def test_reduce():
torch.manual_seed(0)
x = torch.rand(1, 3, 128, 128)
actual = misc.reduce(x, "mean")
desired = torch.mean(x)
ptu.assert_allclose(actual, desired)
actual = misc.reduce(x, "sum")
desired = torch.sum(x)
ptu.assert_allclose(actual, desired)
actual = misc.reduce(x, "none")
desired = x
ptu.assert_allclose(actual, desired)
def value_range_loss(input: torch.Tensor,
min: float = 0.0,
max: float = 1.0,
reduction: str = "mean") -> torch.Tensor:
# TODO: remove sinh call; quadratic, i.e. x * (x-1) is enough
loss = relu(torch.sinh(input - min) * torch.sinh(input - max))
return reduce(loss, reduction)
def total_variation_loss(input: torch.Tensor,
exponent: float = 2.0,
reduction: str = "mean") -> torch.Tensor:
r"""Calculates the total variation loss. See
:class:`pystiche.ops.TotalVariationOperator` for details.
Args:
input: Input image
exponent: Parameter :math:`\beta` . A higher value leads to more smoothed
results. Defaults to ``2.0``.
reduction: Reduction method of the output passed to
:func:`pystiche.misc.reduce`. Defaults to ``"mean"``.
Examples:
>>> import pystiche.loss.functional as F
>>> input = torch.rand(2, 3, 256, 256)
>>> score = F.total_variation_loss(input)
"""
# this ignores the last row and column of the image
grad_vert = input[:, :, :-1, :-1] - input[:, :, 1:, :-1]
grad_horz = input[:, :, :-1, :-1] - input[:, :, :-1, 1:]
grad = pystiche.nonnegsqrt(grad_vert**2.0 + grad_horz**2.0)
loss = grad**exponent
return reduce(loss, reduction)
