图解Swin Transformer
Swin-Transformer网络结构详解
【机器学习】详解 Swin Transformer (SwinT)
论文下载
官方源码下载
学习的话,请下载Image Classification的代码,配置相对简单,其他的配置会很麻烦。如下图所示:
Install :
pytorch安装:感觉pytorch > 1.4版本都没问题的。
2、pip install timm==0.3.2(最新版本也行)
1、pip install Apex
- win 10系统下安装NVIDIA apex
这个我认为windows安装可能会很啃。
1、首先在github下载源码https://github.com/NVIDIA/apex到本地文件夹
2、打开cmd命令窗口,切换到apex所在的文件夹
3、使用命令:python setup.py install 即可完成安装注意事项: 可能会出现的问题:
setuptools有ModuleNotFoundError→更新setuptools
pip install --upgrade setuptools
- linux系统下安装NVIDIA apex
git clone https://github.com/NVIDIA/apex
cd apex
pip install -v --no-cache-dir --global-option="--cpp_ext" --global-option="--cuda_ext" ./
注意:不要用ImageNet数据集:显卡可能会受不了,就是为了学习swin代码对吧,可以自己找一个小的ImageNet的数据集。
首先运行main.py文件,如下图:
再点击main.py配置:
最后在下图Parameters处填入:
--cfg configs/swin_tiny_patch4_window7_224.yaml --data-path imagenet --local_rank 0 --batch-size 2
报错1如下: |
Swin transformer TypeError: __init__() got an unexpected keyword argument ‘t_mul‘
报错2如下: |
from timm.data.transforms import _pil_interp无法导入_pil_interp
pip install timm==0.3.2(最新版本也行)
但是我安装最新版本后:from timm.data.transforms import _pil_interp无法导入_pil_interp,然后我查看了timm.data 中的transforms.py文件,完全就没有定义_pil_interp。完整的timm.data 中的transforms.py文件我在下面也把这个文件_pil_interp代码复制在下面,可以自行补充_pil_interp。
def _pil_interp(method):
if method == 'bicubic':
return Image.BICUBIC
elif method == 'lanczos':
return Image.LANCZOS
elif method == 'hamming':
return Image.HAMMING
if has_interpolation_mode:
_torch_interpolation_to_str = {
InterpolationMode.NEAREST: 'nearest',
InterpolationMode.BILINEAR: 'bilinear',
InterpolationMode.BICUBIC: 'bicubic',
InterpolationMode.BOX: 'box',
InterpolationMode.HAMMING: 'hamming',
InterpolationMode.LANCZOS: 'lanczos',
}
_str_to_torch_interpolation = {b: a for a, b in _torch_interpolation_to_str.items()}
else:
_pil_interpolation_to_torch = {}
_torch_interpolation_to_str = {}
完整的timm.data 中的transforms.py文件:界面如下图
报错3:如下图所示: |
解决办法,删除Swin-Transformer/lr_scheduler.py的第24行‘t_mul=1.,’
下面PPT是对Swin-Transformer做了一个大概的概括,具体细节可以参考第四节代码部分。 |
注意:本人代码部分是按照Debug顺序进行编写的。并不是按照一个一个模块去分开讲解的,所以大家看起来可能会很按难受。这里推荐一篇博客图解Swin Transformer,是按照每个结构分开单独编写,容易理解,思路清晰。
准备工作 |
1、首先我们打开main.py和swin_transformer.py文件。
2、然后在swin_transformer.py中找到class SwinTransformer(nn.Module):类,在其def forward_features(self, x):下第一行插入断点,那么下面我们就开始一步一步debug吧。
def forward_features(self, x):
print(x.shape) # [2, 3, 224, 224], batch_size = 2
x = self.patch_embed(x) # 详解在3.1节
print(x.shape)
if self.ape: # self.ape = False不用考虑
x = x + self.absolute_pos_embed
x = self.pos_drop(x) # 就是一个Droupout层
print(x.shape) # [2, 3136, 96]
for layer in self.layers:
x = layer(x)
print(x.shape)
x = self.norm(x) # B L C
print(x.shape)
x = self.avgpool(x.transpose(1, 2)) # B C 1
print(x.shape)
x = torch.flatten(x, 1)
print(x.shape)
return x
在输入开始的时候,做了一个Patch Embedding,将图片切成一个个图块,并嵌入到Embedding。
在每个Stage里,由Patch Merging和多个Block组成。
其中Patch Merging模块主要在每个Stage一开始降低图片分辨率。
而Block具体结构如右图所示,主要是LayerNorm,MLP,Window Attention 和 Shifted Window Attention组成 (为了方便讲解,我会省略掉一些参数)
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 # [2, 3, 224, 224]
# 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]})."
# proj是先卷积,再flatten(2)把三四列变成一列(即56*56=3136)
x = self.proj(x).flatten(2).transpose(1, 2) # B Ph*Pw C
# x = torch.Size([2, 3136, 96])
# 56*56 = 3136个patch=tokens
# 每个patch或tokens的向量维度为96
print(x.shape) #4 3136 96 其中3136就是 224/4 * 224/4 相当于有这么长的序列,其中每个元素是96维向量
if self.norm is not None:
x = self.norm(x)
print(x.shape)
return x
其实只要看forward就行了。
x = self.proj(x).flatten(2).transpose(1, 2) # B Ph*Pw C
这一行就是把原图[2, 3, 224, 224]转化为3136个patch,每个patch的维度等于96。
class SwinTransformerBlock(nn.Module):
def forward(self, x):
H, W = self.input_resolution
B, L, C = x.shape
assert L == H * W, "input feature has wrong size"
shortcut = x # x = [2,3136,96]
x = self.norm1(x)
x = x.view(B, H, W, C) # (2, 56, 56, 96)
# cyclic shift
# 在第一次我们是W-MSA,没有滑动窗口,所以self.shift_size > 0 =False
if self.shift_size > 0:
shifted_x = torch.roll(x, shifts=(-self.shift_size, -self.shift_size), dims=(1, 2))
else:
shifted_x = x
# partition windows
x_windows = window_partition(shifted_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
# W-MSA/SW-MSA
attn_windows = self.attn(x_windows, mask=self.attn_mask) # nW*B, window_size*window_size, C
# merge windows (把attention后的数据还原成原来输入的shape)
attn_windows = attn_windows.view(-1, self.window_size, self.window_size, C)
shifted_x = window_reverse(attn_windows, self.window_size, H, W) # B H' W' C
# reverse cyclic shift
if self.shift_size > 0:
x = torch.roll(shifted_x, shifts=(self.shift_size, self.shift_size), dims=(1, 2))
print(x.shape)
else:
x = shifted_x
x = x.view(B, H * W, C)
print(x.shape)
# FFN
x = shortcut + self.drop_path(x)
print(x.shape)
x = x + self.drop_path(self.mlp(self.norm2(x)))
print(x.shape)
return x
第一部分:W-MSA部分和窗口的构建 |
x = x.view(B, H, W, C) # (2, 56, 56, 96)
# cyclic shift
# 在第一次我们是W-MSA,没有滑动窗口,所以self.shift_size > 0 =False
if self.shift_size > 0:
shifted_x = torch.roll(x, shifts=(-self.shift_size, -self.shift_size), dims=(1, 2))
else:
shifted_x = x
# partition windows
x_windows = window_partition(shifted_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
第一部分中的window_partition()类:(就是把序列转化为窗口) |
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) # (2,8,7,8,7,96):指把56*56的patch按照7*7的窗口划分
print(x.shape) # (2,8,7,8,7,96)
windows = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(-1, window_size, window_size, C) # window的数量 H/7 * W/7 *batch
print(windows.shape)
# windows=(128, 7, 7, 96)
# 128 = batch_size * 8 * 8 = 128窗口的数量
# 7 = window_size 窗口的大小尺寸,说明每个窗口包含49个patch
return windows
详解self.attn(x_windows, mask=self.attn_mask) |
class WindowAttention(nn.Module):
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)
print(qkv.shape) # torch.Size([3, 128, 3, 49, 32])
q, k, v = qkv[0], qkv[1], qkv[2] # make torchscript happy (cannot use tensor as tuple)
print(q.shape) # torch.Size([128, 3, 49, 32])
print(k.shape) # torch.Size([128, 3, 49, 32])
print(v.shape) # torch.Size([128, 3, 49, 32])
q = q * self.scale # q = [128, 3, 49, 32]
attn = (q @ k.transpose(-2, -1))
print(attn.shape) # torch.Size([128, 3, 49, 49])
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
print(relative_position_bias.shape)
relative_position_bias = relative_position_bias.permute(2, 0, 1).contiguous() # nH, Wh*Ww, Wh*Ww
print(relative_position_bias.shape)
attn = attn + relative_position_bias.unsqueeze(0)
print(attn.shape)
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)
print(attn.shape)
x = (attn @ v).transpose(1, 2).reshape(B_, N, C)
print(x.shape)
x = self.proj(x) # 全连接层,用于整合新信息的
print(x.shape)
x = self.proj_drop(x)
print(x.shape) # 还原成输入的形式[2,3136,96]
return x
qkv.shape = [3, 128, 3, 49, 32]
(1)3:是指Q、K、V三个
(2)128:是指128个windows
(3)3:是指Multi–Head = 3(多头注意力机制)
(4)49:是指每个窗口含有49个patchs,每个窗口的49个patchs之间要相互做self–attention
(5)32:是指经过多头后,每个head分配32个维度。
attn = (q @ k.transpose(-2, -1))
print(attn.shape) # torch.Size([128, 3, 49, 49])
就是正常的计算 α \alpha α相关性。
attn = attn + relative_position_bias.unsqueeze(0)
这是把attention( α \alpha α)和 位置编码进行相加,和ViT中在tokens加上位置编码。(具体如下图所示)
后面的代码和VIT中基本一样,详细请看本人上一篇博客ViT( Vision Transformer)详解。
ViT的Attention公式如下:
A t t e n t i o n ( Q , K , V ) = s o f t m a x ( Q K T d k ) V = s o f t m a x ( α d k ) V Attention(Q,K,V) = softmax(\frac {QK^T}{\sqrt{d_k}}) V= softmax(\frac {\boldsymbol{\alpha}}{\sqrt{d_k}}) V Attention(Q,K,V)=softmax(dkQKT)V=softmax(dkα)V:
Swin–Transformer公式如下:
A t t e n t i o n ( Q , K , V ) = s o f t m a x ( Q K T d k + B ) V = s o f t m a x ( α d k + B ) V Attention(Q,K,V) = softmax(\frac {QK^T}{\sqrt{d_k}} + B) V= softmax(\frac {\boldsymbol{\alpha}}{\sqrt{d_k}} + B) V Attention(Q,K,V)=softmax(dkQKT+B)V=softmax(dkα+B)V:
上面公式的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。后续实验有证明相对位置编码的加入提升了模型性能。
相对位置编码 |
由于论文中并没有详解讲解这个相对位置偏执,所以我自己根据阅读源码做了简单的总结。(主要借鉴了Swin-Transformer网络结构详解这篇博客)如下图,假设输入的feature map高宽都为2,那么首先我们可以构建出每个像素的绝对位置(左下方的矩阵),对于每个像素的绝对位置是使用行号和列号表示的。比如蓝色的像素对应的是第0行第0列所以绝对位置索引是 ( 0 , 0 ) (0,0) (0,0),接下来再看看相对位置索引。首先看下蓝色的像素,在蓝色像素使用q与所有像素k进行匹配过程中,是以蓝色像素为参考点。然后用蓝色像素的绝对位置索引与其他位置索引进行相减,就得到其他位置相对蓝色像素的相对位置索引。例如黄色像素的绝对位置索引是 ( 0 , 1 ) (0,1) (0,1),则它相对蓝色像素的相对位置索引为 ( 0 , 0 ) − ( 0 , 1 ) = ( 0 , − 1 ) (0,0) - (0,1) = (0,-1) (0,0)−(0,1)=(0,−1),这里是严格按照源码中来讲的,请不要杠。那么同理可以得到其他位置相对蓝色像素的相对位置索引矩阵。同样,也能得到相对黄色,红色以及绿色像素的相对位置索引矩阵。接下来将每个相对位置索引矩阵按行展平,并拼接在一起可以得到下面的4x4矩阵 。
代码过程如下:
coords_h = torch.arange(2)
coords_w = torch.arange(2)
coords = torch.meshgrid([coords_h, coords_w])
print(coords)
coords = torch.stack(coords) # 2, Wh, Ww
print("1 1 1 "* 10)
print(coords)
coords_flatten = torch.flatten(coords, 1) # 2, Wh*Ww
print("2 2 2 "* 10)
print(coords_flatten)
relative_coords_first = coords_flatten[:, :, None] # 2, wh*ww, 1
print("3 3 3 "*10)
print(relative_coords_first)
relative_coords_second = coords_flatten[:, None, :] # 2, 1, wh*ww
print("4 4 4 "*10)
print(relative_coords_second)
relative_coords = relative_coords_first - relative_coords_second # 最终得到 2, wh*ww, wh*ww 形状的张量
print("5 5 5 "*10)
print(relative_coords)
relative_coords = relative_coords.permute(1, 2, 0).contiguous() # Wh*Ww, Wh*Ww, 2
print("6 6 6 "*10)
print(relative_coords)
print(relative_coords.shape)
请注意,我这里描述的一直是相对位置索引,并不是相对位置偏执参数。因为后面我们会根据相对位置索引去取对应的参数。比如说黄色像素是在蓝色像素的右边,所以相对蓝色像素的相对位置索引为( 0 , − 1 ) 。绿色像素是在红色像素的右边,所以相对红色像素的相对位置索引为( 0 , − 1 )。可以发现这两者的相对位置索引都是( 0 , − 1 ) ,所以他们使用的相对位置偏执参数都是一样的。其实讲到这基本已经讲完了,但在源码中作者为了方便把二维索引给转成了一维索引。具体这么转的呢,有人肯定想到,简单啊直接把行、列索引相加不就变一维了吗?比如上面的相对位置索引中有( 0 , − 1 ) 和( − 1 , 0 )在二维的相对位置索引中明显是代表不同的位置,但如果简单相加都等于-1那不就出问题了吗?接下来我们看看源码中是怎么做的。首先在原始的相对位置索引上加上M-1(M为窗口的大小,在本示例中M=2),加上之后索引中就不会有负数了。
relative_coords[:, :, 0] += 2 - 1
print("7 7 7 "*10)
print(relative_coords)
relative_coords[:, :, 1] += 2 - 1
print("8 8 8 "*10)
print(relative_coords)
relative_coords[:, :, 0] *= 2 * 2 - 1
print("9 9 9 "*10)
print(relative_coords)
最后将行标和列标进行相加。这样即保证了相对位置关系,而且不会出现上述 0 + ( − 1 ) = ( − 1 ) + 0 0 + (-1) = (-1) + 0 0+(−1)=(−1)+0的问题了,是不是很神奇。
代码过程如下:
relative_position_index = relative_coords.sum(-1) # Wh*Ww, Wh*Ww
print("10 10 10 "* 3)
print(relative_position_index)
刚刚上面也说了,之前计算的是相对位置索引,并不是相对位置偏执参数。真正使用到的可训练参数 B B B是保存在relative position bias table表里的,这个表的长度是等于 ( 2 M − 1 ) × ( 2 M − 1 ) ( 2 M − 1 ) × ( 2 M − 1 ) (2M−1)×(2M−1)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,如下图所示。
以上过程结束,代表Swin–Transformer–Block中的第一部分(W–MSA)结束。返回的x = [2, 3136, 49]。如下图所示:
那么接下来我们要继续执行Swin--Transformer--Block中的第二部分(SW--MSA)。 |
与前一部份的Block的不同之处在于SW-MSA,有个滑动窗口、偏移量等新的东西加入。相同的部分我们就不再代码讲述,下面我们只看不同的部分。
首先我们看到这一部分:
# cyclic shift
if self.shift_size > 0: # self.shift_size = 3,偏移量为3.
shifted_x = torch.roll(x, shifts=(-self.shift_size, -self.shift_size), dims=(1, 2))
else:
shifted_x = x
再看以下Mask部分:可以看到,一直到运行到attn部分,都和前面的W-MSA参数的size是一样的。
... ...
... ...
attn = (q @ k.transpose(-2, -1))
print(attn.shape)
... ...
... ...
attn = attn + relative_position_bias.unsqueeze(0)
print(attn.shape) # torch.Size([128, 3, 49, 49])
print(attn.shape) # torch.Size([128, 3, 49, 49])
# mask部分,mask = self.attn_mask
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)
Shifted Window Attention,前面的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。下面看一下self.shift_size 的定义吧
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。
特征图移位操作
代码里对特征图移位是通过torch.roll来实现的,
shifted_x = torch.roll(x, shifts=(-self.shift_size, -self.shift_size), dims=(1, 2))
下面是示意图:
如果需要reverse cyclic shift(就是还原操作)的话只需把参数shifts设置为对应的正数值。
Attention Mask
我认为这是Swin Transformer的精华,通过设置合理的mask,让Shifted Window Attention在与Window Attention相同的窗口个数下,达到等价的计算结果。
首先我们对Shift Window后的每个窗口都给上index,并且做一个roll操作(window_size=2, shift_size=-1)
我们希望在计算Attention的时候,让具有相同index QK进行计算,而忽略不同index QK计算结果。
最后正确的结果如下图所示:
而要想在原始四个窗口下得到正确的结果,我们就必须给Attention的结果加入一个mask(如上图最右边所示)相关代码如下:
slice(start,end)函数:方法可从已有数组中返回选定的元素,返回一个新数组,包含从start到end(不包含该元素)的数组元素
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))
# h_slices = (slice(0, -7, None) ,slice(7, -3, None) ,slice(-3, None, 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
mask_windows = window_partition(img_mask, self.window_size) # nW, window_size, window_size, 1
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))
意思就是对于attn_mask不为0的部分填充为-100,有什么用呢?想一想softmax函数的计算公式,
当 e − 100 ≈ 0 e^{-100}\approx 0 e−100≈0 , 那么这样是不是等于忽略不同index(指下图中的0,1,2,···,8) QK计算结果。
Downsample(下采样操作):Patch Merging |
注意:这里的Patch Merging下采样操作用的可不是 1 × 1 1 \times 1 1×1卷积进行的下采样。
该模块的作用是在每个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
下面是一个示意图(输入张量N=1, H=W=8, C=1,不包含最后的全连接层调整)
class BasicLayer(nn.Module):
def forward(self, x):
for blk in self.blocks:
if self.use_checkpoint:
x = checkpoint.checkpoint(blk, x)
else:
x = blk(x)
if self.downsample is not None:
x = self.downsample(x)
return x
整体结构SwinTransformer(),最后分类输出层的概述 |
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
"""
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, **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)
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)
@torch.jit.ignore
def no_weight_decay(self):
return {'absolute_pos_embed'}
@torch.jit.ignore
def no_weight_decay_keywords(self):
return {'relative_position_bias_table'}
def forward_features(self, x):
print(x.shape) # [2, 3, 224, 224], batch_size = 2
x = self.patch_embed(x)
print(x.shape)
if self.ape:
x = x + self.absolute_pos_embed
x = self.pos_drop(x)
print(x.shape)
for layer in self.layers:
x = layer(x)
print(x.shape)
x = self.norm(x) # B L C
print(x.shape) # [2, 49, 768]
x = self.avgpool(x.transpose(1, 2)) # B C 1
print(x.shape) # [2, 768, 1]
x = torch.flatten(x, 1)
print(x.shape) # [2, 768]
return x
def forward(self, x):
x = self.forward_features(x)
x = self.head(x) # [2,1000]做imagenet的1000分类
return x
整体流程如下
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_windows × \times ×B, window_size, window_size, C,其中 num_windows = H × \times ×W / 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):
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