbox{'mean';}\\ \mbox{sum}(L), & \mbox{if reduction} = \mbox{'sum'.} \end{array} $$
This is used for measuring the error of a reconstruction in for example
an auto-encoder. Note that the targets t[i] should be numbers
between 0 and 1.
It's possible to trade off recall and precision by adding weights to positive examples.
In the case of multi-label classification the loss can be described as:
$$ \ell_c(x, y) = L_c = \{l_{1,c},\dots,l_{N,c}\}^\top, \quad l_{n,c} = - w_{n,c} \left[ p_c y_{n,c} \cdot \log \sigma(x_{n,c}) + (1 - y_{n,c}) \cdot \log (1 - \sigma(x_{n,c})) \right], $$ 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.
For example, if a dataset contains 100 positive and 300 negative examples of a single class,
then pos_weight for the class should be equal to \(\frac{300}{100}=3\).
The loss would act as if the dataset contains \(3\times 100=300\) positive examples.
Shape
Input: \((N, *)\) where \(*\) means, any number of additional dimensions
Target: \((N, *)\), same shape as the input
Output: scalar. If
reductionis'none', then \((N, *)\), same shape as input.
Examples
if (torch_is_installed()) {
loss <- nn_bce_with_logits_loss()
input <- torch_randn(3, requires_grad = TRUE)
target <- torch_empty(3)$random_(1, 2)
output <- loss(input, target)
output$backward()
target <- torch_ones(10, 64, dtype = torch_float32()) # 64 classes, batch size = 10
output <- torch_full(c(10, 64), 1.5) # A prediction (logit)
pos_weight <- torch_ones(64) # All weights are equal to 1
criterion <- nn_bce_with_logits_loss(pos_weight = pos_weight)
criterion(output, target) # -log(sigmoid(1.5))
}
#> torch_tensor
#> 0.201413
#> [ CPUFloatType{} ]