SE注意力机制，升级SEA注意力机制。

个人感觉SE机制 ，其实在通道数量大的情况下，表现非常好。但是基于RESNET网络的时候，提取边界特征等，效果不是很好。所以提出了SEA注意力机制。

下面是我代码实现,代码进行了注释，大家可以看的懂：

大家点点赞，求求了

class SEA_Block(nn.Module):
    def __init__(self, channel, reduction=16):
        super(SEA_Block, self).__init__()
        self.Max_pool = nn.AdaptiveMaxPool2d(1)
        self.avg_pool = nn.AdaptiveAvgPool2d(1)
        # reduced_channels = channel // reduction
        self.conv_reduce = nn.Conv2d(2 * channel, (2 * channel)//reduction, kernel_size=1, bias=False)
        self.relu = nn.ReLU(inplace=True)

        self.conv_expand_left = nn.Conv2d((2 * channel)//reduction, channel, kernel_size=1, bias=False)
        self.conv_expand_right = nn.Conv2d((2 * channel)//reduction, channel, kernel_size=1, bias=False)
        self.sigmoid = nn.Sigmoid()

    def forward(self, x):
        avg_pool = self.Max_pool(x) # (B, C, 1, 1)
        max_pool = self.avg_pool(x)  # (B, C, 1, 1)

        concat_pool = torch.cat([avg_pool, max_pool], dim=1)  # (B, 2C, 1, 1)

        reduced = self.relu(self.conv_reduce(concat_pool))  # (B, 2C//r, 1, 1)

        left = self.sigmoid(self.conv_expand_left(reduced))  # (B, C, 1, 1)
        right = self.sigmoid(self.conv_expand_right(reduced))  # (B, C, 1, 1)

        left_weighted = x * left  # 左分支权重
        right_right = left_weighted * right

        out = right_right + x
        return out