BiSeNetV2 model

[uyolo]

There are some differences between your implementation and others (PaddleSeg, mmsegmentation, BiSeNet)

For example:
Your implementation

semantic-segmentation/semseg/models/bisenetv2.py

Lines 139 to 181 in 492231c

 class AggregationLayer(nn.Module): 
     def __init__(self) -> None: 
         super().__init__() 
         self.left1 = nn.Sequential( 
             nn.Conv2d(128, 128, 3, 1, 1, groups=128, bias=False), 
             nn.BatchNorm2d(128), 
             nn.Conv2d(128, 128, 1, 1, 0, bias=False) 
         ) 
         self.left2 = nn.Sequential( 
             nn.Conv2d(128, 128, 3, 2, 1, bias=False), 
             nn.BatchNorm2d(128), 
             nn.AvgPool2d(3, 2, 1, ceil_mode=False) 
         ) 
  
         self.right1 = nn.Sequential( 
             nn.Conv2d(128, 128, 3, 1, 1, bias=False), 
             nn.BatchNorm2d(128), 
             nn.Upsample(scale_factor=4), 
             nn.Sigmoid() 
         ) 
         self.right2 = nn.Sequential( 
             nn.Conv2d(128, 128, 3, 1, 1, groups=128, bias=False), 
             nn.BatchNorm2d(128), 
             nn.Conv2d(128, 128, 1, 1, 0, bias=False), 
             nn.Sigmoid() 
         ) 
  
         self.up = nn.Upsample(scale_factor=4)            
         self.conv = ConvModule(128, 128, 3, 1, 1)        
  
  
     def forward(self, x_d, x_s): 
         x1 = self.left1(x_d) 
         x2 = self.left2(x_d) 
         x3 = self.right1(x_s) 
         x4 = self.right2(x_s) 
  
         left = x1 * x3 
         right = x2 * x4 
         right = self.up(right) 
         out = left + right 
  
         return self.conv(out) 

Others
https://github.com/PaddlePaddle/PaddleSeg/blob/13aac35b31b147d2dd9976821ade13dabebff685/paddleseg/models/bisenet.py#L253-L302

https://github.com/CoinCheung/BiSeNet/blob/f2b901599752ce50656d2e50908acecd06f7eb47/lib/models/bisenetv2.py#L224

AggregationLayer only uses one resize op, others have two resize ops