bce_with_logits

bce_with_logits(outputs, targets, size_average=True)

Computes the binary cross entropy between targets and output's logits.

See :class:~torch.nn.BCEWithLogitsLoss for details.

Arguments

outputs -- A torch FloatTensor of arbitrary shape with a 1 dimensional channel axis targets -- A binary torch LongTensor of the same size without the channel axis size_average -- By default, the losses are averaged over observations for each minibatch. However, if the field size_average is set to False, the losses are instead summed for each minibatch. - _Returns_:

A scalar tensor equal to the total loss of the output.

• Examples::

input = autograd.Variable(torch.randn(3), requires_grad=True) target = autograd.Variable(torch.FloatTensor(3).random_(2)) loss = bce_with_logits(input, target) loss.backward()

---

categorical_crossentropy

categorical_crossentropy(outputs, targets, size_average=True)

Computes the categorical crossentropy loss over some outputs and targets according the equation for the ith output

-log(output[target])

and is accumulated with a sum or average over all outputs.

• _Arguments_:

outputs -- The torch FloatTensor output from a model with the shape (N, C) where N is the number of outputs and C is the number of classes. targets -- The torch LongTensor indicies of the ground truth with the shape (N,) where N is the number of outputs and each target t is 0 <= t < C. size_average -- By default, the losses are averaged over observations for each minibatch. However, if the field size_average is set to False, the losses are instead summed for each minibatch. - _Returns_:

A scalar tensor equal to the total loss of the output.