Swin Transformer model代码

以下是Swin Transformer的PyTorch代码实现，包括Swin Transformer的模型定义和训练过程：

import torch
import torch.nn as nn
import torch.nn.functional as F


class SwinBlock(nn.Module):
    """Swin Transformer Block"""
    
    def __init__(self, dim, num_heads, window_size=7, shift_size=0):
        super().__init__()
        self.norm1 = nn.LayerNorm(dim)
        self.attn = nn.MultiheadAttention(dim, num_heads)
        self.norm2 = nn.LayerNorm(dim)
        self.mlp = nn.Sequential(
            nn.Linear(dim, dim * 4),
            nn.GELU(),
            nn.Linear(dim * 4, dim),
        )
        self.window_size = window_size
        self.shift_size = shift_size
        
        if window_size == 1 and shift_size == 0:
            self.window_attn = None
        else:
            self.window_attn = nn.MultiheadAttention(dim, num_heads)
        
    def forward(self, x):
        res = x
        x = self.norm1(x)
        if self.window_attn is not None:
            b, n, d = x.shape
            assert n % self.window_size == 0, "sequence length must be divisible by window size"
            
            x = x.reshape(b, n // self.window_size, self.window_size, d)
            x = x.permute(0, 2, 1, 3)
            x = x.reshape(b * self.window_size, n // self.window_size, d)
            
            window_res = x
            x = self.window_attn(x, x, x)[0]
            
            x = x.reshape(b, self.window_size, n // self.window_size, d)
            x = x.permute(0, 2, 1, 3)
            x = x.reshape(b, n, d)
            
            x += window_res
        
        x = x + self.attn(x, x, x)[0]
        x = res + x
        res = x
        x = self.norm2(x)
        x = x + self.mlp(x)
        x = res + x
        
        if self.shift_size > 0:
            x = F.pad(x, (0, 0, 0, 0, self.shift_size, 0))
            x = x[:, :-self.shift_size, :]
        
        return x


class SwinTransformer(nn.Module):
    """Swin Transformer Model"""
    
    def __init__(self, img_size, patch_size, in_chans, num_classes, embed_dim, depths, num_heads, window_size=7, shift_size=0):
        super().__init__()
        assert img_size % patch_size == 0, "image size must be divisible by patch size"
        num_patches = (img_size // patch_size) ** 2
        patch_dim = in_chans * patch_size ** 2
        self.patch_size = patch_size
        
        # Patch Embeddings
        self.patch_embed = nn.Conv2d(in_chans, embed_dim, kernel_size=patch_size, stride=patch_size)
        self.pos_embed = nn.Parameter(torch.zeros(1, num_patches, embed_dim))
        
        # Stages
        self.stages = nn.ModuleList([
            nn.Sequential(*[
                SwinBlock(embed_dim, num_heads, window_size, shift_size)
                for _ in range(depth)
            ])
            for depth in depths
        ])
        
        # Classifier Head
        self.norm = nn.LayerNorm(embed_dim)
        self.avgpool = nn.AdaptiveAvgPool2d(1)
        self.fc = nn.Linear(embed_dim, num_classes)
        
    def forward(self, x):
        x = self.patch_embed(x)
        b, c, h, w = x.shape
        x = x.flatten(2).transpose(1, 2)
        x = x + self.pos_embed
        x = self.norm(x)
        
        for stage in self.stages:
            x = stage(x)
        
        x = self.avgpool(x.transpose(1, 2).reshape(b, c, h // self.patch_size, w // self.patch_size))
        x = x.flatten(1)
        x = self.fc(x)
        
        return x

训练过程：

import torch.optim as optim

# Define Model
model = SwinTransformer(
    img_size=224,
    patch_size=4,
    in_chans=3,
    num_classes=10,
    embed_dim=96,
    depths=[2, 2, 6, 2],
    num_heads=6,
    window_size=7,
    shift_size=2,
)
model.cuda()

# Define Loss and Optimizer
criterion = nn.CrossEntropyLoss()
optimizer = optim.SGD(model.parameters(), lr=0.01, momentum=0.9)

# Train Model
for epoch in range(num_epochs):
    running_loss = 0.0
    for i, data in enumerate(trainloader, 0):
        inputs, labels = data
        inputs, labels = inputs.cuda(), labels.cuda()

        optimizer.zero_grad()

        outputs = model(inputs)
        loss = criterion(outputs, labels)
        loss.backward()
        optimizer.step()

        running_loss += loss.item()

    print(f"Epoch {epoch+1}: Training Loss = {running_loss/len(trainloader)}")