from __future__ import annotations """ 一键训练脚本(无命令行参数): - 直接修改本文件顶部常量即可 - 数据目录:DATA_DIR - 提示词来自文件名(HASH-xxx_yyy.png => "xxx yyy") 运行方式(推荐 Python 3.10): - py -3.10 train_oneclick.py """ import dataclasses import os import torch from torch.utils.data import DataLoader from blockdiffusion.data.filename_captions import build_caption_items_from_filenames 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.text_unet16 import TextCondConfig, TextCondUNet16 from blockdiffusion.models.unet16 import UNet16Config from blockdiffusion.text.simple_tokenizer import SimpleVocab, VocabConfig from blockdiffusion.train.trainer import Trainer, TrainerConfig from blockdiffusion.utils.checkpoint import load_checkpoint from blockdiffusion.utils.seed import get_device, seed_everything # ========================= # 可修改常量(按需改这里) # ========================= # 数据目录(你提供的路径) DATA_DIR = r"G:\PycharmProjects\modrinth_downloads\Training_v2" # 纹理尺寸(本项目目标 16) IMAGE_SIZE = 16 # 通道数:3=RGB,4=RGBA(需要保留透明就用 4) CHANNELS = 4 # 扩散步数与调度 TIMESTEPS = 1000 BETA_SCHEDULE = "cosine" # "linear" / "cosine" # UNet 规模(16×16 不要太深) BASE_CHANNELS = 64 NUM_RES_BLOCKS = 2 DROPOUT = 0.0 # 文本(提示词)配置 MAX_TEXT_LEN = 32 VOCAB_MAX = 8000 COND_DROP_PROB = 0.1 # CFG 训练丢条件概率(0~1) PRETOKENIZE = False # 大数据集建议 False:启动更快、占用更小 # 透明度提示词(可选): # - 仅在 CHANNELS=4 且“新训练(非续训)”时建议开启 # - 作用:自动根据图片 alpha 是否含透明像素,为文本追加 "transparent"/"opaque" # 这样模型才能学到“透明度与提示词相关”的监督信号 APPEND_ALPHA_TAG = False ALPHA_TAG_TRANSPARENT = "transparent" ALPHA_TAG_OPAQUE = "opaque" # 训练配置(4070Ti 12G:RGBA 建议先从 128/256 试起) BATCH_SIZE = 128 # 数据加载并行:Windows 若不稳定可改回 0;稳定时建议 4~8 NUM_WORKERS = 4 PREFETCH_FACTOR = 4 # NUM_WORKERS>0 时生效 PERSISTENT_WORKERS = True # NUM_WORKERS>0 时生效 LR = 2e-4 WEIGHT_DECAY = 1e-4 TOTAL_STEPS = 700_000 GRAD_CLIP = 1.0 AMP = True EMA_DECAY = 0.9999 GRAD_ACCUM_STEPS = 1 # 不想OOM但想更大有效batch:例如设为 2/4 LR_WARMUP_STEPS = 2000 LR_MIN_RATIO = 0.1 SNR_GAMMA = 5.0 # <=0 关闭;5 通常比较稳 CHANNELS_LAST = True LOG_EVERY = 10 SAVE_EVERY = 5000 SAMPLE_EVERY = 5000 # 输出 OUT_DIR = "outputs" RUN_NAME = "oneclick_filename_caption" # 断点续训: # - 若你明确指定某个 checkpoint,就填 RESUME_PATH # - 若 RESUME_PATH=None 且 AUTO_RESUME=True,会自动扫描 outputs//checkpoints 下最新 step_*.pt 继续训练 RESUME_PATH = None # 例如 r"outputs/oneclick_filename_caption/checkpoints/step_5000.pt" AUTO_RESUME = True # 训练中采样(可选):用这个提示词做固定可视化 SAMPLE_PROMPT = "block" # 改成你想看的关键词;不想要可设为 None SAMPLE_CFG_SCALE = 3.0 # 随机种子 SEED = 42 DEVICE = None # 例如 "cuda:0";None 表示自动 def main() -> None: seed_everything(SEED, deterministic=False) device = get_device(DEVICE) print(f"[init] device={device} channels={CHANNELS} image_size={IMAGE_SIZE}") # RTX 40 系列性能优化(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 # ===== # 续训注意: # - 如果设置了 RESUME_PATH,为保证词表/模型结构不变,本脚本会优先从 checkpoint 读取并复用配置; # - 新增数据会自动被扫描并参与训练;新出现的词如果不在旧词表里,会被编码为 (可继续训练,但不会学到新词语义)。 # ===== def _auto_find_latest_ckpt() -> str | None: ckpt_dir = os.path.join(OUT_DIR, RUN_NAME, "checkpoints") if not os.path.isdir(ckpt_dir): return None best_step = -1 best_path = None for fn in os.listdir(ckpt_dir): if not (fn.startswith("step_") and fn.endswith(".pt")): continue # 形如 step_5000.pt core = fn[len("step_") : -len(".pt")] try: step = int(core) except Exception: continue if step > best_step: best_step = step best_path = os.path.join(ckpt_dir, fn) return best_path effective_resume_path = RESUME_PATH if (effective_resume_path is None) and AUTO_RESUME: effective_resume_path = _auto_find_latest_ckpt() resume_meta = None if effective_resume_path: print(f"[resume] 使用 checkpoint: {effective_resume_path}") ckpt = load_checkpoint(effective_resume_path, map_location="cpu") resume_meta = ckpt.get("meta", {}) or {} # 从 checkpoint 复用关键配置(避免“重建词表”导致继续训练变差/不可用) if resume_meta: data_meta = resume_meta.get("data", {}) or {} diff_meta = resume_meta.get("diffusion", {}) or {} unet_cfg_dict = resume_meta.get("unet_cfg", None) text_cfg_dict = resume_meta.get("text_cfg", None) tok_dict = resume_meta.get("tokenizer", None) cfg_image_size = int(data_meta.get("image_size", IMAGE_SIZE)) cfg_channels = int(data_meta.get("channels", CHANNELS)) cfg_timesteps = int(diff_meta.get("timesteps", TIMESTEPS)) cfg_beta_schedule = str(diff_meta.get("beta_schedule", BETA_SCHEDULE)) if unet_cfg_dict is None or text_cfg_dict is None or tok_dict is None: raise ValueError("RESUME_PATH 指向的 checkpoint 缺少 tokenizer/text_cfg/unet_cfg,无法安全续训") vocab = SimpleVocab.from_dict(tok_dict) text_cfg = TextCondConfig(**text_cfg_dict) unet_cfg = UNet16Config(**unet_cfg_dict) else: cfg_image_size = IMAGE_SIZE cfg_channels = CHANNELS cfg_timesteps = TIMESTEPS cfg_beta_schedule = BETA_SCHEDULE text_cfg = TextCondConfig(max_text_len=MAX_TEXT_LEN, emb_dim=128, dropout=0.0) unet_cfg = UNet16Config( in_channels=cfg_channels, out_channels=cfg_channels, base_channels=BASE_CHANNELS, channel_mults=(1, 2, 2), # 16->8->4 num_res_blocks=NUM_RES_BLOCKS, attn_resolutions=(8, 4), dropout=DROPOUT, time_emb_dim=256, cond_dim=256, ) # 非续训:词表会基于当前数据构建 vocab = None # 扫描目录(只扫描一次):用于训练集;非续训时也用于构建词表 print("[data] 扫描图片并从文件名提取提示词...") cap_items = build_caption_items_from_filenames(DATA_DIR, verbose=True) if vocab is None: # 新训练:可强制加入额外 token,避免采样时提示词变成 而“看起来不遵从” extra_tokens = None append_alpha_tag = bool(APPEND_ALPHA_TAG and (cfg_channels == 4)) if append_alpha_tag: extra_tokens = [ALPHA_TAG_TRANSPARENT, ALPHA_TAG_OPAQUE] vocab = SimpleVocab.build( (it.text for it in cap_items), VocabConfig(max_vocab=VOCAB_MAX, min_freq=1), extra_tokens=extra_tokens, ) else: # 续训:vocab 固定,不能改变;否则 embedding 尺寸不匹配 append_alpha_tag = False dataset = TextureCaptionDataset( items=cap_items, vocab=vocab, image_size=cfg_image_size, channels=cfg_channels, max_text_len=text_cfg.max_text_len, random_flip=True, pretokenize=PRETOKENIZE, append_alpha_tag=append_alpha_tag, alpha_tag_transparent=ALPHA_TAG_TRANSPARENT, alpha_tag_opaque=ALPHA_TAG_OPAQUE, verbose=True, ) print(f"[data] dataset_size={len(dataset)} vocab_size={vocab.size} pretokenize={PRETOKENIZE}") dl_kwargs = dict( batch_size=BATCH_SIZE, shuffle=True, num_workers=NUM_WORKERS, pin_memory=(device.type == "cuda"), drop_last=True, ) if NUM_WORKERS > 0: dl_kwargs["prefetch_factor"] = int(PREFETCH_FACTOR) dl_kwargs["persistent_workers"] = bool(PERSISTENT_WORKERS) dataloader = DataLoader(dataset, **dl_kwargs) print(f"[data] dataloader_workers={NUM_WORKERS}") # 2) 模型(文本条件) model = TextCondUNet16(unet_cfg=unet_cfg, vocab=vocab, text_cfg=text_cfg) # 3) Diffusion betas = get_beta_schedule(cfg_beta_schedule, cfg_timesteps).to(device) diffusion = GaussianDiffusion(betas) # 4) 训练配置 trainer_cfg = TrainerConfig( image_size=cfg_image_size, channels=cfg_channels, batch_size=BATCH_SIZE, num_workers=NUM_WORKERS, lr=LR, weight_decay=WEIGHT_DECAY, total_steps=TOTAL_STEPS, grad_clip=GRAD_CLIP, amp=AMP, ema_decay=EMA_DECAY, cond_drop_prob=COND_DROP_PROB, grad_accum_steps=GRAD_ACCUM_STEPS, lr_warmup_steps=LR_WARMUP_STEPS, lr_min_ratio=LR_MIN_RATIO, snr_gamma=SNR_GAMMA, channels_last=CHANNELS_LAST, log_every=LOG_EVERY, save_every=SAVE_EVERY, sample_every=SAMPLE_EVERY, sample_cfg_scale=SAMPLE_CFG_SCALE, out_dir=OUT_DIR, run_name=RUN_NAME, resume_path=effective_resume_path, ) # 训练中采样 CFG(可选) sample_model_kwargs = None sample_uncond_model_kwargs = None if SAMPLE_PROMPT: tok = torch.tensor([vocab.encode(SAMPLE_PROMPT, max_len=text_cfg.max_text_len)], dtype=torch.long) m = (tok != vocab.pad_id) sample_model_kwargs = {"tokens": tok.to(device), "mask": m.to(device)} 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)} meta = { "model_type": "text_cond", "unet_cfg": dataclasses.asdict(unet_cfg), "diffusion": {"timesteps": cfg_timesteps, "beta_schedule": cfg_beta_schedule}, "data": {"image_size": cfg_image_size, "channels": cfg_channels}, "tokenizer": vocab.to_dict(), "text_cfg": dataclasses.asdict(text_cfg), # 额外信息:方便回溯 "data_dir": os.path.abspath(DATA_DIR), "caption_source": "filename", } 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, ) print("[train] start...") trainer.train() if __name__ == "__main__": main()