from __future__ import annotations import argparse import dataclasses import os from typing import Optional import torch from torch.utils.data import DataLoader from blockdiffusion.data.texture_folder import TextureFolderDataset from blockdiffusion.data.captions import load_captions from blockdiffusion.data.texture_caption_dataset import TextureCaptionDataset from blockdiffusion.diffusion.gaussian_diffusion import GaussianDiffusion from blockdiffusion.diffusion.schedules import get_beta_schedule from blockdiffusion.models.unet16 import UNet16, UNet16Config from blockdiffusion.models.text_unet16 import TextCondConfig, TextCondUNet16 from blockdiffusion.text.simple_tokenizer import SimpleVocab, VocabConfig from blockdiffusion.train.trainer import Trainer, TrainerConfig from blockdiffusion.utils.seed import get_device, seed_everything def parse_args() -> argparse.Namespace: p = argparse.ArgumentParser(description="BlockDiffusion 16×16 纹理训练入口") # 数据 p.add_argument("--data_dir", type=str, required=True, help="纹理图片目录(可含子目录)") p.add_argument("--image_size", type=int, default=16, help="纹理边长(默认 16)") p.add_argument("--channels", type=int, default=3, help="纹理通道数(默认 RGB=3)") p.add_argument("--captions", type=str, default=None, help="图文配对文件(.jsonl/.csv),提供则启用提示词条件训练") p.add_argument("--max_text_len", type=int, default=32, help="文本最大长度(中文按字、英文按词)") p.add_argument("--vocab_max", type=int, default=8000, help="词表最大大小") # 扩散 p.add_argument("--timesteps", type=int, default=1000, help="扩散步数") p.add_argument("--beta_schedule", type=str, default="cosine", choices=["linear", "cosine"], help="噪声调度") # UNet p.add_argument("--base_channels", type=int, default=64, help="UNet 基础通道数") p.add_argument("--num_res_blocks", type=int, default=2, help="每个 level 的 ResBlock 数") p.add_argument("--dropout", type=float, default=0.0, help="ResBlock dropout(一般可为 0)") # 训练 p.add_argument("--batch_size", type=int, default=512) # Windows 下多进程 DataLoader 可能更挑剔,默认给 0 更稳;需要更快可手动调大。 p.add_argument("--num_workers", type=int, default=(0 if os.name == "nt" else 4)) p.add_argument("--lr", type=float, default=2e-4) p.add_argument("--weight_decay", type=float, default=1e-4) p.add_argument("--total_steps", type=int, default=200_000) p.add_argument("--grad_clip", type=float, default=1.0) p.add_argument("--no_amp", action="store_true", help="关闭混合精度") p.add_argument("--ema_decay", type=float, default=0.9999) p.add_argument("--cond_drop_prob", type=float, default=0.1, help="CFG 训练:丢条件概率(0~1)") p.add_argument("--sample_prompt", type=str, default=None, help="训练中采样使用的提示词(仅条件训练时生效)") p.add_argument("--sample_cfg_scale", type=float, default=3.0, help="采样 CFG scale(仅条件训练时生效)") # 输出 p.add_argument("--out_dir", type=str, default="outputs") p.add_argument("--run_name", type=str, default="default") p.add_argument("--resume", type=str, default=None, help="从 checkpoint 恢复训练") # 其他 p.add_argument("--seed", type=int, default=42) p.add_argument("--device", type=str, default=None, help='如 "cuda" / "cuda:0" / "cpu"') return p.parse_args() def main() -> None: args = parse_args() seed_everything(args.seed, deterministic=False) device = get_device(args.device) # 性能优化:适用于 RTX 40 系列(不影响数值正确性) if hasattr(torch, "set_float32_matmul_precision"): torch.set_float32_matmul_precision("high") if device.type == "cuda": torch.backends.cuda.matmul.allow_tf32 = True torch.backends.cudnn.allow_tf32 = True sample_model_kwargs = None sample_uncond_model_kwargs = None # 选择数据集:无条件/条件(captions) if args.captions: cap_items = load_captions(args.captions, data_root=args.data_dir) # 可选:强制加入透明/不透明 token,方便学习 alpha 与提示词关联(需你在 captions 里实际使用这些词) vocab = SimpleVocab.build( (it.text for it in cap_items), VocabConfig(max_vocab=args.vocab_max, min_freq=1), extra_tokens=["transparent", "opaque"], ) dataset = TextureCaptionDataset( items=cap_items, vocab=vocab, image_size=args.image_size, channels=args.channels, max_text_len=args.max_text_len, random_flip=True, ) else: vocab = None dataset = TextureFolderDataset( root_dir=args.data_dir, image_size=args.image_size, channels=args.channels, random_flip=True, ) dataloader = DataLoader( dataset, batch_size=args.batch_size, shuffle=True, num_workers=args.num_workers, pin_memory=(device.type == "cuda"), drop_last=True, ) unet_cfg = UNet16Config( in_channels=args.channels, out_channels=args.channels, base_channels=args.base_channels, channel_mults=(1, 2, 2), # 16->8->4 num_res_blocks=args.num_res_blocks, attn_resolutions=(8, 4), dropout=args.dropout, time_emb_dim=256, cond_dim=256, ) model_type = "uncond" text_cfg = None if args.captions: # 条件模型:TextEncoder 输出维度与 time_emb_dim 对齐 text_cfg = TextCondConfig(max_text_len=args.max_text_len, emb_dim=128, dropout=0.0) model = TextCondUNet16(unet_cfg=unet_cfg, vocab=vocab, text_cfg=text_cfg) model_type = "text_cond" # 训练中采样(可选):提供 sample_prompt 才启用 CFG 采样,否则走无条件采样 if args.sample_prompt: tok = torch.tensor([vocab.encode(args.sample_prompt, max_len=args.max_text_len)], dtype=torch.long) m = (tok != vocab.pad_id) sample_model_kwargs = {"tokens": tok.to(device), "mask": m.to(device)} # 无条件分支:mask 全 0(tokens 可以随意,统一用 pad) un_tok = torch.full_like(tok, fill_value=vocab.pad_id) un_m = torch.zeros_like(m) sample_uncond_model_kwargs = {"tokens": un_tok.to(device), "mask": un_m.to(device)} else: model = UNet16(unet_cfg) betas = get_beta_schedule(args.beta_schedule, args.timesteps).to(device) diffusion = GaussianDiffusion(betas) trainer_cfg = TrainerConfig( image_size=args.image_size, channels=args.channels, batch_size=args.batch_size, num_workers=args.num_workers, lr=args.lr, weight_decay=args.weight_decay, total_steps=args.total_steps, grad_clip=args.grad_clip, amp=not args.no_amp, ema_decay=args.ema_decay, cond_drop_prob=args.cond_drop_prob, sample_cfg_scale=args.sample_cfg_scale, out_dir=args.out_dir, run_name=args.run_name, resume_path=args.resume, ) meta = { "model_type": model_type, "unet_cfg": dataclasses.asdict(unet_cfg), "diffusion": {"timesteps": args.timesteps, "beta_schedule": args.beta_schedule}, "data": {"image_size": args.image_size, "channels": args.channels}, } if args.captions and vocab is not None and text_cfg is not None: meta["tokenizer"] = vocab.to_dict() meta["text_cfg"] = dataclasses.asdict(text_cfg) trainer = Trainer( model=model, diffusion=diffusion, dataloader=dataloader, device=device, cfg=trainer_cfg, meta=meta, sample_model_kwargs=sample_model_kwargs, sample_uncond_model_kwargs=sample_uncond_model_kwargs, ) trainer.train() if __name__ == "__main__": main()