Detection_Experiment/YOLO_FACE_EXP/train_ver2.py

import torch
import torch.nn as nn
import torch.optim as optim
import os
import argparse
from tqdm import tqdm
from datetime import datetime
from torchvision import transforms
from torch.utils.data import DataLoader

# Import Custom Modules
from model import get_face_model, ArcFace
from dataset import get_face_dataloaders
from validation import LFWDataset, validate_lfw

def main():
    # 1. Hyperparameters
    parser = argparse.ArgumentParser(description='Face Recognition Training with LFW Validation')
    parser.add_argument('--data_dir', type=str, default='/home/cuuva/face_exp/datasets', help='Root dataset dir')
    parser.add_argument('--project_dir', type=str, default='./results', help='Save dir')
    parser.add_argument('--epochs', type=int, default=20)
    parser.add_argument('--batch_size', type=int, default=64)
    parser.add_argument('--lr', type=float, default=0.1)
    args = parser.parse_args()

    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
    print(f"Device: {device}")

    # 디렉토리 설정
    current_time = datetime.now().strftime("%y-%m-%d-%H-%M")
    base_save_dir = os.path.join(args.project_dir, current_time)
    best_save_dir = os.path.join(base_save_dir, 'best_model')
    os.makedirs(best_save_dir, exist_ok=True)
    
    print(f"✅ Save path: {base_save_dir}")

    # 2. Train Data Loader
    print("Loading Train Data...")
    try:
        train_loader, num_classes = get_face_dataloaders(args.data_dir, args.batch_size)
    except Exception as e:
        print(f"❌ Train Data Error: {e}")
        return

    # --------------------------------------------------------
    # [수정됨] 3. LFW Validation Loader (경로 로직 개선)
    # --------------------------------------------------------
    print("Loading LFW Data...")
    
    # pairs.txt가 있는 루트 경로: ~/face_exp/datasets/LFW
    lfw_root = os.path.join(args.data_dir, 'LFW')
    pairs_path = os.path.join(lfw_root, 'pairs.txt') 

    # 실제 이미지가 있는 경로 찾기
    # 1순위: ~/face_exp/datasets/LFW/lfw (보여주신 구조)
    # 2순위: ~/face_exp/datasets/LFW (일반적인 구조)
    if os.path.exists(os.path.join(lfw_root, 'lfw')):
        lfw_img_dir = os.path.join(lfw_root, 'lfw')
    else:
        lfw_img_dir = lfw_root
    
    print(f"ℹ️  Pairs path: {pairs_path}")
    print(f"ℹ️  Images path: {lfw_img_dir}")

    if os.path.exists(pairs_path) and os.path.exists(lfw_img_dir):
        lfw_transform = transforms.Compose([
            transforms.Resize((128, 128), interpolation=transforms.InterpolationMode.BICUBIC),
            transforms.ToTensor(),
            transforms.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5])
        ])
        
        # 여기서 lfw_img_dir를 넘겨주는 것이 핵심!
        lfw_dataset = LFWDataset(lfw_img_dir, pairs_path, transform=lfw_transform)
        
        lfw_loader = DataLoader(lfw_dataset, batch_size=64, shuffle=False, num_workers=4, drop_last=False)
        do_validation = True
        print(f"✅ LFW Loaded. Pairs: {len(lfw_dataset)}")
    else:
        print(f"⚠️ Warning: 'pairs.txt' or Image dir not found. Skipping Validation.")
        do_validation = False

    # 4. Model & Loss
    backbone = get_face_model().to(device)
    metric_fc = ArcFace(in_features=128, out_features=num_classes).to(device)
    criterion = nn.CrossEntropyLoss()
    
    optimizer = optim.SGD([
        {'params': backbone.parameters()},
        {'params': metric_fc.parameters()}
    ], lr=args.lr, momentum=0.9, weight_decay=5e-4)
    
    scheduler = optim.lr_scheduler.MultiStepLR(optimizer, milestones=[8, 14, 18], gamma=0.1)

    # 5. Training Loop
    print("🚀 Start Training...")
    
    best_acc = 0.0 

    for epoch in range(args.epochs):
        # --- TRAIN ---
        backbone.train()
        metric_fc.train()
        
        running_loss = 0.0
        pbar = tqdm(train_loader, desc=f"Epoch {epoch+1}/{args.epochs}")
        
        for images, labels in pbar:
            images, labels = images.to(device), labels.to(device)
            optimizer.zero_grad()
            
            features = backbone(images)
            outputs = metric_fc(features, labels)
            loss = criterion(outputs, labels)
            loss.backward()
            optimizer.step()
            
            running_loss += loss.item()
            pbar.set_postfix({'loss': f"{running_loss / (pbar.n + 1):.4f}"})
        
        scheduler.step()

        # --- VALIDATION (LFW) ---
        if do_validation:
            acc, th = validate_lfw(backbone, lfw_loader, device)
            print(f"📊 LFW Acc: {acc*100:.2f}% (Threshold: {th:.2f})")
            
            if acc > best_acc:
                best_acc = acc
                save_path = os.path.join(best_save_dir, "best_backbone.pt")
                torch.save(backbone.state_dict(), save_path)
                print(f"🏆 Best Model Updated! Saved to {save_path}")
        else:
            save_path = os.path.join(base_save_dir, f"backbone_epoch_{epoch+1}.pt")
            torch.save(backbone.state_dict(), save_path)

        torch.save(backbone.state_dict(), os.path.join(base_save_dir, "last_backbone.pt"))

    print("\n🎉 Training Finished!")

if __name__ == "__main__":
    main()