from __future__ import annotations from dataclasses import dataclass from typing import Optional import torch from torch import nn @dataclass(frozen=True) class TextEncoderConfig: """ 文本编码器配置: - 目标是轻量、可训练、可随 checkpoint 一起保存,不依赖外部大模型。 """ vocab_size: int max_len: int = 32 emb_dim: int = 128 out_dim: int = 256 dropout: float = 0.0 class TextEncoder(nn.Module): """ 极简 TextEncoder: - token embedding - masked mean pooling - MLP 投影到 out_dim(通常与 UNet time_emb_dim 对齐) """ def __init__(self, cfg: TextEncoderConfig, pad_id: int = 0): super().__init__() self.cfg = cfg self.pad_id = int(pad_id) self.emb = nn.Embedding(cfg.vocab_size, cfg.emb_dim, padding_idx=self.pad_id) self.proj = nn.Sequential( nn.Dropout(cfg.dropout) if cfg.dropout > 0 else nn.Identity(), nn.Linear(cfg.emb_dim, cfg.out_dim), nn.SiLU(), nn.Linear(cfg.out_dim, cfg.out_dim), ) def forward(self, tokens: torch.Tensor, mask: Optional[torch.Tensor] = None) -> torch.Tensor: """ tokens: (B, L) int64 mask: (B, L) bool/0-1;为 None 时自动根据 pad_id 生成 返回: - cond: (B, out_dim) """ if mask is None: mask = (tokens != self.pad_id) mask_f = mask.to(dtype=self.emb.weight.dtype) x = self.emb(tokens) # (B,L,D) # masked mean denom = mask_f.sum(dim=1, keepdim=True).clamp(min=1.0) pooled = (x * mask_f.unsqueeze(-1)).sum(dim=1) / denom # 若整句为空(mask 全 0),pooled 会是全 0,等价于无条件 return self.proj(pooled)