-
Notifications
You must be signed in to change notification settings - Fork 4
/
loss.py
75 lines (62 loc) · 2.92 KB
/
loss.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
# | ||
# || |_
import torch
from torch.nn import LogSigmoid
from torch.nn.functional import _Reduction
from torch.nn.modules.loss import _WeightedLoss
from torch._jit_internal import weak_module, weak_script_method
nb_scale = 1e-2
logsigmoid = LogSigmoid()
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
class MultiLabelNBLoss(_WeightedLoss):
"""
Shape:
- Input: :math:`(N, C)` where `N` is the batch size and `C` is the number of classes.
- Target: :math:`(N, C)`, same shape as the input.
- Output: scalar. If `reduce` is False, then `(N)`.
"""
__constants__ = ['weight', 'reduction']
def __init__(self, mat, weight=None, size_average=None, reduce=None, reduction='mean', scaling_c=nb_scale):
super(MultiLabelNBLoss, self).__init__(weight, size_average, reduce, reduction)
self.nb_mat = mat
self.scaling_c = scaling_c
@weak_script_method
def forward(self, input, target):
return multilabel_nb_loss(self.nb_mat, input, target, weight=self.weight, reduction=self.reduction, scaling_c=self.scaling_c)