Vit极简原理+pytorch代码

Vit比它爹Transformer步骤要简单的多，需要注意的点也要少得多，最令人兴奋的是它在代码中没有令人头疼的MASK，还有许多简化的操作，容我慢慢道来。

原理

1、打成patch+线性变化

它所解决的核心问题就是如何将图片塞入Transformer，如果每个像素作为输入的话，那么一个小小的224*224的图片的序列长度就会是50176，而nlp的Transformer最初设定长度才是512，并且attention的复杂度是平方级的，这50176令人不敢恭维。Vit无非就是将一张图片打成一个一个的patch，将每个patch作为一个输入，仅此而已。

将图片打成patch可以通过很简单的卷积实现。使用卷积核大小为16*16，步长为16，卷积核数维768。

用维度来说明一下。假如一个224224的图片，每个patch的大小是1616，那么最终只有（224/16）（224/16）=196个输入，每一个输入的dim是1616*3=768,3是rgb通道数。

2、cls token+位置编码

那么Vit怎么进行分类?可以像bert一样在最前面增加一个cls token，最后只用cls token经过Encoder的结果。

位置编码的方法使用的可学习的方法，并且作者炼丹后发现一维，二维，相对位置编码的结果差不多，为了简单使用一维位置编码。

  self.pos_embed = nn.Parameter(torch.zeros(1, num_patches + self.num_tokens, embed_dim))

3、 Encoder+分类

与transformer的Encoder不同，这里是先norm再attention，并且不存在padding。并且还有一点需要注意，在这里的q，k，v通过同一个Linear得到，不像Transformer中需要分别进行Linear得到，因为它不需要解决Decoder中的交叉注意力中kv和q来源不同的情况,这样可以提高我们的训练速度。还有一点值得注意的是，在Vit中的kqv矩阵维度相同，这样可以得到全局感受野但是计算量大。

最后分类只用cls token的输出

代码

我在前人的基础上进一步增加了一些注释，补全了维度变换。

我认为想彻底搞懂Vit光靠看几个视频，几个文章肯定是不行的。强烈建议亲自调试一遍，点步入步过按钮一步一步运行，将每一步的维度变换理解才是重中之重。下面的代码你只需要在最后一行打个断点就能直接调试，希望对大家能有帮助！

"""
original code from rwightman:
https://github.com/rwightman/pytorch-image-models/blob/master/timm/models/vision_transformer.py
"""
from functools import partial
from collections import OrderedDict

import torch
import torch.nn as nn

class PatchEmbed(nn.Module):
    """
    2D Image to Patch Embedding
    """
    def __init__(self, img_size=224, patch_size=16, in_c=3, embed_dim=768, norm_layer=None):
        super().__init__()
        img_size = (img_size, img_size) # [224,224]
        patch_size = (patch_size, patch_size) # [16,16]
        self.img_size = img_size # [224,224]
        self.patch_size = patch_size # [16,16]
        self.grid_size = (img_size[0] // patch_size[0], img_size[1] // patch_size[1]) # [14,14]
        self.num_patches = self.grid_size[0] * self.grid_size[1] # 196
        self.proj = nn.Conv2d(in_c, embed_dim, kernel_size=patch_size, stride=patch_size) # 3，768,（16,16），16
        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] and W == self.img_size[1], \
            f"Input image size ({H}*{W}) doesn't match model ({self.img_size[0]}*{self.img_size[1]})."
        # proj:[B,3,224,224] -> [B,768,14,14]
        # flatten: [B, 768, 14, 14] -> [B, 768, 196]
        # transpose: [B, 768, 196] -> [B, 196, 768]
        x = self.proj(x).flatten(2).transpose(1, 2)
        x = self.norm(x)
        return x

class Attention(nn.Module):
    def __init__(self,
                 dim,   # 输入token的dim 768
                 num_heads=8, # multi-head 12
                 qkv_bias=False, # True
                 qk_scale=None, # 和根号dimk作用相同
                 attn_drop_ratio=0., # dropout率
                 proj_drop_ratio=0.):
        super(Attention, self).__init__()
        self.num_heads = num_heads
        head_dim = dim // num_heads # 768 // 12 = 64
        self.scale = qk_scale or head_dim ** -0.5
        self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias) # qkv经过一个linear得到。 768 --> 2304
        self.attn_drop = nn.Dropout(attn_drop_ratio)
        self.proj = nn.Linear(dim, dim) # 一次新的映射
        self.proj_drop = nn.Dropout(proj_drop_ratio)

    def forward(self, x):
        # [batch_size, num_patches + 1, total_embed_dim]
        B, N, C = x.shape

        # qkv(): -> [batch_size, num_patches + 1, 3 * total_embed_dim]
        # reshape: -> [batch_size, num_patches + 1, 3, num_heads, embed_dim_per_head]
        # permute: -> [3, batch_size, num_heads, num_patches + 1, embed_dim_per_head]
        qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
        # [batch_size, num_heads, num_patches + 1, embed_dim_per_head]
        q, k, v = qkv[0], qkv[1], qkv[2]  # 在Vit中qkv维度相同，都是[B,12,197,64]

        # transpose: -> [batch_size, num_heads, embed_dim_per_head, num_patches + 1]
        # @: multiply -> [batch_size, num_heads, num_patches + 1, num_patches + 1]
        attn = (q @ k.transpose(-2, -1)) * self.scale
        # 按行进行softmax
        attn = attn.softmax(dim=-1)
        attn = self.attn_drop(attn)

        # @: multiply -> [batch_size, num_heads, num_patches + 1, embed_dim_per_head]
        # transpose: -> [batch_size, num_patches + 1, num_heads, embed_dim_per_head]
        # reshape: -> [batch_size, num_patches + 1, total_embed_dim]
        # C就是total_embed_dim
        x = (attn @ v).transpose(1, 2).reshape(B, N, C)
        # 再通过一次映射使其更好的融合
        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=nn.GELU, drop=0.):
        super().__init__()
        out_features = out_features or in_features # 768
        hidden_features = hidden_features or in_features # 3072
        self.fc1 = nn.Linear(in_features, hidden_features) # 768 --> 3072
        self.act = act_layer()
        self.fc2 = nn.Linear(hidden_features, out_features) # 3072 --> 768
        self.drop = nn.Dropout(drop)

    def forward(self, x):
        x = self.fc1(x)
        x = self.act(x)
        x = self.drop(x)
        x = self.fc2(x)
        x = self.drop(x)
        return x

class Block(nn.Module):
    def __init__(self,
                 dim,
                 num_heads,
                 mlp_ratio=4.,
                 qkv_bias=False,
                 qk_scale=None,
                 drop_ratio=0.,
                 attn_drop_ratio=0.,
                 drop_path_ratio=0.,
                 act_layer=nn.GELU,
                 norm_layer=nn.LayerNorm):
        super(Block, self).__init__()
        self.norm1 = norm_layer(dim) # 与Transformer不同，这里先norm
        self.attn = Attention(dim, num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale,
                              attn_drop_ratio=attn_drop_ratio, proj_drop_ratio=drop_ratio)
        # NOTE: drop path for stochastic depth, we shall see if this is better than dropout here
        self.drop_path = DropPath(drop_path_ratio) if drop_path_ratio > 0. else nn.Identity()
        self.norm2 = norm_layer(dim)
        mlp_hidden_dim = int(dim * mlp_ratio) # 隐藏层神经元数
        self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop_ratio) # MLP

    def forward(self, x):
        # 残差连接
        x = x + self.drop_path(self.attn(self.norm1(x)))
        x = x + self.drop_path(self.mlp(self.norm2(x)))
        return x

class VisionTransformer(nn.Module):
    def __init__(self, img_size=224, patch_size=16, in_c=3, num_classes=1000,
                 embed_dim=768, depth=12, num_heads=12, mlp_ratio=4.0, qkv_bias=True,
                 qk_scale=None, representation_size=None, distilled=False, drop_ratio=0.,
                 attn_drop_ratio=0., drop_path_ratio=0., embed_layer=PatchEmbed, norm_layer=None,
                 act_layer=None):
        """
        Args:
            img_size (int, tuple): input image size
            patch_size (int, tuple): patch size
            in_c (int): 输入通道数
            num_classes (int): 多分类数量
            embed_dim (int): embedding后维度
            depth (int): Encoder数量
            num_heads (int): multi-head头数
            mlp_ratio (int): mlp隐藏层维度是输入的多少倍
            qkv_bias (bool): qkv的Linear过程有没有偏置？
            qk_scale (float): 类似根号dimk
            representation_size (Optional[int]): MLP head是否只有一个全连接层，对应Pre-logits，是一个可选项
            distilled (bool): 用于Deit的，与Vit无关
            drop_ratio (float): dropout rate
            attn_drop_ratio (float): attention dropout rate
            drop_path_ratio (float): stochastic depth rate
            embed_layer (nn.Module): Embedding层
            norm_layer: (nn.Module): normalization层
        """
        super(VisionTransformer, self).__init__()
        self.num_classes = num_classes # 1000
        self.num_features = self.embed_dim = embed_dim  # 768
        self.num_tokens = 2 if distilled else 1 # 不管
        norm_layer = norm_layer or partial(nn.LayerNorm, eps=1e-6)
        act_layer = act_layer or nn.GELU
        # 224,16,3,768
        self.patch_embed = embed_layer(img_size=img_size, patch_size=patch_size, in_c=in_c, embed_dim=embed_dim)
        num_patches = self.patch_embed.num_patches # 196

        self.cls_token = nn.Parameter(torch.zeros(1, 1, embed_dim)) # 1,1,768
        self.dist_token = nn.Parameter(torch.zeros(1, 1, embed_dim)) if distilled else None # 不管
        self.pos_embed = nn.Parameter(torch.zeros(1, num_patches + self.num_tokens, embed_dim)) # 位置编码，注意要+1（cls token）
        self.pos_drop = nn.Dropout(p=drop_ratio) # 位置编码后的dropout

        dpr = [x.item() for x in torch.linspace(0, drop_path_ratio, depth)]  # 构建一个dropout的等差序列，但默认为0
        # *可以解引用
        self.blocks = nn.Sequential(*[
            Block(dim=embed_dim, num_heads=num_heads, mlp_ratio=mlp_ratio, qkv_bias=qkv_bias, qk_scale=qk_scale,
                  drop_ratio=drop_ratio, attn_drop_ratio=attn_drop_ratio, drop_path_ratio=dpr[i],
                  norm_layer=norm_layer, act_layer=act_layer)
            for i in range(depth)
        ])
        self.norm = norm_layer(embed_dim)

        # 之前所说的可选项,略过
        if representation_size and not distilled:
            self.has_logits = True
            self.num_features = representation_size
            self.pre_logits = nn.Sequential(OrderedDict([
                ("fc", nn.Linear(embed_dim, representation_size)),
                ("act", nn.Tanh())
            ]))
        else:
            self.has_logits = False
            self.pre_logits = nn.Identity() # 什么也不做

        # 分类，不用管else 768 --> 1000
        self.head = nn.Linear(self.num_features, num_classes) if num_classes > 0 else nn.Identity()

        # 以下与Vit无关 //开始
        self.head_dist = None
        if distilled:
            self.head_dist = nn.Linear(self.embed_dim, self.num_classes) if num_classes > 0 else nn.Identity()
        # 结束//

        # 权重初始化
        nn.init.trunc_normal_(self.pos_embed, std=0.02) # 位置编码
        if self.dist_token is not None:
            nn.init.trunc_normal_(self.dist_token, std=0.02)

        nn.init.trunc_normal_(self.cls_token, std=0.02) # cls token
        self.apply(_init_vit_weights)

    def forward_features(self, x):
        # [B, C, H, W] -> [B, num_patches, embed_dim]
        x = self.patch_embed(x)  # [B, 196, 768]
        # [1, 1, 768] -> [B, 1, 768]
        cls_token = self.cls_token.expand(x.shape[0], -1, -1)
        if self.dist_token is None:
            x = torch.cat((cls_token, x), dim=1)  # [B, 197, 768]
        else:
            x = torch.cat((cls_token, self.dist_token.expand(x.shape[0], -1, -1), x), dim=1)

        x = self.pos_drop(x + self.pos_embed)
        x = self.blocks(x)
        x = self.norm(x)
        if self.dist_token is None:
            return self.pre_logits(x[:, 0])
        else:
            return x[:, 0], x[:, 1]

    def forward(self, x):
        x = self.forward_features(x)
        if self.head_dist is not None:
            x, x_dist = self.head(x[0]), self.head_dist(x[1])
            if self.training and not torch.jit.is_scripting():
                # during inference, return the average of both classifier predictions
                return x, x_dist
            else:
                return (x + x_dist) / 2
        else:
            x = self.head(x)
        return x

def _init_vit_weights(m):
    """
    ViT weight initialization
    :param m: module
    """
    if isinstance(m, nn.Linear):
        nn.init.trunc_normal_(m.weight, std=.01)
        if m.bias is not None:
            nn.init.zeros_(m.bias)
    elif isinstance(m, nn.Conv2d):
        nn.init.kaiming_normal_(m.weight, mode="fan_out")
        if m.bias is not None:
            nn.init.zeros_(m.bias)
    elif isinstance(m, nn.LayerNorm):
        nn.init.zeros_(m.bias)
        nn.init.ones_(m.weight)

def drop_path(x, drop_prob: float = 0., training: bool = False):
    """
    Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks).
    This is the same as the DropConnect impl I created for EfficientNet, etc networks, however,
    the original name is misleading as 'Drop Connect' is a different form of dropout in a separate paper...
    See discussion: https://github.com/tensorflow/tpu/issues/494#issuecomment-532968956 ... I've opted for
    changing the layer and argument names to 'drop path' rather than mix DropConnect as a layer name and use
    'survival rate' as the argument.
    """
    if drop_prob == 0. or not training:
        return x
    keep_prob = 1 - drop_prob
    shape = (x.shape[0],) + (1,) * (x.ndim - 1)  # work with diff dim tensors, not just 2D ConvNets
    random_tensor = keep_prob + torch.rand(shape, dtype=x.dtype, device=x.device)
    random_tensor.floor_()  # binarize
    output = x.div(keep_prob) * random_tensor
    return output

class DropPath(nn.Module):
    """
    Drop paths (Stochastic Depth) per sample  (when applied in main path of residual blocks).
    """
    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)
    
def vit_base_patch16_224(num_classes: int = 1000):
    """
    ViT-Base model (ViT-B/16) from original paper (https://arxiv.org/abs/2010.11929).
    ImageNet-1k weights @ 224x224, source https://github.com/google-research/vision_transformer.
    weights ported from official Google JAX impl:
    链接: https://pan.baidu.com/s/1zqb08naP0RPqqfSXfkB2EA  密码: eu9f
    """
    model = VisionTransformer(img_size=224,
                              patch_size=16,
                              embed_dim=768,
                              depth=12,
                              num_heads=12,
                              representation_size=None,
                              num_classes=num_classes)
    return model

vit_base_patch16_224()

文章仅供学习使用