VIT Vision Transformer | 先从PyTorch代码了解

• 代码来自github

【前言】：看代码的时候，也许会不理解VIT中各种组件的含义，但是这个文章的目的是了解其实现。在之后看论文的时候，可以做到心中有数，而不是一片茫然。

VIT类

初始化

和之前的学习一样，从大模型类开始看起，然后一点一点看小模型类：

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__()
        assert image_size % patch_size == 0, 'Image dimensions must be divisible by the patch size.'
        num_patches = (image_size // patch_size) ** 2
        patch_dim = channels * patch_size ** 2
        assert num_patches > MIN_NUM_PATCHES, f'your number of patches ({num_patches}) is way too small for attention to be effective (at least 16). Try decreasing your patch size'
        assert pool in {'cls', 'mean'}, 'pool type must be either cls (cls token) or mean (mean pooling)'

        self.patch_size = patch_size

        self.pos_embedding = nn.Parameter(torch.randn(1, num_patches + 1, dim))
        self.patch_to_embedding = nn.Linear(patch_dim, 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.Sequential(
            nn.LayerNorm(dim),
            nn.Linear(dim, num_classes)
        )

在实际的调用中，是如下调用的：

model = ViT(
    dim=128,
    image_size=224,
    patch_size=32,
    num_classes=2,
    channels=3,
).to(device)

输入参数讲解：

• image_size：图片的大小；
• patch_size:把图片划分成小的patch，小的patch的尺寸；
• num_classes:这次分类任务的类别总数；
• channels:输入图片的通道数。

VIT类中初始化的组件：

• num_patches:一个图片划分成多少个patch，因为图片224，patch32，所以划分成7x7=49个patches；
• patch_dim:3x32x32，理解为一个patch中的元素个数；

......这样展示是不是非常的麻烦，还要上下来回翻看代码，所以我写成注释的形式

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.):
    # image_size=224,patch_size=32,num_classes=2,channels=3，dim=128
        super().__init__()
        assert image_size % patch_size == 0, 'Image dimensions must be divisible by the patch size.'
        # num_pathes = （224//32）**2 = 7*7=49
        num_patches = (image_size // patch_size) ** 2
        # patch_dim = 3*32*32
        patch_dim = channels * patch_size ** 2
        assert num_patches > MIN_NUM_PATCHES, f'your number of patches ({num_patches}) is way too small for attention to be effective (at least 16). Try decreasing your patch size'
        assert pool in {'cls', 'mean'}, 'pool type must be either cls (cls token) or mean (mean pooling)'
		# self.patch_size = 32
        self.patch_size = patch_size
        # self.pos_embedding是一个形状为（1，50，128）
        self.pos_embedding = nn.Parameter(torch.randn(1, num_patches + 1, dim))
        # self.patch_to_embedding是一个从3*32*32到128映射的线性层
        self.patch_to_embedding = nn.Linear(patch_dim, dim)
        # self.cls_token是一个随机初始化的形状为（1，1，128）这样的变量
        self.cls_token = nn.Parameter(torch.randn(1, 1, dim))
        self.dropout = nn.Dropout(emb_dropout)
        
        # Transformer后面会讲解
        self.transformer = Transformer(dim, depth, heads, dim_head, mlp_dim, dropout)

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

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

forward

现在看VIT的推理过程：

    def forward(self, img, mask = None):
 		# p=32
        p = self.patch_size
        x = rearrange(img, 'b c (h p1) (w p2) -> b (h w) (p1 p2 c)', p1 = p, p2 = p)
        x = self.patch_to_embedding(x) # x.shape=[b,49,128]
        b, n, _ = x.shape # n = 49

        cls_tokens = repeat(self.cls_token, '() n d -> b n d', b = b)
        x = torch.cat((cls_tokens, x), dim=1) # x.shape=[b,50,128]
        x += self.pos_embedding[:, :(n + 1)] # x.shape=[b,50,128]
        x = self.dropout(x) 

        x = self.transformer(x, mask) # x.shape=[b,50,128],mask=None

        x = x.mean(dim = 1) if self.pool == 'mean' else x[:, 0]

        x = self.to_latent(x)
        return self.mlp_head(x)

• 这里的代码用到了from einops import rearrange, repeat，这个库函数，einops是一个库函数，是对张量进行操作的库函数，支持pytorch，TF等。
• einops.rearrange是把输入的img,从[b,3,224,224]的形状改成[b,3,7,32,7,32]的形状，通过矩阵的转置换成[b,7,7,32,32,3]的样子，最后合并成[b,49,32x32x3]
• self.patch_to_embedding,输出的x的形状为[b,49,128];
• einops.repeat是把self.cls_token从[1,1,128]复制成[b,1,128]

现在，我们知道从patch到embedding是用线性层实现的。

transformer

class Transformer(nn.Module):
    def __init__(self, dim, depth, heads, dim_head, mlp_dim, dropout):
        # dim=128,depth=12，heads=8，dim_head=64,mlp_dim=128
        super().__init__()
        self.layers = nn.ModuleList([])
        for _ in range(depth):
            self.layers.append(nn.ModuleList([
                Residual(PreNorm(dim, Attention(dim, heads = heads, dim_head = dim_head, dropout = dropout))),
                Residual(PreNorm(dim, FeedForward(dim, mlp_dim, dropout = dropout)))
            ]))
    def forward(self, x, mask = None):
        for attn, ff in self.layers:
            x = attn(x, mask = mask)
            x = ff(x)
        return x

• self.layers中包含depth组的Attention+FeedForward模块。
• 这里需要记得，输入的x的尺寸为[b,50,128]

Attention

class Attention(nn.Module):
    def __init__(self, dim, heads = 8, dim_head = 64, dropout = 0.):
        super().__init__()
        inner_dim = dim_head *  heads # 64 x 8
        self.heads = heads # 8
        self.scale = dim_head ** -0.5

        self.to_qkv = nn.Linear(dim, inner_dim * 3, bias = False)
        self.to_out = nn.Sequential(
            nn.Linear(inner_dim, dim),
            nn.Dropout(dropout)
        )

    def forward(self, x, mask = None):
        b, n, _, h = *x.shape, self.heads # n=50,h=8
        # self.to_qkv(x)得到的尺寸为[b,50,64x8x3],然后chunk成3份
        # 也就是说，qkv是一个三元tuple,每一份都是[b,50,64x8]的大小
        qkv = self.to_qkv(x).chunk(3, dim = -1)
        # 把每一份从[b,50,64x8]变成[b,8,50,64]的形式
        q, k, v = map(lambda t: rearrange(t, 'b n (h d) -> b h n d', h = h), qkv)
		# 这一步不太好理解，q和k都是[b,8,50,64]的形式，50理解为特征数量，64为特征变量
        # dots.shape=[b,8,50,50]
        dots = torch.einsum('bhid,bhjd->bhij', q, k) * self.scale
        # 不考虑mask这一块的内容
        mask_value = -torch.finfo(dots.dtype).max

        if mask is not None:
            mask = F.pad(mask.flatten(1), (1, 0), value = True)
            assert mask.shape[-1] == dots.shape[-1], 'mask has incorrect dimensions'
            mask = mask[:, None, :] * mask[:, :, None]
            dots.masked_fill_(~mask, mask_value)
            del mask
		# 对[b,8,50,50]的最后一个维度做softmax
        attn = dots.softmax(dim=-1)

		# 这个attn就是计算出来的自注意力值，和v做点乘，out.shape=[b,8,50,64]
        out = torch.einsum('bhij,bhjd->bhid', attn, v)
        # out.shape变成[b,50,8x64]
        out = rearrange(out, 'b h n d -> b n (h d)')
        # out.shape重新变成[b,60,128]
        out =  self.to_out(out)
        return out

综上所属，这个attention其实就是一个自注意力模块，输入的是[b,50,128]，返回的也是[b,50,128]。实现的过程因为使用了torch.einsum所以有些复杂，但是总的来说，和我之前讲过的一篇论文"non-local"模块，是完全一样的。torch.einsum和torch.mm原理相同，只是因为torch.mm不支持高纬度的张量做矩阵乘法。

PreNorm

class PreNorm(nn.Module):
    def __init__(self, dim, fn):
    # dim=128,fn=Attention/FeedForward
        super().__init__()
        self.norm = nn.LayerNorm(dim)
        self.fn = fn
    def forward(self, x, **kwargs):
        return self.fn(self.norm(x), **kwargs)

先对输入的x(x.shape=[b,50,128])做一个layerNormalization层归一化，然后再放到上面的Attention模块中做自注意力。

Residual

class Residual(nn.Module):
    def __init__(self, fn):
        super().__init__()
        self.fn = fn
    def forward(self, x, **kwargs):
        return self.fn(x, **kwargs) + x

一个残差模块罢了。

FeedForward

class FeedForward(nn.Module):
    def __init__(self, dim, hidden_dim, dropout = 0.):
    # dim=128,hidden_dim=128
        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)

就是两个线性层，这里有意思的是GELU()激活函数，这个激活函数可以直接使用torch.nn.GELU()调用，回头有机会再好好讲一下GELU()的原理。

transformer总结

Residual(PreNorm(dim, Attention(dim, heads = heads, dim_head = dim_head, dropout = dropout))),
Residual(PreNorm(dim, FeedForward(dim, mlp_dim, dropout = dropout)))

• 第一个就是，先对输入做layerNormalization，然后放到attention得到attention的结果，然后结果和做layerNormalization之前的输入相加做一个残差链接；
• 第二个就是，x->LayerNormalization->FeedForward线性层->y,然后这个y和输入的x相加，做残差连接。

VIT总结

回顾一下整个流程：

• 一个图片224x224，分成了49个32x32的patch；
• 对这么多的patch做embedding，成49个128向量；
• 再拼接一个cls_tokens，变成50个128向量；
• 再加上pos_embedding，还是50个128向量；
• 这些向量输入到transformer中进行自注意力的特征提取；
• 输出的是50个128向量，然后对这个50个求军职，变成一个128向量；
• 然后线性层把128维变成2维从而完成二分类任务的transformer模型。

问题：我对NLP了解不深入，有没有人可以回答一下这个问题：cls_tokens和pos_embedding的用处是什么？