Compressor/backend/LabelingPreTraining/tspulse/train.py

"""Step 1: unsupervised TSPulse training (masked reconstruction).

Trains a single TSPulse encoder on all normal cycle samples from every part,
then freezes the weights into a single checkpoint file used later for
inference. Progress is printed via tqdm (epochs and batches with live loss).

Usage:
    python3 tspulse/train.py [--epochs 30] [--batch-size 64] [--lr 1e-3]
"""

import argparse
import json
import random
from pathlib import Path

import numpy as np
import torch
from torch.utils.data import DataLoader
from tqdm import tqdm

from tspulse import CycleDataset, TSPulse, build_split_datasets, describe, get_device, make_random_mask
from tspulse.dataset import N_CHANNELS, SEQ_LEN

ROOT = Path(__file__).resolve().parents[1]
CHECKPOINT_DIR = ROOT / "checkpoints"
CHECKPOINT_PATH = CHECKPOINT_DIR / "tspulse_frozen.pt"


def set_seed(seed: int) -> None:
    random.seed(seed)
    np.random.seed(seed)
    torch.manual_seed(seed)


def masked_mse(recon: torch.Tensor, target: torch.Tensor, mask: torch.BoolTensor, patch_size: int) -> torch.Tensor:
    sample_mask = mask.repeat_interleave(patch_size, dim=1)  # (B, L)
    diff = (recon - target) ** 2
    masked_diff = diff[sample_mask]
    if masked_diff.numel() == 0:
        return torch.tensor(0.0, device=recon.device)
    return masked_diff.mean()


def main() -> int:
    parser = argparse.ArgumentParser()
    parser.add_argument("--epochs", type=int, default=30)
    parser.add_argument("--batch-size", type=int, default=64)
    parser.add_argument("--lr", type=float, default=1e-3)
    parser.add_argument("--mask-ratio", type=float, default=0.15)
    parser.add_argument("--dim", type=int, default=128)
    parser.add_argument("--depth", type=int, default=4)
    parser.add_argument("--heads", type=int, default=4)
    parser.add_argument("--seed", type=int, default=42)
    args = parser.parse_args()

    set_seed(args.seed)
    device = get_device()
    print(f"设备: {describe(device)}")

    samples, mean, std = build_split_datasets()
    train_samples = [item for item in samples if item["split"] == "train"]
    val_samples = [item for item in samples if item["split"] == "val"]
    print(f"样本: 训练 {len(train_samples)}, 验证 {len(val_samples)}, 评估 {len(samples) - len(train_samples) - len(val_samples)}")

    train_set = CycleDataset(train_samples, mean, std)
    val_set = CycleDataset(val_samples, mean, std)
    train_loader = DataLoader(train_set, batch_size=args.batch_size, shuffle=True, drop_last=True)
    val_loader = DataLoader(val_set, batch_size=args.batch_size, shuffle=False)

    model = TSPulse(dim=args.dim, depth=args.depth, heads=args.heads).to(device)
    optimizer = torch.optim.AdamW(model.parameters(), lr=args.lr, weight_decay=0.05)
    scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(optimizer, T_max=args.epochs)

    config = {
        "dim": args.dim,
        "depth": args.depth,
        "heads": args.heads,
        "seq_len": SEQ_LEN,
        "n_channels": N_CHANNELS,
        "mask_ratio": args.mask_ratio,
    }

    best_val = float("inf")
    CHECKPOINT_DIR.mkdir(parents=True, exist_ok=True)
    history: list[dict] = []

    for epoch in range(1, args.epochs + 1):
        model.train()
        epoch_loss = 0.0
        seen = 0
        bar = tqdm(train_loader, desc=f"epoch {epoch}/{args.epochs}", leave=False)
        for batch in bar:
            x, _lengths = batch
            x = x.to(device)
            mask = make_random_mask(x.shape[0], model.num_patches, args.mask_ratio, device)
            _, recon = model(x, mask)
            loss = masked_mse(recon, x, mask, model.patch_size)
            optimizer.zero_grad()
            loss.backward()
            optimizer.step()
            epoch_loss += loss.item() * x.shape[0]
            seen += x.shape[0]
            bar.set_postfix(loss=f"{loss.item():.5f}")
        scheduler.step()
        train_loss = epoch_loss / max(seen, 1)

        model.eval()
        val_loss = 0.0
        val_seen = 0
        with torch.no_grad():
            for batch in val_loader:
                x, _lengths = batch
                x = x.to(device)
                mask = make_random_mask(x.shape[0], model.num_patches, args.mask_ratio, device)
                _, recon = model(x, mask)
                val_loss += masked_mse(recon, x, mask, model.patch_size).item() * x.shape[0]
                val_seen += x.shape[0]
        val_loss /= max(val_seen, 1)

        history.append({"epoch": epoch, "train_loss": train_loss, "val_loss": val_loss})
        print(f"epoch {epoch:>3}/{args.epochs}  训练 loss {train_loss:.5f}  验证 loss {val_loss:.5f}")

        if val_loss < best_val:
            best_val = val_loss
            torch.save(
                {
                    "state_dict": model.state_dict(),
                    "config": config,
                    "mean": mean,
                    "std": std,
                    "parts": sorted({item["part"] for item in samples}),
                    "history": history,
                    "device": describe(device),
                },
                CHECKPOINT_PATH,
            )
            print(f"  -> 已保存最佳模型 {CHECKPOINT_PATH}")

    print(f"训练完成: 最佳验证 loss {best_val:.5f}, 模型固化于 {CHECKPOINT_PATH}")
    return 0


if __name__ == "__main__":
    raise SystemExit(main())