pytorch 实现 Restormer 主要模块（多头通道自注意力机制和门控制结构）

        前面的博文读论文：Restormer: Efficient Transformer for High-Resolution Image Restoration 介绍了 Restormer 网络结构的网络技术特点，本文用 pytorch 实现其中的主要网络结构模块。

1. MDTA(Multi-Dconv Head Transposed Attention：多头注意力机制

## Multi-DConv Head Transposed Self-Attention (MDTA)
class Attention(nn.Module):
    def __init__(self, dim, num_heads, bias):
        super(Attention, self).__init__()
        self.num_heads = num_heads  # 注意力头的个数
        self.temperature = nn.Parameter(torch.ones(num_heads, 1, 1))  # 可学习系数
        
        # 1*1 升维
        self.qkv = nn.Conv2d(dim, dim*3, kernel_size=1, bias=bias)
        # 3*3 分组卷积
        self.qkv_dwconv = nn.Conv2d(dim*3, dim*3, kernel_size=3, stride=1, padding=1, groups=dim*3, bias=bias)
        # 1*1 卷积
        self.project_out = nn.Conv2d(dim, dim, kernel_size=1, bias=bias)

    def forward(self, x):
        b,c,h,w = x.shape  # 输入的结构 batch 数，通道数和高宽

        qkv = self.qkv_dwconv(self.qkv(x))
        q,k,v = qkv.chunk(3, dim=1)  #  第 1 个维度方向切分成 3 块
        # 改变 q, k, v 的结构为 b head c (h w)，将每个二维 plane 展平
        q = rearrange(q, 'b (head c) h w -> b head c (h w)', head=self.num_heads)
        k = rearrange(k, 'b (head c) h w -> b head c (h w)', head=self.num_heads)
        v = rearrange(v, 'b (head c) h w -> b head c (h w)', head=self.num_heads)

        q = torch.nn.functional.normalize(q, dim=-1)  # C 维度标准化，这里的 C 与通道维度略有不同
        k = torch.nn.functional.normalize(k, dim=-1)

        attn = (q @ k.transpose(-2, -1)) * self.temperature # @ 是矩阵乘
        attn = attn.softmax(dim=-1)

        out = (attn @ v)  # 注意力图(严格来说不算图)
        
        # 将展平后的注意力图恢复
        out = rearrange(out, 'b head c (h w) -> b (head c) h w', head=self.num_heads, h=h, w=w)
        # 真正的注意力图
        out = self.project_out(out)
        return out

2. GDFN( Gated-Dconv Feed-Forward Network) 

## Gated-Dconv Feed-Forward Network (GDFN)
class FeedForward(nn.Module):
    def __init__(self, dim, ffn_expansion_factor, bias):
        super(FeedForward, self).__init__()
        
        # 隐藏层特征维度等于输入维度乘以扩张因子
        hidden_features = int(dim*ffn_expansion_factor)
        # 1*1 升维
        self.project_in = nn.Conv2d(dim, hidden_features*2, kernel_size=1, bias=bias)
        # 3*3 分组卷积
        self.dwconv = nn.Conv2d(hidden_features*2, hidden_features*2, kernel_size=3, stride=1, padding=1, groups=hidden_features*2, bias=bias)
        # 1*1 降维
        self.project_out = nn.Conv2d(hidden_features, dim, kernel_size=1, bias=bias)

    def forward(self, x):
        x = self.project_in(x)
        x1, x2 = self.dwconv(x).chunk(2, dim=1)  # 第 1 个维度方向切分成 2 块
        x = F.gelu(x1) * x2  # gelu 相当于 relu+dropout
        x = self.project_out(x)
        return x

3. TransformerBlock

## 就是标准的 Transformer 架构
class TransformerBlock(nn.Module):
    def __init__(self, dim, num_heads, ffn_expansion_factor, bias, LayerNorm_type):
        super(TransformerBlock, self).__init__()

        self.norm1 = LayerNorm(dim, LayerNorm_type)  # 层标准化
        self.attn = Attention(dim, num_heads, bias)  # 自注意力
        self.norm2 = LayerNorm(dim, LayerNorm_type)  # 层表转化
        self.ffn = FeedForward(dim, ffn_expansion_factor, bias)  # FFN

    def forward(self, x):
        x = x + self.attn(self.norm1(x))  # 残差
        x = x + self.ffn(self.norm2(x))  # 残差

        return x

4. 测试样例

model = Restormer()
print(model)  # 打印网络结构

x = torch.randn((1, 3, 512, 512))  #随机生成输入图像
x = model(x)  # 送入网络
print(x.shape) # 打印网络输入的图像结构