Vision Transformer是一种加入了self-attention的Sequence to Sequence model。它包括特征提取和分类两部分。
在特征提取部分,VIT所作的工作就是特征提取。特征提取部分在图片中的对应区域是Patch+Position Embedding和Transformer Encoder。
Patch+Position Embedding就是对输入进来的图片进行分块处理,每隔一定区域大小划分图片块,然后将划分后的图片块组合成序列。在获得序列信息之后,传入Transformer Encoder进行特征提取,这是Transformer特有的Mult-head Self-attention结构,通过自注意力机制,关注每个图片块的重要程度。
在分类部分,VIT所做的工作就是将提取到的特征进行分类。在进行特征提取的时候,我们会在图片序列中添加上classtoken,classtoken会作为一个单位的序列信息一起进行特征提取,在提取的过程中,classtoken会与其他的特征进行特征交互,融合其他图片序列的特征。最终,我们利用Mult-head Self-attention结构提取特征后的classtoken进行全连接分类。
Vision Transformer源码链接:VIT源码
Vision Transformer的结构如下图所示。
输入的图片尺寸为,patch size为,则分块的数目为,然后将每一个图片块展开成一维向量,每个向量大小为
设输入图片尺寸为(480,480,3),patch size为16,分块的数目为(480x480)/(16x16)=900块。每个图片块展开成一维向量,向量大小为16x16x3=768。
序列化后就获得了一个(900,768)的特征层。再加上classtoken后就是一个(901,768)的特征层。再给所有特征加上位置信息,这样网络才有区分不同区域的能力。此时生成的pos_Embedding的shape也为(901,768),代表每一个特征的位置信息。添加位置信息的方式就是生成一个(901,768)的参数矩阵,这个参数矩阵是可训练的,再把这个参数矩阵加上(901,768)的特征层就可以了。
图中0*就是casstoken。
self.pos_embed = nn.Parameter(torch.zeros(1, num_patches, num_features))
在上一步获得shape为(901,768)的序列信息后,将序列信息传入Transform Encoder进行特征提取,这是Transformer特有的Mult-head self-attention结果。通过自注意力机制,关注每个图片块的重要程度。
处理sequence最常用的就是RNN,它的输入是一串vector sequence,输出是另一串vector sequence。如果是singe directional的RNN,输出时,默认,,,都已经看过了。如果是bi-directional的RNN,输出时,默认,,,都已经看过了。但RNN有一个问题:不能并行化(hard to parallel)。例如在singe directional的RNN中要 算出,就必须 先看,再看,再看,再看,这个过程很难并行。
而self-attention,它的输入、输出和RNN一样,都是输入一串sequence,输出另外一串sequence,它和bi-directional一样,- 每一个输出都看过了整个的输入sequence。但self-attention与RNN不同的是:它的 - 每一个输出都可以并行化计算。
如上图所示,设input是-,这是一个sequence,-分别与一个矩阵相乘得到embedding,即-,接着这个embedding进入self-attention层,每个-分别乘上三个不同的transformer 矩阵、、,例如得到、、
接着将每个query 都对每个key 做attention,attention就是匹配这2个向量有多接近。比如要对、做attention,就是对这2个向量做scaled inner product,得到,再用、做attention,得到;用、做attention,得到;用、做attention,得到。
做scaled inner product的方法
其中d是q、k的维度,的值会随着维度的增大而增大,所以要除以,相当于一个归一化。
接下来就是对计算得到的所有做softmax操作,得到 。
接着将与相乘,即与相乘,与相乘,与相乘,与相乘,再把结果相加得到。在产生的整个过程中使用了整个sequence的资讯。如果要考虑局部的information,只需要学习相应的,就不带有那个分支的信息了。如果要考虑全局的information,只需要学习相应的,就带有所有分支的信息了。
同样的方法可以求出、、。
以上一连串的计算,self-attention layer做的事情和RNN是一样的,但是它可以并行化计算。
首先输入的embedding是,用乘得到,它的每一列代表一个向量。用乘得到,它的每一列代表一个向量。用乘得到,它的每一列代表一个向量。
输入矩阵分别乘上三个不同的矩阵、、得到3个中间矩阵、、,将转置之后与相乘得到attention矩阵,代表每一个位置两两之间的attention。再将它去softmax操作之后得到,最后将它乘以矩阵得到输出vect或
以2个head为例,由生成进一步乘以2个转移矩阵变为、,由生成 进一步乘以2个转移矩阵变为、, 由生成 进一步乘以2个转移矩阵变为、。接下来,做attention,得到加权和的权重,再与做加权和,得到最终的,同理得到。现在有了,,把它们连接起来,再通过一个transformation矩阵调整维度,使之与维度一致。
Multi-Head Attention 包含多个 Self-Attention 层,首先将输入分别传递到 2个不同的 Self-Attention 中,计算得到 2 个输出结果。得到2个输出矩阵之后,Multi-Head Attention 将它们拼接在一起 (Concat),然后传入一个Linear层,得到 Multi-Head Attention 最终的输出 。可以看到 Multi-Head Attention 输出的矩阵与其输入的矩阵的维度是一样的。
例如在第一步进行图像分割后,我们获得的特征层为(901,768)。在施加多头的时候,我们直接对(901,768)的最后一维度进行分割,比如我们想分割成12个头,那么矩阵的shape就变成(901,12,64)。然后对(901,12,64)进行转置,将12放到前面,获得(12,901,64)的特征层。之后忽略这个12,把它和batchsize维度同等对待,只对901,64进行处理,以上就是注意力机制的过程了。
注意力机制代码如下:
class Attention(nn.Module):
def __init__(self, dim, num_heads=12, qkv_bias=False, attn_drop=0., proj_drop=0.):
super().__init__()
self.num_heads = num_heads
self.scale = (dim // num_heads) ** -0.5
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)
def forward(self, x):
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]
attn = (q @ k.transpose(-2, -1)) * self.scale
attn = attn.softmax(dim=-1)
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
在完成MultiHeadSelfAttention的构建后,我们需要在其后加上两个全连接。就构建了整个TransformerBlock。
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=GELU, drop=0.):
super().__init__()
out_features = out_features or in_features
hidden_features = hidden_features or in_features
drop_probs = (drop, drop)
self.fc1 = nn.Linear(in_features, hidden_features)
self.act = act_layer()
self.drop1 = nn.Dropout(drop_probs[0])
self.fc2 = nn.Linear(hidden_features, out_features)
self.drop2 = nn.Dropout(drop_probs[1])
def forward(self, x):
x = self.fc1(x)
x = self.act(x)
x = self.drop1(x)
x = self.fc2(x)
x = self.drop2(x)
return x
class Block(nn.Module):
def __init__(self, dim, num_heads, mlp_ratio=4., qkv_bias=False, drop=0., attn_drop=0.,
drop_path=0., act_layer=GELU, norm_layer=nn.LayerNorm):
super().__init__()
self.norm1 = norm_layer(dim)
self.attn = Attention(dim, num_heads=num_heads, qkv_bias=qkv_bias, attn_drop=attn_drop, proj_drop=drop)
self.norm2 = norm_layer(dim)
self.mlp = Mlp(in_features=dim, hidden_features=int(dim * mlp_ratio), act_layer=act_layer, drop=drop)
self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
def forward(self, x):
x = x + self.drop_path(self.attn(self.norm1(x)))
#-----------------------------------------------#
# 施加层标准化、多头注意力机制、残差结构
#-----------------------------------------------#
x = x + self.drop_path(self.mlp(self.norm2(x)))
#-----------------------------------------------#
# 施加层标准化、两次全连接、残差结构
#-----------------------------------------------#
return x
整个VIT模型由一个Patch+Position Embedding加上多个TransformerBlock组成。典型的TransforerBlock的数量为12个。
self.blocks = nn.Sequential(
*[
Block(
dim = num_features,
num_heads = num_heads,
mlp_ratio = mlp_ratio,
qkv_bias = qkv_bias,
drop = drop_rate,
attn_drop = attn_drop_rate,
drop_path = dpr[i],
norm_layer = norm_layer,
act_layer = act_layer
)for i in range(depth)#depth == 12
]
)
Vision Transformer的构建代码:
import math
from collections import OrderedDict
from functools import partial
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
#--------------------------------------#
# Gelu激活函数的实现
# 利用近似的数学公式
#--------------------------------------#
class GELU(nn.Module):
def __init__(self):
super(GELU, self).__init__()
def forward(self, x):
return 0.5 * x * (1 + torch.tanh(np.sqrt(2 / np.pi) * (x + 0.044715 * torch.pow(x,3))))
def drop_path(x, drop_prob: float = 0., training: bool = False):
if drop_prob == 0. or not training:
return x
keep_prob = 1 - drop_prob
shape = (x.shape[0],) + (1,) * (x.ndim - 1)
random_tensor = keep_prob + torch.rand(shape, dtype=x.dtype, device=x.device)
random_tensor.floor_()
output = x.div(keep_prob) * random_tensor
return output
class DropPath(nn.Module):
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)
class PatchEmbed(nn.Module):
def __init__(self, input_shape=[480, 480], patch_size=16, in_chans=3, num_features=768, norm_layer=None, flatten=True):
super().__init__()
self.num_patches = (input_shape[0] // patch_size) * (input_shape[1] // patch_size)
self.flatten = flatten
self.proj = nn.Conv2d(in_chans, num_features, kernel_size=patch_size, stride=patch_size)
self.norm = norm_layer(num_features) if norm_layer else nn.Identity()
def forward(self, x):
x = self.proj(x)
if self.flatten:
x = x.flatten(2).transpose(1, 2) # BCHW -> BNC
x = self.norm(x)
return x
#--------------------------------------------------------------------------------------------------------------------#
# Attention机制
# 将输入的特征qkv特征进行划分,首先生成query, key, value。query是查询向量、key是键向量、v是值向量。
# 然后利用 查询向量query 点乘 转置后的键向量key,这一步可以通俗的理解为,利用查询向量去查询序列的特征,获得序列每个部分的重要程度score。
# 然后利用 score 点乘 value,这一步可以通俗的理解为,将序列每个部分的重要程度重新施加到序列的值上去。
#--------------------------------------------------------------------------------------------------------------------#
class Attention(nn.Module):
def __init__(self, dim, num_heads=12, qkv_bias=False, attn_drop=0., proj_drop=0.):
super().__init__()
self.num_heads = num_heads
self.scale = (dim // num_heads) ** -0.5
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)
def forward(self, x):
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]
attn = (q @ k.transpose(-2, -1)) * self.scale
attn = attn.softmax(dim=-1)
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
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=GELU, drop=0.):
super().__init__()
out_features = out_features or in_features
hidden_features = hidden_features or in_features
drop_probs = (drop, drop)
self.fc1 = nn.Linear(in_features, hidden_features)
self.act = act_layer()
self.drop1 = nn.Dropout(drop_probs[0])
self.fc2 = nn.Linear(hidden_features, out_features)
self.drop2 = nn.Dropout(drop_probs[1])
def forward(self, x):
x = self.fc1(x)
x = self.act(x)
x = self.drop1(x)
x = self.fc2(x)
x = self.drop2(x)
return x
class Block(nn.Module):
def __init__(self, dim, num_heads, mlp_ratio=4., qkv_bias=False, drop=0., attn_drop=0.,
drop_path=0., act_layer=GELU, norm_layer=nn.LayerNorm):
super().__init__()
self.norm1 = norm_layer(dim)
self.attn = Attention(dim, num_heads=num_heads, qkv_bias=qkv_bias, attn_drop=attn_drop, proj_drop=drop)
self.norm2 = norm_layer(dim)
self.mlp = Mlp(in_features=dim, hidden_features=int(dim * mlp_ratio), act_layer=act_layer, drop=drop)
self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
def forward(self, x):
x = x + self.drop_path(self.attn(self.norm1(x)))
x = x + self.drop_path(self.mlp(self.norm2(x)))
return x
class Conv2dReLU(nn.Sequential):
def __init__(
self,
in_channels,
out_channels,
kernel_size,
padding=0,
stride=1,
use_batchnorm=True,
):
conv = nn.Conv2d(
in_channels,
out_channels,
kernel_size,
stride=stride,
padding=padding,
bias=not (use_batchnorm),
)
relu = nn.ReLU(inplace=True)
bn = nn.BatchNorm2d(out_channels)
super(Conv2dReLU, self).__init__(conv, bn, relu)
class VisionTransformer(nn.Module):
def __init__(
self, input_shape=[480, 480], patch_size=16, in_chans=3, num_classes=2, num_features=768,
depth=12, num_heads=12, mlp_ratio=4., qkv_bias=True, drop_rate=0.1, attn_drop_rate=0.1, drop_path_rate=0.1,
norm_layer=partial(nn.LayerNorm, eps=1e-6), act_layer=GELU
):
super().__init__()
#-----------------------------------------------#
# 480, 480, 3 -> 196, 768
#-----------------------------------------------#
self.patch_embed = PatchEmbed(input_shape=input_shape, patch_size=patch_size, in_chans=in_chans, num_features=num_features)
num_patches = (480 // patch_size) * (480 // patch_size)
self.num_features = num_features
self.new_feature_shape = [int(input_shape[0] // patch_size), int(input_shape[1] // patch_size)]
self.old_feature_shape = [int(480 // patch_size), int(480 // patch_size)]
#--------------------------------------------------------------------------------------------------------------------#
# classtoken部分是transformer的分类特征。用于堆叠到序列化后的图片特征中,作为一个单位的序列特征进行特征提取。
#
# 在利用步长为16x16的卷积将输入图片划分成14x14的部分后,将14x14部分的特征平铺,一幅图片会存在序列长度为196的特征。
# 此时生成一个classtoken,将classtoken堆叠到序列长度为196的特征上,获得一个序列长度为197的特征。
# 在特征提取的过程中,classtoken会与图片特征进行特征的交互。最终分类时,我们取出classtoken的特征,利用全连接分类。
#--------------------------------------------------------------------------------------------------------------------#
# 196, 768 -> 197, 768
self.cls_token = nn.Parameter(torch.zeros(1, 1, num_features))
#--------------------------------------------------------------------------------------------------------------------#
# 为网络提取到的特征添加上位置信息。
# 以输入图片为480, 480, 3为例,我们获得的序列化后的图片特征为196, 768。加上classtoken后就是197, 768
# 此时生成的pos_Embedding的shape也为197, 768,代表每一个特征的位置信息。
#--------------------------------------------------------------------------------------------------------------------#
# 197, 768 -> 197, 768
self.pos_embed = nn.Parameter(torch.zeros(1, num_patches, num_features))
self.pos_drop = nn.Dropout(p=drop_rate)
#-----------------------------------------------#
# 197, 768 -> 197, 768 12次
#-----------------------------------------------#
dpr = [x.item() for x in torch.linspace(0, drop_path_rate, depth)]
self.blocks = nn.Sequential(
*[
Block(
dim = num_features,
num_heads = num_heads,
mlp_ratio = mlp_ratio,
qkv_bias = qkv_bias,
drop = drop_rate,
attn_drop = attn_drop_rate,
drop_path = dpr[i],
norm_layer = norm_layer,
act_layer = act_layer
)for i in range(depth)
]
)
self.norm = norm_layer(num_features)
self.head = nn.Linear(num_features, num_classes) if num_classes > 0 else nn.Identity()
n_patches = (480 // 16) * (480 // 16)
self.position_embeddings = nn.Parameter(torch.zeros(1, n_patches, 768))
self.conv_more = Conv2dReLU(
768,
2048,
kernel_size=3,
padding=1,
use_batchnorm=True,
)
def forward_features(self, x):
x = self.patch_embed(x)
# cls_token = self.cls_token.expand(x.shape[0], -1, -1)
# x = torch.cat((cls_token, x), dim=1)
#
# cls_token_pe = self.pos_embed[:, 0:1, :]
# img_token_pe = self.pos_embed[:, 1: , :]
# print(img_token_pe.shape)
#
# img_token_pe = img_token_pe.view(1, *self.old_feature_shape, -1).permute(0, 3, 1, 2)
# print(img_token_pe.shape)
# img_token_pe = F.interpolate(img_token_pe, size=self.new_feature_shape, mode='bicubic', align_corners=False)
# print(img_token_pe.shape)
# img_token_pe = img_token_pe.permute(0, 2, 3, 1).flatten(2)
# print(img_token_pe.shape)
# pos_embed = torch.cat([cls_token_pe, img_token_pe], dim=1)
# x = x.flatten(2)
# x = x.transpose(-1, -2)
x = x + self.position_embeddings
x = self.pos_drop(x)
x = self.blocks(x)
x = self.norm(x)
return x
def forward1(self,hidden_states):
B, n_patch, hidden = hidden_states.size() # reshape from (B, n_patch, hidden) to (B, h, w, hidden)
h, w = int(np.sqrt(n_patch)), int(np.sqrt(n_patch))
x = hidden_states.permute(0, 2, 1)
x = x.contiguous().view(B, hidden, h, w)
x = self.conv_more(x)
return x
def forward(self, x):
x = self.forward_features(x)
x = self.forward1(x)
# print(x.shape)
return x
def freeze_backbone(self):
backbone = [self.patch_embed, self.cls_token, self.pos_embed, self.pos_drop, self.blocks[:8]]
for module in backbone:
try:
for param in module.parameters():
param.requires_grad = False
except:
module.requires_grad = False
def Unfreeze_backbone(self):
backbone = [self.patch_embed, self.cls_token, self.pos_embed, self.pos_drop, self.blocks[:8]]
for module in backbone:
try:
for param in module.parameters():
param.requires_grad = True
except:
module.requires_grad = True
def vit(input_shape=[480, 480], pretrained=False, num_classes=2):
model = VisionTransformer(input_shape)
if pretrained:
model.load_state_dict(torch.load("model_data/vit-patch_16.pth"))
if num_classes!=1000:
model.head = nn.Linear(model.num_features, num_classes)
return model