从零搭建Pytorch模型教程 | 搭建Transformer网络

前言 本文介绍了Transformer的基本流程，分块的两种实现方式，Position Emebdding的几种实现方式，Encoder的实现方式，最后分类的两种方式，以及最重要的数据格式的介绍。

from einops import rearrange, repeat
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),
      )

#假设images的shape为[32,200,400,3]
#实现view和reshape的功能
Rearrange(images,'b h w c -> (b h) w c')#shape变为（32*200, 400, 3）
#实现permute的功能
Rearrange(images, 'b h w c -> b c h w')#shape变为（32, 3, 200, 400）
#实现这几个都很难实现的功能
Rearrange(images, 'b h w c -> (b c w) h')#shape变为（32*3*400, 200）

Rearrange('b c (h p1) (w p2) -> b (h w) (p1 p2 c)', p1 = patch_height, p2 = patch_width)

self.proj = nn.Conv2d(in_chans, embed_dim, kernel_size=patch_size, stride=patch_size)

x = self.proj(x).flatten(2).transpose(1, 2)  # B Ph*Pw C

self.pos_embedding = nn.Parameter(torch.randn(1, num_patches + 1, dim))
x += self.pos_embedding[:, :(n + 1)]
#之所以是n+1，是因为ViT中选择随机初始化一个class token，与分块得到的tokens拼接。所以patches的数量为num_patches+1。

from timm.models.layers import trunc_normal_
self.absolute_pos_embed = nn.Parameter(torch.zeros(1, num_patches, embed_dim))
trunc_normal_(self.absolute_pos_embed, std=.02)

self.pos_emb = torch.nn.Embedding(num_positions + 1, dim)

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

       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):
       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.heads), qkv)
       dots = torch.matmul(q, k.transpose(-1, -2)) * self.scale
       attn = self.attend(dots)
       attn = self.dropout(attn)

       out = torch.matmul(attn, v)
       out = rearrange(out, 'b h n d -> b n (h d)')
       return self.to_out(out)

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

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

#生成cls_token部分
from einops import repeat
self.cls_token = nn.Parameter(torch.randn(1, 1, dim))

cls_tokens = repeat(self.cls_token, '1 n d -> b n d', b = b)
x = torch.cat((cls_tokens, x), dim=1)
################################
#分类部分
self.mlp_head = nn.Sequential(
           nn.LayerNorm(dim),
           nn.Linear(dim, num_classes)
      )
x = x.mean(dim = 1) if self.pool == 'mean' else x[:, 0]

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

self.avgpool = nn.AdaptiveAvgPool1d(1)
self.head = nn.Linear(self.num_features, num_classes) if num_classes > 0 else nn.Identity()

x = self.avgpool(x.transpose(1, 2))  # B C 1
x = torch.flatten(x, 1)
x = self.head(x)

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)

       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)'

       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),
      )

       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.Sequential(
           nn.LayerNorm(dim),
           nn.Linear(dim, num_classes)
      )

   def forward(self, img):
       x = self.to_patch_embedding(img)
       b, n, _ = x.shape

       cls_tokens = repeat(self.cls_token, '1 n d -> b n d', b = b)
       x = torch.cat((cls_tokens, x), dim=1)
       x += self.pos_embedding[:, :(n + 1)]
       x = self.dropout(x)
       x = self.transformer(x)
       x = x.mean(dim = 1) if self.pool == 'mean' else x[:, 0]

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

Rearrange('b c (h p1) (w p2) -> b (h w) (p1 p2 c)', p1 = patch_height, p2 = patch_width)

ViT：https://github.com/lucidrains/vit-pytorch/blob/main/vit_pytorch/vit.py
swin: https://github.com/microsoft/Swin-Transformer/blob/main/models/swin_transformer.py
TimeSformer:https://github.com/lucidrains/TimeSformer-pytorch/blob/main/timesformer_pytorch/timesformer_pytorch.py