def get_reduction(m):
result = getattr(m, 'reduction', None)
if result is None:
result = _Reduction.legacy_get_string(getattr(m, 'sizeAverage', None),
True,
emit_warning=False)
assert result is not None
return result
def ms_ssim_loss(input,
target,
max_val,
filter_size=11,
k1=0.01,
k2=0.03,
sigma=1.5,
size_average=None,
reduce=None,
reduction='mean'):
# type: (Tensor, Tensor, float, int, float, float, float, Optional[bool], Optional[bool], str) -> Tensor
r"""ms_ssim_loss(input, target, max_val, filter_size, k1, k2,
sigma, size_average=None, reduce=None, reduction='mean') -> Tensor
Measures the multi-scale structural similarity index (MS-SSIM) error.
See :class:`~torch.nn.MSSSIMLoss` for details.
"""
dim = input.dim()
if dim != 4:
raise ValueError(
'Expected 4 dimensions (got {}) from input'.format(dim))
if input.size() != target.size():
raise ValueError(
'Expected input size ({}) to match target size ({}).'.format(
input.size(0), target.size(0)))
if size_average is not None or reduce is not None:
reduction = _Reduction.legacy_get_string(size_average, reduce)
_, channel, _, _ = input.size()
kernel = _fspecial_gaussian(filter_size, channel, sigma)
weights = torch.tensor([0.0448, 0.2856, 0.3001, 0.2363, 0.1333])
weights = weights.unsqueeze(-1).unsqueeze(-1)
levels = weights.size(0)
mssim = []
mcs = []
for _ in range(levels):
ssim, cs = _ssim(input, target, max_val, k1, k2, channel, kernel)
ssim = ssim.mean((2, 3))
cs = cs.mean((2, 3))
mssim.append(ssim)
mcs.append(cs)
input = avg_pool2d(input, (2, 2))
target = avg_pool2d(target, (2, 2))
mssim = torch.stack(mssim)
mcs = torch.stack(mcs)
p1 = mcs**weights
p2 = mssim**weights
ret = torch.prod(p1[:-1], 0) * p2[-1]
if reduction != 'none':
ret = torch.mean(ret) if reduction == 'mean' else torch.sum(ret)
return ret
def multilabel_nb_loss(nb_mat,
input,
target,
weight=None,
size_average=None,
reduce=None,
reduction='mean',
scaling_c=nb_scale):
# type: (Tensor, Tensor, Optional[Tensor], Optional[bool], Optional[bool], str) -> Tensor
# multilabel_nb_loss(input, target, weight=None, size_average=None) -> Tensor
if size_average is not None or reduce is not None:
reduction = _Reduction.legacy_get_string(size_average, reduce)
# An example of target labels: [person, horse, car] -> one-hot encoding
# For each class present in the true label, look up P(all labels|class).
# This corresponds to a full column in nb_mat (The printed, transposed
# version has them as rows instead).
# The loss for a particular class x decreases when there are occurrences
# of other classes c. Therefore a high value of P(x|c) will contribute
# to decreased loss for class x.
loss = -(target * logsigmoid(input) + (1 - target) * logsigmoid(-input))
# For each sample, find the class indices (where target is 1).
# If there is more than one class, update the loss via this equation:
# loss = loss - nb_mat[c,x], where c is found in company of x.
for row in range(target.shape[0]):
class_indices = target[row].nonzero()
if len(class_indices) > 1:
for x in class_indices:
for c in class_indices:
if not torch.equal(x, c):
loss[row, x] -= nb_mat[c, x] * scaling_c
if weight is not None:
loss = loss * weight
loss = loss.sum(dim=1) / input.size(1) # only return N loss values
if reduction == 'none':
ret = loss
elif reduction == 'mean':
ret = loss.mean()
elif reduction == 'sum':
ret = loss.sum()
else:
ret = input
raise ValueError(reduction + " is not valid")
torch.set_printoptions(profile="default")
return ret
def ssim_loss(input,
target,
max_val,
filter_size=11,
k1=0.01,
k2=0.03,
sigma=1.5,
size_average=None,
reduce=None,
reduction='mean'):
# type: (Tensor, Tensor, float, int, float, float, float, Optional[bool], Optional[bool], str) -> Tensor
r"""ssim_loss(input, target, max_val, filter_size, k1, k2,
sigma, size_average=None, reduce=None, reduction='mean') -> Tensor
Measures the structural similarity index (SSIM) error.
See :class:`~torch.nn.SSIMLoss` for details.
"""
if input.size() != target.size():
raise ValueError(
'Expected input size ({}) to match target size ({}).'.format(
input.size(0), target.size(0)))
if input.device != target.device:
raise RuntimeError(
f'The input device {input.device} and target device {target.device} do not match.'
)
dim = input.dim()
if dim == 2: # Expand dims if the inputs have too few of them. This does not copy data.
input = input.expand(1, 1, input.dim(-2), input.dim(-1))
target = target.expand(1, 1, target.dim(-2), target.dim(-1))
elif dim == 3:
input = input.expand(1, input.dim(-3), input.dim(-2), input.dim(-1))
target = target.expand(1, target.dim(-3), target.dim(-2),
target.dim(-1))
elif dim != 4:
raise ValueError('Expected 2, 3, or 4 dimensions (got {})'.format(dim))
if size_average is not None or reduce is not None:
reduction = _Reduction.legacy_get_string(size_average, reduce)
channel = input.size(1)
kernel = _fspecial_gaussian(filter_size, channel, sigma, input.device)
ret, _ = _ssim(input, target, max_val, k1, k2, channel, kernel)
if reduction != 'none':
ret = torch.mean(ret) if reduction == 'mean' else torch.sum(ret)
return ret
def ssim_loss(input,
target,
max_val,
filter_size=11,
k1=0.01,
k2=0.03,
sigma=1.5,
kernel=None,
size_average=None,
reduce=None,
reduction='mean'):
# type: (Tensor, Tensor, float, int, float, float, float, Tensor, Optional[bool], Optional[bool], str) -> Tensor
r"""ssim_loss(input, target, max_val, filter_size, k1, k2,
sigma, kernel=None, size_average=None, reduce=None, reduction='mean') -> Tensor
Measures the structural similarity index (SSIM) error.
See :class:`~torch.nn.SSIMLoss` for details.
"""
if input.size() != target.size():
raise ValueError(
'Expected input size ({}) to match target size ({}).'.format(
input.size(0), target.size(0)))
if size_average is not None or reduce is not None:
reduction = _Reduction.legacy_get_string(size_average, reduce)
dim = input.dim()
if dim == 2:
input = input.expand(1, 1, -1, -1)
target = target.expand(1, 1, -1, -1)
elif dim == 3:
input = input.expand(1, -1, -1, -1)
target = target.expand(1, -1, -1, -1)
elif dim != 4:
raise ValueError('Expected 2, 3, or 4 dimensions (got {})'.format(dim))
if kernel is None:
kernel = _fspecial_gaussian(filter_size, sigma)
kernel = kernel.to(device=input.device)
ret = _ssim(input, target, max_val, k1, k2, kernel)[0]
if reduction != 'none':
ret = torch.mean(ret) if reduction == 'mean' else torch.sum(ret)
return ret
def __init__(self,
weight=None,
size_average=True,
batch_average=True,
ignore_index=255,
cuda=False,
reduce=None):
super(SegmentationCELosses, self).__init__()
# if size_average is not None or reduce is not None:
self.reduction = _Reduction.legacy_get_string(size_average, reduce)
# else:
# self.reduction = reduction
self.ignore_index = ignore_index
self.weight = weight
self.size_average = size_average
self.batch_average = batch_average
self.cuda = cuda
def ms_ssim_loss(input,
target,
max_val,
filter_size=11,
k1=0.01,
k2=0.03,
sigma=1.5,
size_average=None,
reduce=None,
reduction='mean'):
# type: (Tensor, Tensor, float, int, float, float, float, Optional[bool], Optional[bool], str) -> Tensor
r"""ms_ssim_loss(input, target, max_val, filter_size, k1, k2,
sigma, size_average=None, reduce=None, reduction='mean') -> Tensor
Measures the multi-scale structural similarity index (MS-SSIM) error.
See :class:`~torch.nn.MSSSIMLoss` for details.
"""
if input.size() != target.size():
raise ValueError(
'Expected input size ({}) to match target size ({}).'.format(
input.size(0), target.size(0)))
if input.device != target.device:
raise RuntimeError(
f'The input device {input.device} and target device {target.device} do not match.'
)
dim = input.dim()
if dim == 2: # Expand does not copy data.
input = input.expand(1, 1, input.dim(-2), input.dim(-1))
target = target.expand(1, 1, target.dim(-2), target.dim(-1))
elif dim == 3:
input = input.expand(1, input.dim(-3), input.dim(-2), input.dim(-1))
target = target.expand(1, target.dim(-3), target.dim(-2),
target.dim(-1))
elif dim != 4:
raise ValueError('Expected 2, 3, or 4 dimensions (got {})'.format(dim))
if size_average is not None or reduce is not None:
reduction = _Reduction.legacy_get_string(size_average, reduce)
channel = input.size(1)
kernel = _fspecial_gaussian(filter_size, channel, sigma, input.device)
# It would be nice if this did not have to be initialized every time that the function was called.
# Copying from host (CPU) to device (GPU) is a major bottleneck.
weights = torch.tensor([0.0448, 0.2856, 0.3001, 0.2363, 0.1333],
device=input.device)
weights = weights.unsqueeze(-1).unsqueeze(-1)
levels = weights.size(0)
mssim = list()
mcs = list()
for _ in range(levels):
ssim, cs = _ssim(input, target, max_val, k1, k2, channel, kernel)
ssim = ssim.mean((2, 3))
cs = cs.mean((2, 3))
mssim.append(ssim)
mcs.append(cs)
input = avg_pool2d(input, (2, 2))
target = avg_pool2d(target, (2, 2))
mssim = torch.stack(mssim)
mcs = torch.stack(mcs)
p1 = mcs**weights
p2 = mssim**weights
ret = torch.prod(p1[:-1], 0) * p2[-1]
if reduction != 'none':
ret = torch.mean(ret) if reduction == 'mean' else torch.sum(ret)
return ret
