from __future__ import annotations import os from dataclasses import dataclass from typing import Any, Dict, Optional import torch from torch import nn from contextlib import contextmanager from torch.utils.data import DataLoader from blockdiffusion.diffusion.gaussian_diffusion import GaussianDiffusion from blockdiffusion.train.guidance import CFGuidanceModel from blockdiffusion.utils.checkpoint import load_checkpoint, save_checkpoint from blockdiffusion.utils.ema import EMA from blockdiffusion.utils.image_io import save_tensor_grid def _make_grad_scaler(enabled: bool, device_type: str): """ 兼容不同 PyTorch 版本的 AMP GradScaler: - 新版推荐 torch.amp.GradScaler(device, ...) - 旧版使用 torch.cuda.amp.GradScaler(...) """ enabled = bool(enabled and device_type == "cuda") try: from torch.amp import GradScaler as _GS # PyTorch 2.x if not enabled: return _GS(enabled=False) # 有的版本是 GradScaler(device='cuda', ...),有的接受位置参数 try: return _GS("cuda", enabled=True) except TypeError: return _GS(enabled=True) except Exception: from torch.cuda.amp import GradScaler as _GS # 兼容旧版本 return _GS(enabled=enabled) @contextmanager def _autocast(device_type: str, enabled: bool): """ 兼容不同 PyTorch 版本的 autocast: - torch.amp.autocast(device_type=..., enabled=...) - torch.cuda.amp.autocast(enabled=...) """ enabled = bool(enabled and device_type == "cuda") try: from torch.amp import autocast as _AC with _AC(device_type=device_type, enabled=enabled): yield except Exception: from torch.cuda.amp import autocast as _AC with _AC(enabled=enabled): yield @dataclass class TrainerConfig: """ 训练器配置(尽量给出稳定且适合 16×16 的默认值)。 """ # 数据 image_size: int = 16 channels: int = 3 batch_size: int = 512 num_workers: int = 4 # 训练 lr: float = 2e-4 weight_decay: float = 1e-4 total_steps: int = 200_000 grad_clip: float = 1.0 amp: bool = True cond_drop_prob: float = 0.1 # 仅在有条件训练时生效(CFG 训练) grad_accum_steps: int = 1 # 梯度累积:提高有效 batch(质量更稳),也可用于控制显存 lr_warmup_steps: int = 2000 # 学习率 warmup 步数(0 表示关闭) lr_min_ratio: float = 0.1 # 余弦退火最低 lr 比例(相对 lr) snr_gamma: float = 5.0 # SNR 加权 MSE 的 gamma(<=0 表示关闭) channels_last: bool = True # Conv 的内存布局优化,通常能略提速 # EMA ema_decay: float = 0.9999 # 日志/保存 log_every: int = 100 save_every: int = 5_000 sample_every: int = 5_000 sample_batch: int = 64 sample_use_ddim: bool = True sample_ddim_steps: int = 50 sample_ddim_eta: float = 0.0 sample_cfg_scale: float = 3.0 out_dir: str = "outputs" run_name: str = "default" resume_path: Optional[str] = None class Trainer: """ 训练器只负责: - dataloader - 优化器/AMP/EMA - checkpoint 与训练过程采样 扩散公式与采样过程由 GaussianDiffusion 负责,保证解耦。 """ def __init__( self, model: nn.Module, diffusion: GaussianDiffusion, dataloader: DataLoader, device: torch.device, cfg: TrainerConfig, meta: Optional[dict] = None, sample_model_kwargs: Optional[Dict[str, Any]] = None, sample_uncond_model_kwargs: Optional[Dict[str, Any]] = None, ): self.model = model.to(device) self.diffusion = diffusion self.dataloader = dataloader self.device = device self.cfg = cfg self.meta = meta or {} self.sample_model_kwargs = sample_model_kwargs self.sample_uncond_model_kwargs = sample_uncond_model_kwargs # 可选:channels_last 提升卷积吞吐(对小图也常有帮助) if cfg.channels_last and device.type == "cuda": self.model = self.model.to(memory_format=torch.channels_last) self.optimizer = torch.optim.AdamW(self.model.parameters(), lr=cfg.lr, weight_decay=cfg.weight_decay) use_amp = bool(cfg.amp and device.type == "cuda") self.scaler = _make_grad_scaler(enabled=use_amp, device_type=device.type) self.ema = EMA(self.model, decay=cfg.ema_decay) self.step = 0 self._prepare_dirs() if cfg.resume_path: self._resume(cfg.resume_path) # 学习率调度(warmup + cosine decay),提升稳定性/质量 self.base_lr = float(cfg.lr) self.min_lr = float(cfg.lr) * float(cfg.lr_min_ratio) def _prepare_dirs(self) -> None: self.run_dir = os.path.join(self.cfg.out_dir, self.cfg.run_name) self.ckpt_dir = os.path.join(self.run_dir, "checkpoints") self.sample_dir = os.path.join(self.run_dir, "samples") os.makedirs(self.ckpt_dir, exist_ok=True) os.makedirs(self.sample_dir, exist_ok=True) def _resume(self, path: str) -> None: ckpt = load_checkpoint(path, map_location="cpu") self.model.load_state_dict(ckpt["model"], strict=True) if "ema" in ckpt: self.ema.load_state_dict(ckpt["ema"]) # 续训:确保 EMA shadow 跟随当前训练设备 self.ema.to(self.device) if "optimizer" in ckpt: self.optimizer.load_state_dict(ckpt["optimizer"]) # 续训:optimizer state 里的 tensor 默认仍在 CPU,需要搬到当前设备 for state in self.optimizer.state.values(): for kk, vv in list(state.items()): if torch.is_tensor(vv): state[kk] = vv.to(self.device) if "scaler" in ckpt: self.scaler.load_state_dict(ckpt["scaler"]) self.step = int(ckpt.get("step", 0)) def _save(self) -> None: path = os.path.join(self.ckpt_dir, f"step_{self.step}.pt") payload = { "step": self.step, "meta": self.meta, "trainer_cfg": self.cfg.__dict__, "model": self.model.state_dict(), "ema": self.ema.state_dict(), "optimizer": self.optimizer.state_dict(), "scaler": self.scaler.state_dict(), } save_checkpoint(path, payload) @torch.no_grad() def _sample(self) -> None: # 采样时使用 EMA 参数,避免训练噪声导致的质量抖动 tmp = {k: v.detach().clone() for k, v in self.model.state_dict().items()} self.ema.copy_to(self.model) self.model.eval() shape = (self.cfg.sample_batch, self.cfg.channels, self.cfg.image_size, self.cfg.image_size) sampler_model: nn.Module = self.model if self.sample_model_kwargs is not None and self.sample_uncond_model_kwargs is not None: sampler_model = CFGuidanceModel( base_model=self.model, cond_kwargs=self.sample_model_kwargs, uncond_kwargs=self.sample_uncond_model_kwargs, scale=self.cfg.sample_cfg_scale, ) if self.cfg.sample_use_ddim: x = self.diffusion.ddim_sample_loop( sampler_model, shape=shape, device=self.device, steps=self.cfg.sample_ddim_steps, eta=self.cfg.sample_ddim_eta, model_kwargs=None, progress=False, ) else: x = self.diffusion.p_sample_loop(sampler_model, shape=shape, device=self.device, model_kwargs=None, progress=False) # [-1,1] -> [0,1] 保存 x = (x.clamp(-1, 1) + 1) * 0.5 out_path = os.path.join(self.sample_dir, f"step_{self.step}.png") save_tensor_grid(x, out_path, nrow=int(self.cfg.sample_batch**0.5), padding=1) # 恢复训练参数 self.model.load_state_dict(tmp, strict=True) self.model.train() def train(self) -> None: self.model.train() data_iter = iter(self.dataloader) while self.step < self.cfg.total_steps: # 梯度累积:每 grad_accum_steps 次 forward 才做一次 optimizer step accum_steps = max(1, int(self.cfg.grad_accum_steps)) self.optimizer.zero_grad(set_to_none=True) total_loss = 0.0 for micro in range(accum_steps): try: batch = next(data_iter) except StopIteration: data_iter = iter(self.dataloader) batch = next(data_iter) model_kwargs: Dict[str, Any] = {} if isinstance(batch, (tuple, list)) and len(batch) == 3: x0, tokens, mask = batch model_kwargs["tokens"] = tokens.to(self.device, non_blocking=True) model_kwargs["mask"] = mask.to(self.device, non_blocking=True) else: x0 = batch # channels_last:数据也转换同布局 x0 = x0.to(self.device, non_blocking=True) if self.cfg.channels_last and self.device.type == "cuda": x0 = x0.contiguous(memory_format=torch.channels_last) b = x0.shape[0] t = torch.randint(0, self.diffusion.num_timesteps, (b,), device=self.device, dtype=torch.long) with _autocast(device_type=self.device.type, enabled=self.scaler.is_enabled()): loss = self.diffusion.training_losses( self.model, x0, t, model_kwargs=model_kwargs, cond_drop_prob=self.cfg.cond_drop_prob, snr_gamma=float(self.cfg.snr_gamma), )["loss"] # 累积时按比例缩放,保持有效学习率不变 loss = loss / accum_steps self.scaler.scale(loss).backward() total_loss += float(loss.detach().item()) # 学习率 warmup + cosine decay(基于全局 step) if self.cfg.lr_warmup_steps and self.cfg.lr_warmup_steps > 0 and self.step < self.cfg.lr_warmup_steps: lr = self.min_lr + (self.base_lr - self.min_lr) * (float(self.step + 1) / float(self.cfg.lr_warmup_steps)) else: # cosine 从 warmup 结束后开始 warm = max(0, int(self.cfg.lr_warmup_steps)) prog = (self.step - warm) / max(1.0, float(self.cfg.total_steps - warm)) prog = min(max(prog, 0.0), 1.0) import math lr = self.min_lr + 0.5 * (self.base_lr - self.min_lr) * (1.0 + math.cos(math.pi * prog)) for pg in self.optimizer.param_groups: pg["lr"] = lr if self.cfg.grad_clip and self.cfg.grad_clip > 0: self.scaler.unscale_(self.optimizer) torch.nn.utils.clip_grad_norm_(self.model.parameters(), self.cfg.grad_clip) self.scaler.step(self.optimizer) self.scaler.update() self.ema.update(self.model) self.step += 1 if self.cfg.log_every > 0 and (self.step % self.cfg.log_every == 0): # total_loss 已按 accum_steps 平均后的值(每 micro 先 /accum) print(f"[step {self.step}] loss={total_loss:.6f} lr={lr:.2e}") if self.cfg.save_every > 0 and (self.step % self.cfg.save_every == 0): self._save() if self.cfg.sample_every > 0 and (self.step % self.cfg.sample_every == 0): self._sample()