Python v0.3.0

KLDivLoss

class dragon.vm.torch.nn.KLDivLoss(
  size_average=None,
  reduce=None,
  reduction='mean',
  log_target=False
)[source]

Compute the Kullback-Leibler divergence.

Examples:

m = torch.nn.KLDivLoss()
eps = 1e-12  # Epsilon to avoid log(0)
# Compute KL(P || Q)
q = torch.tensor([0.0, 0.1, 0.2, 0.3, 1.0])
p = torch.tensor([0.0, 0.3, 0.2, 0.1, 0.9])
loss = m(torch.log(torch.clamp(q, eps)), torch.clamp(p, eps))

See also

torch.nn.functional.kl_div(…)

__init__

KLDivLoss.__init__(
  size_average=None,
  reduce=None,
  reduction='mean',
  log_target=False
)[source]

Create a KDivLoss module.

Parameters:
  • size_average (bool, optional) – True to set the reduction to ‘mean’.
  • reduce (bool, optional) – True to set the reduction to ‘sum’ or ‘mean’.
  • reduction ({'none', 'batchmean', 'mean', 'sum'}, optional) – The reduce method.
  • log_target (bool, optional, default=False) – The flag indicating whether target is passed in log space.