Shaun ZhouandClaude Opus 5
141cb907f2
移除"训练必须有 Nvidia GPU"的硬性守卫，改为按后端能力分派。
- utils/device: resolve_device("auto") 按 cuda > mps > cpu 挑最快的可用后端；
显式指定不可用的后端时直接报错，不会悄悄退回 CPU——训练慢几十倍却不吭声是最坏的情况
- AMP 随后端取默认值：CUDA 开、MPS 关。MPS 上 autocast 数值正常但实测更慢
（tiny_conv @128²：fp32 76 ms/步、fp16 90 ms、bf16 108 ms），需要时 amp: true 可强开
- PeakMemory 统一显存峰值统计：CUDA 读内置计数器，MPS 无峰值 API 改用逐 step 采样
（是下界，日志与产出里已标注）
- profile_inference 支持 MPS，并提示 torch.profiler 没有 MPS 的 ProfilerActivity，
算子明细只反映 CPU 侧下发耗时
- train.py / eval.py 新增 --device，配置默认改为 auto
- check_env 改为列出全部可用后端与 auto 的选择结果
实测：tiny_conv 训练在 MPS 上比 CPU 快约 32 倍。
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
105 lines
3.8 KiB
Python
Executable File
105 lines
3.8 KiB
Python
Executable File
#!/usr/bin/env python
|
||
"""训练入口。
|
||
|
||
用法:
|
||
python train.py --config configs/train/baseline.yaml
|
||
python train.py --config configs/train/baseline.yaml --set train.epochs=10 train.optimizer.lr=5e-4
|
||
python train.py --config configs/train/baseline.yaml --device mps
|
||
|
||
设备由配置的 `device` 决定，默认 `auto`：按 cuda > mps > cpu 挑最快的可用后端。
|
||
显式指定但不可用时直接报错，不会悄悄退回 CPU。
|
||
"""
|
||
|
||
from __future__ import annotations
|
||
|
||
import argparse
|
||
import sys
|
||
from pathlib import Path
|
||
|
||
from datasets import build_dataloader, build_dataset, load_classes
|
||
from engine import Trainer
|
||
from models import build_model
|
||
from utils import check_label_mode, load_config, seed_everything, setup_logging
|
||
from utils.device import DeviceNotAvailableError, describe_device, resolve_device
|
||
|
||
|
||
def parse_args() -> argparse.Namespace:
|
||
parser = argparse.ArgumentParser(description="薄膜缺陷分类定位模型训练")
|
||
parser.add_argument("--config", required=True, help="yaml 配置路径")
|
||
parser.add_argument("--set", nargs="*", default=[], metavar="KEY=VALUE", help="覆盖配置项")
|
||
parser.add_argument("--work-dir", default=None, help="覆盖配置中的 work_dir")
|
||
parser.add_argument("--device", default=None, help="覆盖配置中的 device：auto / cuda / mps / cpu")
|
||
parser.add_argument("--resume", default=None, help="从检查点恢复训练")
|
||
return parser.parse_args()
|
||
|
||
|
||
def main() -> int:
|
||
args = parse_args()
|
||
cfg = load_config(args.config, args.set)
|
||
|
||
work_dir = Path(args.work_dir or cfg.get("work_dir", "runs/default"))
|
||
work_dir.mkdir(parents=True, exist_ok=True)
|
||
logger = setup_logging(work_dir / "train.log")
|
||
|
||
try:
|
||
device = resolve_device(args.device or cfg.get("device", "auto"))
|
||
except DeviceNotAvailableError as exc:
|
||
logger.error("%s", exc)
|
||
return 1
|
||
logger.info("训练设备: %s", describe_device(device))
|
||
|
||
seed_everything(int(cfg.get("seed", 42)), bool(cfg.get("deterministic", False)))
|
||
|
||
classes = load_classes(cfg["classes"])
|
||
# label_mode 由 yaml 决定，模型与数据集必须一致，这里统一下发。
|
||
label_mode = check_label_mode(str(cfg.get("label_mode", "multilabel")))
|
||
model_cfg = dict(cfg["model"])
|
||
model_cfg.setdefault("num_classes", len(classes))
|
||
model_cfg.setdefault("label_mode", label_mode)
|
||
model = build_model(model_cfg).to(device)
|
||
logger.info(
|
||
"模型 %s 输出 stride=%d，参数量 %.2fM",
|
||
model_cfg["name"],
|
||
model.out_stride,
|
||
sum(p.numel() for p in model.parameters()) / 1e6,
|
||
)
|
||
|
||
loaders = {}
|
||
for split in ("train", "val"):
|
||
split_cfg = cfg["data"].get(split)
|
||
if not split_cfg:
|
||
loaders[split] = None
|
||
continue
|
||
dataset = build_dataset(
|
||
split_cfg["dataset"], classes=classes, stride=model.out_stride, label_mode=label_mode
|
||
)
|
||
loaders[split] = build_dataloader(dataset, **dict(split_cfg.get("loader", {})))
|
||
logger.info("%s 集样本数: %d", split, len(dataset))
|
||
|
||
trainer = Trainer(
|
||
model=model,
|
||
train_loader=loaders["train"],
|
||
val_loader=loaders["val"],
|
||
cfg=cfg.get("train", {}),
|
||
device=device,
|
||
work_dir=work_dir,
|
||
full_config=cfg.to_dict(),
|
||
)
|
||
|
||
if args.resume:
|
||
from engine import load_checkpoint
|
||
|
||
state = load_checkpoint(args.resume, model, trainer.optimizer, map_location=device)
|
||
trainer.start_epoch = int(state.get("epoch", 0)) + 1
|
||
trainer.best_metric = float(state.get("best_metric") or float("-inf"))
|
||
logger.info("已从 %s 恢复，起始 epoch=%d", args.resume, trainer.start_epoch)
|
||
|
||
cfg.dump(work_dir / "config.yaml")
|
||
best = trainer.fit()
|
||
logger.info("训练结束，最佳 %s=%.4f", trainer.monitor, best)
|
||
return 0
|
||
|
||
|
||
if __name__ == "__main__":
|
||
sys.exit(main())
|