from __future__ import annotations from dataclasses import dataclass from typing import Optional import torch from torch import nn from blockdiffusion.models.unet16 import UNet16, UNet16Config from blockdiffusion.text.simple_tokenizer import SimpleVocab from blockdiffusion.text.text_encoder import TextEncoder, TextEncoderConfig @dataclass(frozen=True) class TextCondConfig: """ 文本条件配置(用于 checkpoint/meta 保存)。 """ max_text_len: int = 32 emb_dim: int = 128 dropout: float = 0.0 class TextCondUNet16(nn.Module): """ 文本条件 UNet: - TextEncoder 将 tokens/mask 编码为 cond 向量 - cond 向量注入到 UNet 的 time embedding(加和) - tokens/mask 为空或 mask 全 0 时等价于无条件 """ def __init__( self, unet_cfg: UNet16Config, vocab: SimpleVocab, text_cfg: TextCondConfig, ): super().__init__() self.unet_cfg = unet_cfg self.text_cfg = text_cfg self.vocab = vocab # 仅用于保存 meta;不注册为 buffer/module te_cfg = TextEncoderConfig( vocab_size=vocab.size, max_len=text_cfg.max_text_len, emb_dim=text_cfg.emb_dim, out_dim=unet_cfg.time_emb_dim, dropout=text_cfg.dropout, ) self.text_encoder = TextEncoder(te_cfg, pad_id=vocab.pad_id) self.unet = UNet16(unet_cfg) def forward( self, x: torch.Tensor, t: torch.Tensor, tokens: Optional[torch.Tensor] = None, mask: Optional[torch.Tensor] = None, ) -> torch.Tensor: cond = None if tokens is not None: cond = self.text_encoder(tokens, mask=mask) return self.unet(x, t, cond=cond)