File size: 5,894 Bytes
2cd560a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import math
import mmcv
import torch
import torch.nn as nn
import torch.nn.functional as F

from ..builder import LOSSES


@LOSSES.register_module()
class RLELoss_poseur_old(nn.Module):
    ''' RLE Regression Loss
    '''

    def __init__(self, OUTPUT_3D=False, use_target_weight=True, size_average=True):
        super(RLELoss_poseur_old, self).__init__()
        self.size_average = size_average
        self.amp = 1 / math.sqrt(2 * math.pi)

    def logQ(self, gt_uv, pred_jts, sigma):
        return torch.log(sigma / self.amp) + torch.abs(gt_uv - pred_jts) / (math.sqrt(2) * sigma + 1e-9)

    def forward(self, output, target_uv, target_uv_weight):

        pred_jts = output.pred_jts
        sigma = output.sigma
        gt_uv = target_uv.reshape(pred_jts.shape)
        gt_uv_weight = target_uv_weight.reshape(pred_jts.shape)



        nf_loss = output.nf_loss * gt_uv_weight[:, :, :1]
        # print(gt_uv.min(), gt_uv.max())

        residual = True
        if residual:
            Q_logprob = self.logQ(gt_uv, pred_jts, sigma) * gt_uv_weight
            loss = nf_loss + Q_logprob

        if self.size_average and gt_uv_weight.sum() > 0:
            return loss.sum() / len(loss)
        else:
            return loss.sum()

@LOSSES.register_module()
class RLELoss_poseur(nn.Module):
    ''' RLE Regression Loss
    '''

    def __init__(self, OUTPUT_3D=False, use_target_weight=True, size_average=True):
        super(RLELoss_poseur, self).__init__()
        self.size_average = size_average
        self.amp = 1 / math.sqrt(2 * math.pi)

    def logQ(self, gt_uv, pred_jts, sigma):
        return torch.log(sigma / self.amp) + torch.abs(gt_uv - pred_jts) / (math.sqrt(2) * sigma + 1e-9)

    def forward(self, output, target_uvd, target_uvd_weight):

        pred_jts = output.pred_jts
        sigma = output.sigma
        gt_uv = target_uvd.reshape(pred_jts.shape)
        gt_uv_weight = target_uvd_weight.reshape(pred_jts.shape)

        # nf_loss = output.nf_loss * gt_uv_weight[:, :, :1]
        nf_loss = output.nf_loss * gt_uv_weight

        residual = True
        if residual:
            Q_logprob = self.logQ(gt_uv, pred_jts, sigma) * gt_uv_weight
            loss = nf_loss + Q_logprob

        if self.size_average and gt_uv_weight.sum() > 0:
            return loss.sum() / len(loss)
        else:
            return loss.sum()

@LOSSES.register_module()
class RLEOHKMLoss(nn.Module):
    ''' RLE Regression Loss
    '''

    def __init__(self, OUTPUT_3D=False, use_target_weight=True, size_average=True, topk=8, 
                    ori_weight = 1.0, ohkm_weight = 0.0):
        super(RLEOHKMLoss, self).__init__()
        self.size_average = size_average
        self.amp = 1 / math.sqrt(2 * math.pi)
        self.topk = topk
        self.ori_weight = ori_weight
        self.ohkm_weight = ohkm_weight
        self.neg_inf = -float("Inf")

    def logQ(self, gt_uv, pred_jts, sigma):
        return torch.log(sigma / self.amp) + torch.abs(gt_uv - pred_jts) / (math.sqrt(2) * sigma + 1e-9)

    def ohkm(self, loss, weight):
        # mask = weight == 0
        loss_value = loss.clone().detach()
        loss_value[weight == 0] = self.neg_inf
        _, topk_idx = torch.topk(
            loss_value, k=self.topk, dim=1, sorted=False)
        tmp_loss = torch.gather(loss, 1, topk_idx)
        tmp_weight = torch.gather(weight, 1, topk_idx)
        # tmp_loss[tmp_loss==-float("Inf")] = 0
        tmp_loss = tmp_loss * tmp_weight
        tmp_loss = tmp_loss.flatten(start_dim=1).sum(dim = 1)
        # tmp_weight = tmp_weight.flatten(start_dim=1).sum(dim = 1)
        # tmp_loss = tmp_loss / tmp_weight

        return tmp_loss.mean()

    def ori(self, loss, weight):
        # mask = weight == 0
        loss = loss * weight
        loss = loss.flatten(start_dim=1).sum(dim = 1)
        # weight = weight.flatten(start_dim=1).sum(dim = 1)

        return loss.mean()

    def forward(self, output, target_uv, target_uv_weight):

        pred_jts = output.pred_jts
        sigma = output.sigma
        gt_uv = target_uv.reshape(pred_jts.shape)
        gt_uv_weight = target_uv_weight.reshape(pred_jts.shape)

        # gt_uv_weight = gt_uv_weight[:, :, :1] 
        nf_loss = output.nf_loss
        q_loss = self.logQ(gt_uv, pred_jts, sigma)

        # nf_loss_ohkm = self.ohkm(nf_loss, gt_uv_weight)
        # q_loss_ohkm = self.ohkm(q_loss, gt_uv_weight)

        ori_loss = nf_loss + q_loss
        ohkm_loss = self.ohkm(ori_loss, gt_uv_weight)
        ori_loss = self.ori(ori_loss, gt_uv_weight)

        loss = self.ori_weight * ori_loss + self.ohkm_weight * ohkm_loss
        return loss #TODO mean?


        # nf_loss = output.nf_loss * gt_uv_weight


        # Q_logprob = self.logQ(gt_uv, pred_jts, sigma) * gt_uv_weight
        # loss = nf_loss + Q_logprob

        # return loss.sum() / len(loss)


@LOSSES.register_module()
class RLELoss3D(nn.Module):
    ''' RLE Regression Loss 3D
    '''

    def __init__(self, OUTPUT_3D=False, size_average=True):
        super(RLELoss3D, self).__init__()
        self.size_average = size_average
        self.amp = 1 / math.sqrt(2 * math.pi)

    def logQ(self, gt_uv, pred_jts, sigma):
        return torch.log(sigma / self.amp) + torch.abs(gt_uv - pred_jts) / (math.sqrt(2) * sigma + 1e-9)

    def forward(self, output, labels):
        nf_loss = output.nf_loss
        pred_jts = output.pred_jts
        sigma = output.sigma
        gt_uv = labels['target_uvd'].reshape(pred_jts.shape)
        gt_uv_weight = labels['target_uvd_weight'].reshape(pred_jts.shape)
        nf_loss = nf_loss * gt_uv_weight

        residual = True
        if residual:
            Q_logprob = self.logQ(gt_uv, pred_jts, sigma) * gt_uv_weight
            loss = nf_loss + Q_logprob

        if self.size_average and gt_uv_weight.sum() > 0:
            return loss.sum() / len(loss)
        else:
            return loss.sum()