File size: 8,048 Bytes
9bf4bd7
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
# Copyright (c) OpenMMLab. All rights reserved.
from typing import Optional, Union

import torch
import torch.nn as nn

from mmocr.registry import MODELS


@MODELS.register_module()
class MaskedBalancedBCEWithLogitsLoss(nn.Module):
    """This loss combines a Sigmoid layers and a masked balanced BCE loss in
    one single class. It's AMP-eligible.

    Args:
        reduction (str, optional): The method to reduce the loss.
            Options are 'none', 'mean' and 'sum'. Defaults to 'none'.
        negative_ratio (float or int, optional): Maximum ratio of negative
            samples to positive ones. Defaults to 3.
        fallback_negative_num (int, optional): When the mask contains no
            positive samples, the number of negative samples to be sampled.
            Defaults to 0.
        eps (float, optional): Eps to avoid zero-division error.  Defaults to
            1e-6.
    """

    def __init__(self,
                 reduction: str = 'none',
                 negative_ratio: Union[float, int] = 3,
                 fallback_negative_num: int = 0,
                 eps: float = 1e-6) -> None:
        super().__init__()
        assert reduction in ['none', 'mean', 'sum']
        assert isinstance(negative_ratio, (float, int))
        assert isinstance(fallback_negative_num, int)
        assert isinstance(eps, float)
        self.eps = eps
        self.negative_ratio = negative_ratio
        self.reduction = reduction
        self.fallback_negative_num = fallback_negative_num
        self.loss = nn.BCEWithLogitsLoss(reduction=reduction)

    def forward(self,
                pred: torch.Tensor,
                gt: torch.Tensor,
                mask: Optional[torch.Tensor] = None) -> torch.Tensor:
        """Forward function.

        Args:
            pred (torch.Tensor): The prediction in any shape.
            gt (torch.Tensor): The learning target of the prediction in the
                same shape as pred.
            mask (torch.Tensor, optional): Binary mask in the same shape of
                pred, indicating positive regions to calculate the loss. Whole
                region will be taken into account if not provided. Defaults to
                None.

        Returns:
            torch.Tensor: The loss value.
        """

        assert pred.size() == gt.size() and gt.numel() > 0
        if mask is None:
            mask = torch.ones_like(gt)
        assert mask.size() == gt.size()

        positive = (gt * mask).float()
        negative = ((1 - gt) * mask).float()
        positive_count = int(positive.sum())
        if positive_count == 0:
            negative_count = min(
                int(negative.sum()), self.fallback_negative_num)
        else:
            negative_count = min(
                int(negative.sum()), int(positive_count * self.negative_ratio))

        assert gt.max() <= 1 and gt.min() >= 0
        loss = self.loss(pred, gt)
        positive_loss = loss * positive
        negative_loss = loss * negative

        negative_loss, _ = torch.topk(negative_loss.view(-1), negative_count)

        balance_loss = (positive_loss.sum() + negative_loss.sum()) / (
            positive_count + negative_count + self.eps)

        return balance_loss


@MODELS.register_module()
class MaskedBalancedBCELoss(MaskedBalancedBCEWithLogitsLoss):
    """Masked Balanced BCE loss.

    Args:
        reduction (str, optional): The method to reduce the loss.
            Options are 'none', 'mean' and 'sum'. Defaults to 'none'.
        negative_ratio (float or int): Maximum ratio of negative
            samples to positive ones. Defaults to 3.
        fallback_negative_num (int): When the mask contains no
            positive samples, the number of negative samples to be sampled.
            Defaults to 0.
        eps (float): Eps to avoid zero-division error.  Defaults to
            1e-6.
    """

    def __init__(self,
                 reduction: str = 'none',
                 negative_ratio: Union[float, int] = 3,
                 fallback_negative_num: int = 0,
                 eps: float = 1e-6) -> None:
        super().__init__()
        assert reduction in ['none', 'mean', 'sum']
        assert isinstance(negative_ratio, (float, int))
        assert isinstance(fallback_negative_num, int)
        assert isinstance(eps, float)
        self.eps = eps
        self.negative_ratio = negative_ratio
        self.reduction = reduction
        self.fallback_negative_num = fallback_negative_num
        self.loss = nn.BCELoss(reduction=reduction)

    def forward(self,
                pred: torch.Tensor,
                gt: torch.Tensor,
                mask: Optional[torch.Tensor] = None) -> torch.Tensor:
        """Forward function.

        Args:
            pred (torch.Tensor): The prediction in any shape.
            gt (torch.Tensor): The learning target of the prediction in the
                same shape as pred.
            mask (torch.Tensor, optional): Binary mask in the same shape of
                pred, indicating positive regions to calculate the loss. Whole
                region will be taken into account if not provided. Defaults to
                None.

        Returns:
            torch.Tensor: The loss value.
        """

        assert pred.max() <= 1 and pred.min() >= 0
        return super().forward(pred, gt, mask)


@MODELS.register_module()
class MaskedBCEWithLogitsLoss(nn.Module):
    """This loss combines a Sigmoid layers and a masked BCE loss in one single
    class. It's AMP-eligible.

    Args:
        eps (float): Eps to avoid zero-division error.  Defaults to
            1e-6.
    """

    def __init__(self, eps: float = 1e-6) -> None:
        super().__init__()
        assert isinstance(eps, float)
        self.eps = eps
        self.loss = nn.BCEWithLogitsLoss(reduction='none')

    def forward(self,
                pred: torch.Tensor,
                gt: torch.Tensor,
                mask: Optional[torch.Tensor] = None) -> torch.Tensor:
        """Forward function.

        Args:
            pred (torch.Tensor): The prediction in any shape.
            gt (torch.Tensor): The learning target of the prediction in the
                same shape as pred.
            mask (torch.Tensor, optional): Binary mask in the same shape of
                pred, indicating positive regions to calculate the loss. Whole
                region will be taken into account if not provided. Defaults to
                None.

        Returns:
            torch.Tensor: The loss value.
        """

        assert pred.size() == gt.size() and gt.numel() > 0
        if mask is None:
            mask = torch.ones_like(gt)
        assert mask.size() == gt.size()

        assert gt.max() <= 1 and gt.min() >= 0
        loss = self.loss(pred, gt)

        return (loss * mask).sum() / (mask.sum() + self.eps)


@MODELS.register_module()
class MaskedBCELoss(MaskedBCEWithLogitsLoss):
    """Masked BCE loss.

    Args:
        eps (float): Eps to avoid zero-division error.  Defaults to
            1e-6.
    """

    def __init__(self, eps: float = 1e-6) -> None:
        super().__init__()
        assert isinstance(eps, float)
        self.eps = eps
        self.loss = nn.BCELoss(reduction='none')

    def forward(self,
                pred: torch.Tensor,
                gt: torch.Tensor,
                mask: Optional[torch.Tensor] = None) -> torch.Tensor:
        """Forward function.

        Args:
            pred (torch.Tensor): The prediction in any shape.
            gt (torch.Tensor): The learning target of the prediction in the
                same shape as pred.
            mask (torch.Tensor, optional): Binary mask in the same shape of
                pred, indicating positive regions to calculate the loss. Whole
                region will be taken into account if not provided. Defaults to
                None.

        Returns:
            torch.Tensor: The loss value.
        """

        assert pred.max() <= 1 and pred.min() >= 0

        return super().forward(pred, gt, mask)