前言

Transformer在NLP领域大放异彩,而实际上NLP(Natural Language Processing,自然语言处理)领域技术的发展都要先于CV(Computer Vision,计算机视觉),那么如何将Transformer这类模型也能适用到图像数据上呢?
在2017年Transformer发布后,历经3年时间,Vision Transformer于2020年问世。与Transformer相同,Vision Transformer也是由Google Brain和Google Research团队开发,然而并不是同一批人(除了Jakob Uszkoreit)。
值得一提的是,Vision Transformer并不是第一个将Transformer应用到CV上的,因为这些巨头的存在(如Google,FaceBook),论文的名气也自然会更大,而且从如今ViT的泛用程度来看也是,大家对其认可度更高纷纷follow。和这些巨头庞大资源比,高校产出的论文光芒显得黯淡了许多。而在大模型时代更是如此,都是“大力出奇迹”的结果。可大模型大数据训练就是AI的最终形态了吗,我觉得不然……或许在AI真正具有“智能”时,深度学习的模型也并不需要这么大吧,因为人脑正是有了联想推理才能拥有知识和技能,而不完全单靠记忆。


一、Vision Transformer架构介绍

在这里插入图片描述

1. Patch Embedding

Patch Embeding可以说是ViT的创新亮点,不过也不复杂,就是将图片按patch进行切分然后通过线性映射层将其转化为和文本向量相同的一维张量。这个操作相当于把一张二维的图片展平成一维向量。而如何将图片按patch映射成向量,做法就分为两种:1)一种是将patch简单粗暴地展平,然后通过Linear全连接层映射;2)另一种也是如今被推崇的方法,那就是先对patch进行卷积操作,然后映射为一维向量,此处可参考后文的代码。

2. Multi-Head Attention

与原Transfomrer相同,这里不再介绍,详情可以去阅读我上篇博客Pytorch从零开始实现Transformer (from scratch)

3. Transformer Block

如图,(a) 是最初Transformer的Encoder结构图, (b)则是ViT的。可以明显看出,Transformer是在multi-head attention和feedforward模块后进行残差操作(即Add)和Norm(标准化),而ViT则是在这些模块前使用Norm操作。

Feed Forward

ViT的Feed Forward模块使用两层全连接层(Linear)和GeLU激活函数。而Transformer使用的是ReLu激活函数。
GeLu于2016年被提出,见于Bridging Nonlinearities and Stochastic Regularizers with Gaussian Error Linear Units,后来经过论文修改改名为“Gaussian Error Linear Units (GELUs)”。论文给出了ReLu和GeLu的图示:
在这里插入图片描述
ReLu确实好用,但缺点也很明显,其在输入值小于0时都会输出0,这样“一刀切”的策略势必会丢掉信息,累计error。因此后来出现了GeLu、LeakyReLu等一系列激活函数来解决神经元”死亡“问题,让输入值小于0时输出不总是0。


二、预备知识

本节的两个操作都是为了方便编程人员更好对tensor进行操作,且让代码更具可读性。

1. Einsum

Einsum即爱因斯坦和,可以更直观地操作张量计算而不需要知道其内部代码如何实现,einsum支持pytorch、TensorFlow和numpy,在pytorch中可用torch.einsum调用。推荐观看Youtube视频:https://www.youtube.com/watch?v=pkVwUVEHmfI 进行学习。
接下来用一些例子便于读者了解einsum的作用(例子来自上面提到的youtube视频):

import torch

x = torch.tensor([[1,2,3],[4,5,6]])

"""张量的排列"""
torch.einsum("ij->ji",x)
#tensor([[1, 4],
#        [2, 5],
#        [3, 6]])

"""张量求和"""
torch.einsum("ij->",x)
#tensor(21)

"""列求和"""
torch.einsum("ij->j",x)
#tensor([5, 7, 9])

"""行求和"""
torch.einsum("ij->i",x)
#tensor([ 6, 15])

"""矩阵与向量间乘法"""
y = torch.tensor([[7,8,9]]) #定义向量y,y.size()为(1,3)
torch.einsum("ij,kj->ik",x,y)
#tensor([[ 50],
#        [122]])
"""解释:
[[1,2,3],[4,5,6]] x [[7,8,9]] = [[1x7+2x8+3x9],[4x7+5x8+6x9]] = [[50],[122]]
"""

"""矩阵与矩阵乘法"""
torch.einsum("ij,kj->ik",x,x)
#tensor([[14, 32],
#        [32, 77]])

"""矩阵点积"""
torch.einsum("ij,ij->",x,x)
#tensor(91)
"""解释:
矩阵每一位数按位相乘最后相加:1x1+2x2+3x3+4x4+5x5+6x6 = 91
"""

"""Hadamard乘积 (element-wise multiplication)"""
torch.einsum("ij,ij->ij",x,x) # 即上一步矩阵点积不做求和运算
#tensor([[ 1,  4,  9],
#        [16, 25, 36]])

"""外积"""
a = torch.tensor([1,2,3])
b = torch.tensor([1,3,5])
torch.einsum("i,j->ij",a,b)
#tensor([[ 1,  3,  5],
#        [ 2,  6, 10],
#        [ 3,  9, 15]])
"""解释:
也可以看做a的转置矩阵乘b,即(3,1)x(1,3)=(3,3),所以得到(3,3)的矩阵
"""

"""Batch矩阵乘法"""
c = torch.tensor([[[0,0,0,0],[1,1,1,1]],
				[[2,2,2,2],[3,3,3,3]]]) # torch.Size([2, 2, 4])
d = torch.tensor([[[0,0,0],[1,1,1],[2,2,2],[3,3,3]],
				[[1,1,1],[2,2,2],[3,3,3],[4,4,4]]]) # torch.Size([2, 4, 3])
torch.einsum("ijk,ikl->ijl",c,d)
#tensor([[[ 0,  0,  0],
#         [ 6,  6,  6]],
#        [[20, 20, 20],
#         [30, 30, 30]]]) #torch.Size([2, 2, 3])

"""矩阵对角线(Diagonal)"""
x = torch.tensor([[1,2,3],[4,5,6],[7,8,9]])
torch.einsum("ii->i",x)
#tensor([1, 5, 9])

"""矩阵的迹(Trace)"""
torch.einsum("ii->",x)
#tensor(15)

2. Einops

是大牛Alex Rogozhnikov受Einsum启发所开发的一个库,主要用于张量的变形等操作。其写了一篇论文并做了ICLR2022的Oral论文报告,视频和论文链接为Einops: Clear and Reliable Tensor Manipulations with Einstein-like Notation,github开源项目在https://github.com/arogozhnikov/einops

pip install einops

einops的github项目已经很完备很清晰,这里就不班门弄斧了,大家自行去学习吧。
接下来欣赏一下einops项目中的demo展示吧:
在这里插入图片描述


三、Vision Transformer代码实现

这次代码并不是直接取用某一份代码,而是参考包括Pytorch官方的代码库、网上博客、github项目综合出的一份Vision Transformer代码,尽可能还原ViT又兼顾代码可读性以便读者学习理解。此处引用比ViT原论文更加具体的ViT模型图:
ViT流程图
此图出自论文Vision Transformers for Remote Sensing Image Classification

0. 导入库

import torch
import torch.nn as nn
from einops import rearrange, repeat
from einops.layers.torch import Rearrange

1. Patch Embedding

class PatchEmbedding(nn.Module):
    def __init__(self, embed_size=768, patch_size=16, channels=3, img_size=224):
        super(PatchEmbedding, self).__init__()
        self.patch_size = patch_size
        # Version 1.0
        # self.patch_projection = nn.Sequential(
        #     Rearrange("b c (h h1) (w w1) -> b (h w) (h1 w1 c)", h1=patch_size, w1=patch_size),
        #     nn.Linear(patch_size * patch_size * channels, embed_size)
        # )

        # Version 2.0
        self.patch_projection = nn.Sequential(
            nn.Conv2d(channels, embed_size, kernel_size=(patch_size, patch_size), stride=(patch_size, patch_size)),
            Rearrange("b e (h) (w) -> b (h w) e"),
        )
        self.cls_token = nn.Parameter(torch.randn(1, 1, embed_size))
        self.positions = nn.Parameter(torch.randn((img_size // patch_size) ** 2 + 1, embed_size))

    def forward(self, x):
        batch_size = x.shape[0]
        x = self.patch_projection(x)
        # prepend the cls token to the input
        cls_tokens = repeat(self.cls_token, "() n e -> b n e", b=batch_size)
        x = torch.cat([cls_tokens, x], dim=1)
        # add position embedding
        x += self.positions
        return x

2. Residual & Norm

class Residual(nn.Module):
    def __init__(self, fn):
        super(Residual, self).__init__()
        self.fn = fn

    def forward(self, x, **kwargs):
        return self.fn(x, **kwargs) + x


class PreNorm(nn.Module):
    def __init__(self, dim, fn):
        super(PreNorm, self).__init__()
        self.norm = nn.LayerNorm(dim)
        self.fn = fn

    def forward(self, x, **kwargs):
        return self.fn(self.norm(x), **kwargs)

3. Multi-Head Attention & FeedForward

class FeedForward(nn.Module):
    def __init__(self, dim, hidden_dim, dropout=0.):
        super(FeedForward, self).__init__()
        self.mlp = nn.Sequential(
            nn.Linear(dim, hidden_dim),
            nn.GELU(),
            nn.Dropout(dropout),
            nn.Linear(hidden_dim, dim),
            nn.Dropout(dropout),
        )

    def forward(self, x):
        return self.mlp(x)


class MultiHeadAttention(nn.Module):
    def __init__(self, embed_dim=768, n_heads=8, dropout=0.):
        """
        Args:
            embed_dim: dimension of embeding vector output
            n_heads: number of self attention heads
        """
        super(MultiHeadAttention, self).__init__()

        self.embed_dim = embed_dim  # 768 dim
        self.n_heads = n_heads  # 8
        self.head_dim = self.embed_dim // self.n_heads  # 768/8 = 96. each key,query,value will be of 96d
        self.scale = self.head_dim ** -0.5

        self.attn_drop = nn.Dropout(dropout)
        # key,query and value matrixes
        self.to_qkv = nn.Linear(self.embed_dim, self.embed_dim * 3, bias=False)
        self.to_out = nn.Sequential(
            nn.Linear(self.embed_dim, self.embed_dim),
            nn.Dropout(dropout)
        )

    def forward(self, x):
        """
        Args:
           x : a unified vector of key query value
        Returns:
           output vector from multihead attention
        """
        qkv = self.to_qkv(x).chunk(3, dim=-1)
        q, k, v = map(lambda t: rearrange(t, "b n (h d) -> b h n d", h=self.n_heads), qkv)

        dots = torch.einsum('bhid,bhjd->bhij', q, k) * self.scale
        attn = dots.softmax(dim=-1)
        attn = self.attn_drop(attn)
        out = torch.einsum('bhij,bhjd->bhid', attn, v)
        out = rearrange(out, "b h n d -> b n (h d)")

        out = self.to_out(out)
        return out

4. Transformer Encoder

class Transformer(nn.Module):
    def __init__(self, dim=768, depth=12, n_heads=8, mlp_expansions=4, dropout=0.):
        super(Transformer, self).__init__()
        self.layers = nn.ModuleList([])
        for _ in range(depth):
            self.layers.append(nn.ModuleList([
                Residual(PreNorm(dim, MultiHeadAttention(dim, n_heads, dropout))),
                Residual(FeedForward(dim, dim * mlp_expansions, dropout))
            ]))

    def forward(self, x):
        for attn, ff in self.layers:
            x = attn(x)
            x = ff(x)
        return x

6. Vision Transformer

class VisionTransformer(nn.Module):
    def __init__(self, dim=768,
                 patch_size=16,
                 channels=3,
                 img_size=224,
                 depth=12,
                 n_heads=8,
                 mlp_expansions=4,
                 dropout=0.,
                 num_classes=0,
                 global_pool='avg'):
        super(VisionTransformer, self).__init__()
        assert global_pool in ('avg', 'token')
        self.global_pool = global_pool
        self.patch_embedding = PatchEmbedding(dim, patch_size, channels, img_size)
        self.transformer = Transformer(dim, depth, n_heads, mlp_expansions, dropout)
        self.mlp_head = nn.Sequential(
            nn.LayerNorm(dim),
            nn.Linear(dim, num_classes)
        ) if num_classes > 0 else nn.Identity()

    def forward(self, img):
        x = self.patch_embedding(img)
        x = self.transformer(x)
        x = x[:, 1:].mean(dim=1) if self.global_pool == 'avg' else x[:, 0]
        x = self.mlp_head(x)
        return x

7. Test Code

if __name__ == '__main__':
    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
    images = torch.randn((16, 3, 224, 224)).to(device)
    vit = VisionTransformer(num_classes=4, global_pool="token").to(device)
    output = vit(images)
    print(output)
    torch.save(vit.state_dict(), "model.pth")

模型参数量计算

1. 卷积核参数量计算

对于二维卷积层,其参数量由输入通道数(C)、卷积核的大小(KxK)、卷积核的数量或者说输出通道数(F)、偏置项的数量等因素决定。计算公式为:
( K × K × C + 1 ) × F (K \times K \times C + 1)\times F (K×K×C+1)×F,其中1为偏置项。

2. 全连接层参数量计算

对于某一层全连接层的参数量只由其输入维度和输出维度(是否带偏置项)决定,将全连接层理解为一个映射函数,假设输入为矩阵A(维度为HxW),输出为矩阵C(维度为HxH),那么一层全连接层参数量就来自其所代表的矩阵B根据矩阵乘法其维度应为WxH,即Linear(W,H),输入维度W,输出维度也是H。计算公式易得:
W × H + H × 1 W \times H + H\times 1 W×H+H×1,其中1代表偏置项,需要输出维度个偏置项。

3. ViT参数量计算

模块/变量名计算过程参数量
PatchEmbedding c o n v 2 d + c l s _ t o k e n + p o s t i t i o n s conv2d + cls\_token + postitions conv2d+cls_token+postitions742656
conv2d ( 16 × 16 × 3 + 1 ) × 768 (16\times 16\times 3 + 1)\times 768 (16×16×3+1)×768590592
cls_token 1 × 1 × 768 1\times1\times768 1×1×768768
postitions ( ( 224 ÷ 16 ) 2 + 1 ) × 768 ((224\div 16)^2+1)\times768 ((224÷16)2+1)×768151296
Feedforward ( 768 × ( 768 × 4 ) + ( 768 × 4 ) ) + ( ( 768 × 4 ) × 768 + 768 ) (768\times(768\times4)+(768\times4)) + ((768\times4)\times768+768) (768×(768×4)+(768×4))+((768×4)×768+768)4722432
MultiHeadAttention t o _ q k v + t o _ o u t to\_qkv + to\_out to_qkv+to_out2360064
to_qkv 768 × ( 768 × 3 ) 768\times(768\times3) 768×(768×3)1769472
to_out 768 × 768 + 768 768\times768+768 768×768+768590592
Transformer 12 × ( F e e d f o r w a r d + M u l t i H e a d A t t e n t i o n ) 12\times(Feedforward+MultiHeadAttention) 12×(Feedforward+MultiHeadAttention)84989952
ViT T r a n s f o r m e r + P a t c h E m b e d d i n g + m l p _ h e a d Transformer+PatchEmbedding+mlp\_head Transformer+PatchEmbedding+mlp_head85735684
mlp_head 768 × n u m _ c l a s s e s + n u m _ c l a s s e s ,本文设置 n u m _ c l a s s e s 为 4 768\times num\_classes+num\_classes,本文设置num\_classes为4 768×num_classes+num_classes,本文设置num_classes43076

最终参数量为 85735684 × 4 ( B ) = 342942736 ( B ) 85735684\times 4(B) = 342942736(B) 85735684×4(B)=342942736(B)为什么要乘以4字节呢?
因为这些参数权重默认为float32保存,需要用到32bits即4Bytes,最终通过换算得,
342942736 ( B ) ÷ 1024 ÷ 1024 = 327.055679321 ( M B ) 342942736(B)\div 1024\div 1024 = 327.055679321(MB) 342942736(B)÷1024÷1024=327.055679321(MB)
因为我们在Test code有保存模型权重为model.pth文件,可以查看model.pth属性来验证计算是否准确。
在这里插入图片描述
在字节数上有所偏差,但足以表明计算过程大致是正确的! 偏差可能原因是model.pth不止要保存权重,还会附带一些其他信息,所以实际文件大小会比参数量要略大。


总结

日志

参考文献

https://www.youtube.com/watch?v=pkVwUVEHmfI
https://blog.csdn.net/weixin_41041772/article/details/123296659
https://iclr.cc/virtual/2022/oral/6603
https://theaisummer.com/vision-transformer/
https://medium.com/mlearning-ai/vision-transformers-from-scratch-pytorch-a-step-by-step-guide-96c3313c2e0c
https://www.kaggle.com/code/hannes82/vision-transformer-trained-from-scratch-pytorch
https://towardsdatascience.com/implementing-visualttransformer-in-pytorch-184f9f16f632
https://github.com/FrancescoSaverioZuppichini/ViT

Logo

科技之力与好奇之心,共建有温度的智能世界

更多推荐