Image Segmentation
Transformers
PyTorch
upernet
Inference Endpoints
File size: 5,531 Bytes
b13b124
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import torch
import torch.nn as nn
import torch.nn.functional as F
from mmcv.cnn import ConvModule

from mmseg.ops import resize
from ..builder import HEADS
from .decode_head import BaseDecodeHead


class ACM(nn.Module):
    """Adaptive Context Module used in APCNet.

    Args:
        pool_scale (int): Pooling scale used in Adaptive Context
            Module to extract region fetures.
        fusion (bool): Add one conv to fuse residual feature.
        in_channels (int): Input channels.
        channels (int): Channels after modules, before conv_seg.
        conv_cfg (dict | None): Config of conv layers.
        norm_cfg (dict | None): Config of norm layers.
        act_cfg (dict): Config of activation layers.
    """

    def __init__(self, pool_scale, fusion, in_channels, channels, conv_cfg,
                 norm_cfg, act_cfg):
        super(ACM, self).__init__()
        self.pool_scale = pool_scale
        self.fusion = fusion
        self.in_channels = in_channels
        self.channels = channels
        self.conv_cfg = conv_cfg
        self.norm_cfg = norm_cfg
        self.act_cfg = act_cfg
        self.pooled_redu_conv = ConvModule(
            self.in_channels,
            self.channels,
            1,
            conv_cfg=self.conv_cfg,
            norm_cfg=self.norm_cfg,
            act_cfg=self.act_cfg)

        self.input_redu_conv = ConvModule(
            self.in_channels,
            self.channels,
            1,
            conv_cfg=self.conv_cfg,
            norm_cfg=self.norm_cfg,
            act_cfg=self.act_cfg)

        self.global_info = ConvModule(
            self.channels,
            self.channels,
            1,
            conv_cfg=self.conv_cfg,
            norm_cfg=self.norm_cfg,
            act_cfg=self.act_cfg)

        self.gla = nn.Conv2d(self.channels, self.pool_scale**2, 1, 1, 0)

        self.residual_conv = ConvModule(
            self.channels,
            self.channels,
            1,
            conv_cfg=self.conv_cfg,
            norm_cfg=self.norm_cfg,
            act_cfg=self.act_cfg)

        if self.fusion:
            self.fusion_conv = ConvModule(
                self.channels,
                self.channels,
                1,
                conv_cfg=self.conv_cfg,
                norm_cfg=self.norm_cfg,
                act_cfg=self.act_cfg)

    def forward(self, x):
        """Forward function."""
        pooled_x = F.adaptive_avg_pool2d(x, self.pool_scale)
        # [batch_size, channels, h, w]
        x = self.input_redu_conv(x)
        # [batch_size, channels, pool_scale, pool_scale]
        pooled_x = self.pooled_redu_conv(pooled_x)
        batch_size = x.size(0)
        # [batch_size, pool_scale * pool_scale, channels]
        pooled_x = pooled_x.view(batch_size, self.channels,
                                 -1).permute(0, 2, 1).contiguous()
        # [batch_size, h * w, pool_scale * pool_scale]
        affinity_matrix = self.gla(x + resize(
            self.global_info(F.adaptive_avg_pool2d(x, 1)), size=x.shape[2:])
                                   ).permute(0, 2, 3, 1).reshape(
                                       batch_size, -1, self.pool_scale**2)
        affinity_matrix = F.sigmoid(affinity_matrix)
        # [batch_size, h * w, channels]
        z_out = torch.matmul(affinity_matrix, pooled_x)
        # [batch_size, channels, h * w]
        z_out = z_out.permute(0, 2, 1).contiguous()
        # [batch_size, channels, h, w]
        z_out = z_out.view(batch_size, self.channels, x.size(2), x.size(3))
        z_out = self.residual_conv(z_out)
        z_out = F.relu(z_out + x)
        if self.fusion:
            z_out = self.fusion_conv(z_out)

        return z_out


@HEADS.register_module()
class APCHead(BaseDecodeHead):
    """Adaptive Pyramid Context Network for Semantic Segmentation.

    This head is the implementation of
    `APCNet <https://openaccess.thecvf.com/content_CVPR_2019/papers/\
    He_Adaptive_Pyramid_Context_Network_for_Semantic_Segmentation_\
    CVPR_2019_paper.pdf>`_.

    Args:
        pool_scales (tuple[int]): Pooling scales used in Adaptive Context
            Module. Default: (1, 2, 3, 6).
        fusion (bool): Add one conv to fuse residual feature.
    """

    def __init__(self, pool_scales=(1, 2, 3, 6), fusion=True, **kwargs):
        super(APCHead, self).__init__(**kwargs)
        assert isinstance(pool_scales, (list, tuple))
        self.pool_scales = pool_scales
        self.fusion = fusion
        acm_modules = []
        for pool_scale in self.pool_scales:
            acm_modules.append(
                ACM(pool_scale,
                    self.fusion,
                    self.in_channels,
                    self.channels,
                    conv_cfg=self.conv_cfg,
                    norm_cfg=self.norm_cfg,
                    act_cfg=self.act_cfg))
        self.acm_modules = nn.ModuleList(acm_modules)
        self.bottleneck = ConvModule(
            self.in_channels + len(pool_scales) * self.channels,
            self.channels,
            3,
            padding=1,
            conv_cfg=self.conv_cfg,
            norm_cfg=self.norm_cfg,
            act_cfg=self.act_cfg)

    def forward(self, inputs):
        """Forward function."""
        x = self._transform_inputs(inputs)
        acm_outs = [x]
        for acm_module in self.acm_modules:
            acm_outs.append(acm_module(x))
        acm_outs = torch.cat(acm_outs, dim=1)
        output = self.bottleneck(acm_outs)
        output = self.cls_seg(output)
        return output