理论部分： 论文笔记：Swin Transformer: Hierarchical Vision Transformer using Shifted Windows_UQI-LIUWJ的博客-CSDN博客
源码部分：Swin-Transformer/models at main · microsoft/Swin-Transformer (github.com)
class SwinTransformer(nn.Module):| img_size | 输入的图片的大小 （int，或者int的tuple） 默认224 |
| patch_size | patch的大小 （int，或者int的tuple） 默认4 |
| in_chans | 输入图片的channel数 默认3 |
| num_classes | 图片分类的类别数 默认1000 |
| embed_dim | patch embedding的维数 默认96 |
| depths | 各swin-transformer层的深度 （int的tuple） |
| num_heads | 各swin-transformer层的attention的头数 （int的tuple） |
| window_size | 窗口大小 （窗口内的点进行attention） |
| mlp_ratio | mlp隐藏层维度：embedding层维度 |
| qkv_bias | QKV是否有bias |
| drop_rate | dropout rate |
| attn_drop_rate | attention的drop rate |
| drop_path_rate | stochastic depth的p值大小 |
| norm_layer | 进行何种规范化 |
| ape | 是否加绝对位置positional encodding |
| patch_norm | 是否在patch embedding之后进行normalization |
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
#####################将像素级图片转成patch级图片的类初始类#######################
# 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)
#用于将图片转换成一个一个patch
num_patches = self.patch_embed.num_patches
#一张图片中有几个patch
patches_resolution = self.patch_embed.patches_resolution
#一张图片基于patch的分辨率
self.patches_resolution = patches_resolution
#############################################################################
########################### 绝对位置编码 #####################################
# 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)
#1*num_patches*embed_dim维度，每一个embed_dim都代表了一个patch的绝对位置的向量
#############################################################################
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
#层数越深stochastic depth的p（不激活这一层的概率）越大
###############################搭建swin-transformer################################
# 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)),
#随着层数的推进，维度翻倍，图片分辨率（大小）减半
depth=depths[i_layer],
#不同层swin-transformer需要不同的block数量
num_heads=num_heads[i_layer],
#不同层swin-transformer需要不同的注意力头数量
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,
#除了最后一层，其他的都需要PatchMerge（类似于CNN的池化）
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)
#对每个子模组进行初始化对每个子模组分别递归地进行初始化
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)
#将图片转成patch级别分辨率，其中每个patch有emb_dim维
#B,Patch_H*Patch_W,C
if self.ape:
x = x + self.absolute_pos_embed
#绝对位置编码
#B,Patch_H*Patch_W,C
x = self.pos_drop(x)
#Dropout
for layer in self.layers:
x = layer(x)
#依次传入不同的layer层
x = self.norm(x)
#LayerNorm
# batch_size, length, dim
x = self.avgpool(x.transpose(1, 2))
#平均池化 # batch_size,dim,1
// 每张图片每个dimension取平均池化，就是这个dimension平均的feature
x = torch.flatten(x, 1)
#[batch_size,dim]
#每张图片有dim个特征，每个特征是这张图片各个patch在这一dimension的平均值
return xdef forward(self, x):
x = self.forward_features(x)
#swin transformer 学习特征
x = self.head(x)
#全连接层进行分类
return x将像素级图片转化成patch级图片
class PatchEmbed(nn.Module): 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)
#224——>（224，224）
patch_size = to_2tuple(patch_size)
#4——>（4，4）
patches_resolution = [img_size[0] // patch_size[0], img_size[1] // patch_size[1]]
#一张图片基于patch的分辨率
self.img_size = img_size
self.patch_size = patch_size
self.patches_resolution = patches_resolution
self.num_patches = patches_resolution[0] * patches_resolution[1]
#一张图片有几个patch
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)
'''
卷积核是patch_size*patch_size,stride是patch_size
——>每个patch*patch*in_chans的部分，通过proj，变成1*1*embed_dim
'''
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
//batch_size,channel_num,height,width
# 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)
'''
每个patch里面的内容进行卷积
将每个patch_size*patch_size的内容变成1*1的内容
proj——> B,emb_dim,Patch_H,Patch_W
flatten(2)——>B,emb_dim,Patch_H*Patch_W
transpose(1,2)——>B,Patch_H*Patch_W,emb_dim
'''
if self.norm is not None:
x = self.norm(x)
return x
一个stage的swin transformer层
class BasicLayer(nn.Module):
""" A basic Swin Transformer layer for one stage.
| dim | 输入channel的数量 |
| input_resolution | 输入的分辨率 |
| depth | block的数量 |
| num_heads | attention头的数量 |
| window_size | window的大小，window_size*window_sizw的内容进行attention |
| mlp_ratio | mlp隐藏层维度：embedding层维度 |
| qkv_bias | QKV是否有bias |
| drop | dropout rate |
| attn_drop | attention的drop rate |
| drop_path | stochastic depth的p |
def __init__(self, dim, input_resolution, depth, num_heads, window_size,
mlp_ratio=4., qkv_bias=True, qk_scale=None, drop=0., attn_drop=0.,
drop_path=0., norm_layer=nn.LayerNorm, downsample=None, use_checkpoint=False,
fused_window_process=False):
super().__init__()
self.dim = dim
self.input_resolution = input_resolution
self.depth = depth
self.use_checkpoint = use_checkpoint
# build blocks
self.blocks = nn.ModuleList([
SwinTransformerBlock(dim=dim,
input_resolution=input_resolution,
num_heads=num_heads,
window_size=window_size,
shift_size=0 if (i % 2 == 0) else window_size // 2,
#由于一个window attention加一个shift window attention是一个swin-transformer块
#所以这里需要根据奇偶判断shift_size是windos_size的一半还是0
mlp_ratio=mlp_ratio,
qkv_bias=qkv_bias,
qk_scale=qk_scale,
drop=drop,
attn_drop=attn_drop,
drop_path=drop_path[i] if isinstance(drop_path, list) else drop_path,
norm_layer=norm_layer,
fused_window_process=fused_window_process)
for i in range(depth)])
# patch merging layer
if downsample is not None:
self.downsample = downsample(input_resolution, dim=dim, norm_layer=norm_layer)
else:
self.downsample = None
#除了最后一层，其他的都需要PatchMerge（类似于CNN的池化） def forward(self, x):
///x:B,Patch_H*Patch_W,C
for blk in self.blocks:
if self.use_checkpoint:
x = checkpoint.checkpoint(blk, x)
else:
x = blk(x)
#依次送入这个basic block 里面的每个swin-transformer block
if self.downsample is not None:
x = self.downsample(x)
#除非最后一层，否则都进行PatchEmerging
return x| dim | 输入channel的数量 |
| input_resolution | 输入的分辨率 |
| num_heads | attention头的数量 |
| window_size | window的大小，window_size*window_sizw的内容进行attention |
| shitf_size | 是否需要滑动窗口，偶数层不用奇数层用 |
| mlp_ratio | mlp隐藏层维度：embedding层维度 |
| qkv_bias | QKV是否有bias |
| drop | dropout rate |
| attn_drop | attention的drop rate |
| drop_path | stochastic depth的p |
def __init__(self, dim, input_resolution, num_heads, window_size=7, shift_size=0,
mlp_ratio=4., qkv_bias=True, qk_scale=None, drop=0., attn_drop=0., drop_path=0.,
act_layer=nn.GELU, norm_layer=nn.LayerNorm,
fused_window_process=False):
super().__init__()
self.dim = dim
self.input_resolution = input_resolution
self.num_heads = num_heads
self.window_size = window_size
self.shift_size = shift_size
self.mlp_ratio = mlp_ratio
if min(self.input_resolution) <= self.window_size:
# if window size is larger than input resolution, we don't partition windows
self.shift_size = 0
self.window_size = min(self.input_resolution)
#如果当前图片的分辨率大小比window size小，那么将window size设置成图片的分辨率大小。同时不进行shift window
assert 0 <= self.shift_size < self.window_size, "shift_size must in 0-window_size"
#默认shift window比windows size 小
self.norm1 = norm_layer(dim)
self.attn = WindowAttention(
dim,
window_size=to_2tuple(self.window_size),
num_heads=num_heads,
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(dim)
mlp_hidden_dim = int(dim * mlp_ratio)
self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)
##################################滑动窗口WSA########################### 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) #B,H,W,C——>B*num_W, window_size, window_size, C #num_W表示可以划分成几个窗口 mask_windows = mask_windows.view(-1, self.window_size * self.window_size) #B*num_W*C，window_size*window_size attn_mask = mask_windows.unsqueeze(1) - mask_windows.unsqueeze(2) #B*num_W*C，window_size*window_size，window_size*window_size attn_mask = attn_mask.masked_fill(attn_mask != 0, float(-100.0)).masked_fill(attn_mask == 0, float(0.0)) else: attn_mask = None ############################################################################## self.register_buffer("attn_mask", attn_mask) self.fused_window_process = fused_window_process
不同数字的cnt对应的是下面这九块
对于滑动窗口这一部分，我们举个例子：H=W=6，window_size=2，shift_size=1
img_mask
'''
tensor([[[[0.],[0.],[0.],[0.],[1.],[2.]],
[[0.],[0.],[0.],[0.],[1.],[2.]],
[[0.],[0.],[0.],[0.],[1.],[2.]],
[[0.],[0.],[0.],[0.],[1.],[2.]],
[[3.],[3.],[3.],[3.],[4.],[5.]],
[[6.],[6.],[6.],[6.],[7.],[8.]]]])
'''mask_windows
'''
tensor([[0., 0., 0., 0.],
[0., 0., 0., 0.],
[1., 2., 1., 2.],
[0., 0., 0., 0.],
[0., 0., 0., 0.],
[1., 2., 1., 2.],
[3., 3., 6., 6.],
[3., 3., 6., 6.],
[4., 5., 7., 8.]])
''' mask_windows.unsqueeze(1) - mask_windows.unsqueeze(2)
'''
我们记：
A=[[ 0., 0., 0., 0.],
[ 0., 0., 0., 0.],
[ 0., 0., 0., 0.],
[ 0., 0., 0., 0.]]
B=[[ 0., 1., 0., 1.],
[-1., 0., -1., 0.],
[ 0., 1., 0., 1.],
[-1., 0., -1., 0.]]
C=[[ 0., 0., 3., 3.],
[ 0., 0., 3., 3.],
[-3., -3., 0., 0.],
[-3., -3., 0., 0.]]
D=[[ 0., 1., 3., 4.],
[-1., 0., 2., 3.],
[-3., -2., 0., 1.],
[-4., -3., -1., 0.]]
结果是[A,A,B,A,A,B,C,C,D]
'''
atten_mask
'''
我们记：
A[[ 0., 0., 0., 0.],
[ 0., 0., 0., 0.],
[ 0., 0., 0., 0.],
[ 0., 0., 0., 0.]]
B=[[ 0., -100., 0., -100.],
[-100., 0., -100., 0.],
[ 0., -100., 0., -100.],
[-100., 0., -100., 0.]]
C=[[ 0., -100., 0., -100.],
[-100., 0., -100., 0.],
[ 0., -100., 0., -100.],
[-100., 0., -100., 0.]]
D=[[ 0., -100., -100., -100.],
[-100., 0., -100., -100.],
[-100., -100., 0., -100.],
[-100., -100., -100., 0.]]
结果是[A,A,B,A,A,B,C,C,D]
'''A,B,C,D分别对应Window 0，1，2，3
def forward(self, x):
#x:B,Patch_H*Patch_W,C
H, W = self.input_resolution
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)
#x:B,Patch_H,Patch_W,C
#############################（如果需要的话）滑动窗口############################
# cyclic shift
if self.shift_size > 0:
if not self.fused_window_process:
shifted_x = torch.roll(x, shifts=(-self.shift_size, -self.shift_size), dims=(1, 2))
#向右下方横纵各平移shift_size
#（最左边和最上面的翻折下来）
# partition windows
x_windows = window_partition(shifted_x, self.window_size)
# num_Window*B, window_size, window_size, C
else:
x_windows = WindowProcess.apply(x, B, H, W, C, -self.shift_size, self.window_size)
else:
shifted_x = x
# partition windows
x_windows = window_partition(shifted_x, self.window_size)
# num_Window*B, window_size, window_size, C
x_windows = x_windows.view(-1, self.window_size * self.window_size, C)
# num_Window*B, window_size*window_size, C
############################################################################
############################（滑动）窗口attention#################
# W-MSA/SW-MSA
attn_windows = self.attn(x_windows, mask=self.attn_mask)
#根据是否是滑动窗口attention，来进行窗口attention/滑动窗口attention
# num_Window*B, window_size*window_size, C
# merge windows
attn_windows = attn_windows.view(-1, self.window_size, self.window_size, C)
# num_Window*B, window_size，window_size, C
# reverse cyclic shift
if self.shift_size > 0:
if not self.fused_window_process:
shifted_x = window_reverse(attn_windows, self.window_size, H, W)
# B H W C
#从window级别视角转换回patch级别视角
x = torch.roll(shifted_x, shifts=(self.shift_size, self.shift_size), dims=(1, 2))
#将向右下方平移后的矩阵平移回去
else:
x = WindowProcessReverse.apply(attn_windows, B, H, W, C, self.shift_size, self.window_size)
else:
shifted_x = window_reverse(attn_windows, self.window_size, H, W)
#如果没有滑动窗口，只要从window级别视角转换回patch级别视角即可
x = shifted_x
######################################################################
x = x.view(B, H * W, C)
x = shortcut + self.drop_path(x)
#每一个SwinTransformerBlock做完后（window att/shift window att），都进行一次stochastic depth
# FFN
x = x + self.drop_path(self.mlp(self.norm2(x)))
#这论文的模型图里没有说明，但应该也是一个stochastic depth的操作
return x
逐window的attention
| dim | 输入channel的数量 |
| num_heads | attention头的数量 |
| window_size | window的大小，window_size*window_sizw的内容进行attention |
| qkv_bias | QKV是否有bias |
| attn_drop | attention的drop rate |
| proj_drop | 输出层的droprate |
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
#由于需要保持维度，所以每个window attention输入输出的维度都是dim
#由于window attention有num_heads个头，所以每个头的dim就是dim//num_heads
self.scale = qk_scale or head_dim ** -0.5
#############################相对位置编码#################################
# 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, n_heads
# 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)
#这里和分别写三个dim——>dim的q，k，v Linear function是异曲同工的
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)relative_position_bias_table是一个可更新的Parameter
我们现在假定窗口大小是2*2
| (0,0) | (0,1) |
| (1,0) | (1,1) |
| 4 (0,0)—>(0,0) | 2 (0,0)—>(0,1) | 1 (0,0)—>(1,0) | 0 (0,0)—>(1,1) |
| 5 (0,1)——(0,0) | 4 (0,1)—>(0,1) | 3 (0,1)—>(1,0) | 1 (1,0)—>(1,1) |
| 7 (1,0)——>(0,0) | 6 (1,0)—>(0.1） | 4 (2,2)—>(2,2) | 2 (1,0)—>(1,1) |
| 8 (1,1)—>(0,0) | 7 (1,1)—>(0,1) | 5 (1,1)—>(1,0) | 4 (3,3)—>(3,3) |
coords_h = torch.arange(window_size[0])
coords_w = torch.arange(window_size[1])
coords_h,coords_w
'''
(tensor([0, 1]), tensor([0, 1]))
'''
coords = torch.stack(torch.meshgrid([coords_h, coords_w]))
coords
'''
(tensor([[0, 0],
[1, 1]]),
tensor([[0, 1],
[0, 1]]))
'''
coords_flatten = torch.flatten(coords, 1)
coords_flatten
'''
tensor([[0, 0, 1, 1],
[0, 1, 0, 1]])
'''
relative_coords = coords_flatten[:, :, None] - coords_flatten[:, None, :]
relative_coords
'''
每个window和另一个window之间相对位置，一共4*4个，所以这里是4*4的矩阵
第一个4*4矩阵是相对位置的纵轴；第二个4*4是相对位置的横轴
tensor([[[ 0, 0, -1, -1],
[ 0, 0, -1, -1],
[ 1, 1, 0, 0],
[ 1, 1, 0, 0]],
[[ 0, -1, 0, -1],
[ 1, 0, 1, 0],
[ 0, -1, 0, -1],
[ 1, 0, 1, 0]]])
'''relative_coords = relative_coords.permute(1, 2, 0).contiguous()
relative_coords
'''
每一行是一个相对位置索引
tensor([[[ 0, 0],
[ 0, -1],
[-1, 0],
[-1, -1]],
[[ 0, 1],
[ 0, 0],
[-1, 1],
[-1, 0]],
[[ 1, 0],
[ 1, -1],
[ 0, 0],
[ 0, -1]],
[[ 1, 1],
[ 1, 0],
[ 0, 1],
[ 0, 0]]])
'''
relative_coords[:, :, 0] += window_size[0] - 1
relative_coords[:, :, 1] += window_size[1] - 1
relative_coords[:, :, 0] *= 2 * window_size[1] - 1
'''
tensor([[[3, 1],
[3, 0],
[0, 1],
[0, 0]],
[[3, 2],
[3, 1],
[0, 2],
[0, 1]],
[[6, 1],
[6, 0],
[3, 1],
[3, 0]],
[[6, 2],
[6, 1],
[3, 2],
[3, 1]]])
'''
relative_position_index = relative_coords.sum(-1)
relative_position_index
'''
tensor([[4, 3, 1, 0],
[5, 4, 2, 1],
[7, 6, 4, 3],
[8, 7, 5, 4]])
''' 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
#num_Window*B, window_size*window_size, C
qkv = self.qkv(x).reshape(B_, N, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
#3，B_,self.num_heads,N,C // self.num_heads
#3，num_Window*B,self.num_heads,window_size*window_size,C // self.num_heads
q, k, v = qkv[0], qkv[1], qkv[2]
# make torchscript happy (cannot use tensor as tuple)
#num_Window*B,self.num_heads,window_size*window_size,C // self.num_heads
q = q * self.scale
attn = (q @ k.transpose(-2, -1))
#Q,K内积
#B_,self.num_heads,N,N
#num_Window*B,self.num_heads,window_size*window_size,window_size*window_size
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)
# window_size*window_size,window_size*window_size,self.num_heads
relative_position_bias = relative_position_bias.permute(2, 0, 1).contiguous()
# self.num_heads,window_size*window_size,window_size*window_size
attn = attn + relative_position_bias.unsqueeze(0)
#Batch张图片中每个window都加上这个relative position
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)
#B_ // nW,num_Window,self.num_heads,window_size*window_size,window_size*window_size
#mask：#nW，window_size*window_size，window_size*window_size
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将patch级别的图片划分成窗口级别
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)
比如原来是(1,56,56,3),窗口大小为7，可分成8*8个窗口
那么返回维度是(64,7,7,3)
"""
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
举例
a=torch.arange(16).reshape(1,4,4,1)
print(a)
'''
tensor([[[[ 0],[ 1],[ 2],[ 3]],
[[ 4],[ 5],[ 6],[ 7]],
[[ 8],[ 9],[10],[11]],
[[12],[13],[14],[15]]]])
'''
w=window_partition(a,2)
w
'''
tensor([[[[ 0],[ 1]],
[[ 4],[ 5]]],
[[[ 2], [ 3]],
[[ 6],[ 7]]],
[[[ 8],[ 9]],
[[12],[13]]],
[[[10],[11]],
[[14],[15]]]])
'''把窗口级别的还原成patch级别
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))
#(H * W / window_size / window_size)就是num_windows
x = windows.view(B, H // window_size, W // window_size, window_size, window_size, -1)
# B, n_patch_H,n_patch_W,window_size,window_size,C
x = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(B, H, W, -1)
#(B,H,W,C)
return x举例
w
'''
tensor([[[[ 0],[ 1]],
[[ 4],[ 5]]],
[[[ 2], [ 3]],
[[ 6],[ 7]]],
[[[ 8],[ 9]],
[[12],[13]]],
[[[10],[11]],
[[14],[15]]]])
'''
window_reverse(w,2,4,4)
'''
tensor([[[[ 0],[ 1],[ 2],[ 3]],
[[ 4],[ 5],[ 6],[ 7]],
[[ 8],[ 9],[10],[11]],
[[12],[13],[14],[15]]]])
'''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
B,H,W,C——>
"""
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