Vision Transformer（Pytorch版）代码阅读
前言
因为Google Research官方的Vision Transformer源码是tensorflow版本，而笔者平时多用pytorch，所以在github上找了作者rwightman版本的代码：rwightman/pytorch-image-models/timm/models/vision_transformer.py
Vision Transformer介绍博客：论文阅读笔记：Vision Transformer
下面的代码介绍以vit_base_patch16_224(ViT-B/16：patch_size=16, img_size=224)为例。
VIT Model
原文中模型由三个模块组成：
· Linear Projection of Flattened Patches
· Transformer Encoder
· MLP Head
对应代码中的三个模块：
· patch embedding layer
· Block
· Representation layer + Classifier head
Linear Projection of Flattened Patches
如图，Linear Projection of Flattened Patches的实现的通过一个kernel_size=stride=16的卷积加上一个flatten实现的。他的功能是将
244
×
244
×
3
244×244×3
244×244×3 的的2D Image转换为
196
×
768
196×768
196×768 的Patch Embedding。具体代码及注释如下：
class PatchEmbed(nn.Module):
"""
2D Image to Patch Embedding
"""
def __init__(self, img_size=224, patch_size=16, in_c=3, embed_dim=768, norm_layer=None):
super().__init__()
'''
image_size = (244,244)
patch_size = (16,16)
gird_size = (244/16,244/16)=(14,14)
num_patches = 14 * 14 = 196
'''
img_size = (img_size, img_size)
patch_size = (patch_size, patch_size)
self.img_size = img_size
self.patch_size = patch_size
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]
'''
使用大小为16，stride为16的卷积核实现embeding，
输出14*14大小，通道为768（768 = 16*16*3，相当于将每个patch部分转换为1维向量）的patch
'''
self.proj = nn.Conv2d(in_c, embed_dim, kernel_size=patch_size, stride=patch_size)
'''
如果norm_layer为true则使用layerNorm，这里作者没有使用，
所以self.norm = nn.Identity()，对输入不做任何改变直接输出
'''
self.norm = norm_layer(embed_dim) if norm_layer else nn.Identity()
def forward(self, x):
B, C, H, W = x.shape
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]})."
'''
self.proj(x):[B,3,244,244]->[B,768,14,14]
flatten(2):[B,768,14,14]->[B,768,14*14]=[B,768,196]
transpose(1, 2):[B,768,196]->[B,196,768]
self.norm(x)不对输入做处理直接输出
'''
x = self.proj(x).flat1ten(2).transpose(1, 2)
x = self.norm(x)
return x
Transformer Encoder
Transformer Encoder由Attention、MLP和DropPath代码组成，其结构图如下：
Multi-Head Attention
关于 Multi-Head Attention 的结构图和详细介绍可查看博文，论文阅读笔记：Attention Is All You Need。
Attention具体代码及注释如下：
class Attention(nn.Module):
def __init__(self,
dim, # 输入token的dim 768
num_heads=8,
qkv_bias=False,
qk_scale=None,
attn_drop_ratio=0.,
proj_drop_ratio=0.):
super(Attention, self).__init__()
'''
num_heads = 12
head_dim = 768 // 12 = 64 （Attention is all you need论文中提到的dk=dv=dmodel/h）
scale = 64 ^ -0.5 = 1/8（Attention is all you need论文中Scaled Dot-Product Attention提到的公式Attention(Q,K,V)中的根号dk分之一）
qkv:将输入线性映射到q,k,v
proj：Attention is all you need论文中Multi-Head Attention最后的Linear
'''
self.num_heads = num_heads
head_dim = dim // num_heads
self.scale = qk_scale or head_dim ** -0.5
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):
'''
B = batch_size
N = 197
C = 768
'''
B, N, C = x.shape
'''
qkv(x) : [B,197,768] -> [B,197,768*3]
reshape : [B,197,768*3] -> [B,197,3,12,64] (3分别代表qkv，12个head，每个head为64维向量)
permute：[B,197,3,12,64] -> [3,B,12,197,64]
'''
qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
'''
q,k,v = [B,12,197,64]
'''
q, k, v = qkv[0], qkv[1], qkv[2] # make torchscript happy (cannot use tensor as tuple)
'''
K.transpose(-2, -1) : [B,12,197,64] = [B,12,64,197]
q @ K.transpose(-2, -1) : [B,12,197,64] @ [B,12,64,197] = [B,12,197,197]
attn : [B,12,197,197]
attn.softmax(dim=-1)对最后一个维度（即每一行）进行softmax处理
'''
attn = (q @ k.transpose(-2, -1)) * self.scale
attn = attn.softmax(dim=-1)
attn = self.attn_drop(attn)
'''
attn @ v = [B,12,197,197] @ [B,12,197,64] = [B,12,197,64]
transpose(1, 2) : [B,197,12,64]
reshape : [B,197,768]
'''
x = (attn @ v).transpose(1, 2).reshape(B, N, C)
x = self.proj(x)
x = self.proj_drop(x)
return x
MLP
MLP结构和代码都很简单，就是全连接加激活函数加dropout，这里的激活函数用的GELU:
G E L U ( x ) = 0.5 x ( 1 + t a n h [ 2 π ( x + 0.044715 x 3 ) ] ) GELU(x)=0.5x(1+tanh[\frac{2}{π}(x+0.044715x^3)]) GELU(x)=0.5x(1+tanh[π2?(x+0.044715x3)])
MLP模块代码如下：
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, 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
DropPath
在Transformer Encoder中代码使用DropPath代替论文中的Dropout，具体代码及注释如下：
def drop_path(x, drop_prob: float = 0., training: bool = False):
'''
x.shape : [B,197,768]
'''
if drop_prob == 0. or not training:
return x
keep_prob = 1 - drop_prob
'''
shape = [B,1,1]
即将X的第一维度保留，其他维度改为1
'''
shape = (x.shape[0],) + (1,) * (x.ndim - 1) # work with diff dim tensors, not just 2D ConvNets
'''
生成形状为shape的随机张量并加上keep_prob
'''
random_tensor = keep_prob + torch.rand(shape, dtype=x.dtype, device=x.device)
'''
将随机张量向下取整，一部分为0，一部分为1
'''
random_tensor.floor_() # binarize
'''
将x除以keep_prob再乘上随机张量，一部分变成0，一部分保留
'''
output = x.div(keep_prob) * random_tensor
return output
class DropPath(nn.Module):
"""
Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks).
"""
def __init__(self, drop_prob=None):
super(DropPath, self).__init__()
self.drop_prob = drop_prob
def forward(self, x):
return drop_path(x, self.drop_prob, self.training)
MLP Head
原文中关于MLP Head的代码：
# 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)
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()
这里的代码也很简单，就不做过多注释了，代码中distilled = False，所以:
self.pre_logits = nn.Sequential(nn.Linear,(embed_dim, representation_size)nn.Tanh())
self.head = nn.Linear(self.num_features, num_classes)
MLPHead(x) = self.head(self.pre_logits(x[:, 0]))
VisionTransformer
ViT-B/16整体网络结构如下图：
ViT-B/16模型使用的图像输入尺寸为 224×224×3，patch尺寸为16×16×3，每个patch embed的维度为768，transformer encoder block的个数为12， Multi-Head Attention的head个数为12，最后两个参数看调用模型时的参数设置，representation_size为pre_logits中全连接层节点个数，num_classes为预测的总分类数。
def vit_base_patch16_224_in21k(num_classes: int = 21843, has_logits: bool = True):
model = VisionTransformer(img_size=224,
patch_size=16,
embed_dim=768,
depth=12,
num_heads=12,
representation_size=768 if has_logits else None,
num_classes=num_classes)
return model
VisionTransformer具体代码及注释如下：
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): 输入图像尺寸
patch_size (int, tuple): patch 尺寸
in_c (int): 输入通道
num_classes (int): 分类数
embed_dim (int): patchembed 维度
depth (int): transformer encoder 模块（ Block 模块）个数
num_heads (int): Multi-Head Attention 中的 head 个数
mlp_ratio (int): MLP 隐藏层和 embed_dim 的比例
qkv_bias (bool): 是否使用 qkv 偏置（即使用 Linear 将输入映射到 qkv 时，Linear是否使用 bias ）
qk_scale (float): qk缩放比例，默认使用根号 dim_k 分之一
representation_size (Optional[int]): pre-logits 中的全连接节点个数，如果是 None 则不要 pre-logits (MLP Head 中只有一个全连接层)
distilled (bool): 是否使用 DeiT 模型（基于知识蒸馏的transformer）,在 VIT 中默认为 False
drop_ratio (float): dropout概率
attn_drop_ratio (float): attention 中的 dropout 概率
drop_path_ratio (float): attention 中的 droppath 概率
embed_layer (nn.Module): patch embedding 层
norm_layer: (nn.Module): normalization 层
"""
super(VisionTransformer, self).__init__()
self.num_classes = num_classes
'''
self.num_features = self.embed_dim = 768
self.num_tokens = 1
norm_layer = nn.LayerNorm(eps=1e-6)
act_layer = nn.GELU
'''
self.num_features = self.embed_dim = embed_dim # num_features for consistency with other models
self.num_tokens = 2 if distilled else 1
norm_layer = norm_layer or partial(nn.LayerNorm, eps=1e-6)
act_layer = act_layer or nn.GELU
'''
构建patch embeding layer
num_patches = (224/16) * (224/16) = 196
'''
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
'''
构建可学习参数:
self.cls_token : [1,1,768] 分类token
self.dist_token : None
self.pos_embed : [1,197,768] 位置编码
'''
self.cls_token = nn.Parameter(torch.zeros(1, 1, embed_dim))
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)
'''
构建首项为0，长度为depth的等差数列，且每一项小于drop_path_ratio
也就是说 传入 Block 的 droppath 概率是递增的。
代码这里是让 drop_path_ratio 默认等于0
最后利用参数构建 depth(12) 层 block 层
并把 LayerNorm(embed_dim) 赋值给self.norm
'''
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)
'''
构建 pre_logits :
1.全连接层:输入embed_dim(768),输出representation_size(768)
2.激活函数:Tanh
'''
# 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()
'''
构建分类器:
self.num_features = 768
'''
# Classifier head(s)
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()
'''
初始化pos_embed、cls_token
初始化网络其他层的权重
'''
# Weight init
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):
'''
self.patch_embed(x) : [B,3,244,244] -> [B,196,768]
合并 cls_token:
self.cls_token : [1,1,768]
cls_token : [B,1,768]
x = torch.cat((cls_token, x), dim=1) : [B,197,768]
'''
x = self.patch_embed(x)
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)
else:
x = torch.cat((cls_token, self.dist_token.expand(x.shape[0], -1, -1), x), dim=1)
'''
加上位置编码:
x = x + self.pos_embed : [B,197,768]
经过 Attention blocks 和 LayerNorm : [B,197,768]
最后返回分类 token 并传入 pre_logits:
return self.pre_logits(x[:, 0]) : [B,768]
'''
x = self.pos_drop(x + self.pos_embed)
x = self.blocks(x)
x = self.norm(x)
if self.dist_token is None:
return self.pre_logits(x[:, 0])
else:
return x[:, 0], x[:, 1]
def forward(self, x):
'''
self.forward_features(x) : [B,3,244,244] -> [B,768]
x = self.head(x) : [B,768] -> [B,num_classes]
'''
x = self.forward_features(x)
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
def _init_vit_weights(m):
"""
ViT weight initialization
:param m: module
"""
if isinstance(m, nn.Linear):
nn.init.trunc_normal_(m.weight, std=.01)
if m.bias is not None:
nn.init.zeros_(m.bias)
elif isinstance(m, nn.Conv2d):
nn.init.kaiming_normal_(m.weight, mode="fan_out")
if m.bias is not None:
nn.init.zeros_(m.bias)
elif isinstance(m, nn.LayerNorm):
nn.init.zeros_(m.bias)
nn.init.ones_(m.weight)
上述代码的distilled参数所涉及的 DeiT models 代码中并没有使用，论文中也没有提到，如有疑惑可查看ViT和DeiT的原理与使用。