qing
d17677bade
- export_onnx: opset 17→18，dynamic_axes 改 dynamic_shapes 消除弃用警告 - predict: action_stats 改为 torch tensor，消除 NumPy 2.0 __array_wrap__ 警告 - train/pretrain: 每 epoch 保存 last.pt，新增 --resume 断点续训练 - model: MobileNetEncoder 下载失败时回退到已缓存权重 - _set.sh: 安装时预缓存 MobileNetV2 权重，失败给出明确指引 - test: 新增 TestResumeCheckpoint（2 用例），总计 58 测试全绿
316 lines
12 KiB
Python
316 lines
12 KiB
Python
"""actpred 预训练入口：从 .mt 脚本学习动作动力学（无需图像）
|
||
|
||
用法:
|
||
bash _run.sh pretrain --mt-dir ../bot4cj/scripts --epochs 200
|
||
|
||
流程:
|
||
1. 解析 .mt 文件中的 @Action 矩阵 → 动作序列
|
||
2. 滑动窗口采样: action_history[H] → target action（下一行）
|
||
3. 用零图像占位训练完整 ActionPredictor（保持 checkpoint 兼容性）
|
||
4. 保存 checkpoint，格式与 train.py 一致，可直接用于推理或微调
|
||
|
||
数据增强:
|
||
--noise-std 控制高斯噪声标准差。因 .mt 脚本数据量有限，
|
||
适度噪声（如 0.5~2.0）可有效防止过拟合。
|
||
"""
|
||
|
||
import json
|
||
import os
|
||
import random
|
||
|
||
import numpy as np
|
||
import torch
|
||
from torch.utils.data import Dataset, DataLoader
|
||
from tqdm import tqdm
|
||
|
||
from config import Config, build_argparser, apply_overrides, validate_config_values
|
||
from model import ActionPredictor
|
||
from mt_parser import parse_mt_directory
|
||
|
||
|
||
class MTActionDataset(Dataset):
|
||
"""从 .mt @Action 矩阵构建的滑动窗口数据集（无图像）
|
||
|
||
每个样本: (action_history[history_len, action_dim], target[action_dim])
|
||
"""
|
||
|
||
def __init__(self, mt_dir: str, history_len: int, action_dim: int,
|
||
file_indices: list = None, files: list = None):
|
||
self.history_len = history_len
|
||
self.action_dim = action_dim
|
||
self.samples = []
|
||
|
||
# 复用调用方已解析的结果（避免 train/val 各自重复解析同一目录）
|
||
all_files = files if files is not None else parse_mt_directory(mt_dir)
|
||
if file_indices is not None:
|
||
all_files = [all_files[i] for i in file_indices]
|
||
|
||
for path, matrix in all_files:
|
||
n_cols = matrix.shape[1]
|
||
if n_cols < action_dim:
|
||
matrix = np.pad(
|
||
matrix,
|
||
((0, 0), (0, action_dim - n_cols)),
|
||
mode="constant",
|
||
)
|
||
elif n_cols > action_dim:
|
||
matrix = matrix[:, :action_dim]
|
||
|
||
for i in range(len(matrix) - history_len):
|
||
history = matrix[i : i + history_len]
|
||
target = matrix[i + history_len]
|
||
self.samples.append((history.copy(), target.copy()))
|
||
|
||
print(f"MTActionDataset: {len(self.samples)} samples "
|
||
f"from {len(all_files)} files "
|
||
f"(history_len={history_len}, action_dim={action_dim})")
|
||
|
||
def __len__(self):
|
||
return len(self.samples)
|
||
|
||
def __getitem__(self, idx):
|
||
history, target = self.samples[idx]
|
||
return {
|
||
"action_history": torch.from_numpy(history.copy()),
|
||
"target": torch.from_numpy(target.copy()),
|
||
}
|
||
|
||
|
||
def pretrain_one_epoch(model, loader, optimizer, device, image_size, frame_len,
|
||
max_grad_norm=None, noise_std=0.0):
|
||
model.train()
|
||
total_loss = 0.0
|
||
count = 0
|
||
zero_images = None
|
||
for batch in tqdm(loader, desc=" pretrain", leave=False):
|
||
history = batch["action_history"].to(device)
|
||
target = batch["target"].to(device)
|
||
if noise_std > 0:
|
||
history = history + torch.randn_like(history) * noise_std
|
||
target = target + torch.randn_like(target) * noise_std
|
||
if zero_images is None or zero_images.size(0) != history.size(0):
|
||
zero_images = torch.zeros(
|
||
history.size(0), frame_len, 3, image_size, image_size,
|
||
device=device,
|
||
)
|
||
optimizer.zero_grad()
|
||
pred = model(zero_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() * history.size(0)
|
||
count += history.size(0)
|
||
return total_loss / max(count, 1)
|
||
|
||
|
||
@torch.no_grad()
|
||
def validate(model, loader, device, image_size, frame_len):
|
||
model.eval()
|
||
total_loss = 0.0
|
||
count = 0
|
||
zero_images = None
|
||
for batch in tqdm(loader, desc=" val ", leave=False):
|
||
history = batch["action_history"].to(device)
|
||
target = batch["target"].to(device)
|
||
if zero_images is None or zero_images.size(0) != history.size(0):
|
||
zero_images = torch.zeros(
|
||
history.size(0), frame_len, 3, image_size, image_size,
|
||
device=device,
|
||
)
|
||
pred = model(zero_images, history)
|
||
loss = torch.nn.functional.mse_loss(pred, target)
|
||
total_loss += loss.item() * history.size(0)
|
||
count += history.size(0)
|
||
if count == 0:
|
||
return float("inf")
|
||
return total_loss / count
|
||
|
||
|
||
def main():
|
||
parser = build_argparser("actpred pre-training from .mt scripts")
|
||
parser.add_argument("--mt-dir", type=str, required=True,
|
||
help=".mt 脚本目录（如 ../bot4cj/scripts）")
|
||
parser.add_argument("--noise-std", type=float, default=0.0,
|
||
help="高斯噪声标准差（数据增强，建议 0.5~2.0）")
|
||
args = parser.parse_args()
|
||
|
||
cfg = apply_overrides(Config(), args)
|
||
validate_config_values(cfg.data, cfg.model, cfg.train)
|
||
cfg.data.data_dir = args.mt_dir
|
||
|
||
os.makedirs(cfg.train.checkpoint_dir, exist_ok=True)
|
||
|
||
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}")
|
||
|
||
# 按文件级划分 train/val，避免滑动窗口样本重叠导致的数据泄漏
|
||
all_files = parse_mt_directory(args.mt_dir)
|
||
if len(all_files) == 0:
|
||
raise ValueError(
|
||
f"数据集为空。请检查 mt-dir={args.mt_dir} "
|
||
f"及 .mt 文件中 @Action 行数 > history_len={cfg.data.history_len}"
|
||
)
|
||
|
||
n_files = len(all_files)
|
||
file_indices = list(range(n_files))
|
||
rng = np.random.RandomState(cfg.train.seed)
|
||
rng.shuffle(file_indices)
|
||
val_file_count = max(1, int(n_files * cfg.train.val_split))
|
||
val_file_indices = file_indices[:val_file_count]
|
||
train_file_indices = file_indices[val_file_count:]
|
||
|
||
train_ds = MTActionDataset(
|
||
args.mt_dir, cfg.data.history_len, cfg.data.action_dim,
|
||
file_indices=train_file_indices, files=all_files,
|
||
)
|
||
val_ds = MTActionDataset(
|
||
args.mt_dir, cfg.data.history_len, cfg.data.action_dim,
|
||
file_indices=val_file_indices, files=all_files,
|
||
)
|
||
|
||
print(f"Files: {len(train_file_indices)} train, {len(val_file_indices)} val")
|
||
print(f"Samples: {len(train_ds)} train, {len(val_ds)} val")
|
||
if len(train_ds) == 0:
|
||
raise ValueError("训练集为空（样本数不足）")
|
||
if len(val_ds) == 0:
|
||
print("WARNING: 验证集为空，模型选择禁用")
|
||
|
||
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, pretrained=False).to(device)
|
||
n_params = sum(p.numel() for p in model.parameters())
|
||
print(f"Model params: {n_params:,}")
|
||
print(f"Config: action_dim={cfg.data.action_dim}, "
|
||
f"history_len={cfg.data.history_len}, "
|
||
f"frame_len={cfg.data.frame_len}, "
|
||
f"backbone={cfg.model.cnn_backbone}, "
|
||
f"attn={cfg.model.attn_layers}x{cfg.model.attn_heads}h"
|
||
f"@{cfg.model.attn_dim}")
|
||
print(f"Pre-training: noise_std={args.noise_std}")
|
||
|
||
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, "pretrain_best.pt")
|
||
inference_path = os.path.join(
|
||
cfg.train.checkpoint_dir, "pretrain_best_inference.pt"
|
||
)
|
||
last_path = os.path.join(cfg.train.checkpoint_dir, "pretrain_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} "
|
||
f"lr={optimizer.param_groups[0]['lr']:.6f}")
|
||
train_loss = pretrain_one_epoch(
|
||
model, train_loader, optimizer, device,
|
||
cfg.data.image_size, cfg.data.frame_len,
|
||
cfg.train.max_grad_norm, args.noise_std,
|
||
)
|
||
val_loss = validate(model, val_loader, device,
|
||
cfg.data.image_size, cfg.data.frame_len)
|
||
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,
|
||
)
|
||
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} "
|
||
f"(patience={cfg.train.patience})")
|
||
break
|
||
|
||
# 每 epoch 保存 last.pt，支持 --resume 断点续训练
|
||
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"\nPre-training done. Best val_loss={best_val_loss:.6f}")
|
||
|
||
if best_val_loss == float("inf"):
|
||
torch.save(
|
||
{
|
||
"model_state_dict": model.state_dict(),
|
||
"config": cfg.to_dict(),
|
||
},
|
||
inference_path,
|
||
)
|
||
print("WARNING: saved final weights (no valid validation loss)")
|
||
else:
|
||
print(f"Best pre-trained model: {inference_path}")
|
||
print("Fine-tune: bash _run.sh train --data-dir ./data "
|
||
f"--init-from {inference_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()
|