存一串代码（简化的动态稀疏视觉Transformer的PyTorch代码）
import torch
import torch.nn as nn
import torch.nn.functional as F class DynamicSparseAttention(nn.Module): def __init__(self, dim, num_heads=8, dropout=0.1): super().__init__() self.num_heads = num_heads self.head_dim = dim // num_heads self.scale = self.head_dim ** -0.5 self.qkv = nn.Linear(dim, dim * 3, bias=False) self.attn_drop = nn.Dropout(dropout) self.proj = nn.Linear(dim, dim) self.proj_drop = nn.Dropout(dropout) def forward(self, x): B, N, C = x.shape qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, self.head_dim).permute(2, 0, 3, 1, 4) q, k, v = qkv.unbind(0) attn = (q @ k.transpose(-2, -1)) * self.scale attn = attn.softmax(dim=-1) attn = self.attn_drop(attn) x = (attn @ v).transpose(1, 2).reshape(B, N, C) x = self.proj(x) x = self.proj_drop(x) return x class HierarchicalRoutingBlock(nn.Module): def __init__(self, dim, num_heads=8, mlp_ratio=4., dropout=0.1): super().__init__() self.norm1 = nn.LayerNorm(dim) self.attn = DynamicSparseAttention(dim, num_heads, dropout) self.norm2 = nn.LayerNorm(dim) self.mlp = nn.Sequential( nn.Linear(dim, int(dim * mlp_ratio)), nn.GELU(), nn.Dropout(dropout), nn.Linear(int(dim * mlp_ratio), dim), nn.Dropout(dropout) ) def forward(self, x): x = x + self.attn(self.norm1(x)) x = x + self.mlp(self.norm2(x)) return x class DynamicSparseVisionTransformer(nn.Module): def __init__(self, img_size=224, patch_size=16, num_classes=1000, dim=768, num_heads=8, depth=12, mlp_ratio=4., dropout=0.1): super().__init__() self.patch_embed = nn.Conv2d(3, dim, kernel_size=patch_size, stride=patch_size) self.pos_embed = nn.Parameter(torch.zeros(1, (img_size // patch_size) ** 2, dim)) self.dropout = nn.Dropout(dropout) self.blocks = nn.ModuleList([HierarchicalRoutingBlock(dim, num_heads, mlp_ratio, dropout) for _ in range(depth)]) self.norm = nn.LayerNorm(dim) self.head = nn.Linear(dim, num_classes) if num_classes > 0 else nn.Identity() def forward(self, x): x = self.patch_embed(x).flatten(2).transpose(1, 2) x = x + self.pos_embed x = self.dropout(x) for blk in self.blocks: x = blk(x) x = self.norm(x) x = x[:, 0] x = self.head(x) return x # 使用
model = DynamicSparseVisionTransformer()
x = torch.randn(1, 3, 224, 224)
output = model(x)
print(output.shape)
代码解释
DynamicSparseAttention：实现动态稀疏注意力模块。
HierarchicalRoutingBlock：实现层次化路由块，包含注意力模块和多层感知机。
DynamicSparseVisionTransformer：实现完整的动态稀疏视觉Transformer模型，包括补丁嵌入、位置嵌入、层次化路由块和分类头。
