train_hybrid.py

"""

Training Script for TransMIL + Query2Label Hybrid Model



Supports:

- End-to-end training with ResNet-50 backbone

- Mixed precision training (AMP) for memory efficiency

- Gradient accumulation for larger effective batch size

- Gradient checkpointing for ResNet-50

- AsymmetricLoss for multi-label imbalance

- Multi-label evaluation metrics (mAP, per-class AP, F1)

"""

import sys
#sys.path.append('query2labels/lib/models')
#sys.path.append('/XYFS01/HDD_POOL/sysu_gbli2/sysu_gbli2xy_1/chenshiyu/ThyroidAgent/ThyroidRegion/HintsVer3/query2labels/lib/')

import os
import argparse
import yaml
from pathlib import Path
from datetime import datetime
import json

import torch
import torch.nn as nn
import torch.optim as optim
from torch.cuda.amp import autocast, GradScaler
from torch.utils.tensorboard import SummaryWriter
import numpy as np
from tqdm import tqdm
from sklearn.metrics import average_precision_score, f1_score

# Import model and dataset
from models.transmil_q2l import TransMIL_Query2Label_E2E
from thyroid_dataset import create_dataloaders

# Import AsymmetricLoss
try:
    from models.aslloss import AsymmetricLossOptimized
except ImportError:
    print("Warning: Could not import AsymmetricLoss.")
    AsymmetricLossOptimized = None
'''

try:

    #from aslloss import AsymmetricLossOptimized

    from models.aslloss import AsymmetricLossOptimized

except ImportError:

    print("Warning: Could not import AsymmetricLoss from query2labels.")

    print("Make sure query2labels/lib/models/aslloss.py is in Python path.")

    AsymmetricLossOptimized = None



'''
# ============================================================================
# Metrics
# ============================================================================

def compute_multilabel_metrics(preds, targets, threshold=0.5):
    """

    Compute multi-label classification metrics.



    Args:

        preds: [N, num_class] numpy array of probabilities

        targets: [N, num_class] numpy array of binary labels

        threshold: Classification threshold for F1 score



    Returns:

        dict with mAP, per-class AP, F1 scores

    """
    metrics = {}

    # Mean Average Precision (mAP)
    aps = []
    for i in range(targets.shape[1]):
        if targets[:, i].sum() > 0:  # Skip classes with no positive samples
            ap = average_precision_score(targets[:, i], preds[:, i])
            aps.append(ap)
        else:
            aps.append(np.nan)

    metrics['mAP'] = np.nanmean(aps)
    metrics['per_class_AP'] = aps

    # F1 Score at threshold
    preds_binary = (preds >= threshold).astype(int)
    f1_micro = f1_score(targets, preds_binary, average='micro', zero_division=0)
    f1_macro = f1_score(targets, preds_binary, average='macro', zero_division=0)
    f1_samples = f1_score(targets, preds_binary, average='samples', zero_division=0)

    metrics['F1_micro'] = f1_micro
    metrics['F1_macro'] = f1_macro
    metrics['F1_samples'] = f1_samples

    return metrics


# ============================================================================
# Training Functions
# ============================================================================

def train_epoch(model, dataloader, criterion, optimizer, scaler, device, config, epoch):
    """

    Train for one epoch with gradient accumulation and mixed precision.



    Args:

        model: TransMIL_Query2Label_E2E model

        dataloader: Training dataloader

        criterion: AsymmetricLoss

        optimizer: AdamW optimizer

        scaler: GradScaler for AMP

        device: torch.device

        config: Config dict

        epoch: Current epoch number



    Returns:

        Average loss for epoch

    """
    model.train()

    total_loss = 0.0
    accumulation_steps = config['training']['gradient_accumulation_steps']
    use_amp = config['training']['use_amp']

    # Progress bar
    pbar = tqdm(dataloader, desc=f"Epoch {epoch}")

    optimizer.zero_grad()

    for i, batch in enumerate(pbar):
        images = batch['images'].to(device)  # [B*N_total, 3, H, W]
        labels = batch['labels'].to(device)  # [B, num_class]
        num_instances_per_case = batch['num_instances_per_case']  # [B]

        # Mixed precision forward pass
        if use_amp:
            with autocast():
                logits = model(images, num_instances_per_case)
                loss = criterion(logits, labels)
                loss = loss / accumulation_steps  # Scale loss for accumulation
        else:
            logits = model(images, num_instances_per_case)
            loss = criterion(logits, labels)
            loss = loss / accumulation_steps

        # Backward pass
        if use_amp:
            scaler.scale(loss).backward()
        else:
            loss.backward()

        # Optimizer step every accumulation_steps
        if (i + 1) % accumulation_steps == 0:
            if use_amp:
                scaler.step(optimizer)
                scaler.update()
            else:
                optimizer.step()
            optimizer.zero_grad()

        # Track loss
        total_loss += loss.item() * accumulation_steps
        pbar.set_postfix({'loss': loss.item() * accumulation_steps})

    return total_loss / len(dataloader)


@torch.no_grad()
def validate(model, dataloader, criterion, device, config):
    """

    Validate model with multi-label metrics.



    Args:

        model: TransMIL_Query2Label_E2E model

        dataloader: Validation dataloader

        criterion: AsymmetricLoss

        device: torch.device

        config: Config dict



    Returns:

        dict with loss and metrics (mAP, F1, etc.)

    """
    model.eval()

    total_loss = 0.0
    all_preds = []
    all_targets = []

    for batch in tqdm(dataloader, desc="Validating"):
        images = batch['images'].to(device)
        labels = batch['labels'].to(device)
        num_instances_per_case = batch['num_instances_per_case']

        # Forward pass
        logits = model(images, num_instances_per_case)
        loss = criterion(logits, labels)

        # Sigmoid for multi-label probabilities
        preds = torch.sigmoid(logits)

        # Store predictions and targets
        all_preds.append(preds.cpu().numpy())
        all_targets.append(labels.cpu().numpy())

        total_loss += loss.item()

    # Concatenate all batches
    all_preds = np.concatenate(all_preds, axis=0)
    all_targets = np.concatenate(all_targets, axis=0)

    # Compute metrics
    metrics = compute_multilabel_metrics(all_preds, all_targets)
    metrics['loss'] = total_loss / len(dataloader)

    return metrics


# ============================================================================
# Main Training Loop
# ============================================================================

def train(config, resume_from=None):
    """

    Main training function.



    Args:

        config: Config dictionary from YAML

        resume_from: Optional checkpoint path to resume training

    """
    # Setup device
    device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
    print(f"\nUsing device: {device}")
    if torch.cuda.is_available():
        print(f"GPU: {torch.cuda.get_device_name(0)}")
        print(f"Memory: {torch.cuda.get_device_properties(0).total_memory / 1e9:.2f} GB")

    # Create save directory
    save_dir = Path(config['training']['save_dir'])
    save_dir.mkdir(parents=True, exist_ok=True)

    # Create tensorboard writer
    log_dir = save_dir / 'logs' / datetime.now().strftime('%Y%m%d_%H%M%S')
    writer = SummaryWriter(log_dir)

    # Save config
    with open(save_dir / 'config.yaml', 'w') as f:
        yaml.dump(config, f)

    # Create dataloaders
    print("\nCreating dataloaders...")
    train_loader, val_loader, test_loader = create_dataloaders(config)

    # Create model
    print("\nCreating model...")
    model = TransMIL_Query2Label_E2E(
        num_class=config['model']['num_class'],
        hidden_dim=config['model']['hidden_dim'],
        nheads=config['model']['nheads'],
        num_decoder_layers=config['model']['num_decoder_layers'],
        pretrained_resnet=config['model']['pretrained_resnet'],
        use_checkpointing=config['training']['gradient_checkpointing'],
        use_ppeg=config['model'].get('use_ppeg', False)
    )
    model = model.to(device)

    # Print model stats
    total_params = sum(p.numel() for p in model.parameters())
    trainable_params = sum(p.numel() for p in model.parameters() if p.requires_grad)
    print(f"Total parameters: {total_params:,}")
    print(f"Trainable parameters: {trainable_params:,}")

    # Create optimizer
    optimizer = optim.AdamW(
        model.parameters(),
        lr=config['training']['lr'],
        weight_decay=config['training']['weight_decay']
    )

    # Create scheduler
    scheduler_type = config['training'].get('scheduler', 'cosine')
    if scheduler_type == 'cosine':
        scheduler = optim.lr_scheduler.CosineAnnealingLR(
            optimizer,
            T_max=config['training']['epochs'],
            eta_min=1e-6
        )
    elif scheduler_type == 'onecycle':
        scheduler = optim.lr_scheduler.OneCycleLR(
            optimizer,
            max_lr=config['training']['lr'],
            epochs=config['training']['epochs'],
            steps_per_epoch=len(train_loader)
        )
    else:
        scheduler = None

    # Create loss function
    if AsymmetricLossOptimized is not None:
        criterion = AsymmetricLossOptimized(
            gamma_neg=config['training']['gamma_neg'],
            gamma_pos=config['training']['gamma_pos'],
            clip=config['training']['clip'],
            eps=1e-5
        )
    else:
        # Fallback to BCEWithLogitsLoss
        print("Warning: Using BCEWithLogitsLoss instead of AsymmetricLoss")
        criterion = nn.BCEWithLogitsLoss()

    # Mixed precision scaler
    scaler = GradScaler() if config['training']['use_amp'] else None

    # Resume from checkpoint if specified
    start_epoch = 0
    best_map = 0.0

    if resume_from is not None and Path(resume_from).exists():
        print(f"\nResuming from {resume_from}")
        checkpoint = torch.load(resume_from, map_location=device)
        model.load_state_dict(checkpoint['model_state_dict'])
        optimizer.load_state_dict(checkpoint['optimizer_state_dict'])
        start_epoch = checkpoint['epoch'] + 1
        best_map = checkpoint.get('best_map', 0.0)
        if scheduler is not None and 'scheduler_state_dict' in checkpoint:
            scheduler.load_state_dict(checkpoint['scheduler_state_dict'])
        print(f"Resumed from epoch {start_epoch}, best mAP: {best_map:.4f}")

    # Training loop
    print(f"\nStarting training for {config['training']['epochs']} epochs...")
    print("="*80)

    for epoch in range(start_epoch, config['training']['epochs']):
        # Train
        train_loss = train_epoch(model, train_loader, criterion, optimizer, scaler, device, config, epoch)

        # Validate
        val_metrics = validate(model, val_loader, criterion, device, config)

        # Update scheduler
        if scheduler is not None:
            if scheduler_type == 'onecycle':
                pass  # OneCycleLR updates per step, not per epoch
            else:
                scheduler.step()

        # Log metrics
        current_lr = optimizer.param_groups[0]['lr']
        writer.add_scalar('Loss/train', train_loss, epoch)
        writer.add_scalar('Loss/val', val_metrics['loss'], epoch)
        writer.add_scalar('Metrics/mAP', val_metrics['mAP'], epoch)
        writer.add_scalar('Metrics/F1_micro', val_metrics['F1_micro'], epoch)
        writer.add_scalar('Metrics/F1_macro', val_metrics['F1_macro'], epoch)
        writer.add_scalar('LR', current_lr, epoch)

        # Print epoch summary
        print(f"\nEpoch {epoch}/{config['training']['epochs']}")
        print(f"  Train Loss: {train_loss:.4f}")
        print(f"  Val Loss: {val_metrics['loss']:.4f}")
        print(f"  mAP: {val_metrics['mAP']:.4f}")
        print(f"  F1 (micro): {val_metrics['F1_micro']:.4f}")
        print(f"  F1 (macro): {val_metrics['F1_macro']:.4f}")
        print(f"  LR: {current_lr:.6f}")

        # Save checkpoint
        is_best = val_metrics['mAP'] > best_map
        if is_best:
            best_map = val_metrics['mAP']

        if (epoch + 1) % config['training']['save_freq'] == 0 or is_best:
            checkpoint = {
                'epoch': epoch,
                'model_state_dict': model.state_dict(),
                'optimizer_state_dict': optimizer.state_dict(),
                'scheduler_state_dict': scheduler.state_dict() if scheduler is not None else None,
                'train_loss': train_loss,
                'val_metrics': val_metrics,
                'best_map': best_map,
                'config': config
            }

            # Save latest checkpoint
            torch.save(checkpoint, save_dir / 'checkpoint_latest.pth')

            # Save best checkpoint
            if is_best:
                torch.save(checkpoint, save_dir / 'checkpoint_best.pth')
                print(f"  ✓ Saved best model (mAP: {best_map:.4f})")

            # Save periodic checkpoint
            if (epoch + 1) % config['training']['save_freq'] == 0:
                torch.save(checkpoint, save_dir / f'checkpoint_epoch_{epoch}.pth')

    print("\n" + "="*80)
    print(f"Training completed! Best mAP: {best_map:.4f}")
    print(f"Checkpoints saved to: {save_dir}")

    writer.close()

    # Final test evaluation
    print("\nEvaluating on test set...")
    test_metrics = validate(model, test_loader, criterion, device, config)
    print(f"\nTest Results:")
    print(f"  mAP: {test_metrics['mAP']:.4f}")
    print(f"  F1 (micro): {test_metrics['F1_micro']:.4f}")
    print(f"  F1 (macro): {test_metrics['F1_macro']:.4f}")

    # Save test results
    with open(save_dir / 'test_results.json', 'w') as f:
        json.dump({k: float(v) if not isinstance(v, list) else v
                   for k, v in test_metrics.items()}, f, indent=2)


# ============================================================================
# Main
# ============================================================================

def main():
    parser = argparse.ArgumentParser(description='Train TransMIL + Query2Label Hybrid Model')
    parser.add_argument('--config', type=str, default='hybrid_model/config.yaml',
                        help='Path to config file')
    parser.add_argument('--resume', type=str, default=None,
                        help='Path to checkpoint to resume from')
    args = parser.parse_args()

    # Load config
    with open(args.config, 'r') as f:
        config = yaml.safe_load(f)

    print("="*80)
    print("TransMIL + Query2Label Hybrid Model Training")
    print("="*80)
    print(f"\nConfig: {args.config}")
    if args.resume:
        print(f"Resume from: {args.resume}")

    # Train
    train(config, resume_from=args.resume)


if __name__ == "__main__":
    main()