Attention Refinement Module (ARM) to refine the features of each stage. Args: in_channels (int): The number of input channels. out_channels (int): The number of output channels. Returns: x_out (torch.Tensor): Feature map of Attention Refinement Module.
| 82 | |
| 83 | |
| 84 | class AttentionRefinementModule(BaseModule): |
| 85 | """Attention Refinement Module (ARM) to refine the features of each stage. |
| 86 | |
| 87 | Args: |
| 88 | in_channels (int): The number of input channels. |
| 89 | out_channels (int): The number of output channels. |
| 90 | Returns: |
| 91 | x_out (torch.Tensor): Feature map of Attention Refinement Module. |
| 92 | """ |
| 93 | |
| 94 | def __init__(self, |
| 95 | in_channels, |
| 96 | out_channel, |
| 97 | conv_cfg=None, |
| 98 | norm_cfg=dict(type='BN'), |
| 99 | act_cfg=dict(type='ReLU'), |
| 100 | init_cfg=None): |
| 101 | super(AttentionRefinementModule, self).__init__(init_cfg=init_cfg) |
| 102 | self.conv_layer = ConvModule( |
| 103 | in_channels=in_channels, |
| 104 | out_channels=out_channel, |
| 105 | kernel_size=3, |
| 106 | stride=1, |
| 107 | padding=1, |
| 108 | conv_cfg=conv_cfg, |
| 109 | norm_cfg=norm_cfg, |
| 110 | act_cfg=act_cfg) |
| 111 | self.atten_conv_layer = nn.Sequential( |
| 112 | nn.AdaptiveAvgPool2d((1, 1)), |
| 113 | ConvModule( |
| 114 | in_channels=out_channel, |
| 115 | out_channels=out_channel, |
| 116 | kernel_size=1, |
| 117 | bias=False, |
| 118 | conv_cfg=conv_cfg, |
| 119 | norm_cfg=norm_cfg, |
| 120 | act_cfg=None), nn.Sigmoid()) |
| 121 | |
| 122 | def forward(self, x): |
| 123 | x = self.conv_layer(x) |
| 124 | x_atten = self.atten_conv_layer(x) |
| 125 | x_out = x * x_atten |
| 126 | return x_out |
| 127 | |
| 128 | |
| 129 | class ContextPath(BaseModule): |