Vision Transformer 代码实现

论文链接：An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale

最近开始恶补CV了（指->新建文件夹）。作为CV Transformer的开山大作，首先要学习的就是ViT（Vision Transformer）模型。有很多大佬都已经解读过这篇工作了，我比较推荐李沐大神团队的视频，不仅能了解模型架构，还能学到些论文写作技巧。也参考了我的b站导师 （代码是基于timm的） 本文代码基于vit-pytorch，该项目包含了ViT和诸多变体。需要预训练模型的话可以参考timm的实现(pytorch-image-models)

---

ViT的架构

ViT的结构如上图，我们按照流程一步步讲解。大概来说，ViT分为这几个步骤。1.图片分块和映射；2.Transformer；3.线性层输出。原论文给出了3种不同大小的模型：Base、Large、Huge，他们的Transformer层数，head不同，中间维度、线性层维度也不相同，具体数值请参考原论文。文章代码基于Base/16

ViT-Base

先来看最终的总体架构，具体步骤我已经做了注释，解释下形参

• image_size：图片尺寸，int或者tuple皆可，例如224，或者(224, 224)，长宽不一定要一样大
• path_size：分块path尺寸，int或tuple，默认为16，需要确保image_size能被path_size整除
• num_classes：分类数，int
• dim：Transformer隐层维度，对于Base来说是768，Large为1024
• depth：Transformer个数，Base为12
• head：多头的个数，Base=12
• mlp_dim：Transformer中的FeedForward中第一个线性层升维后的维度，默认为768*4，先升维4倍再降维回去
• pool：默认'cls，选取CLS token作为输出，可选'mean'，在patch维度做平均池化
• channel：图片输入的特征维度，RGB图像为3，灰度图为1

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)  # 在这个项目中没有限定图片的尺寸
        patch_height, patch_width = pair(patch_size)  # 默认为16

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

        # num patches -> (224 / 16) = 14, 14 * 14 = 196
        num_patches = (image_height // patch_height) * (image_width // patch_width)
        # path dim -> 3 * 16 * 16 = 768，和Bert-base一致
        patch_dim = channels * patch_height * patch_width  
        assert pool in {'cls', 'mean'}, 'pool type must be either cls (cls token) or mean (mean pooling)'  # 输出选cls token还是做平均池化

        # 步骤一：图像分块与映射。首先将图片分块，然后接一个线性层做映射
        self.to_patch_embedding = nn.Sequential(
            Rearrange('b c (h p1) (w p2) -> b (h w) (p1 p2 c)', p1 = patch_height, p2 = patch_width),
            nn.Linear(patch_dim, dim),
        )

        # pos_embedding：位置编码；cls_token：在序列最前面插入一个cls token作为分类输出
        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)

        # 步骤二：Transformer Encoder结构来提特征
        self.transformer = Transformer(dim, depth, heads, dim_head, mlp_dim, dropout)

        self.pool = pool
        self.to_latent = nn.Identity()  # 这一步上面都没做

        # 线性层输出
        self.mlp_head = nn.Sequential(
            nn.LayerNorm(dim),
            nn.Linear(dim, num_classes)
        )

    def forward(self, img):
    ...

1. 分块与映射

分块这里调用了einops这个库，可以很方便直观地进行维度变换。简单起见，下面按照默认的patch 宽高=16的设置，所以在行和列上一共有224 / 16 = 14个patch。这里的操作就是把一个batch的，shape为(batch, 3, 224, 224)的图片，首先划分为(batch, 3, ( 14 x 16), (14 x 16))，按照块给他展平，操作和动图里的一样，最后的输出就是(batch, ( 14 x 14), ( 16 x 16 x 3 = 768))。除掉第一维的batch，单独拎一个path来看，其特征维度是768，就是每一个像素铺平后的长度(RGB图像有3个特征维度)。铺平后接上一个linear做线性映射，维度也是768。

from einops.layers.torch import Rearrange

self.to_patch_embedding = nn.Sequential(
            Rearrange('b c (h p1) (w p2) -> b (h w) (p1 p2 c)', p1 = patch_height, p2 = patch_width),
            nn.Linear(patch_dim, dim),
        )

# forward:
x = self.to_patch_embedding(img)

接下来要加上位置编码。文章采用的是一维位置编码，作者也测试了2维和相对位置编码，结果都差不多。一维的意思就是把展平后的图片看成一个序列，第二维的14 * 14 =196看成NLP里token的长度，只考虑这个长度。ViT的位置编码也可以理解为一种可以训练的绝对位置编码(因为取的是embedding中前196+1个，然后对应位置和x相加)。

论文的附录图10

然后在拼上一个CLS token

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)

# forward:
b, n, _ = x.shape
# 1 x 1 x 768的CLS token重复至 batch x 1 x 768
cls_tokens = repeat(self.cls_token, '() n d -> b n d', b = b)
x = torch.cat((cls_tokens, x), dim=1)
x += self.pos_embedding[:, :(n + 1)]  # 因为多了个CLS token所以要n+1
x = self.dropout(x)
# x.shape -> (batch, 196 + 1, 768)

2. Transformer Encoder

ViT框架里就这么一步

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

Transformer的流程我简单画了一下，这一块我直接在代码里注释了下，如果有不懂的建议大家先去看一下Transformer的代码和原理，ViT和Bert一样只有encoder的部分，要简单一些。

Self-Attention总结起来就是 ���(�,�,�)=�������(�⋅����)�

class PreNorm(nn.Module):
    def __init__(self, dim, fn):
        super().__init__()
        self.norm = nn.LayerNorm(dim)
        self.fn = fn
    def forward(self, x, **kwargs):
        # 先LN
        return self.fn(self.norm(x), **kwargs)

class FeedForward(nn.Module):
    def __init__(self, dim, hidden_dim, dropout = 0.):
        super().__init__()
        self.net = 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.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  # 论文里的\sqrt{d_k}

        self.attend = nn.Softmax(dim = -1)
        # Q K V一块处理了，有的工作是用了3个维度为dim的Linear层分别得到的
        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):
        # 输入 x -> (batch, 197, 768)即(batch, num_patch + 1, hid_dims)
        # 按照最后一维（特征维度)分成3块，分别对应QKV
        # chunk后是一个tuple，即(q, k, v)
        qkv = self.to_qkv(x).chunk(3, dim = -1)
        # q, k, v都做维度变换，(batch, 197, 768) -> (batch, 12, 197, 768 / 12 = 64)
        # 12是head的数量，目的是做**多头**注意力机制
        q, k, v = map(lambda t: rearrange(t, 'b n (h d) -> b h n d', h = self.heads), qkv)
        # Q @ K^T  /  \sqrt{d_k}
        dots = torch.matmul(q, k.transpose(-1, -2)) * self.scale

        attn = self.attend(dots)  # softmax

        out = torch.matmul(attn, v)
        # 把多头拼回去 -> (batch, 197, 768)
        out = rearrange(out, 'b h n d -> b n (h d)')
        return self.to_out(out)

class Transformer(nn.Module):
    def __init__(self, dim, depth, heads, dim_head, mlp_dim, dropout = 0.):
        super().__init__()
        self.layers = nn.ModuleList([])
        for _ in range(depth):
            self.layers.append(nn.ModuleList([
                PreNorm(dim, Attention(dim, heads = heads, dim_head = dim_head, dropout = dropout)),
                PreNorm(dim, FeedForward(dim, mlp_dim, dropout = dropout))
            ]))
            #  先LN再做Att或者Forward
    def forward(self, x):
        for attn, ff in self.layers:  # attn为Multi-head Attention，ff就是FeedForward
            x = attn(x) + x  # 残差连接，图片中的边线
            x = ff(x) + x
        return x

3.线性层输出

self.pool = pool
self.to_latent = nn.Identity()
 # 线性层输出
 self.mlp_head = nn.Sequential(
    nn.LayerNorm(dim),
    nn.Linear(dim, num_classes)
    )
# forward
x = x.mean(dim = 1) if self.pool == 'mean' else x[:, 0]
 x = self.to_latent(x)
return self.mlp_head(x)

这一块没什么好说的，一般直接用CLS token作为输出。为什么cls可以作为输出呢？主要还是因为Transformer的自注意机制，每个地方都能兼顾全局，而不是CNN那种的局部信息。

ViT-timm实现

从论文中我们知道如果训练的图片数据集规模没有特别大的话，ViT是不如CNN的。在实际项目中我们更多用到的是微调预训练模型，上边的库没有预训练模型，所以提供预训练模型的timm的使用率会更高。timm的实现要更复杂，因为源码里包含了DEIT（ViT+蒸馏），源代码里还有很多初始化、加载预训练权重等操作。这里就只看模型的架构，说一下和上边不同的地方。

总体架构部分

class VisionTransformer(nn.Module):
    """ Vision Transformer
    A PyTorch impl of : `An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale`
        - https://arxiv.org/abs/2010.11929
    Includes distillation token & head support for `DeiT: Data-efficient Image Transformers`
        - https://arxiv.org/abs/2012.12877
    """

    def __init__(self, img_size=224, patch_size=16, in_chans=3, num_classes=1000, embed_dim=768, depth=12,
                 num_heads=12, mlp_ratio=4., qkv_bias=True, representation_size=None, distilled=False,
                 drop_rate=0., attn_drop_rate=0., drop_path_rate=0., embed_layer=PatchEmbed, norm_layer=None,
                 act_layer=None, weight_init=''):
        """
        Args:
            img_size (int, tuple): input image size
            patch_size (int, tuple): patch size
            in_chans (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
            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_rate (float): dropout rate
            attn_drop_rate (float): attention dropout rate
            drop_path_rate (float): stochastic depth rate
            embed_layer (nn.Module): patch embedding layer
            norm_layer: (nn.Module): normalization layer
            weight_init: (str): weight init scheme
        """

首先如果要用timm的预训练权重，图片尺寸必须为224x224或者384x384，使用方法如下

import timm
model = timm.from_pretrained('vit_base_patch16_224', num_classes=xx)

形参上多了mlp_ratio，对应mlp_dim(4 x 768的4)；

以及qkv_bias=True的可选操作;

self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)  # Attention模块
# forward
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)

distilled=False是为DEIT准备的

embed_layer=PatchEmbed：实现如下

class PatchEmbed(nn.Module):
    """ 2D Image to Patch Embedding
    """
    def __init__(self, img_size=224, patch_size=16, in_chans=3, embed_dim=768, norm_layer=None, flatten=True):
        super().__init__()
        img_size = to_2tuple(img_size)
        patch_size = to_2tuple(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])  # 224 / 16 = 14，即(14, 14)
        self.num_patches = self.grid_size[0] * self.grid_size[1]  # 196
        self.flatten = flatten

        self.proj = nn.Conv2d(in_chans, embed_dim, kernel_size=patch_size, stride=patch_size)
        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], f"Input image height ({H}) doesn't match model ({self.img_size[0]}).")
        _assert(W == self.img_size[1], f"Input image width ({W}) doesn't match model ({self.img_size[1]}).")
        x = self.proj(x)
        if self.flatten:
            x = x.flatten(2).transpose(1, 2)  # BCHW -> BNC
        x = self.norm(x)  # 默认是没有norm的
        return x

从之前的线性层替换为2维卷积了，卷积大小和stride都为patch size，即16

经过卷积，输入(batch, 3, 224, 224) -> (batch, 768, 14, 14)，非常取巧的操作啊，花里胡哨的，和线性层是一样的效果

之后在第三个维度展平：(batch, 768, 14, 14) flatten -> (batch, 768, 196) transpose(1, 2) -> (batch, 196, 768)

其他地方在写法上有不同，但方法是基本一样的，对了，timm还在Transformer那部分加入了drop_path