model/train.py

#!/usr/bin/env python3
"""
Unified Training Script – YOLOv11 + CNN-BiGRU
Based on: Nature Scientific Reports (Nov 2025)

Usage:
    # Train YOLOv11 detector only
    python train.py yolo --data dataset/data.yaml --epochs 100

    # Train CNN-BiGRU severity model (requires sequence data)
    python train.py bigru --data severity_sequences/ --epochs 50

    # Train both sequentially
    python train.py all --data dataset/data.yaml --bigru-data severity_sequences/
"""

import os
import sys
import shutil
import logging
import argparse
from pathlib import Path
from datetime import datetime

import torch
import yaml

logging.basicConfig(
    level=logging.INFO,
    format="%(asctime)s [%(levelname)s] %(message)s",
    handlers=[logging.StreamHandler(), logging.FileHandler("training.log")],
)
logger = logging.getLogger("train")


# ═══════════════════════════════════════════════════════════════════════════
# YOLOv11 Training
# ═══════════════════════════════════════════════════════════════════════════
def train_yolo(args):
    from yolo_detection import YOLOv11Detector

    logger.info("=" * 60)
    logger.info("  YOLOv11 Road Anomaly Detection – Training")
    logger.info("=" * 60)

    # GPU info
    if torch.cuda.is_available():
        name = torch.cuda.get_device_properties(0).name
        vram = torch.cuda.get_device_properties(0).total_memory / (1024 ** 3)
        logger.info("GPU: %s  (%.1f GB)", name, vram)
    else:
        logger.info("Training on CPU (this will be slow)")

    # Resolve data.yaml
    data_yaml = str(Path(args.data).resolve())
    logger.info("Dataset config: %s", data_yaml)

    # Determine batch size from VRAM
    batch = args.batch
    if batch == 0:
        # Auto-select based on GPU VRAM
        #   RTX 2050 (4 GB) → batch 4 @ 416px
        #   RTX 3060 (8 GB) → batch 8
        #   RTX 3090+ (20+ GB) → batch 16
        if torch.cuda.is_available():
            vram_gb = torch.cuda.get_device_properties(0).total_memory / (1024 ** 3)
            if vram_gb >= 20:
                batch = 16
            elif vram_gb >= 8:
                batch = 8
            else:
                batch = 4
        else:
            batch = 2
        logger.info("Auto batch size: %d  (VRAM=%.1f GB)", batch,
                     vram_gb if torch.cuda.is_available() else 0)

    detector = YOLOv11Detector(
        model_path=args.model,
        img_size=args.imgsz,
    )

    results = detector.train(
        data_yaml=data_yaml,
        epochs=args.epochs,
        batch=batch,
        optimizer=args.optimizer,
        lr0=args.lr,
        weight_decay=args.weight_decay,
        warmup_epochs=args.warmup,
        mosaic=0.5,
        cache=args.cache,
        amp=not args.no_amp,
        workers=args.workers,
        project=args.project,
        name=args.name,
        resume=args.resume,
    )

    # Copy best.pt to project root for easy access
    best_src = Path(args.project) / args.name / "weights" / "best.pt"
    if best_src.exists():
        best_dst = Path("runs/best.pt")
        best_dst.parent.mkdir(parents=True, exist_ok=True)
        shutil.copy2(best_src, best_dst)
        logger.info("✅ Best model → %s", best_dst)

    # Export
    if args.export:
        for fmt in args.export:
            try:
                detector.export(format=fmt, half=(fmt == "engine"))
                logger.info("✅ Exported → %s", fmt)
            except Exception as e:
                logger.warning("Export %s failed: %s", fmt, e)

    return results


# ═══════════════════════════════════════════════════════════════════════════
# CNN-BiGRU Training
# ═══════════════════════════════════════════════════════════════════════════
def train_bigru(args):
    from cnn_bigru import CNNBiGRU, AnomalySequenceDataset, BiGRUTrainer
    from torch.utils.data import DataLoader, random_split

    logger.info("=" * 60)
    logger.info("  CNN-BiGRU Severity Prediction – Training")
    logger.info("=" * 60)

    # Load dataset
    dataset = AnomalySequenceDataset(
        root=args.bigru_data,
        seq_len=args.seq_len,
        patch_size=64,
    )

    # Split 80/20
    n_val = max(1, int(len(dataset) * 0.2))
    n_train = len(dataset) - n_val
    train_ds, val_ds = random_split(dataset, [n_train, n_val])

    train_loader = DataLoader(
        train_ds, batch_size=args.bigru_batch, shuffle=True,
        num_workers=args.workers, pin_memory=True,
    )
    val_loader = DataLoader(
        val_ds, batch_size=args.bigru_batch, shuffle=False,
        num_workers=args.workers, pin_memory=True,
    )

    logger.info("Train sequences: %d  |  Val sequences: %d", n_train, n_val)

    # Create model
    model = CNNBiGRU(
        in_channels=3,
        hidden_size=128,
        num_gru_layers=2,
        num_severity_classes=4,
    )

    trainer = BiGRUTrainer(
        model=model,
        lr=args.bigru_lr,
        weight_decay=1e-4,
    )

    history = trainer.fit(
        train_loader=train_loader,
        val_loader=val_loader,
        epochs=args.bigru_epochs,
        save_dir=args.bigru_save_dir,
        patience=args.patience,
    )

    # Copy best to project root
    best_src = Path(args.bigru_save_dir) / "best_bigru.pth"
    if best_src.exists():
        best_dst = Path("runs/best_bigru.pth")
        shutil.copy2(best_src, best_dst)
        logger.info("✅ Best BiGRU → %s", best_dst)

    return history


# ═══════════════════════════════════════════════════════════════════════════
# CLI
# ═══════════════════════════════════════════════════════════════════════════
def build_parser():
    parser = argparse.ArgumentParser(
        description="Train YOLOv11 + CNN-BiGRU Road Anomaly Detection System",
    )
    sub = parser.add_subparsers(dest="mode", required=True)

    # ---- yolo ----
    p_yolo = sub.add_parser("yolo", help="Train YOLOv11 detector")
    p_yolo.add_argument("--data", required=True, help="data.yaml path")
    p_yolo.add_argument("--model", default="yolo11n.pt",
                        help="Base model (yolo11n/s/m/l/x.pt)")
    p_yolo.add_argument("--epochs", type=int, default=100)
    p_yolo.add_argument("--batch", type=int, default=0,
                        help="Batch size (0 = auto from VRAM)")
    p_yolo.add_argument("--imgsz", type=int, default=416)
    p_yolo.add_argument("--optimizer", default="AdamW")
    p_yolo.add_argument("--lr", type=float, default=0.001)
    p_yolo.add_argument("--weight-decay", type=float, default=0.0005)
    p_yolo.add_argument("--warmup", type=float, default=3.0)
    p_yolo.add_argument("--cache", default="disk",
                        help="'ram', 'disk', or '' for none")
    p_yolo.add_argument("--no-amp", action="store_true")
    p_yolo.add_argument("--workers", type=int, default=4)
    p_yolo.add_argument("--project", default="road_anomaly")
    p_yolo.add_argument("--name", default="yolov11_road_detection")
    p_yolo.add_argument("--resume", action="store_true")
    p_yolo.add_argument("--export", nargs="*", default=[],
                        help="Export formats after training (onnx, engine, tflite)")

    # ---- bigru ----
    p_bigru = sub.add_parser("bigru", help="Train CNN-BiGRU severity model")
    p_bigru.add_argument("--bigru-data", required=True,
                         help="Root dir with sequences/ + labels.csv")
    p_bigru.add_argument("--seq-len", type=int, default=8)
    p_bigru.add_argument("--bigru-batch", type=int, default=8)
    p_bigru.add_argument("--bigru-epochs", type=int, default=50)
    p_bigru.add_argument("--bigru-lr", type=float, default=1e-3)
    p_bigru.add_argument("--bigru-save-dir", default="bigru_checkpoints")
    p_bigru.add_argument("--patience", type=int, default=10)
    p_bigru.add_argument("--workers", type=int, default=4)

    # ---- all ----
    p_all = sub.add_parser("all", help="Train YOLO then BiGRU")
    # Inherit all args from both
    p_all.add_argument("--data", required=True)
    p_all.add_argument("--model", default="yolo11n.pt")
    p_all.add_argument("--epochs", type=int, default=100)
    p_all.add_argument("--batch", type=int, default=0)
    p_all.add_argument("--imgsz", type=int, default=416)
    p_all.add_argument("--optimizer", default="AdamW")
    p_all.add_argument("--lr", type=float, default=0.001)
    p_all.add_argument("--weight-decay", type=float, default=0.0005)
    p_all.add_argument("--warmup", type=float, default=3.0)
    p_all.add_argument("--cache", default="disk")
    p_all.add_argument("--no-amp", action="store_true")
    p_all.add_argument("--workers", type=int, default=4)
    p_all.add_argument("--project", default="road_anomaly")
    p_all.add_argument("--name", default="yolov11_road_detection")
    p_all.add_argument("--resume", action="store_true")
    p_all.add_argument("--export", nargs="*", default=[])
    p_all.add_argument("--bigru-data", default=None)
    p_all.add_argument("--seq-len", type=int, default=8)
    p_all.add_argument("--bigru-batch", type=int, default=8)
    p_all.add_argument("--bigru-epochs", type=int, default=50)
    p_all.add_argument("--bigru-lr", type=float, default=1e-3)
    p_all.add_argument("--bigru-save-dir", default="bigru_checkpoints")
    p_all.add_argument("--patience", type=int, default=10)

    return parser


def main():
    parser = build_parser()
    args = parser.parse_args()

    print()
    print("🚗  ROAD ANOMALY DETECTION – YOLOv11 + CNN-BiGRU")
    print("    Based on Nature Scientific Reports (Nov 2025)")
    print(f"    Started: {datetime.now():%Y-%m-%d %H:%M:%S}")
    print()

    if args.mode == "yolo":
        train_yolo(args)

    elif args.mode == "bigru":
        train_bigru(args)

    elif args.mode == "all":
        # Phase 1: YOLO
        logger.info("═══ Phase 1/2: YOLOv11 Training ═══")
        train_yolo(args)

        # Phase 2: BiGRU (if data provided)
        if args.bigru_data:
            logger.info("═══ Phase 2/2: CNN-BiGRU Training ═══")
            train_bigru(args)
        else:
            logger.info(
                "Skipping BiGRU training – provide --bigru-data to enable."
            )

    print()
    print("🎯  Training pipeline complete!")
    print(f"    Finished: {datetime.now():%Y-%m-%d %H:%M:%S}")
    print()


if __name__ == "__main__":
    main()