P2P-Net/model.py at main · yangxh11/P2P-Net · GitHub

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
import torch.nn as nn
import torch
import numpy as np
import torch.nn.functional as F
from anchors import generate_default_anchor_maps, hard_nms
from clustering import PartsResort


class PMG(nn.Module):
    def __init__(self, model, feature_size, num_ftrs, classes_num, topn):
        super(PMG, self).__init__()

        self.backbone = model
        self.num_ftrs = num_ftrs
        self.topn = topn
        self.im_sz = 448
        self.pad_side = 224
        self.PR = PartsResort(self.topn, self.num_ftrs//2)

        self.proposal_net = ProposalNet(self.num_ftrs)
        _, edge_anchors, _ = generate_default_anchor_maps()
        self.edge_anchors = (edge_anchors+self.pad_side).astype(np.int)

        # mlp for regularization
        self.reg_mlp1 = nn.Sequential(
            nn.Linear(self.num_ftrs//2 * self.topn, self.num_ftrs//2),
            nn.ELU(inplace=True),
            nn.Linear(self.num_ftrs//2, self.num_ftrs//2)
        )
        self.reg_mlp2 = nn.Sequential(
            nn.Linear(self.num_ftrs//2 * self.topn, self.num_ftrs//2),
            nn.ELU(inplace=True),
            nn.Linear(self.num_ftrs//2, self.num_ftrs//2)
        )
        self.reg_mlp3 = nn.Sequential(
            nn.Linear(self.num_ftrs//2 * self.topn, self.num_ftrs//2),
            nn.ELU(inplace=True),
            nn.Linear(self.num_ftrs//2, self.num_ftrs//2)
        )

        # stage 1
        self.conv_block1 = nn.Sequential(
            BasicConv(self.num_ftrs//4, feature_size, kernel_size=1, stride=1, padding=0, relu=True),
            BasicConv(feature_size, self.num_ftrs//2, kernel_size=3, stride=1, padding=1, relu=True),
            nn.AdaptiveMaxPool2d(1)
        )
        self.classifier1 = nn.Sequential(
            nn.BatchNorm1d(self.num_ftrs//2),
            nn.Linear(self.num_ftrs//2, feature_size),
            nn.BatchNorm1d(feature_size),
            nn.ELU(inplace=True),
            nn.Dropout(0.5),
            nn.Linear(feature_size, classes_num),
        )

        # stage 2
        self.conv_block2 = nn.Sequential(
            BasicConv(self.num_ftrs//2, feature_size, kernel_size=1, stride=1, padding=0, relu=True),
            BasicConv(feature_size, self.num_ftrs//2, kernel_size=3, stride=1, padding=1, relu=True),
            nn.AdaptiveMaxPool2d(1)
        )
        self.classifier2 = nn.Sequential(
            nn.BatchNorm1d(self.num_ftrs//2),
            nn.Linear(self.num_ftrs//2, feature_size),
            nn.BatchNorm1d(feature_size),
            nn.ELU(inplace=True),
            nn.Dropout(0.5),
            nn.Linear(feature_size, classes_num),
        )

        # stage 3
        self.conv_block3 = nn.Sequential(
            BasicConv(self.num_ftrs, feature_size, kernel_size=1, stride=1, padding=0, relu=True),
            BasicConv(feature_size, self.num_ftrs//2, kernel_size=3, stride=1, padding=1, relu=True),
            nn.AdaptiveMaxPool2d(1)
        )
        self.classifier3 = nn.Sequential(
            nn.BatchNorm1d(self.num_ftrs//2),
            nn.Linear(self.num_ftrs//2, feature_size),
            nn.BatchNorm1d(feature_size),
            nn.ELU(inplace=True),
            nn.Dropout(0.5),
            nn.Linear(feature_size, classes_num),
        )

        # concat features from different stages
        self.classifier_concat = nn.Sequential(
            nn.BatchNorm1d(self.num_ftrs//2 * 3),
            nn.Linear(self.num_ftrs//2 * 3, feature_size),
            nn.BatchNorm1d(feature_size),
            nn.ELU(inplace=True),
            nn.Linear(feature_size, classes_num),
        )

    def forward(self, x, is_train=True):
        _, _, f1, f2, f3 = self.backbone(x)

        batch = x.shape[0]
        rpn_score = self.proposal_net(f3.detach())
        all_cdds = [np.concatenate((x.reshape(-1, 1),
                    self.edge_anchors.copy(),
                    np.arange(0, len(x)).reshape(-1, 1)),
                    axis=1) for x in rpn_score.data.cpu().numpy()]
        top_n_cdds = np.array([hard_nms(x, self.topn, iou_thresh=0.25) for x in all_cdds])
        top_n_index = top_n_cdds[:, :, -1].astype(np.int)
        top_n_index = torch.from_numpy(top_n_index).long().to(x.device)
        top_n_prob = torch.gather(rpn_score, dim=1, index=top_n_index)

        # re-input salient parts
        part_imgs = torch.zeros([batch, self.topn, 3, 224, 224]).to(x.device)
        x_pad = F.pad(x, (self.pad_side, self.pad_side, self.pad_side, self.pad_side), mode='constant', value=0)
        for i in range(batch):
            for j in range(self.topn):
                [y0, x0, y1, x1] = top_n_cdds[i, j, 1:5].astype(np.int)
                part_imgs[i:i + 1, j] = F.interpolate(x_pad[i:i + 1, :, y0:y1, x0:x1],
                                                        size=(224, 224), mode='bilinear',
                                                        align_corners=True)

        part_imgs = part_imgs.view(batch*self.topn, 3, 224, 224)
        _, _, f1_part, f2_part, f3_part = self.backbone(part_imgs.detach())
        f1_part = self.conv_block1(f1_part).view(batch*self.topn, -1)
        f2_part = self.conv_block2(f2_part).view(batch*self.topn, -1)
        f3_part = self.conv_block3(f3_part).view(batch*self.topn, -1)
        yp1 = self.classifier1(f1_part)
        yp2 = self.classifier2(f2_part)
        yp3 = self.classifier3(f3_part)
        yp4 = self.classifier_concat(torch.cat((f1_part, f2_part, f3_part), -1))


        # resort parts
        feature_points = f3_part.view(batch, self.topn, -1)
        parts_order = self.PR.classify(feature_points.data.cpu().numpy(), is_train)
        parts_order = torch.from_numpy(parts_order).long().to(x.device)
        parts_order = parts_order.unsqueeze(2).expand(batch, self.topn, self.num_ftrs//2)

        f1_points = torch.gather(f1_part.view(batch, self.topn, -1), dim=1, index=parts_order)
        f1_m = self.reg_mlp1(f1_points.view(batch, -1))
        f2_points = torch.gather(f2_part.view(batch, self.topn, -1), dim=1, index=parts_order)
        f2_m = self.reg_mlp2(f2_points.view(batch, -1))
        f3_points = torch.gather(f3_part.view(batch, self.topn, -1), dim=1, index=parts_order)
        f3_m = self.reg_mlp3(f3_points.view(batch, -1))

        # stage-wise classification
        f1 = self.conv_block1(f1).view(batch, -1)
        f2 = self.conv_block2(f2).view(batch, -1)
        f3 = self.conv_block3(f3).view(batch, -1)
        y1 = self.classifier1(f1)
        y2 = self.classifier2(f2)
        y3 = self.classifier3(f3)
        y4 = self.classifier_concat(torch.cat((f1, f2, f3), -1))

        return y1, y2, y3, y4, yp1, yp2, yp3, yp4, top_n_prob, f1_m, f1, f2_m, f2, f3_m, f3


class BasicConv(nn.Module):
    def __init__(self, in_planes, out_planes, kernel_size, stride=1, padding=0, dilation=1, groups=1, relu=True, bn=True, bias=False):
        super(BasicConv, self).__init__()
        self.out_channels = out_planes
        self.conv = nn.Conv2d(in_planes, out_planes, kernel_size=kernel_size,
                              stride=stride, padding=padding, dilation=dilation, groups=groups, bias=bias)
        self.bn = nn.BatchNorm2d(out_planes, eps=1e-5,
                                 momentum=0.01, affine=True) if bn else None
        self.relu = nn.ReLU() if relu else None

    def forward(self, x):
        x = self.conv(x)
        if self.bn is not None:
            x = self.bn(x)
        if self.relu is not None:
            x = self.relu(x)
        return x


class ProposalNet(nn.Module):
    def __init__(self, depth):
        super(ProposalNet, self).__init__()
        self.down1 = nn.Conv2d(depth, 128, 3, 1, 1)
        self.down2 = nn.Conv2d(128, 128, 3, 2, 1)
        self.down3 = nn.Conv2d(128, 128, 3, 2, 1)
        self.ReLU = nn.ReLU()
        self.tidy1 = nn.Conv2d(128, 6, 1, 1, 0)
        self.tidy2 = nn.Conv2d(128, 6, 1, 1, 0)
        self.tidy3 = nn.Conv2d(128, 9, 1, 1, 0)
        # proposals: 14x14x6, 7x7x6, 4x4x9

    def forward(self, x):
        batch_size = x.size(0)
        d1 = self.ReLU(self.down1(x))
        d2 = self.ReLU(self.down2(d1))
        d3 = self.ReLU(self.down3(d2))
        t1 = self.tidy1(d1).view(batch_size, -1)
        t2 = self.tidy2(d2).view(batch_size, -1)
        t3 = self.tidy3(d3).view(batch_size, -1)
        return torch.cat((t1, t2, t3), dim=1)