qing
316758d80e
两个问题（本轮归一化编译链路上实测暴露）： 1. 统计量只按 targets 计算：targets 中恒定的列 std 落到 1e-8 地板值，而 history 窗口可能含该列的其他取值（.mt 脚本前几行）， z-score 爆炸（实测 (90-45)/1e-8=4.5e9）→ act_tokens ±1e9 → DFC optimize 报 NegativeSlopeExponentNonFixable（desired shift 23 > 15 data bits）。pretrain_mt.py 与 dataset.py 的统计量改为 history∪target 联合计算，修复后校准数据 max|z| = 4.77。 2. action_stats_path 存成含目录的相对路径（"./checkpoints/ action_stats.npy"），推理端（predict/export_deploy/gen_calibration） 按 checkpoint 目录解析时二次拼接成 checkpoints/./checkpoints， 导出直接 FileNotFoundError。train.py/pretrain_mt.py 改存裸文件名 "action_stats.npy"（stats 与 checkpoint 同目录，语义一致）。 新增回归测试 test_stats_cover_history_values（history 独有取值不 得被 1e-8 地板值打爆），全量 77 通过。 归一化重编译全链路通过：token 量级 |img|=0.53 |act|=2.93 健康， optimize 通过，连通性检查 5/5，HEF 2.02MB 已部署 Pi 待 _npu_test 复验。
212 lines
8.6 KiB
Python
212 lines
8.6 KiB
Python
import json
|
|
import os
|
|
import random
|
|
import numpy as np
|
|
import torch
|
|
from torch.utils.data import DataLoader, Subset
|
|
from tqdm import tqdm
|
|
|
|
from config import Config, build_argparser, apply_overrides, validate_config_values
|
|
from model import ActionPredictor
|
|
from dataset import ActionDataset
|
|
|
|
|
|
def split_episodes(data_dir: str, val_ratio: float, seed: int = 42):
|
|
n = ActionDataset.get_episode_count(data_dir)
|
|
indices = list(range(n))
|
|
rng = np.random.RandomState(seed)
|
|
rng.shuffle(indices)
|
|
val_count = max(1, int(n * val_ratio))
|
|
return indices[val_count:], indices[:val_count]
|
|
|
|
|
|
def train_one_epoch(model, loader, optimizer, device, max_grad_norm: float = None):
|
|
model.train()
|
|
total_loss = 0.0
|
|
count = 0
|
|
for batch in tqdm(loader, desc=" train", leave=False):
|
|
images = batch["images"].to(device)
|
|
history = batch["action_history"].to(device)
|
|
target = batch["target"].to(device)
|
|
optimizer.zero_grad()
|
|
pred = model(images, history)
|
|
loss = torch.nn.functional.mse_loss(pred, target)
|
|
loss.backward()
|
|
if max_grad_norm:
|
|
torch.nn.utils.clip_grad_norm_(model.parameters(), max_grad_norm)
|
|
optimizer.step()
|
|
total_loss += loss.item() * images.size(0)
|
|
count += images.size(0)
|
|
return total_loss / max(count, 1)
|
|
|
|
|
|
@torch.no_grad()
|
|
def validate(model, loader, device):
|
|
model.eval()
|
|
total_loss = 0.0
|
|
count = 0
|
|
for batch in tqdm(loader, desc=" val ", leave=False):
|
|
images = batch["images"].to(device)
|
|
history = batch["action_history"].to(device)
|
|
target = batch["target"].to(device)
|
|
pred = model(images, history)
|
|
loss = torch.nn.functional.mse_loss(pred, target)
|
|
total_loss += loss.item() * images.size(0)
|
|
count += images.size(0)
|
|
if count == 0:
|
|
return float("inf") # 空验证集不参与比较，避免假 best
|
|
return total_loss / count
|
|
|
|
|
|
def main():
|
|
parser = build_argparser("actpred training")
|
|
args = parser.parse_args()
|
|
cfg = apply_overrides(Config(), args)
|
|
validate_config_values(cfg.data, cfg.model, cfg.train)
|
|
|
|
os.makedirs(cfg.train.checkpoint_dir, exist_ok=True)
|
|
|
|
# 全局随机种子，保证训练可复现
|
|
random.seed(cfg.train.seed)
|
|
np.random.seed(cfg.train.seed)
|
|
torch.manual_seed(cfg.train.seed)
|
|
if torch.cuda.is_available():
|
|
torch.cuda.manual_seed_all(cfg.train.seed)
|
|
|
|
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
|
print(f"Device: {device}")
|
|
|
|
train_indices, val_indices = split_episodes(
|
|
cfg.data.data_dir, cfg.train.val_split, seed=cfg.train.seed)
|
|
print(f"Episodes: {len(train_indices)} train, {len(val_indices)} val")
|
|
|
|
train_ds = ActionDataset(
|
|
cfg.data.data_dir, cfg.data.history_len, cfg.data.action_dim,
|
|
cfg.data.image_width, cfg.data.image_height, cfg.data.normalize_actions,
|
|
cfg.data.action_stats_path, train_indices, cfg.data.frame_len)
|
|
if cfg.data.normalize_actions and not cfg.data.action_stats_path:
|
|
# 归一化统计量必须只用训练集计算，再供验证集复用，否则 val loss 不可信
|
|
stats_path = os.path.join(cfg.train.checkpoint_dir, "action_stats.npy")
|
|
np.save(stats_path, train_ds.get_action_stats())
|
|
print(f"Action stats saved to: {stats_path}")
|
|
# 只存裸文件名：推理端按 checkpoint 目录解析相对路径
|
|
cfg.data.action_stats_path = "action_stats.npy"
|
|
val_ds = ActionDataset(
|
|
cfg.data.data_dir, cfg.data.history_len, cfg.data.action_dim,
|
|
cfg.data.image_width, cfg.data.image_height, cfg.data.normalize_actions,
|
|
cfg.data.action_stats_path, val_indices, cfg.data.frame_len)
|
|
|
|
print(f"Samples: {len(train_ds)} train, {len(val_ds)} val")
|
|
if len(train_ds) == 0:
|
|
raise ValueError(
|
|
f"训练集为空（0 样本）。请检查 data_dir={cfg.data.data_dir} "
|
|
f"及 episode 数据完整性。")
|
|
if len(val_ds) == 0:
|
|
print("WARNING: validation set is empty (episodes too short/few); "
|
|
"model selection is disabled, last epoch will be saved.")
|
|
|
|
loader_gen = torch.Generator()
|
|
loader_gen.manual_seed(cfg.train.seed)
|
|
train_loader = DataLoader(train_ds, batch_size=cfg.train.batch_size,
|
|
shuffle=True, num_workers=cfg.train.num_workers,
|
|
generator=loader_gen)
|
|
val_loader = DataLoader(val_ds, batch_size=cfg.train.batch_size,
|
|
shuffle=False, num_workers=cfg.train.num_workers)
|
|
|
|
model = ActionPredictor(cfg.model, cfg.data).to(device)
|
|
n_params = sum(p.numel() for p in model.parameters())
|
|
print(f"Model params: {n_params:,}")
|
|
|
|
if args.init_from:
|
|
init_ckpt = torch.load(args.init_from, map_location="cpu", weights_only=True)
|
|
missing, unexpected = model.load_state_dict(
|
|
init_ckpt["model_state_dict"], strict=False)
|
|
if missing:
|
|
print(f" init-from missing keys: {missing}")
|
|
if unexpected:
|
|
print(f" init-from unexpected keys: {unexpected}")
|
|
print(f"Loaded pretrained weights from: {args.init_from}")
|
|
|
|
optimizer = torch.optim.AdamW(model.parameters(), lr=cfg.train.lr,
|
|
weight_decay=cfg.train.weight_decay)
|
|
scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(optimizer, T_max=cfg.train.epochs)
|
|
|
|
best_val_loss = float("inf")
|
|
patience_counter = 0
|
|
start_epoch = 1
|
|
best_path = os.path.join(cfg.train.checkpoint_dir, "best.pt")
|
|
last_path = os.path.join(cfg.train.checkpoint_dir, "last.pt")
|
|
|
|
if args.resume:
|
|
resume_ckpt = torch.load(args.resume, map_location=device, weights_only=True)
|
|
model.load_state_dict(resume_ckpt["model_state_dict"])
|
|
optimizer.load_state_dict(resume_ckpt["optimizer_state_dict"])
|
|
scheduler.load_state_dict(resume_ckpt["scheduler_state_dict"])
|
|
start_epoch = resume_ckpt["epoch"] + 1
|
|
best_val_loss = resume_ckpt.get("best_val_loss", float("inf"))
|
|
patience_counter = resume_ckpt.get("patience_counter", 0)
|
|
print(f"Resumed from {args.resume} at epoch {start_epoch} "
|
|
f"(best_val_loss={best_val_loss:.6f})")
|
|
|
|
for epoch in range(start_epoch, cfg.train.epochs + 1):
|
|
print(f"\nEpoch {epoch}/{cfg.train.epochs} lr={optimizer.param_groups[0]['lr']:.6f}")
|
|
train_loss = train_one_epoch(model, train_loader, optimizer, device,
|
|
cfg.train.max_grad_norm)
|
|
val_loss = validate(model, val_loader, device)
|
|
scheduler.step()
|
|
print(f" train_loss={train_loss:.6f} val_loss={val_loss:.6f}")
|
|
|
|
if val_loss < best_val_loss:
|
|
best_val_loss = val_loss
|
|
patience_counter = 0
|
|
torch.save({
|
|
"epoch": epoch,
|
|
"model_state_dict": model.state_dict(),
|
|
"optimizer_state_dict": optimizer.state_dict(),
|
|
"val_loss": val_loss,
|
|
"config": cfg.to_dict(),
|
|
}, best_path)
|
|
inference_path = os.path.join(cfg.train.checkpoint_dir, "best_inference.pt")
|
|
torch.save({
|
|
"model_state_dict": model.state_dict(),
|
|
"config": cfg.to_dict(),
|
|
}, inference_path)
|
|
print(f" -> saved best (val_loss={val_loss:.6f})")
|
|
else:
|
|
patience_counter += 1
|
|
if patience_counter >= cfg.train.patience:
|
|
print(f"\nEarly stopping at epoch {epoch} (patience={cfg.train.patience})")
|
|
break
|
|
|
|
# 每 epoch 保存 last.pt，支持 --resume 断点续训
|
|
torch.save({
|
|
"epoch": epoch,
|
|
"model_state_dict": model.state_dict(),
|
|
"optimizer_state_dict": optimizer.state_dict(),
|
|
"scheduler_state_dict": scheduler.state_dict(),
|
|
"best_val_loss": best_val_loss,
|
|
"patience_counter": patience_counter,
|
|
"config": cfg.to_dict(),
|
|
}, last_path)
|
|
|
|
print(f"\nTraining done. Best val_loss={best_val_loss:.6f}")
|
|
|
|
if best_val_loss == float("inf"):
|
|
# 验证集为空时无 best 可选，保底保存最终权重
|
|
torch.save({
|
|
"model_state_dict": model.state_dict(),
|
|
"config": cfg.to_dict(),
|
|
}, os.path.join(cfg.train.checkpoint_dir, "best_inference.pt"))
|
|
print("WARNING: saved final weights as best_inference.pt (no valid validation loss)")
|
|
else:
|
|
print(f"Best model saved to: {best_path}")
|
|
|
|
if args.save_config:
|
|
with open(args.save_config, "w") as f:
|
|
json.dump(cfg.to_dict(), f, indent=2, ensure_ascii=False)
|
|
print(f"Config saved to: {args.save_config}")
|
|
|
|
|
|
if __name__ == "__main__":
|
|
main()
|