def mrf_loss(
input: torch.Tensor,
target: torch.Tensor,
eps: float = 1e-8,
reduction: str = "mean",
) -> torch.Tensor:
"""Calculates the MRF loss. See :class:`pystiche.ops.MRFOperator` for details.
Args:
input: Input tensor.
target: Target tensor.
eps: Small value to avoid zero division. Defaults to ``1e-8``.
reduction: Reduction method of the output passed to
:func:`pystiche.misc.reduce`. Defaults to ``"mean"``.
Examples:
>>> input = torch.rand(256, 64, 3, 3)
>>> target = torch.rand(256, 64, 3, 3)
>>> score = F.mrf_loss(input, target)
"""
with torch.no_grad():
similarity = pystiche.cosine_similarity(input, target, eps=eps)
idcs = torch.argmax(similarity, dim=1)
target = torch.index_select(target, dim=0, index=idcs)
return mse_loss(input, target, reduction=reduction)
def test_cosine_similarity_batched_input():
torch.manual_seed(0)
x1 = torch.rand(2, 1, 256)
x2 = torch.rand(2, 1, 256)
eps = 1e-6
actual = pystiche.cosine_similarity(x1, x2, eps=eps, batched_input=True)
expected = F.cosine_similarity(x1, x2, dim=2, eps=eps).unsqueeze(2)
ptu.assert_allclose(actual, expected, rtol=1e-6)
def test_cosine_similarity():
torch.manual_seed(0)
input = torch.rand(1, 256)
target = torch.rand(1, 256)
eps = 1e-6
actual = pystiche.cosine_similarity(input, target, eps=eps)
expected = F.cosine_similarity(input, target, dim=1, eps=eps).unsqueeze(1)
ptu.assert_allclose(actual, expected, rtol=1e-6)
def test_shape(self):
torch.manual_seed(0)
input = torch.rand(2, 3, 4, 5)
target = torch.rand(2, 3, 4, 5)
actual = pystiche.cosine_similarity(input, target,
batched_input=True).size()
expected = (2, 3, 3)
assert actual == expected
def mrf_loss(
input: torch.Tensor,
target: torch.Tensor,
eps: float = 1e-8,
reduction: str = "mean",
batched_input: Optional[bool] = None,
) -> torch.Tensor:
r"""Calculates the MRF loss. See :class:`pystiche.loss.MRFLoss` for details.
Args:
input: Input of shape :math:`B \times S_1 \times N_1 \times \dots \times N_D`.
target: Target of shape :math:`B \times S_2 \times N_1 \times \dots \times N_D`.
eps: Small value to avoid zero division. Defaults to ``1e-8``.
reduction: Reduction method of the output passed to
:func:`pystiche.misc.reduce`. Defaults to ``"mean"``.
batched_input: If ``False``, treat the first dimension of the inputs as sample
dimension, i.e. :math:`S \times N_1 \times \dots \times N_D`. Defaults to
``True``. See :func:`pystiche.cosine_similarity` for details.
Examples:
>>> import pystiche.loss.functional as F
>>> input = torch.rand(1, 256, 64, 3, 3)
>>> target = torch.rand(1, 128, 64, 3, 3)
>>> score = F.mrf_loss(input, target, batched_input=True)
"""
if batched_input is None:
msg = (
"The default value of batched_input has changed "
"from False to True in version 1.0.0. "
"To suppress this warning, pass the wanted behavior explicitly.")
warnings.warn(msg, UserWarning)
batched_input = True
with torch.no_grad():
similarity = pystiche.cosine_similarity(input,
target,
eps=eps,
batched_input=batched_input)
index = torch.argmax(similarity, dim=-1)
index = index.view(*index.shape,
*[1] * (target.ndim - index.ndim)).expand(
*[-1] * index.ndim, *target.shape[index.ndim:])
target = torch.gather(target, 1 if batched_input else 0, index)
return mse_loss(input, target, reduction=reduction)
def mrf_loss(
input: torch.Tensor,
target: torch.Tensor,
eps: float = 1e-8,
reduction: str = "mean",
batched_input: Optional[bool] = None,
) -> torch.Tensor:
r"""Calculates the MRF loss. See :class:`pystiche.ops.MRFOperator` for details.
Args:
input: Input of shape :math:`S_1 \times N_1 \times \dots \times N_D`.
target: Target of shape :math:`S_2 \times N_1 \times \dots \times N_D`.
eps: Small value to avoid zero division. Defaults to ``1e-8``.
reduction: Reduction method of the output passed to
:func:`pystiche.misc.reduce`. Defaults to ``"mean"``.
batched_input: If ``True``, treat the first dimension of the inputs as batch
dimension, i.e. :math:`B \times S \times N_1 \times \dots \times N_D`.
Defaults to ``False``. See :func:`pystiche.cosine_similarity` for details.
Note:
The default value of ``batched_input`` will change from ``False`` to ``True``
in the future.
Examples:
>>> input = torch.rand(256, 64, 3, 3)
>>> target = torch.rand(256, 64, 3, 3)
>>> score = F.mrf_loss(input, target)
"""
with torch.no_grad():
similarity = pystiche.cosine_similarity(input,
target,
eps=eps,
batched_input=batched_input)
idcs = torch.argmax(similarity, dim=1)
target = torch.index_select(target, dim=0, index=idcs)
return mse_loss(input, target, reduction=reduction)
def test_cosine_similarity_shape():
torch.manual_seed(0)
input = torch.rand(2, 3, 4, 5)
target = torch.rand(2, 3, 4, 5)
assert pystiche.cosine_similarity(input, target).size() == (2, 2)
def test_cosine_similarity_future_warning():
x1 = torch.empty(1, 2)
x2 = torch.empty(1, 2)
with pytest.warns(FutureWarning):
pystiche.cosine_similarity(x1, x2)
def test_batched_input_not_specified(self):
input = torch.rand(2, 1, 256)
target = torch.rand(2, 1, 256)
with pytest.warns(UserWarning):
pystiche.cosine_similarity(input, target)
