Grid-R-CNN/mmdet/models/mask_heads/grid_head.py at 23ef988911b3b68fe8d5b21225995d6b392ee7b6 · STVIR/Grid-R-CNN · 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
import mmcv
import numpy as np
import pycocotools.mask as mask_util
import torch
import torch.nn as nn
import functools
import torch.nn.functional as F
from torch.distributed import get_world_size, get_rank
import random

from ..registry import HEADS
from ..utils import ConvModule
from ..utils import build_norm_layer
from mmdet.core import grid_target, random_jitter, bbox2roi

@HEADS.register_module
class GridHead(nn.Module):

    def __init__(self,
                 num_convs=8,
                 roi_feat_size=14,
                 in_channels=256,
                 conv_kernel_size=3,
                 conv_out_channels=256,
                 deconv_kernel = 4,
                 num_grids = 9,
                 class_agnostic=False,
                 conv_cfg=None,
                 norm_cfg=None):
        super(GridHead, self).__init__()
        self.num_convs = num_convs
        self.roi_feat_size = roi_feat_size  # WARN: not used and reserved
        self.in_channels = in_channels
        self.conv_kernel_size = conv_kernel_size
        self.conv_out_channels = conv_out_channels
        self.class_agnostic = class_agnostic
        self.conv_cfg = conv_cfg
        self.norm_cfg = norm_cfg
        self.num_grids = num_grids
        self.test_mode = False

        self.convs = []
        for i in range(self.num_convs):
            in_channels = (
                self.in_channels if i == 0 else self.conv_out_channels)
            strides = 2 if i==0 else 1
            padding = (self.conv_kernel_size - 1) // 2
            self.convs.append(
                nn.Sequential(nn.Conv2d(in_channels,self.conv_out_channels,self.conv_kernel_size,strides,padding),
                nn.GroupNorm(36,self.conv_out_channels),
                nn.ReLU(inplace=True)))
        self.convs = nn.Sequential(*self.convs)

        planes = self.conv_out_channels
        self.single_plane = self.conv_out_channels // num_grids

        self.updeconv1 = nn.ConvTranspose2d(planes,planes,kernel_size=deconv_kernel,stride=2,
                                            padding=(deconv_kernel - 2) // 2, groups=num_grids)
        self.sbn1 = nn.GroupNorm(num_grids,planes)
        self.updeconv2 = nn.ConvTranspose2d(planes, num_grids, kernel_size=deconv_kernel, stride=2,
                                            padding=(deconv_kernel - 2) // 2, groups=num_grids)

        self.neighborpoint = ((1, 3), (0, 2, 4), (1, 5), (0, 4, 6), (1, 3, 5, 7), (2, 4, 8), (3, 7), (4, 6, 8), (5, 7))
        self.num_edges = functools.reduce(lambda x, y: x + y, map(lambda x: len(x), self.neighborpoint))
        self.firstOrderConvs = []
        self.secondOrderConvs = []
        for _point in self.neighborpoint:
            _foc = [nn.Sequential(nn.Conv2d(self.single_plane, self.single_plane, 5, 1, 2,groups=self.single_plane),
                                            nn.Conv2d(self.single_plane,self.single_plane,1,1,0)) for _idx in range(len(_point))]
            _soc = [nn.Sequential(nn.Conv2d(self.single_plane, self.single_plane, 5, 1, 2,groups=self.single_plane),
                                            nn.Conv2d(self.single_plane,self.single_plane,1,1,0)) for _idx in range(len(_point))]
            self.firstOrderConvs.append(nn.Sequential(*_foc))
            self.secondOrderConvs.append(nn.Sequential(*_soc))

        self.firstOrderConvs = nn.Sequential(*self.firstOrderConvs)
        self.secondOrderConvs = nn.Sequential(*self.secondOrderConvs)

    def init_weights(self):
        for m in self.modules():
            if isinstance(m, nn.Conv2d) or isinstance(m, nn.Linear):
                nn.init.kaiming_normal_(m.weight.data)
                if m.bias is not None:
                    nn.init.constant_(m.bias, 0)
        for m in self.modules():
            if isinstance(m, nn.ConvTranspose2d):
                nn.init.normal_(m.weight.data, std=0.001)
                if m.bias is not None:
                    m.bias.data.zero_()
        nn.init.constant_(self.updeconv2.bias,-np.log(0.99/0.01))

    def forward(self, x):
        x = self.convs(x)

        first_order_x = [None] * self.num_grids
        for _idx, _point_idx in enumerate(self.neighborpoint):
            first_order_x[_idx] = x[:, _idx * self.single_plane:(_idx + 1) * self.single_plane]
            for _iidx, _neighbor_idx in enumerate(_point_idx):
                first_order_x[_idx] = first_order_x[_idx] + self.firstOrderConvs[_idx][_iidx](
                    x[:, _neighbor_idx * self.single_plane:(_neighbor_idx + 1) * self.single_plane])

        second_order_x = [None] * self.num_grids
        for _idx, _point_idx in enumerate(self.neighborpoint):
            second_order_x[_idx] = x[:, _idx * self.single_plane:(_idx + 1) * self.single_plane]
            for _iidx, _neighbor_idx in enumerate(_point_idx):
                second_order_x[_idx] = second_order_x[_idx] + self.secondOrderConvs[_idx][_iidx](
                    first_order_x[_neighbor_idx])

        x2 = torch.cat(second_order_x, dim=1)
        x2 = self.updeconv1(x2)
        x2 = nn.functional.relu(self.sbn1(x2),inplace=True)
        x2 = self.updeconv2(x2)

        if not self.test_mode:
            x1 = x
            x1 = self.updeconv1(x1)
            x1 = nn.functional.relu(self.sbn1(x1),inplace=True)
            x1 = self.updeconv2(x1)
        else:
            x1 = x2

        return x1, x2

    def get_target(self, sampling_results, rcnn_train_cfg):
        grid_targets = grid_target(sampling_results,rcnn_train_cfg)
        return grid_targets

    def loss(self, grid_pred1, grid_pred2, grid_targets):
        loss = dict()
        grid_loss = F.binary_cross_entropy_with_logits(grid_pred1, grid_targets) + F.binary_cross_entropy_with_logits(grid_pred2,grid_targets)
        grid_loss = grid_loss * 15

        loss['loss_grid'] = grid_loss
        return loss

    def get_bboxes(self, det_bboxes, grid_pred, img_meta):
        assert(det_bboxes.shape[0] == grid_pred.shape[0])
        det_bboxes = det_bboxes.cpu()
        cls_scores = det_bboxes[:,[4]]
        det_bboxes = det_bboxes[:,:4]
        grid_pred = torch.sigmoid(grid_pred).cpu()

        #expand pos_bboxes
        widths = det_bboxes[:,2] - det_bboxes[:,0]
        heights = det_bboxes[:,3] - det_bboxes[:,1]
        x1 = det_bboxes[:,0] - widths / 2
        y1 = det_bboxes[:,1] - heights / 2
        x2 = det_bboxes[:,2] + widths / 2
        y2 = det_bboxes[:,3] + heights / 2

        R,C,H,W = grid_pred.shape
        grid_pred = grid_pred.view(R*C,H*W)
        pred_scores,pred_position = grid_pred.max(dim=1)

        xs = pred_position % W
        ys = pred_position // W
        base = (0,14,28)
        for i in range(9):
            xs[i::9] = xs[i::9] + base[i//3]
            ys[i::9] = ys[i::9] + base[i%3]
        pred_scores,xs,ys = tuple(map(lambda x:x.view(R,C),[pred_scores,xs,ys]))

        grid_points = (xs.float() + 0.5) / (2*W) * widths.view(-1,1) * 2 + x1.view(-1,1), \
                        (ys.float() + 0.5) / (2*H) * heights.view(-1,1) * 2 + y1.view(-1,1)

        res_dets_x1 = (grid_points[0][:,[0,1,2]] * pred_scores[:,[0,1,2]]).sum(dim=1,keepdim=True) / (pred_scores[:,[0,1,2]].sum(dim=1,keepdim=True))
        res_dets_y1 = (grid_points[1][:,[0,3,6]] * pred_scores[:,[0,3,6]]).sum(dim=1,keepdim=True) / (pred_scores[:,[0,3,6]].sum(dim=1,keepdim=True))
        res_dets_x2 = (grid_points[0][:,[6,7,8]] * pred_scores[:,[6,7,8]]).sum(dim=1,keepdim=True) / (pred_scores[:,[6,7,8]].sum(dim=1,keepdim=True))
        res_dets_y2 = (grid_points[1][:,[2,5,8]] * pred_scores[:,[2,5,8]]).sum(dim=1,keepdim=True) / (pred_scores[:,[2,5,8]].sum(dim=1,keepdim=True))

        det_res = torch.cat([res_dets_x1,res_dets_y1,res_dets_x2,res_dets_y2,cls_scores],dim=1)
        det_res[:,[0,2]].clamp_(min=0,max=img_meta[0]['img_shape'][1]-1)
        det_res[:,[1,3]].clamp_(min=0,max=img_meta[0]['img_shape'][0]-1)

        return det_res