1 摘要

1.1 核心

通过将图像切成patch线形层编码成token特征编码的方法，用transformer的encoder来做图像分类

2 模型架构

2.1 概览

2.2 对应CV的特定修改和相关理解

解决问题：

1. transformer输入限制: 由于自注意力+backbone，算法复杂度为o(n²)，token长度一般要<512才足够运算
   解决：a) 将图片转为token输入 b) 将特征图转为token输入 c)√ 切patch转为token输入
2. transformer无先验知识：卷积存在平移不变性(同特征同卷积核同结果)和局部相似性(相邻特征相似结果)，
   而transformer无卷积核概念，只有整个编解码器，需要从头学
   解决：大量数据训练
3. cv的各种自注意力机制需要复杂工程实现：
   解决：直接用整个transformer模块
4. 分类head：
   解决：直接沿用transformer cls token
5. position编码：
   解决：1D编码

pipeline：
224x224输入切成16x16patch进行位置编码和线性编码后增加cls token 一起输入的encoder encoder中有L个selfattention模块
输出的cls token为目标类别

3 代码

如果理解了transformer，看完这个结构感觉真的很简单，这篇论文也只是开山之作，没有特别复杂的结构，所以想到代码里看看。

import torch
from torch import nn

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

# helpers

def pair(t):
    return t if isinstance(t, tuple) else (t, t)

# classes

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

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

class Attention(nn.Module):
    def __init__(self, dim, heads = 8, dim_head = 64, dropout = 0.):
        super().__init__()
        inner_dim = dim_head *  heads
        project_out = not (heads == 1 and dim_head == dim)

        self.heads = heads
        self.scale = dim_head ** -0.5

        self.norm = nn.LayerNorm(dim)

        self.attend = nn.Softmax(dim = -1)
        self.dropout = nn.Dropout(dropout)

        # linear(1024 , 3072)
        self.to_qkv = nn.Linear(dim, inner_dim * 3, bias = False)

        self.to_out = nn.Sequential(
            nn.Linear(inner_dim, dim),
            nn.Dropout(dropout)
        ) if project_out else nn.Identity()

    def forward(self, x):
        # [1, 65, 1024]
        x = self.norm(x)
        # [1, 65, 1024]
        qkv = self.to_qkv(x).chunk(3, dim = -1)
        # self.to_qkv(x)                [1, 65, 3072]
        # self.to_qkv(x).chunk(3,-1)    [3, 1, 65, 1024]
        q, k, v = map(lambda t: rearrange(t, 'b n (h d) -> b h n d', h = self.heads), qkv)
        # q,k,v                         [1, 65, 1024] -> [1, 16, 65, 64]
        # 把 65个1024的特征分为 heads个65个d维的特征 然后每个heads去分别有自己要处理的隐藏层，对不同的特征建立不同学习能力
        dots = torch.matmul(q, k.transpose(-1, -2)) * self.scale
        # [1, 16, 65, 64] * [1, 16, 64, 65] -> [1, 16, 65, 65]
        # scale 保证在softmax前所有的值都不太大

        attn = self.attend(dots)
        # softmax [1, 16, 65, 65]
        
        attn = self.dropout(attn)
        # dropout [1, 16, 65, 65]
        
        out = torch.matmul(attn, v)
        # out [1, 16, 65, 64]
        
        out = rearrange(out, 'b h n d -> b n (h d)')
        # out [1, 65, 1024]
        
        return self.to_out(out)
        # out [1, 65, 1024]
        

class Transformer(nn.Module):
    def __init__(self, dim, depth, heads, dim_head, mlp_dim, dropout = 0.):
        super().__init__()
        self.norm = nn.LayerNorm(dim)
        self.layers = nn.ModuleList([])
        for _ in range(depth):
            self.layers.append(nn.ModuleList([
                Attention(dim, heads = heads, dim_head = dim_head, dropout = dropout),
                FeedForward(dim, mlp_dim, dropout = dropout)
            ]))

    def forward(self, x):
        # [1, 65, 1024]
        for attn, ff in self.layers:
            # [1, 65, 1024]
            x = attn(x) + x
            # [1, 65, 1024]
            x = ff(x) + x

        # [1, 65, 1024]
        return self.norm(x)
        # shape不会改变

class ViT(nn.Module):
    def __init__(self, *, image_size, patch_size, num_classes, dim, depth, heads, mlp_dim, pool = 'cls', channels = 3, dim_head = 64, dropout = 0., emb_dropout = 0.):
        super().__init__()
        image_height, image_width = pair(image_size)
        patch_height, patch_width = pair(patch_size)

        assert image_height % patch_height == 0 and image_width % patch_width == 0, 'Image dimensions must be divisible by the patch size.'

        num_patches = (image_height // patch_height) * (image_width // patch_width)
        patch_dim = channels * patch_height * patch_width
        assert pool in {'cls', 'mean'}, 'pool type must be either cls (cls token) or mean (mean pooling)'

        # num_patches   64
        # patch_dim     3072
        # dim           1024
        self.to_patch_embedding = nn.Sequential(
            #Rearrange是einops中的一个方法
            # einops：灵活和强大的张量操作，可读性强和可靠性好的代码。支持numpy、pytorch、tensorflow等。
            # 代码中Rearrage的意思是将传入的image（3，224，224），按照（3，（h,p1）,(w,p2))也就是224=hp1,224 = wp2，接着把shape变成b (h w) (p1 p2 c)格式的，这样把图片分成了每个patch并且将patch拉长，方便下一步的全连接层
            # 还有一种方法是采用窗口为16*16，stride 16的卷积核提取每个patch，然后再flatten送入全连接层。
            Rearrange('b c (h p1) (w p2) -> b (h w) (p1 p2 c)', p1 = patch_height, p2 = patch_width),
            nn.LayerNorm(patch_dim),
            nn.Linear(patch_dim, dim),
            nn.LayerNorm(dim),
        )

        self.pos_embedding = nn.Parameter(torch.randn(1, num_patches + 1, dim))
        self.cls_token = nn.Parameter(torch.randn(1, 1, dim))
        self.dropout = nn.Dropout(emb_dropout)

        self.transformer = Transformer(dim, depth, heads, dim_head, mlp_dim, dropout)

        self.pool = pool
        self.to_latent = nn.Identity()

        self.mlp_head = nn.Linear(dim, num_classes)

    def forward(self, img):
        # 1. [1, 3, 256, 256]       输入img
        x = self.to_patch_embedding(img)
        # 2. [1, 64, 1024]          patch embd
        b, n, _ = x.shape
        # 3. [1, 1, 1024]           cls_tokens
        cls_tokens = repeat(self.cls_token, '1 1 d -> b 1 d', b = b)
        # 4. [1, 65, 1024]          cat [cls_tokens, x]
        x = torch.cat((cls_tokens, x), dim=1)
        # 5. [1, 65, 1024]          add [x] [pos_embedding]
        x += self.pos_embedding[:, :(n + 1)]
        # 6. [1, 65, 1024]          dropout
        x = self.dropout(x)
        # 7. [1, 65, 1024]          N * transformer
        x = self.transformer(x)
        # 8. [1,1024]               cls_x output
        x = x.mean(dim = 1) if self.pool == 'mean' else x[:, 0]
        # 9. [1,1024]               cls_x output mean
        x = self.to_latent(x)
        # 10.[1,1024]               nn.Identity()不改变输入和输出 占位层
        return self.mlp_head(x)
        # 11.[1,cls]                mlp_cls_head

4 总结

multihead和我原有的理解偏差修正。
我以为的是QKV会有N块相同的copy()，每一份去做后续的linear等操作。
代码里是直接用linear将QKV分为一整个大块，用permute/rearrange的操作切成了N块，f(Q,K)之后再恢复成一整个大块，很强。