Vision Transformer结构解析
ViT简介
Vision Transformer。transformer于2017年的Attention is all your need提出,该模型最大的创新点就是将transformer应用于cv任务。
论文题目:An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale
论文链接:https://arxiv.org/pdf/2010.11929.pdf
代码地址:https://github.com/google-research/vision_transformer
ViT模型整体结构图如下:
ViT三种不同尺寸模型的参数对比:
![Panda](https://i-blog.csdnimg.cn/blog_migrate/03e35cce39436237d1500efb1451541d.png)
ViT参数量
vit_base_patch16_224,num_classes=5, has_logits=False
Total params: 85,650,437
Trainable params: 85,650,437
Non-trainable params: 0
----------------------------------------------------------------
Input size (MB): 0.57
Forward/backward pass size (MB): 408.54
Params size (MB): 326.73
Estimated Total Size (MB): 735.84
----------------------------------------------------------------
vit_base_patch32_224,num_classes=5, has_logits=False
Total params: 87,419,909
Trainable params: 87,419,909
Non-trainable params: 0
----------------------------------------------------------------
Input size (MB): 0.57
Forward/backward pass size (MB): 95.61
Params size (MB): 333.48
Estimated Total Size (MB): 429.66
----------------------------------------------------------------
vit_large_patch32_224,num_classes=5, has_logits=False
Total params: 305,463,301
Trainable params: 305,463,301
Non-trainable params: 0
----------------------------------------------------------------
Input size (MB): 0.57
Forward/backward pass size (MB): 253.01
Params size (MB): 1165.25
Estimated Total Size (MB): 1418.84
vit_huge_patch14_224,num_classes=5, has_logits=False
Total params: 630,440,965
Trainable params: 630,440,965
Non-trainable params: 0
----------------------------------------------------------------
Input size (MB): 0.57
Forward/backward pass size (MB): 2358.66
Params size (MB): 2404.94
Estimated Total Size (MB): 4764.17
----------------------------------------------------------------
网络结构打印,以常用的【vit_base_patch32_224,num_classes=5】为例:
----------------------------------------------------------------
Layer (type) Output Shape Param #
================================================================
Conv2d-1 [-1, 768, 7, 7] 2,360,064
Identity-2 [-1, 49, 768] 0
PatchEmbed-3 [-1, 49, 768] 0
Dropout-4 [-1, 50, 768] 0
LayerNorm-5 [-1, 50, 768] 1,536
Linear-6 [-1, 50, 2304] 1,771,776
Dropout-7 [-1, 12, 50, 50] 0
Linear-8 [-1, 50, 768] 590,592
Dropout-9 [-1, 50, 768] 0
Attention-10 [-1, 50, 768] 0
Identity-11 [-1, 50, 768] 0
LayerNorm-12 [-1, 50, 768] 1,536
Linear-13 [-1, 50, 3072] 2,362,368
GELU-14 [-1, 50, 3072] 0
Dropout-15 [-1, 50, 3072] 0
Linear-16 [-1, 50, 768] 2,360,064
Dropout-17 [-1, 50, 768] 0
Mlp-18 [-1, 50, 768] 0
Identity-19 [-1, 50, 768] 0
Block-20 [-1, 50, 768] 0
LayerNorm-21 [-1, 50, 768] 1,536
Linear-22 [-1, 50, 2304] 1,771,776
Dropout-23 [-1, 12, 50, 50] 0
Linear-24 [-1, 50, 768] 590,592
Dropout-25 [-1, 50, 768] 0
Attention-26 [-1, 50, 768] 0
Identity-27 [-1, 50, 768] 0
LayerNorm-28 [-1, 50, 768] 1,536
Linear-29 [-1, 50, 3072] 2,362,368
GELU-30 [-1, 50, 3072] 0
Dropout-31 [-1, 50, 3072] 0
Linear-32 [-1, 50, 768] 2,360,064
Dropout-33 [-1, 50, 768] 0
Mlp-34 [-1, 50, 768] 0
Identity-35 [-1, 50, 768] 0
Block-36 [-1, 50, 768] 0
LayerNorm-37 [-1, 50, 768] 1,536
Linear-38 [-1, 50, 2304] 1,771,776
Dropout-39 [-1, 12, 50, 50] 0
Linear-40 [-1, 50, 768] 590,592
Dropout-41 [-1, 50, 768] 0
Attention-42 [-1, 50, 768] 0
Identity-43 [-1, 50, 768] 0
LayerNorm-44 [-1, 50, 768] 1,536
Linear-45 [-1, 50, 3072] 2,362,368
GELU-46 [-1, 50, 3072] 0
Dropout-47 [-1, 50, 3072] 0
Linear-48 [-1, 50, 768] 2,360,064
Dropout-49 [-1, 50, 768] 0
Mlp-50 [-1, 50, 768] 0
Identity-51 [-1, 50, 768] 0
Block-52 [-1, 50, 768] 0
LayerNorm-53 [-1, 50, 768] 1,536
Linear-54 [-1, 50, 2304] 1,771,776
Dropout-55 [-1, 12, 50, 50] 0
Linear-56 [-1, 50, 768] 590,592
Dropout-57 [-1, 50, 768] 0
Attention-58 [-1, 50, 768] 0
Identity-59 [-1, 50, 768] 0
LayerNorm-60 [-1, 50, 768] 1,536
Linear-61 [-1, 50, 3072] 2,362,368
GELU-62 [-1, 50, 3072] 0
Dropout-63 [-1, 50, 3072] 0
Linear-64 [-1, 50, 768] 2,360,064
Dropout-65 [-1, 50, 768] 0
Mlp-66 [-1, 50, 768] 0
Identity-67 [-1, 50, 768] 0
Block-68 [-1, 50, 768] 0
LayerNorm-69 [-1, 50, 768] 1,536
Linear-70 [-1, 50, 2304] 1,771,776
Dropout-71 [-1, 12, 50, 50] 0
Linear-72 [-1, 50, 768] 590,592
Dropout-73 [-1, 50, 768] 0
Attention-74 [-1, 50, 768] 0
Identity-75 [-1, 50, 768] 0
LayerNorm-76 [-1, 50, 768] 1,536
Linear-77 [-1, 50, 3072] 2,362,368
GELU-78 [-1, 50, 3072] 0
Dropout-79 [-1, 50, 3072] 0
Linear-80 [-1, 50, 768] 2,360,064
Dropout-81 [-1, 50, 768] 0
Mlp-82 [-1, 50, 768] 0
Identity-83 [-1, 50, 768] 0
Block-84 [-1, 50, 768] 0
LayerNorm-85 [-1, 50, 768] 1,536
Linear-86 [-1, 50, 2304] 1,771,776
Dropout-87 [-1, 12, 50, 50] 0
Linear-88 [-1, 50, 768] 590,592
Dropout-89 [-1, 50, 768] 0
Attention-90 [-1, 50, 768] 0
Identity-91 [-1, 50, 768] 0
LayerNorm-92 [-1, 50, 768] 1,536
Linear-93 [-1, 50, 3072] 2,362,368
GELU-94 [-1, 50, 3072] 0
Dropout-95 [-1, 50, 3072] 0
Linear-96 [-1, 50, 768] 2,360,064
Dropout-97 [-1, 50, 768] 0
Mlp-98 [-1, 50, 768] 0
Identity-99 [-1, 50, 768] 0
Block-100 [-1, 50, 768] 0
LayerNorm-101 [-1, 50, 768] 1,536
Linear-102 [-1, 50, 2304] 1,771,776
Dropout-103 [-1, 12, 50, 50] 0
Linear-104 [-1, 50, 768] 590,592
Dropout-105 [-1, 50, 768] 0
Attention-106 [-1, 50, 768] 0
Identity-107 [-1, 50, 768] 0
LayerNorm-108 [-1, 50, 768] 1,536
Linear-109 [-1, 50, 3072] 2,362,368
GELU-110 [-1, 50, 3072] 0
Dropout-111 [-1, 50, 3072] 0
Linear-112 [-1, 50, 768] 2,360,064
Dropout-113 [-1, 50, 768] 0
Mlp-114 [-1, 50, 768] 0
Identity-115 [-1, 50, 768] 0
Block-116 [-1, 50, 768] 0
LayerNorm-117 [-1, 50, 768] 1,536
Linear-118 [-1, 50, 2304] 1,771,776
Dropout-119 [-1, 12, 50, 50] 0
Linear-120 [-1, 50, 768] 590,592
Dropout-121 [-1, 50, 768] 0
Attention-122 [-1, 50, 768] 0
Identity-123 [-1, 50, 768] 0
LayerNorm-124 [-1, 50, 768] 1,536
Linear-125 [-1, 50, 3072] 2,362,368
GELU-126 [-1, 50, 3072] 0
Dropout-127 [-1, 50, 3072] 0
Linear-128 [-1, 50, 768] 2,360,064
Dropout-129 [-1, 50, 768] 0
Mlp-130 [-1, 50, 768] 0
Identity-131 [-1, 50, 768] 0
Block-132 [-1, 50, 768] 0
LayerNorm-133 [-1, 50, 768] 1,536
Linear-134 [-1, 50, 2304] 1,771,776
Dropout-135 [-1, 12, 50, 50] 0
Linear-136 [-1, 50, 768] 590,592
Dropout-137 [-1, 50, 768] 0
Attention-138 [-1, 50, 768] 0
Identity-139 [-1, 50, 768] 0
LayerNorm-140 [-1, 50, 768] 1,536
Linear-141 [-1, 50, 3072] 2,362,368
GELU-142 [-1, 50, 3072] 0
Dropout-143 [-1, 50, 3072] 0
Linear-144 [-1, 50, 768] 2,360,064
Dropout-145 [-1, 50, 768] 0
Mlp-146 [-1, 50, 768] 0
Identity-147 [-1, 50, 768] 0
Block-148 [-1, 50, 768] 0
LayerNorm-149 [-1, 50, 768] 1,536
Linear-150 [-1, 50, 2304] 1,771,776
Dropout-151 [-1, 12, 50, 50] 0
Linear-152 [-1, 50, 768] 590,592
Dropout-153 [-1, 50, 768] 0
Attention-154 [-1, 50, 768] 0
Identity-155 [-1, 50, 768] 0
LayerNorm-156 [-1, 50, 768] 1,536
Linear-157 [-1, 50, 3072] 2,362,368
GELU-158 [-1, 50, 3072] 0
Dropout-159 [-1, 50, 3072] 0
Linear-160 [-1, 50, 768] 2,360,064
Dropout-161 [-1, 50, 768] 0
Mlp-162 [-1, 50, 768] 0
Identity-163 [-1, 50, 768] 0
Block-164 [-1, 50, 768] 0
LayerNorm-165 [-1, 50, 768] 1,536
Linear-166 [-1, 50, 2304] 1,771,776
Dropout-167 [-1, 12, 50, 50] 0
Linear-168 [-1, 50, 768] 590,592
Dropout-169 [-1, 50, 768] 0
Attention-170 [-1, 50, 768] 0
Identity-171 [-1, 50, 768] 0
LayerNorm-172 [-1, 50, 768] 1,536
Linear-173 [-1, 50, 3072] 2,362,368
GELU-174 [-1, 50, 3072] 0
Dropout-175 [-1, 50, 3072] 0
Linear-176 [-1, 50, 768] 2,360,064
Dropout-177 [-1, 50, 768] 0
Mlp-178 [-1, 50, 768] 0
Identity-179 [-1, 50, 768] 0
Block-180 [-1, 50, 768] 0
LayerNorm-181 [-1, 50, 768] 1,536
Linear-182 [-1, 50, 2304] 1,771,776
Dropout-183 [-1, 12, 50, 50] 0
Linear-184 [-1, 50, 768] 590,592
Dropout-185 [-1, 50, 768] 0
Attention-186 [-1, 50, 768] 0
Identity-187 [-1, 50, 768] 0
LayerNorm-188 [-1, 50, 768] 1,536
Linear-189 [-1, 50, 3072] 2,362,368
GELU-190 [-1, 50, 3072] 0
Dropout-191 [-1, 50, 3072] 0
Linear-192 [-1, 50, 768] 2,360,064
Dropout-193 [-1, 50, 768] 0
Mlp-194 [-1, 50, 768] 0
Identity-195 [-1, 50, 768] 0
Block-196 [-1, 50, 768] 0
LayerNorm-197 [-1, 50, 768] 1,536
Identity-198 [-1, 768] 0
Linear-199 [-1, 5] 3,845
================================================================
ViT三大模块
ViT主要包含三大模块:PatchEmbed、多层Transformer Encoder、MLP(FFN),下面用结构图和代码解析这第三大模块。
ViT图像预处理模块——PatchEmbed
VIT划分patches的原理:
输入图像尺寸(224x224x3),按16x16的大小进行划分,共(224x224) / (16x16) = 196个patches,每个patch的维度为(16x16x3),为满足Transformer的需求,对每个patch进行投影,[16, 16, 3]->[768],这样就将原始的[224, 224, 3]转化为[196, 768]。
![Panda](https://i-blog.csdnimg.cn/blog_migrate/9691c648db2012b49fbfeafacaf286f1.png)
代码实现如下:
class PatchEmbed(nn.Module):
"""
2D Image to Patch Embedding,二维图像patch Embedding
"""
def __init__(self, img_size=224, patch_size=16, in_c=3, embed_dim=768, norm_layer=None):
super().__init__()
img_size = (img_size, img_size) # 图片尺寸224*224
patch_size = (patch_size, patch_size) #下采样倍数,一个grid cell包含了16*16的图片信息
self.img_size = img_size
self.patch_size = patch_size
# grid_size是经过patchembed后的特征层的尺寸
self.grid_size = (img_size[0] // patch_size[0], img_size[1] // patch_size[1])
self.num_patches = self.grid_size[0] * self.grid_size[1] #path个数 14*14=196
# 通过一个卷积,完成patchEmbed
self.proj = nn.Conv2d(in_c, embed_dim, kernel_size=patch_size, stride=patch_size)
# 如果使用了norm层,如BatchNorm2d,将通道数传入,以进行归一化,否则进行恒等映射
self.norm = norm_layer(embed_dim) if norm_layer else nn.Identity()
def forward(self, x):
B, C, H, W = x.shape #batch,channels,heigth,weigth
# 输入图片的尺寸要满足既定的尺寸
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: [B, C, H, W] -> [B, C, H,W] , [B,3,224,224]-> [B,768,14,14]
# flatten: [B, C, H, W] -> [B, C, HW] , [B,768,14,14]-> [B,768,196]
# transpose: [B, C, HW] -> [B, HW, C] , [B,768,196]-> [B,196,768]
x = self.proj(x).flatten(2).transpose(1, 2)
x = self.norm(x)
return x
多层Transformer Encoder模块
该模块的主要结构是Muti-head Attention,也就是self-attention,它能够使得网络看到全局的信息,而不是CNN的局部感受野。
self-attention计算公式:
Q
=
x
W
,
K
=
x
W
K
,
V
=
x
W
V
Q=xW_,K=xW_K,V=xW_V
Q=xW,K=xWK,V=xWV
x
o
u
t
=
s
o
f
t
m
a
x
(
Q
K
T
h
)
∗
V
∗
W
o
+
x
x_{out}=softmax(\frac{QK^T}{\sqrt{h}})*V*W_o+x
xout=softmax(hQKT)∗V∗Wo+x
self-attention的结构示例如下:
![Panda](https://i-blog.csdnimg.cn/blog_migrate/3060e596f0a3fe76be930e96bc974f10.png)
class Attention(nn.Module):
"""
muti-head attention模块,也是transformer最主要的操作
"""
def __init__(self,
dim, # 输入token的dim,768
num_heads=8, #muti-head的head个数,实例化时base尺寸的vit默认为12
qkv_bias=False,
qk_scale=None,
attn_drop_ratio=0.,
proj_drop_ratio=0.):
super(Attention, self).__init__()
self.num_heads = num_heads
head_dim = dim // num_heads #平均每个head的维度
self.scale = qk_scale or head_dim ** -0.5 #进行query操作时,缩放因子
# qkv矩阵相乘操作,dim * 3使得一次性进行qkv操作
self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
self.attn_drop = nn.Dropout(attn_drop_ratio)
self.proj = nn.Linear(dim, dim)
self.proj_drop = nn.Dropout(proj_drop_ratio)
def forward(self, x):
# [batch_size, num_patches + 1, total_embed_dim] 如 [batch,197,768]
B, N, C = x.shape # N:197 , C:768
# qkv进行注意力操作,reshape进行muti-head的维度分配,permute维度调换以便后续操作
# qkv(): -> [batch_size, num_patches + 1, 3 * total_embed_dim] 如 [b,197,2304]
# reshape: -> [batch_size, num_patches + 1, 3, num_heads, embed_dim_per_head] 如 [b,197,3,12,64]
# permute: -> [3, batch_size, num_heads, num_patches + 1, embed_dim_per_head]
qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
# qkv的维度相同,[batch_size, num_heads, num_patches + 1, embed_dim_per_head]
q, k, v = qkv[0], qkv[1], qkv[2] # make torchscript happy (cannot use tensor as tuple)
# transpose: -> [batch_size, num_heads, embed_dim_per_head, num_patches + 1]
# @: multiply -> [batch_size, num_heads, num_patches + 1, num_patches + 1]
attn = (q @ k.transpose(-2, -1)) * self.scale #矩阵相乘操作
attn = attn.softmax(dim=-1) #每一patch进行softmax操作
attn = self.attn_drop(attn)
# [b,12,197,197]@[b,12,197,64] -> [b,12,197,64]
# @: multiply -> [batch_size, num_heads, num_patches + 1, embed_dim_per_head]
# 维度交换 transpose: -> [batch_size, num_patches + 1, num_heads, embed_dim_per_head]
# reshape: -> [batch_size, num_patches + 1, total_embed_dim]
x = (attn @ v).transpose(1, 2).reshape(B, N, C)
x = self.proj(x) #经过一层FC
x = self.proj_drop(x) #Dropout
return x
MLP(FFN)模块
一个MLP模块的结构如下:
![Panda](https://i-blog.csdnimg.cn/blog_migrate/45055ef047d3f0e9c0e16f2c2883bfa6.png)
class Mlp(nn.Module):
"""
MLP as used in Vision Transformer, MLP-Mixer and related networks
"""
def __init__(self, in_features, hidden_features=None, out_features=None,
act_layer=nn.GELU, # GELU是更加平滑的relu
drop=0.):
super().__init__()
out_features = out_features or in_features #如果out_features不存在,则为in_features
hidden_features = hidden_features or in_features #如果hidden_features不存在,则为in_features
self.fc1 = nn.Linear(in_features, hidden_features) # fc层1
self.act = act_layer() #激活
self.fc2 = nn.Linear(hidden_features, out_features) # fc层2
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
基本的Transformer模块
由Self-attention和MLP可以组合成Transformer的基本模块。Transformer的基本模块还使用了残差连接结构。
一个Transformer Block的结构如下:
![Panda](https://i-blog.csdnimg.cn/blog_migrate/a87a92fadbd435f04faa70e4ab02411e.png)
class Block(nn.Module):
"""
基本的Transformer模块
"""
def __init__(self,
dim,
num_heads,
mlp_ratio=4.,
qkv_bias=False,
qk_scale=None,
drop_ratio=0.,
attn_drop_ratio=0.,
drop_path_ratio=0.,
act_layer=nn.GELU,
norm_layer=nn.LayerNorm):
super(Block, self).__init__()
self.norm1 = norm_layer(dim) #norm层
self.attn = Attention(dim, num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale,
attn_drop_ratio=attn_drop_ratio, proj_drop_ratio=drop_ratio)
# NOTE: drop path for stochastic depth, we shall see if this is better than dropout here
# 代码使用了DropPath,而不是原版的dropout
self.drop_path = DropPath(drop_path_ratio) if drop_path_ratio > 0. else nn.Identity()
self.norm2 = norm_layer(dim) #norm层
mlp_hidden_dim = int(dim * mlp_ratio) #隐藏层维度扩张后的通道数
# 多层感知机
self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop_ratio)
def forward(self, x):
x = x + self.drop_path(self.attn(self.norm1(x))) # attention后残差连接
x = x + self.drop_path(self.mlp(self.norm2(x))) # mlp后残差连接
return x
Vision Transformer类的实现
class VisionTransformer(nn.Module):
def __init__(self, img_size=224, patch_size=16, in_c=3, num_classes=1000,
embed_dim=768, depth=12, num_heads=12, mlp_ratio=4.0, qkv_bias=True,
qk_scale=None, representation_size=None, distilled=False, drop_ratio=0.,
attn_drop_ratio=0., drop_path_ratio=0., embed_layer=PatchEmbed, norm_layer=None,
act_layer=None):
"""
Args:
img_size (int, tuple): input image size
patch_size (int, tuple): patch size
in_c (int): number of input channels
num_classes (int): number of classes for classification head
embed_dim (int): embedding dimension
depth (int): depth of transformer
num_heads (int): number of attention heads
mlp_ratio (int): ratio of mlp hidden dim to embedding dim
qkv_bias (bool): enable bias for qkv if True
qk_scale (float): override default qk scale of head_dim ** -0.5 if set
representation_size (Optional[int]): enable and set representation layer (pre-logits) to this value if set
distilled (bool): model includes a distillation token and head as in DeiT models
drop_ratio (float): dropout rate
attn_drop_ratio (float): attention dropout rate
drop_path_ratio (float): stochastic depth rate
embed_layer (nn.Module): patch embedding layer
norm_layer: (nn.Module): normalization layer
"""
super(VisionTransformer, self).__init__()
self.num_classes = num_classes #分类类别数量
self.num_features = self.embed_dim = embed_dim # num_features for consistency with other models
self.num_tokens = 2 if distilled else 1 #distilled在vit中没有使用到
norm_layer = norm_layer or partial(nn.LayerNorm, eps=1e-6) #层归一化
act_layer = act_layer or nn.GELU #激活函数
self.patch_embed = embed_layer(img_size=img_size, patch_size=patch_size, in_c=in_c, embed_dim=embed_dim)
num_patches = self.patch_embed.num_patches
self.cls_token = nn.Parameter(torch.zeros(1, 1, embed_dim)) #[1,1,768],以0填充
self.dist_token = nn.Parameter(torch.zeros(1, 1, embed_dim)) if distilled else None
self.pos_embed = nn.Parameter(torch.zeros(1, num_patches + self.num_tokens, embed_dim))
self.pos_drop = nn.Dropout(p=drop_ratio)
# 按照block数量等间距设置drop率
dpr = [x.item() for x in torch.linspace(0, drop_path_ratio, depth)] # stochastic depth decay rule
self.blocks = nn.Sequential(*[
Block(dim=embed_dim, num_heads=num_heads, mlp_ratio=mlp_ratio, qkv_bias=qkv_bias, qk_scale=qk_scale,
drop_ratio=drop_ratio, attn_drop_ratio=attn_drop_ratio, drop_path_ratio=dpr[i],
norm_layer=norm_layer, act_layer=act_layer)
for i in range(depth)
])
self.norm = norm_layer(embed_dim) # layer_norm
# Representation layer
if representation_size and not distilled:
self.has_logits = True
self.num_features = representation_size
self.pre_logits = nn.Sequential(OrderedDict([
("fc", nn.Linear(embed_dim, representation_size)),
("act", nn.Tanh())
]))
else:
self.has_logits = False
self.pre_logits = nn.Identity()
# Classifier head(s),分类头,self.num_features=768
self.head = nn.Linear(self.num_features, num_classes) if num_classes > 0 else nn.Identity()
self.head_dist = None
if distilled:
self.head_dist = nn.Linear(self.embed_dim, self.num_classes) if num_classes > 0 else nn.Identity()
# Weight init,权重初始化
nn.init.trunc_normal_(self.pos_embed, std=0.02)
if self.dist_token is not None:
nn.init.trunc_normal_(self.dist_token, std=0.02)
nn.init.trunc_normal_(self.cls_token, std=0.02)
self.apply(_init_vit_weights)
def forward_features(self, x):
# [B, C, H, W] -> [B, num_patches, embed_dim]
x = self.patch_embed(x) # [B, 196, 768]
# cls_token类别token [1, 1, 768] -> [B, 1, 768],扩张为batch个cls_token
cls_token = self.cls_token.expand(x.shape[0], -1, -1)
if self.dist_token is None:
x = torch.cat((cls_token, x), dim=1) # [B, 196, 768]-> [B, 197, 768],维度1上的cat
else:
x = torch.cat((cls_token, self.dist_token.expand(x.shape[0], -1, -1), x), dim=1)
x = self.pos_drop(x + self.pos_embed) #添加位置嵌入信息
x = self.blocks(x) #通过attention堆叠模块(12个)
x = self.norm(x) #layer_norm
if self.dist_token is None:
return self.pre_logits(x[:, 0]) #返回第一层特征,即为分类值
else:
return x[:, 0], x[:, 1]
def forward(self, x):
# 分类头
x = self.forward_features(x) # 经过att操作,但是没有进行分类头的前传
if self.head_dist is not None:
x, x_dist = self.head(x[0]), self.head_dist(x[1])
if self.training and not torch.jit.is_scripting():
# during inference, return the average of both classifier predictions
return x, x_dist
else:
return (x + x_dist) / 2
else:
x = self.head(x)
return x
Transformer知识点
论文:Attention Is All You Need
论文地址:https://arxiv.org/pdf/1706.03762.pdf
Transformer由Attention和Feed Forward Neural Network(也称FFN)组成,其中Attention包含self Attention与Mutil-Head Attention。
网络结构
![Panda](https://i-blog.csdnimg.cn/blog_migrate/f27a9eb84e83ba57fba646357eda6cd3.png)
attention和multi-head-attention结构:
![Panda](https://i-blog.csdnimg.cn/blog_migrate/423456bfbe289bf42681b7558d338f42.png)
计算过程:
![Panda](https://i-blog.csdnimg.cn/blog_migrate/74bcf67b014145dfc3dd2494f2e3b7cd.png)
计算复杂度对比
![Panda](https://i-blog.csdnimg.cn/blog_migrate/f86bb26870478fcdbec44ce42aa18a54.png)
Transformer的参数量和计算量
值得注意的是,模型具体的延时 (Latency) 指标也很重要,相同 FLOPs 的两个模型,它们的延时可能会差出很远来。因为 FLOPs 只考虑模型总的计算量,而不考虑内存访问成本 (memory access cost, MAC) 和并行度 (degree of parallelism)。
transformer模型由L个相同的层组成,每个层分为两部分:self-attention块和MLP块。
为表示方便,下面统一记transformer模型的层数为 l l l,隐藏层维度为 h h h,注意力头数为 a a a。词表大小为 V V V,训练数据的批次大小为 b b b,序列长度为 s s s。
Transformer的参数量
self-attention块的模型参数有Q、K、V的权重矩阵Wq、Wk、Wv和偏置,输出权重矩阵Wo和偏置,4个权重矩阵的形状为 [ h , h ] [h,h] [h,h],4个偏置的形状为 [ h ] [h] [h]。所以self- attention块的参数量为 4 h 2 + 4 h 4h^{2}+4h 4h2+4h。
MLP块由2个线性层组成,一般地,第一个线性层是先将维度从 h h h映射到 4 h 4h 4h,第二个线性层再将维度从 4 h 4h 4h映射到 h h h。第一个线性层的权重矩阵W1的形状为 [ h , 4 h ] [h,4h] [h,4h],偏置的形状为 [ 4 h ] [4h] [4h]。第二个线性层权重矩阵W2的形状为 [ 4 h , h ] [4h,h] [4h,h],偏置形状为 [ h ] [h] [h]。MLP块的参数量为 8 h 2 + 5 h 8h^{2}+5h 8h2+5h。
self-attention块和MLP块各有一个layer normalization,包含了2个可训练模型参数:缩放参数 γ \gamma γ和平移参数 β \beta β,形状都是 [ h ] [h] [h]。2个layer normalization的参数量为 4 h 4h 4h。
所以,每个transformer层的参数量为 4 h 2 + 4 h + 8 h 2 + 5 h + 4 h = 12 h 2 + 13 h 4h^{2}+4h+ 8h^{2}+5h+4h=12h^{2}+13h 4h2+4h+8h2+5h+4h=12h2+13h。
除此之外,词嵌入矩阵的参数量也较多,词向量维度通常等于隐藏层维度 h h h,词嵌入矩阵的参数量为 V h Vh Vh。最后的输出层的权重矩阵通常与词嵌入矩阵是参数共享的。
综上, l l l层transformer模型的可训练模型参数量为 l ( 12 h 2 + 13 h ) + V h l(12h^{2}+13h)+Vh l(12h2+13h)+Vh。当隐藏维度 h h h较大时,可以忽略一次项,模型参数量近似为 12 l h 2 12lh^{2} 12lh2。
Transformer的显存占用
推理的显存占用明显低于训练阶段。
1.推理阶段:
2
Φ
2\Phi
2Φ bytes
显存主要来源于模型参数,如果使用float16进行推理,由于float16数据类型的元素占2个bytes,设模型参数量为
Φ
\Phi
Φ,则显存为
2
Φ
2\Phi
2Φ bytes。此外,输入数据也需要放到GPU上,还有一些中间结果(推理过程中的中间结果用完会尽快释放掉),不过这部分占用的显存很小,可以忽略。
2.训练阶段:
20
Φ
20\Phi
20Φ bytes
占用显存的大头主要分为四部分:模型参数、前向计算过程中产生的中间激活、后向传递计算得到的梯度、优化器状态。
训练大模型时通常会采用AdamW优化器,并用混合精度训练来加速训练,基于这个前提分析显存占用。
在一次训练迭代中,每个可训练模型参数都会对应1个梯度,并对应2个优化器状态(Adam优化器梯度的一阶动量和二阶动量)。设模型参数量为
Φ
\Phi
Φ,那么梯度的元素数量为
Φ
\Phi
Φ,AdamW优化器的元素数量为
2
Φ
2\Phi
2Φ。float16数据类型的元素占2个bytes,float32数据类型的元素占4个bytes。在混合精度训练中,会使用float16的模型参数进行前向传递和后向传递,计算得到float16的梯度;在优化器更新模型参数时,会使用float32的优化器状态、float32的梯度、float32的模型参数来更新模型参数。因此,对于每个可训练模型参数,占用了weight、gradients、Adam states分别
(
2
+
4
)
+
(
2
+
4
)
+
(
4
+
4
)
=
20
(2+4)+(2+4)+(4+4)=20
(2+4)+(2+4)+(4+4)=20 bytes。
Transformer的计算量FLOPs
FLOPs,floating point operations,表示浮点数运算次数,衡量了计算量的大小。
计算量主要包括self-attention块、MLP、logits三部分的计算。