(即插即用模块-Attention部分) 五十六、(2023 WACV) Skip Attention 跳跃注意力

paper：Attention Attention Everywhere: Monocular Depth Prediction with Skip Attention

Code：https://github.com/ashutosh1807/PixelFormer

---

1、Skip Attention

现有方法在融合编码器和解码器特征时，通常使用特征图拼接以及卷积操作，但卷积核权重固定，限制了语义信息的流动，导致深度预测不准确。为此，这篇论文提出一种 跳跃注意力（Skip Attention），Skip Attention 旨在解决这个问题，通过窗口化交叉注意力机制，有效地融合编码器和解码器特征，提高深度预测的准确性。

Skip Attention 模块通过利用窗口化交叉注意力，计算像素查询之间的自相似度，并关注更长的范围，从而有效地融合编码器和解码器特征。

对于输入X，Skip Attention 的实现过程：

1. 特征图处理：首先，对编码器特征图和解码器特征图进行卷积操作，确保它们的通道数一致。使用 MLP 层分别从编码器特征图和解码器特征图中提取查询矩阵 (Q)、键矩阵 (K) 和值矩阵 (V)。
2. 窗口划分：然后，将查询矩阵、键矩阵和值矩阵划分为大小为 WxW 的窗口，其中 W 是预定义的窗口大小。
3. 窗口内交叉注意力：其次，对每个窗口内的查询矩阵和键矩阵进行自注意力计算，得到注意力权重。将注意力权重与值矩阵相乘，得到每个查询的加权值。使用 softmax 函数将注意力权重归一化，使其总和为 1。
4. 窗口重组：将每个窗口内的加权值重新排列，恢复到原始的像素级位置。这一步确保了注意力计算结果能够正确地反映每个像素的特征。
5. 多头注意力机制：将像素查询分解为多个头，并对每个头分别进行注意力计算。在这之后，再将多头注意力计算结果进行拼接，并使用 MLP 层进行特征聚合。
6. 输出：最后，将注意力计算结果与原始像素查询进行残差连接，并使用 LayerNorm 层进行归一化。

相比于基于卷积的跳过连接，Skip Attention 模块可以更好地融合具有长距离依赖关系的解码器特征，从而提高深度预测的准确性。除此之外，Skip Attention 模块可以有效地利用全局上下文信息，从而提高模型的泛化能力。

Skip Attention 结构图：

2、代码实现

import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.utils.checkpoint as checkpoint
import numpy as np
from timm.models.layers import DropPath, to_2tuple, trunc_normal_


class Mlp(nn.Module):
    """ Multilayer perceptron."""

    def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU, drop=0.):
        super().__init__()
        out_features = out_features or in_features
        hidden_features = hidden_features or in_features
        self.fc1 = nn.Linear(in_features, hidden_features)
        self.act = act_layer()
        self.fc2 = nn.Linear(hidden_features, out_features)
        self.drop = nn.Dropout(drop)

    def forward(self, x):
        x = self.fc1(x)
        x = self.act(x)
        x = self.drop(x)
        x = self.fc2(x)
        x = self.drop(x)
        return x


def window_partition(x, window_size):
    B, H, W, C = x.shape
    x = x.view(B, H // window_size, window_size, W // window_size, window_size, C)
    windows = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(-1, window_size, window_size, C)
    return windows


def window_reverse(windows, window_size, H, W):
    B = int(windows.shape[0] / (H * W / window_size / window_size))
    x = windows.view(B, H // window_size, W // window_size, window_size, window_size, -1)
    x = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(B, H, W, -1)
    return x


class WindowAttention(nn.Module):
    def __init__(self, dim, window_size, num_heads, v_dim, qkv_bias=True, qk_scale=None, attn_drop=0., proj_drop=0.):

        super().__init__()
        self.dim = dim
        self.window_size = window_size  # Wh, Ww
        self.num_heads = num_heads
        head_dim = dim // num_heads
        self.scale = qk_scale or head_dim ** -0.5

        # define a parameter table of relative position bias
        self.relative_position_bias_table = nn.Parameter(
            torch.zeros((2 * window_size[0] - 1) * (2 * window_size[1] - 1), num_heads))  # 2*Wh-1 * 2*Ww-1, nH

        # get pair-wise relative position index for each token inside the window
        coords_h = torch.arange(self.window_size[0])
        coords_w = torch.arange(self.window_size[1])
        coords = torch.stack(torch.meshgrid([coords_h, coords_w]))  # 2, Wh, Ww
        coords_flatten = torch.flatten(coords, 1)  # 2, Wh*Ww
        relative_coords = coords_flatten[:, :, None] - coords_flatten[:, None, :]  # 2, Wh*Ww, Wh*Ww
        relative_coords = relative_coords.permute(1, 2, 0).contiguous()  # Wh*Ww, Wh*Ww, 2
        relative_coords[:, :, 0] += self.window_size[0] - 1  # shift to start from 0
        relative_coords[:, :, 1] += self.window_size[1] - 1
        relative_coords[:, :, 0] *= 2 * self.window_size[1] - 1
        relative_position_index = relative_coords.sum(-1)  # Wh*Ww, Wh*Ww
        self.register_buffer("relative_position_index", relative_position_index)

        self.kv = nn.Linear(dim, dim * 2, bias=qkv_bias)
        self.q = nn.Linear(dim, dim, bias=qkv_bias)
        self.attn_drop = nn.Dropout(attn_drop)
        self.proj = nn.Linear(v_dim, v_dim)
        self.proj_drop = nn.Dropout(proj_drop)

        trunc_normal_(self.relative_position_bias_table, std=.02)
        self.softmax = nn.Softmax(dim=-1)

    def forward(self, x, v, mask=None):
        B_, N, C = x.shape
        q = self.q(x).view(B_, N, self.num_heads, -1).transpose(1, 2)
        kv = self.kv(v).reshape(B_, N, 2, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
        k, v = kv[0], kv[1]  # make torchscript happy (cannot use tensor as tuple)

        q = q * self.scale
        attn = (q @ k.transpose(-2, -1))

        relative_position_bias = self.relative_position_bias_table[self.relative_position_index.view(-1)].view(
            self.window_size[0] * self.window_size[1], self.window_size[0] * self.window_size[1], -1)  # Wh*Ww,Wh*Ww,nH
        relative_position_bias = relative_position_bias.permute(2, 0, 1).contiguous()  # nH, Wh*Ww, Wh*Ww
        attn = attn + relative_position_bias.unsqueeze(0)

        if mask is not None:
            nW = mask.shape[0]
            attn = attn.view(B_ // nW, nW, self.num_heads, N, N) + mask.unsqueeze(1).unsqueeze(0)
            attn = attn.view(-1, self.num_heads, N, N)
            attn = self.softmax(attn)
        else:
            attn = self.softmax(attn)

        attn = self.attn_drop(attn)

        x = (attn @ v).transpose(1, 2).reshape(B_, N, C)
        x = self.proj(x)
        x = self.proj_drop(x)
        return x


class SAMBLOCK(nn.Module):
    def __init__(self,
                 dim,
                 num_heads,
                 v_dim,
                 window_size=7,
                 mlp_ratio=4.,
                 qkv_bias=True,
                 qk_scale=None,
                 drop=0.,
                 attn_drop=0.,
                 drop_path=0.,
                 norm_layer=nn.LayerNorm,
                 ):
        super().__init__()
        self.window_size = window_size
        self.dim = dim
        self.num_heads = num_heads
        self.v_dim = v_dim
        self.window_size = window_size
        self.mlp_ratio = mlp_ratio
        act_layer = nn.GELU
        norm_layer = nn.LayerNorm

        self.norm1 = norm_layer(dim)
        self.normv = norm_layer(dim)
        self.attn = WindowAttention(
            dim, window_size=to_2tuple(self.window_size), num_heads=num_heads, v_dim=v_dim,
            qkv_bias=qkv_bias, qk_scale=qk_scale, attn_drop=attn_drop, proj_drop=drop)

        self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
        self.norm2 = norm_layer(v_dim)
        mlp_hidden_dim = int(v_dim * mlp_ratio)
        self.mlp = Mlp(in_features=v_dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)

    def forward(self, x, v, H, W):
        B, L, C = x.shape
        assert L == H * W, "input feature has wrong size"

        shortcut = x
        x = self.norm1(x)
        x = x.view(B, H, W, C)

        shortcut_v = v
        v = self.normv(v)
        v = v.view(B, H, W, C)

        # pad feature maps to multiples of window size
        pad_l = pad_t = 0
        pad_r = (self.window_size - W % self.window_size) % self.window_size
        pad_b = (self.window_size - H % self.window_size) % self.window_size
        x = F.pad(x, (0, 0, pad_l, pad_r, pad_t, pad_b))
        v = F.pad(v, (0, 0, pad_l, pad_r, pad_t, pad_b))
        _, Hp, Wp, _ = x.shape

        # partition windows
        x_windows = window_partition(x, self.window_size)  # nW*B, window_size, window_size, C
        x_windows = x_windows.view(-1, self.window_size * self.window_size, C)  # nW*B, window_size*window_size, C
        v_windows = window_partition(v, self.window_size)  # nW*B, window_size, window_size, C
        v_windows = v_windows.view(-1, self.window_size * self.window_size,
                                   v_windows.shape[-1])  # nW*B, window_size*window_size, C

        # W-MSA/SW-MSA
        attn_windows = self.attn(x_windows, v_windows, mask=None)  # nW*B, window_size*window_size, C

        # merge windows
        attn_windows = attn_windows.view(-1, self.window_size, self.window_size, self.v_dim)
        x = window_reverse(attn_windows, self.window_size, Hp, Wp)  # B H' W' C

        if pad_r > 0 or pad_b > 0:
            x = x[:, :H, :W, :].contiguous()

        x = x.view(B, H * W, self.v_dim)

        # FFN
        x = self.drop_path(x) + shortcut
        x = x + self.drop_path(self.mlp(self.norm2(x)))

        return x, H, W


class SAM(nn.Module):
    def __init__(self,
                 input_dim=96,
                 embed_dim=96,
                 v_dim=64,
                 window_size=7,
                 num_heads=4,
                 patch_size=4,
                 in_chans=3,
                 norm_layer=nn.LayerNorm,
                 patch_norm=True):
        super().__init__()

        self.embed_dim = embed_dim

        if input_dim != embed_dim:
            self.proj_e = nn.Conv2d(input_dim, embed_dim, 3, padding=1)
        else:
            self.proj_e = None

        if v_dim != embed_dim:
            self.proj_q = nn.Conv2d(v_dim, embed_dim, 3, padding=1)
        elif embed_dim % v_dim == 0:
            self.proj_q = None
        self.proj = nn.Conv2d(embed_dim, embed_dim, 3, padding=1)

        v_dim = embed_dim
        self.sam_block = SAMBLOCK(
            dim=embed_dim,
            num_heads=num_heads,
            v_dim=v_dim,
            window_size=window_size,
            mlp_ratio=4.,
            qkv_bias=True,
            qk_scale=None,
            drop=0.,
            attn_drop=0.,
            drop_path=0.,
            norm_layer=norm_layer)

        layer = norm_layer(embed_dim)
        layer_name = 'norm_sam'
        self.add_module(layer_name, layer)

    def forward(self, e, q):
        if self.proj_q is not None:
            q = self.proj_q(q)
        if self.proj_e is not None:
            e = self.proj_e(e)
        e_proj = e
        q_proj = q

        Wh, Ww = q.size(2), q.size(3)
        q = q.flatten(2).transpose(1, 2)
        e = e.flatten(2).transpose(1, 2)

        q_out, H, W = self.sam_block(q, e, Wh, Ww)
        norm_layer = getattr(self, f'norm_sam')
        q_out = norm_layer(q_out)
        q_out = q_out.view(-1, H, W, self.embed_dim).permute(0, 3, 1, 2).contiguous()

        return q_out + e_proj + q_proj


if __name__ == '__main__':
    x = torch.randn(4, 96, 128, 128)
    y = torch.randn(4, 96, 128, 128)
    model = SAM(v_dim=96)
    output = model(x, y)
    print(output.shape)