简介
主页:https://github. com/microsoft/Swin-Transformer.
Swin Transformer 是 2021 ICCV最佳论文,屠榜了各大CV任务,性能优于DeiT、ViT和EfficientNet等主干网络,已经替代经典的CNN架构,成为了计算机视觉领域通用的backbone。
Swin Transformer 基于了ViT模型的思想,创新性的引入了滑动窗口机制,让模型能够学习到跨窗口的信息,同时也。同时通过下采样层,使得模型能够处理超分辨率的图片,节省计算量以及能够关注全局和局部的信息
ViT 开启了transformer在视觉领域的征途,但是transformer从自然语言领域应用到计算机视觉领域还有两大挑战:
- 视觉实体的方差较大,例如同一个物体,拍摄角度不同,转化为二进制后的图片就会具有很大的差异。同时在不同场景下视觉 Transformer 性能未必很好。
- 图像分辨率高,像素点多,如果采用ViT模型,自注意力的计算量会与像素的平方成正比。
针对上述两个问题,论文中提出了一种基于滑动窗口机制,具有层级设计(下采样层) 的 Swin Transformer。
其中滑窗操作包括不重叠的 local window,和重叠的 cross-window。将注意力计算限制在一个窗口(window size固定)中,一方面能引入 CNN 卷积操作的局部性,另一方面能大幅度节省计算量,它只和窗口数量成线性关系。
通过下采样的层级设计,能够逐渐增大感受野,从而使得注意力机制也能够注意到全局的特征
从上图可知,Swin Transformer 思想是实现 ViT 到类似卷积模式的转变,这样的结构模式能适用于各类视觉任务,真正成为视觉领域通用的backbone。
实现原理
模型整体采取了层次化的设计
- 在输入开始的时候做了一个Patch Partition,即ViT中的Patch Embedding操作,通过Patch_size为4的卷积层将图片切成一个个Patch,并嵌入到Embedding,将embedding_size 转变为48(可以将 CV 中图片的通道数理解为NLP中token的词嵌入长度)
- 第一行Stage中通过Linear Embedding 调整通道数为 C
- 在后3个Stage均由Patch Merging 和 多个 Swin Transformer Block组成
- Patch Merging 模块主要在每个Stage一开始降低图片分辨率,进行下采样操作
- Swin Transformer Block如上图右边所示,主要是LayerNorm,Window Attention,Shifted Window Attention和MLP组成
Patch Merging 总是在两个Swin Transformer Block之间执行下采样,最后一个Stage不需要下采样操作,之间通过后续的全连接层与 target label 计算损失。
class SwinTransformer(nn.Module):
r""" Swin Transformer
A PyTorch impl of : `Swin Transformer: Hierarchical Vision Transformer using Shifted Windows` -
https://arxiv.org/pdf/2103.14030
Args:
img_size (int | tuple(int)): Input image size. Default 224
patch_size (int | tuple(int)): Patch size. Default: 4
in_chans (int): Number of input image channels. Default: 3
num_classes (int): Number of classes for classification head. Default: 1000
embed_dim (int): Patch embedding dimension. Default: 96
depths (tuple(int)): Depth of each Swin Transformer layer.
num_heads (tuple(int)): Number of attention heads in different layers.
window_size (int): Window size. Default: 7
mlp_ratio (float): Ratio of mlp hidden dim to embedding dim. Default: 4
qkv_bias (bool): If True, add a learnable bias to query, key, value. Default: True
qk_scale (float): Override default qk scale of head_dim ** -0.5 if set. Default: None
drop_rate (float): Dropout rate. Default: 0
attn_drop_rate (float): Attention dropout rate. Default: 0
drop_path_rate (float): Stochastic depth rate. Default: 0.1
norm_layer (nn.Module): Normalization layer. Default: nn.LayerNorm.
ape (bool): If True, add absolute position embedding to the patch embedding. Default: False
patch_norm (bool): If True, add normalization after patch embedding. Default: True
use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False
fused_window_process (bool, optional): If True, use one kernel to fused window shift & window partition for acceleration, similar for the reversed part. Default: False
"""
def __init__(self, img_size=224, patch_size=4, in_chans=3, num_classes=1000,
embed_dim=96, depths=[2, 2, 6, 2], num_heads=[3, 6, 12, 24],
window_size=7, mlp_ratio=4., qkv_bias=True, qk_scale=None,
drop_rate=0., attn_drop_rate=0., drop_path_rate=0.1,
norm_layer=nn.LayerNorm, ape=False, patch_norm=True,
use_checkpoint=False, fused_window_process=False, **kwargs):
super().__init__()
self.num_classes = num_classes
self.num_layers = len(depths)
self.embed_dim = embed_dim
self.ape = ape
self.patch_norm = patch_norm
self.num_features = int(embed_dim * 2 ** (self.num_layers - 1))
self.mlp_ratio = mlp_ratio
# split image into non-overlapping patches
self.patch_embed = PatchEmbed(
img_size=img_size, patch_size=patch_size, in_chans=in_chans, embed_dim=embed_dim,
norm_layer=norm_layer if self.patch_norm else None)
num_patches = self.patch_embed.num_patches
patches_resolution = self.patch_embed.patches_resolution
self.patches_resolution = patches_resolution
# absolute position embedding
if self.ape:
self.absolute_pos_embed = nn.Parameter(torch.zeros(1, num_patches, embed_dim))
trunc_normal_(self.absolute_pos_embed, std=.02)
self.pos_drop = nn.Dropout(p=drop_rate)
# stochastic depth
dpr = [x.item() for x in torch.linspace(0, drop_path_rate, sum(depths))] # stochastic depth decay rule
# build layers
self.layers = nn.ModuleList()
for i_layer in range(self.num_layers):
layer = BasicLayer(dim=int(embed_dim * 2 ** i_layer),
input_resolution=(patches_resolution[0] // (2 ** i_layer),
patches_resolution[1] // (2 ** i_layer)),
depth=depths[i_layer],
num_heads=num_heads[i_layer],
window_size=window_size,
mlp_ratio=self.mlp_ratio,
qkv_bias=qkv_bias, qk_scale=qk_scale,
drop=drop_rate, attn_drop=attn_drop_rate,
drop_path=dpr[sum(depths[:i_layer]):sum(depths[:i_layer + 1])],
norm_layer=norm_layer,
downsample=PatchMerging if (i_layer < self.num_layers - 1) else None,
use_checkpoint=use_checkpoint,
fused_window_process=fused_window_process)
self.layers.append(layer)
self.norm = norm_layer(self.num_features)
self.avgpool = nn.AdaptiveAvgPool1d(1)
self.head = nn.Linear(self.num_features, num_classes) if num_classes > 0 else nn.Identity()
self.apply(self._init_weights)
def _init_weights(self, m):
if isinstance(m, nn.Linear):
trunc_normal_(m.weight, std=.02)
if isinstance(m, nn.Linear) and m.bias is not None:
nn.init.constant_(m.bias, 0)
elif isinstance(m, nn.LayerNorm):
nn.init.constant_(m.bias, 0)
nn.init.constant_(m.weight, 1.0)
def forward_features(self, x):
x = self.patch_embed(x)
if self.ape:
x = x + self.absolute_pos_embed
x = self.pos_drop(x)
for layer in self.layers:
x = layer(x)
x = self.norm(x) # B L C
x = self.avgpool(x.transpose(1, 2)) # B C 1
x = torch.flatten(x, 1)
return x
def forward(self, x):
x = self.forward_features(x)
x = self.head(x)
return x
def flops(self):
flops = 0
flops += self.patch_embed.flops()
for i, layer in enumerate(self.layers):
flops += layer.flops()
flops += self.num_features * self.patches_resolution[0] * self.patches_resolution[1] // (2 ** self.num_layers)
flops += self.num_features * self.num_classes
return flops
- ViT 在输入会给 embedding 进行位置编码。而 Swin-T 这里则是作为一个可选项(self.ape),Swin-T 是在计算 Attention 的时候做了一个相对位置编码,我认为这是这篇论文设计最巧妙的地方。
- ViT 会单独加上一个可学习参数,作为分类的 token。而 Swin-T 则是直接做平均(avgpool),输出分类,有点类似 CNN 最后的全局平均池化层。
Patch Embedding
在进入 Block 前,需要将图片切分成多个 patch,然后嵌入向量,具体做法是对原始图片裁成多个 window_size * window_size 的窗口大小,然后进行嵌入。即通过二维卷积层,设置 stride = kernel_size = window_size,设定输出通道来确定嵌入向量的大小。最后将 H,W 维度展开,并移动到第一维度。
class PatchEmbed(nn.Module):
r""" Image to Patch Embedding
Args:
img_size (int): Image size. Default: 224.
patch_size (int): Patch token size. Default: 4.
in_chans (int): Number of input image channels. Default: 3.
embed_dim (int): Number of linear projection output channels. Default: 96.
norm_layer (nn.Module, optional): Normalization layer. Default: None
"""
def __init__(self, img_size=224, patch_size=4, in_chans=3, embed_dim=96, norm_layer=None):
super().__init__()
img_size = to_2tuple(img_size)
patch_size = to_2tuple(patch_size)
patches_resolution = [img_size[0] // patch_size[0], img_size[1] // patch_size[1]]
self.img_size = img_size
self.patch_size = patch_size
self.patches_resolution = patches_resolution
self.num_patches = patches_resolution[0] * patches_resolution[1]
self.in_chans = in_chans
self.embed_dim = embed_dim
self.proj = nn.Conv2d(in_chans, embed_dim, kernel_size=patch_size, stride=patch_size)
if norm_layer is not None:
self.norm = norm_layer(embed_dim)
else:
self.norm = None
def forward(self, x):
B, C, H, W = x.shape
# FIXME look at relaxing size constraints
assert H == self.img_size[0] and W == self.img_size[1], \
f"Input image size ({H}*{W}) doesn't match model ({self.img_size[0]}*{self.img_size[1]})."
x = self.proj(x).flatten(2).transpose(1, 2) # B Ph*Pw C
if self.norm is not None:
x = self.norm(x)
return x
def flops(self):
Ho, Wo = self.patches_resolution
flops = Ho * Wo * self.embed_dim * self.in_chans * (self.patch_size[0] * self.patch_size[1])
if self.norm is not None:
flops += Ho * Wo * self.embed_dim
return flops
Patch Merging
在每个 Stage 开始前做降采样,用于缩小分辨率,调整通道数进而形成层次化的设计,同时也能节省一定运算量。
在 CNN 中,则是在每个 Stage 开始前用stride=2的卷积/池化层来降低分辨率。
每次降采样是两倍,因此在行方向和列方向上,间隔 2 选取元素。
然后拼接在一起作为一整个张量,最后展开。此时通道维度会变成原先的 4 倍(因为 H,W 各缩小 2 倍),此时再通过一个全连接层再调整通道维度为原来的两倍。
class PatchMerging(nn.Module):
r""" Patch Merging Layer.
Args:
input_resolution (tuple[int]): Resolution of input feature.
dim (int): Number of input channels.
norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm
"""
def __init__(self, input_resolution, dim, norm_layer=nn.LayerNorm):
super().__init__()
self.input_resolution = input_resolution
self.dim = dim
self.reduction = nn.Linear(4 * dim, 2 * dim, bias=False)
self.norm = norm_layer(4 * dim)
def forward(self, x):
"""
x: B, H*W, C
"""
H, W = self.input_resolution
B, L, C = x.shape
assert L == H * W, "input feature has wrong size"
assert H % 2 == 0 and W % 2 == 0, f"x size ({H}*{W}) are not even."
x = x.view(B, H, W, C)
x0 = x[:, 0::2, 0::2, :] # B H/2 W/2 C
x1 = x[:, 1::2, 0::2, :] # B H/2 W/2 C
x2 = x[:, 0::2, 1::2, :] # B H/2 W/2 C
x3 = x[:, 1::2, 1::2, :] # B H/2 W/2 C
x = torch.cat([x0, x1, x2, x3], -1) # B H/2 W/2 4*C
x = x.view(B, -1, 4 * C) # B H/2*W/2 4*C
x = self.norm(x)
x = self.reduction(x)
return x
def extra_repr(self) -> str:
return f"input_resolution={self.input_resolution}, dim={self.dim}"
def flops(self):
H, W = self.input_resolution
flops = H * W * self.dim
flops += (H // 2) * (W // 2) * 4 * self.dim * 2 * self.dim
return flops
Window Partition/Reverse
window partition函数是用于对张量划分窗口,指定窗口大小。将原本的张量从 N H W C, 划分成 num_windowsB, window_size, window_size, C,其中 num_windows = HW / window_size*window_size,即窗口的个数。而window reverse函数则是对应的逆过程。这两个函数会在后面的Window Attention用到。
def window_partition(x, window_size):
"""
Args:
x: (B, H, W, C)
window_size (int): window size
Returns:
windows: (num_windows*B, window_size, window_size, C)
"""
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):
"""
Args:
windows: (num_windows*B, window_size, window_size, C)
window_size (int): Window size
H (int): Height of image
W (int): Width of image
Returns:
x: (B, H, W, C)
"""
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
Window Attention
传统的 Transformer 都是基于全局来计算注意力的,因此计算复杂度十分高。而 Swin Transformer 则将注意力的计算限制在每个窗口内,进而减少了计算量。
主要区别是在原始计算 Attention 的公式中的 Q,K 时加入了相对位置编码 B
class WindowAttention(nn.Module):
r""" Window based multi-head self attention (W-MSA) module with relative position bias.
It supports both of shifted and non-shifted window.
Args:
dim (int): Number of input channels.
window_size (tuple[int]): The height and width of the window.
num_heads (int): Number of attention heads.
qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True
qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set
attn_drop (float, optional): Dropout ratio of attention weight. Default: 0.0
proj_drop (float, optional): Dropout ratio of output. Default: 0.0
"""
def __init__(self, dim, window_size, num_heads, 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.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
self.attn_drop = nn.Dropout(attn_drop)
self.proj = nn.Linear(dim, dim)
self.proj_drop = nn.Dropout(proj_drop)
trunc_normal_(self.relative_position_bias_table, std=.02)
self.softmax = nn.Softmax(dim=-1)
相关位置编码
Q,K,V.shape=[numWindwosB, num_heads, window_sizewindow_size, head_dim]
window_size*window_size 即 NLP 中token的个数
head_dim =
E
m
b
e
d
d
i
n
g
d
i
m
n
u
m
h
e
a
d
s
\frac{Embedding_dim}{num_heads}
numheadsEmbeddingdim即 NLP 中token的词嵌入向量的维度
Q K T QK^T QKT计算出来的Attention张量的形状为[numWindowsB, num_heads, Q_tokens, K_tokens],其中Q_tokens=K_tokens=window_sizewindow_size
以 window_size = 2 为例
第 i 行表示第 i 个 token 的query对所有token的key的attention。
对于 Attention 张量来说,以不同元素为原点,其他元素的坐标也是不同的
由于最终我们希望使用一维的位置坐标 x+y 代替二维的位置坐标(x,y),为了避免 (1,2) (2,1) 两个坐标转为一维时均为3,我们之后对相对位置索引进行了一些线性变换,使得能通过一维的位置坐标唯一映射到一个二维的位置坐标,详细可以通过代码部分进行理解。
利用torch.arange和torch.meshgrid函数生成对应的坐标
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)
"""
(tensor([[0, 0],
[1, 1]]),
tensor([[0, 1],
[0, 1]]))
"""
# 堆叠起来,展开为一个二维向量
coords = torch.stack(coords) # 2, Wh, Ww
coords_flatten = torch.flatten(coords, 1) # 2, Wh*Ww
"""
tensor([[0, 0, 1, 1],
[0, 1, 0, 1]])
"""
# 利用广播机制,分别在第一维,第二维,插入一个维度,进行广播相减,得到 2, wh*ww, wh*ww的张量
relative_coords_first = coords_flatten[:, :, None] # 2, wh*ww, 1
relative_coords_second = coords_flatten[:, None, :] # 2, 1, wh*ww
relative_coords = relative_coords_first - relative_coords_second # 最终得到 2, wh*ww, wh*ww 形状的张量
因为采取的是相减,所以得到的索引是从负数开始的,加上偏移量,让其从 0 开始。
relative_coords = relative_coords.permute(1, 2, 0).contiguous() # Wh*Ww, Wh*Ww, 2
relative_coords[:, :, 0] += self.window_size[0] - 1
relative_coords[:, :, 1] += self.window_size[1] - 1
需要将其展开成一维偏移量。而对于 (1,2)和(2,1)这两个坐标。在二维上是不同的,但是通过将 x,y 坐标相加转换为一维偏移的时候,他的偏移量是相等的。
对其中做了个乘法操作,以进行区分
relative_coords[:, :, 0] *= 2 * self.window_size[1] - 1
然后再最后一维上进行求和,展开成一个一维坐标,并注册为一个不参与网络学习的变量
relative_position_index = relative_coords.sum(-1) # Wh*Ww, Wh*Ww
self.register_buffer("relative_position_index", relative_position_index)
之前计算的是相对位置索引,并不是相对位置偏置参数。真正使用到的可训练参数
β
^
\hat{\beta}
β^是保存在 relative position bias table 表里的,这个表的长度是等于 (2M−1) × (2M−1) (在二维位置坐标中线性变化乘以2M-1导致)的。那么上述公式中的相对位置偏执参数 B是根据上面的相对位置索引表根据查relative position bias table表得到的。
接着Window Attention代码
def forward(self, x, mask=None):
"""
Args:
x: input features with shape of (num_windows*B, N, C)
mask: (0/-inf) mask with shape of (num_windows, Wh*Ww, Wh*Ww) or None
"""
B_, N, C = x.shape
qkv = self.qkv(x).reshape(B_, N, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
q, k, v = qkv[0], qkv[1], qkv[2] # 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
def extra_repr(self) -> str:
return f'dim={self.dim}, window_size={self.window_size}, num_heads={self.num_heads}'
def flops(self, N):
# calculate flops for 1 window with token length of N
flops = 0
# qkv = self.qkv(x)
flops += N * self.dim * 3 * self.dim
# attn = (q @ k.transpose(-2, -1))
flops += self.num_heads * N * (self.dim // self.num_heads) * N
# x = (attn @ v)
flops += self.num_heads * N * N * (self.dim // self.num_heads)
# x = self.proj(x)
flops += N * self.dim * self.dim
return flops
- 首先输入张量形状为 [numWindows*B, window_size * window_size, C]
- 然后经过self.qkv这个全连接层后,进行 reshape,调整轴的顺序,得到形状为[3, numWindowsB, num_heads, window_sizewindow_size, c//num_heads],并分配给q,k,v
- 根据公式,我们对q乘以一个scale缩放系数,然后与k(为了满足矩阵乘要求,需要将最后两个维度调换)进行相乘。得到形状为[numWindowsB,num_heads, window_sizewindow_size, window_size*window_size]的attn张量
- 之前我们针对位置编码设置了个形状为(2window_size-12window_size-1,numHeads)的可学习变量。我们用计算得到的相对编码位置索引self.relative_position_index.vew(-1)选取,得到形状为(window_sizewindow_size,window_size*window_size, numHeads)的编码,再permute(2,0,1)后加到attn张量上
- 暂不考虑 mask 的情况,剩下就是跟 transformer 一样的 softmax,dropout,与V矩阵乘,再经过一层全连接层和dropout
Shifted Window Attention
Window Attention 是在每个窗口下计算注意力的,为了更好的和其他 window 进行信息交互,Swin Transformer 还引入了 shifted window 操作。
左边是没有重叠的 Window Attention,而右边则是将窗口进行移位的 Shift Window Attention。可以看到移位后的窗口包含了原本相邻窗口的元素。但这也引入了一个新问题,即 window 的个数翻倍了,由原本四个窗口变成了 9 个窗口。
在实际代码里,我们是通过对特征图移位,并给 Attention 设置 mask 来间接实现的。能在保持原有的 window 个数下,最后的计算结果等价。
Shifted Window step
代码里对特征图移位是通过 torch.roll 来实现的
torch.roll(input, shifts, dims=None) → Tensor
shifts的值为正数相当于向下挤牙膏,挤出的牙膏又从顶部塞回牙膏里面;shifts的值为负数相当于向上挤牙膏,挤出的牙膏又从底部塞回牙膏里面
以 4x4 矩阵 a 为例
a 矩阵中的 ( 1 )代表 A 区域,( 2,3,4 ) 代表 C 区域,( 5,9,13 )代表 B区域,首先将第一行挤到最后一行,如下图矩阵 b
然后再将第一列挤到最后一列,如下图矩阵 b
如果需要reverse cyclic shift的话只需把参数shifts设置为对应的正数值
Attention Mask
通过 roll 操作,我们确实把9块归为了4块,但是 cyclic shift 中,A 从左上角 移动到了 右下角,显然,直接对 cyclic shift 4块进行计算会破坏原有的语义信息,为此,这里使用了 mask 操作。
上图展示 cyclic shift 后 特征图,拿到 window后,执行
Q
K
T
QK^T
QKT ,就是将Q
K
T
K^T
KT 分别展平然后对应元素相乘,根据这一过程,可以得到如上图所示不同 Windows 的 Mask,-100的紫色区域表示遮掩,紫色部分是不同块的运算结果,应该丢弃
具体代码在 SwinTransformerBlock中
if self.shift_size > 0:
# calculate attention mask for SW-MSA
H, W = self.input_resolution
img_mask = torch.zeros((1, H, W, 1)) # 1 H W 1
h_slices = (slice(0, -self.window_size),
slice(-self.window_size, -self.shift_size),
slice(-self.shift_size, None))
w_slices = (slice(0, -self.window_size),
slice(-self.window_size, -self.shift_size),
slice(-self.shift_size, None))
cnt = 0
for h in h_slices:
for w in w_slices:
img_mask[:, h, w, :] = cnt
cnt += 1
mask_windows = window_partition(img_mask, self.window_size) # nW, window_size, window_size, 1
mask_windows = mask_windows.view(-1, self.window_size * self.window_size)
attn_mask = mask_windows.unsqueeze(1) - mask_windows.unsqueeze(2)
attn_mask = attn_mask.masked_fill(attn_mask != 0, float(-100.0)).masked_fill(attn_mask == 0, float(0.0))
else:
attn_mask = None
从上图代码中,可以得到如下的Mask
tensor([[[[[ 0., 0., 0., 0.],
[ 0., 0., 0., 0.],
[ 0., 0., 0., 0.],
[ 0., 0., 0., 0.]]],
[[[ 0., -100., 0., -100.],
[-100., 0., -100., 0.],
[ 0., -100., 0., -100.],
[-100., 0., -100., 0.]]],
[[[ 0., 0., -100., -100.],
[ 0., 0., -100., -100.],
[-100., -100., 0., 0.],
[-100., -100., 0., 0.]]],
[[[ 0., -100., -100., -100.],
[-100., 0., -100., -100.],
[-100., -100., 0., -100.],
[-100., -100., -100., 0.]]]]])
在上面的 window attention 模块的前向代码中,使用mask掩膜
if mask is not None:
nW = mask.shape[0] # 一张图被分为多少个windows eg:[4,49,49]
attn = attn.view(B_ // nW, nW, self.num_heads, N, N) + mask.unsqueeze(1).unsqueeze(0) # torch.Size([128, 4, 12, 49, 49]) torch.Size([1, 4, 1, 49, 49])
attn = attn.view(-1, self.num_heads, N, N)
attn = self.softmax(attn)
else:
attn = self.softmax(attn)
将 mask 加到 attention 的计算结果,并进行 softmax。mask 的值设置为 - 100,softmax 后就会忽略掉对应的值。
W-MSA和MSA的复杂度对比
在原论文中,作者提出的基于滑动窗口操作的 W-MSA 能大幅度减少计算量。那么两者的计算量和算法复杂度大概是如何的呢,论文中给出了一下两个公式进行对比。
- h:feature map的高度
- w:feature map的宽度
- C:feature map的通道数(也可以称为embedding size的大小)
- M:window_size的大小
MSA模块的计算量
首先对于feature map中每一个token(一共有 h w 个token,通道数为C),记作 X h w × C X^{hw\times C} Xhw×C,需要通过三次线性变换 V q , W k , W v V_q,W_k,W_v Vq,Wk,Wv ,产生对应的q,k,v向量,记作 Q h w × C , K h w × C , V h w × C Q^{hw \times C},K^{hw \times C},V^{hw \times C} Qhw×C,Khw×C,Vhw×C(通道数为C)。
根据矩阵运算的计算量公式可以得到运算量为
3
h
w
C
2
3hwC^2
3hwC2
忽略除以
d
\sqrt{d}
d 以及softmax的计算量,根据根据矩阵运算的计算量公式可得
h
w
C
×
h
w
+
h
w
2
×
C
hwC \times hw + hw^2 \times C
hwC×hw+hw2×C,即
2
(
h
w
2
)
C
2(hw^2)C
2(hw2)C
最终再通过一个Linear层输出,计算量为
h
w
C
2
hwC^2
hwC2。因此整体的计算量为
4
h
w
C
2
+
2
(
h
w
)
2
C
4hwC^2+2(hw)^2C
4hwC2+2(hw)2C
W-MSA模块的计算量
对于W-MSA模块,首先会将feature map根据window_size分成
h
w
M
2
\frac{hw}{M^2}
M2hw的窗口,每个窗口的宽高均为 M,然后在每个窗口进行MSA的运算。因此,可以利用上面MSA的计算量公式,将 h = M, w = M 带入,可以得到一个窗口的计算量为
4
M
2
C
2
+
2
M
4
C
4M^2C^2 + 2M^4C
4M2C2+2M4C,又因为有
h
w
M
2
\frac{hw}{M^2}
M2hw个窗口
整体流程