trainer/train_gpu.py

"""
VERIDEX — GPU Trainer v4.0 (RTX 3050 Optimized)
=================================================
RTX 3050 ki specially optimize chesina trainer.
70,000 images support chestundi.

DATASET STRUCTURE (mandatory):
  trainer/
    dataset/
      train/
        real/   <- Training real images     (70% = ~24,500)
        fake/   <- Training fake images     (70% = ~24,500)
      val/
        real/   <- Validation real images   (15% = ~5,250)
        fake/   <- Validation fake images   (15% = ~5,250)
      test/
        real/   <- Final test real images   (15% = ~5,250)
        fake/   <- Final test fake images   (15% = ~5,250)

Quick split for 70k images (35k real + 35k fake):
  train/ = 49,000 images (24,500 real + 24,500 fake)
  val/   = 10,500 images ( 5,250 real +  5,250 fake)
  test/  = 10,500 images ( 5,250 real +  5,250 fake)

USAGE (Windows CMD - VERIDEX/trainer folder lo):
  venv_gpu\\Scripts\\activate
  python train_gpu.py

OUTPUT:
  weights/efficientnet_deepfake.pth       <- model weights
  weights/efficientnet_b4_meta.json       <- metadata + test results

After training complete:
  Copy both files to VERIDEX/backend/weights/
"""

import os, json, time
import torch
import torch.nn as nn
import torch.optim as optim
from torch.utils.data import DataLoader, Dataset
import torchvision.transforms as T
from PIL import Image
import timm
from tqdm import tqdm

# ── RTX 3050 Config (4GB VRAM optimized) ──────────────────────
CFG = {
    "dataset_dir":  "dataset",
    "weights_dir":  "weights",
    "output_name":  "efficientnet_deepfake",
    "img_size":     256,      # 224 kante better clarity, 320 kante faster
    "batch_size":   24,       # 256px tho 24 RTX 3050 ki safe
    "epochs":       25,
    "lr":           2e-4,
    "weight_decay": 1e-4,
    "num_workers":  0,        # Windows lo MUST be 0
    "fake_label":   0,        # 0=fake, 1=real (marchipoyoddu!)
    "early_stop":   6,
    "use_amp":      True,     # Mixed Precision (RTX 3050 support)
    "grad_accum":   1,        # effective batch = 24
}

SUPPORTED = {".jpg", ".jpeg", ".png", ".webp", ".bmp", ".tiff"}

os.makedirs(CFG["weights_dir"], exist_ok=True)
os.makedirs("training_logs",    exist_ok=True)


# ── Dataset Class ─────────────────────────────────────────────
class DeepfakeDataset(Dataset):
    def __init__(self, samples, transform):
        self.samples   = samples
        self.transform = transform

    def __len__(self):
        return len(self.samples)

    def __getitem__(self, idx):
        path, label = self.samples[idx]
        try:
            img = Image.open(path).convert("RGB")
        except Exception:
            img = Image.new("RGB", (CFG["img_size"], CFG["img_size"]))
        return self.transform(img), label


# ── Load split folder ─────────────────────────────────────────
def load_split(split_name):
    """
    split_name: 'train', 'val', or 'test'
    dataset/<split_name>/real/ and dataset/<split_name>/fake/ nundi load
    """
    real_dir = os.path.join(CFG["dataset_dir"], split_name, "real")
    fake_dir = os.path.join(CFG["dataset_dir"], split_name, "fake")

    for d in [real_dir, fake_dir]:
        if not os.path.exists(d):
            print(f"\n  ERROR: Folder not found: {d}")
            print(f"  Create cheyyi: trainer/dataset/{split_name}/real/")
            print(f"                 trainer/dataset/{split_name}/fake/")
            raise SystemExit(1)

    samples = []
    for label_int, folder in [(1, real_dir), (CFG["fake_label"], fake_dir)]:
        files = [
            f for f in os.listdir(folder)
            if os.path.splitext(f)[1].lower() in SUPPORTED
            and not f.startswith(".")   # .gitkeep etc skip
        ]
        for f in files:
            samples.append((os.path.join(folder, f), label_int))

    real_c = sum(1 for _, l in samples if l == 1)
    fake_c = sum(1 for _, l in samples if l == 0)

    print(f"   {split_name:5s}  |  real: {real_c:7,}  fake: {fake_c:7,}  total: {len(samples):7,}")

    if len(samples) == 0:
        print(f"\n  ERROR: {split_name}/ lo images lev! Images add chesii malli run cheyyi.")
        raise SystemExit(1)

    if real_c > 0 and fake_c > 0:
        ratio = max(real_c, fake_c) / min(real_c, fake_c)
        if ratio > 1.5:
            print(f"   WARNING: {split_name} imbalanced ({ratio:.1f}x) — real:fake 1:1 better!")

    return samples


# ── Transforms ────────────────────────────────────────────────
def get_transforms():
    mean = [0.485, 0.456, 0.406]
    std  = [0.229, 0.224, 0.225]
    sz   = CFG["img_size"]

    train_tf = T.Compose([
        T.Resize((sz + 40, sz + 40)),
        T.RandomCrop((sz, sz)),
        T.RandomHorizontalFlip(0.5),
        T.RandomVerticalFlip(0.05),
        T.ColorJitter(brightness=0.25, contrast=0.25, saturation=0.15, hue=0.08),
        T.RandomGrayscale(p=0.05),
        T.RandomRotation(10),
        T.GaussianBlur(kernel_size=3, sigma=(0.1, 1.0)),
        T.ToTensor(),
        T.Normalize(mean, std),
    ])
    eval_tf = T.Compose([
        T.Resize((sz, sz)),
        T.ToTensor(),
        T.Normalize(mean, std),
    ])
    return train_tf, eval_tf


# ── Model ─────────────────────────────────────────────────────
def build_model(device):
    print("\n  Building EfficientNet-B4...")
    import torch.backends.cudnn as cudnn
    cudnn.benchmark = True
    cudnn.deterministic = False
    model = timm.create_model("efficientnet_b4", pretrained=True)
    model.classifier = nn.Sequential(
        nn.Dropout(0.4),
        nn.Linear(model.num_features, 512),
        nn.GELU(),
        nn.BatchNorm1d(512),
        nn.Dropout(0.3),
        nn.Linear(512, 2),
    )
    model = model.to(device)
    params = sum(p.numel() for p in model.parameters() if p.requires_grad)
    print(f"  Trainable params: {params:,}")
    return model


# ── Test Set Evaluation ───────────────────────────────────────
def evaluate_test(model, test_dl, device, scaler):
    model.eval()
    correct = total = 0
    tp = fp = tn = fn = 0

    with torch.no_grad():
        for imgs, labels in tqdm(test_dl, desc="  Test eval", ncols=65, leave=False):
            imgs   = imgs.to(device, non_blocking=True)
            labels = labels.to(device, non_blocking=True)

            if scaler:
                with torch.cuda.amp.autocast():
                    out = model(imgs)
            else:
                out = model(imgs)

            preds   = out.argmax(1)
            correct += (preds == labels).sum().item()
            total   += len(labels)

            for p, l in zip(preds.cpu(), labels.cpu()):
                p, l = int(p), int(l)
                if p == 1 and l == 1: tp += 1
                if p == 1 and l == 0: fp += 1
                if p == 0 and l == 0: tn += 1
                if p == 0 and l == 1: fn += 1

    acc       = correct / max(total, 1)
    precision = tp / max(tp + fp, 1)
    recall    = tp / max(tp + fn, 1)
    f1        = 2 * precision * recall / max(precision + recall, 1e-9)

    return {
        "test_accuracy":  round(acc, 4),
        "precision":      round(precision, 4),
        "recall":         round(recall, 4),
        "f1_score":       round(f1, 4),
        "true_positive":  tp,
        "false_positive": fp,
        "true_negative":  tn,
        "false_negative": fn,
    }


# ── Main Training ─────────────────────────────────────────────
def train():
    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

    print("=" * 65)
    print("  VERIDEX Trainer v4.0  |  RTX 3050 GPU Optimized")
    print("=" * 65)

    if device.type == "cuda":
        gpu_name = torch.cuda.get_device_name(0)
        vram_gb  = torch.cuda.get_device_properties(0).total_memory / 1e9
        print(f"\n  GPU    : {gpu_name}")
        print(f"  VRAM   : {vram_gb:.1f} GB")
        print(f"  AMP    : {'ON (faster!)' if CFG['use_amp'] else 'OFF'}")
        if vram_gb < 3.5:
            print(f"\n  Low VRAM! batch_size 12 -> 6 reduce chestunna...")
            CFG["batch_size"] = 6
    else:
        print(f"\n  WARNING: GPU not found! CPU mode (slow)")
        print(f"  nvidia-smi run chesii CUDA check cheyyi")
        CFG["use_amp"] = False

    print(f"\n  Config:")
    print(f"  Image size  : {CFG['img_size']}x{CFG['img_size']}")
    print(f"  Batch size  : {CFG['batch_size']}  (effective: {CFG['batch_size'] * CFG['grad_accum']})")
    print(f"  Epochs      : {CFG['epochs']}")
    print(f"  Device      : {device}")

    # Load datasets
    print(f"\n  Dataset loading...")
    print(f"  {'split':5s}  |  {'real':>12}  {'fake':>12}  {'total':>12}")
    print(f"  {'-'*55}")
    train_samples = load_split("train")
    val_samples   = load_split("val")
    test_samples  = load_split("test")
    grand_total   = len(train_samples) + len(val_samples) + len(test_samples)
    print(f"  {'-'*55}")
    print(f"  TOTAL                                     {grand_total:>12,}")

    # DataLoaders
    train_tf, eval_tf = get_transforms()
    train_ds = DeepfakeDataset(train_samples, train_tf)
    val_ds   = DeepfakeDataset(val_samples,   eval_tf)
    test_ds  = DeepfakeDataset(test_samples,  eval_tf)

    train_dl = DataLoader(train_ds, batch_size=CFG["batch_size"], shuffle=True,
                          num_workers=CFG["num_workers"], pin_memory=(device.type == "cuda"))
    val_dl   = DataLoader(val_ds,   batch_size=CFG["batch_size"], shuffle=False,
                          num_workers=CFG["num_workers"], pin_memory=(device.type == "cuda"))
    test_dl  = DataLoader(test_ds,  batch_size=CFG["batch_size"], shuffle=False,
                          num_workers=CFG["num_workers"], pin_memory=(device.type == "cuda"))

    # Model + optimizer
    model     = build_model(device)
    optimizer = optim.AdamW(model.parameters(), lr=CFG["lr"], weight_decay=CFG["weight_decay"])
    scheduler = optim.lr_scheduler.CosineAnnealingLR(optimizer, T_max=CFG["epochs"], eta_min=1e-6)
    criterion = nn.CrossEntropyLoss(label_smoothing=0.05)
    scaler    = torch.cuda.amp.GradScaler() if CFG["use_amp"] and device.type == "cuda" else None

    best_val_acc = 0.0
    no_improve   = 0
    log_path     = f"training_logs/train_{int(time.time())}.csv"
    log_lines    = ["epoch,train_loss,train_acc,val_loss,val_acc,lr"]
    start_time   = time.time()

    print(f"\n  Training started!\n")
    print(f"  {'Epoch':>5}  {'TrnLoss':>8}  {'TrnAcc':>7}  {'ValLoss':>8}  {'ValAcc':>7}  {'Note'}")
    print(f"  {'-'*65}")

    for epoch in range(1, CFG["epochs"] + 1):
        ep_start = time.time()

        # TRAIN
        model.train()
        t_loss = t_correct = t_total = 0
        optimizer.zero_grad()

        for step, (imgs, labels) in enumerate(
            tqdm(train_dl, desc=f"  Ep{epoch:02d} TRAIN", leave=False, ncols=60)
        ):
            imgs   = imgs.to(device, non_blocking=True)
            labels = labels.to(device, non_blocking=True)

            if scaler:
                with torch.cuda.amp.autocast():
                    out  = model(imgs)
                    loss = criterion(out, labels) / CFG["grad_accum"]
                scaler.scale(loss).backward()
                if (step + 1) % CFG["grad_accum"] == 0:
                    scaler.unscale_(optimizer)
                    torch.nn.utils.clip_grad_norm_(model.parameters(), 1.0)
                    scaler.step(optimizer)
                    scaler.update()
                    optimizer.zero_grad()
            else:
                out  = model(imgs)
                loss = criterion(out, labels) / CFG["grad_accum"]
                loss.backward()
                if (step + 1) % CFG["grad_accum"] == 0:
                    torch.nn.utils.clip_grad_norm_(model.parameters(), 1.0)
                    optimizer.step()
                    optimizer.zero_grad()

            t_loss    += loss.item() * CFG["grad_accum"] * len(imgs)
            t_correct += (out.argmax(1) == labels).sum().item()
            t_total   += len(imgs)

        # VALIDATION
        model.eval()
        v_loss = v_correct = v_total = 0
        with torch.no_grad():
            for imgs, labels in tqdm(
                val_dl, desc=f"  Ep{epoch:02d} VAL  ", leave=False, ncols=60
            ):
                imgs   = imgs.to(device, non_blocking=True)
                labels = labels.to(device, non_blocking=True)
                if scaler:
                    with torch.cuda.amp.autocast():
                        out  = model(imgs)
                        loss = criterion(out, labels)
                else:
                    out  = model(imgs)
                    loss = criterion(out, labels)
                v_loss    += loss.item() * len(imgs)
                v_correct += (out.argmax(1) == labels).sum().item()
                v_total   += len(imgs)

        scheduler.step()
        current_lr = scheduler.get_last_lr()[0]

        tl = t_loss / max(t_total, 1)
        ta = t_correct / max(t_total, 1)
        vl = v_loss   / max(v_total, 1)
        va = v_correct / max(v_total, 1)

        elapsed = time.time() - ep_start
        eta_min = elapsed * (CFG["epochs"] - epoch) / 60

        log_lines.append(f"{epoch},{tl:.4f},{ta:.4f},{vl:.4f},{va:.4f},{current_lr:.6f}")

        if va > best_val_acc:
            best_val_acc = va
            no_improve   = 0
            wpath = os.path.join(CFG["weights_dir"], f"{CFG['output_name']}.pth")
            torch.save(model.state_dict(), wpath)
            note = f"BEST! saved  (ETA {eta_min:.0f}m)"
        else:
            no_improve += 1
            note = f"no improve {no_improve}/{CFG['early_stop']}  (ETA {eta_min:.0f}m)"
            if no_improve >= CFG["early_stop"]:
                print(f"  {epoch:5d}  {tl:8.4f}  {ta:7.1%}  {vl:8.4f}  {va:7.1%}  Early stop!")
                break

        print(f"  {epoch:5d}  {tl:8.4f}  {ta:7.1%}  {vl:8.4f}  {va:7.1%}  {note}")

    # Save log
    with open(log_path, "w") as f:
        f.write("\n".join(log_lines))

    # FINAL TEST EVALUATION
    print(f"\n  {'='*65}")
    print(f"  Final TEST set evaluation (unseen data)...")
    wpath = os.path.join(CFG["weights_dir"], f"{CFG['output_name']}.pth")
    model.load_state_dict(torch.load(wpath, map_location=device))
    res = evaluate_test(model, test_dl, device, scaler)

    print(f"\n  Test Accuracy  : {res['test_accuracy']:.1%}")
    print(f"  Precision      : {res['precision']:.1%}")
    print(f"  Recall         : {res['recall']:.1%}")
    print(f"  F1 Score       : {res['f1_score']:.1%}")
    print(f"\n  Confusion Matrix:")
    print(f"  True  Positive : {res['true_positive']:,}")
    print(f"  False Positive : {res['false_positive']:,}")
    print(f"  True  Negative : {res['true_negative']:,}")
    print(f"  False Negative : {res['false_negative']:,}")

    # Save metadata
    meta = {
        "model":          "EfficientNet-B4",
        "trained_on":     "RTX 3050 GPU",
        "fake_label":     CFG["fake_label"],
        "img_size":       CFG["img_size"],
        "best_val_acc":   round(best_val_acc, 4),
        "epochs_trained": epoch,
        "dataset_size": {
            "train": len(train_samples),
            "val":   len(val_samples),
            "test":  len(test_samples),
            "total": grand_total,
        },
        "test_results": res,
    }
    with open(os.path.join(CFG["weights_dir"], "efficientnet_b4_meta.json"), "w") as f:
        json.dump(meta, f, indent=2)

    total_time = (time.time() - start_time) / 60
    print(f"\n  {'='*65}")
    print(f"  TRAINING COMPLETE!")
    print(f"  Best val accuracy   : {best_val_acc:.1%}")
    print(f"  Final test accuracy : {res['test_accuracy']:.1%}")
    print(f"  F1 Score            : {res['f1_score']:.1%}")
    print(f"  Total time          : {total_time:.0f} minutes")
    print(f"\n  NEXT STEP:")
    print(f"  1. trainer\\weights\\efficientnet_deepfake.pth  --> copy")
    print(f"  2. trainer\\weights\\efficientnet_b4_meta.json  --> copy")
    print(f"  3. VERIDEX\\backend\\weights\\ lo paste cheyyi")
    print(f"  4. Backend restart: uvicorn main:app --reload --port 8000")
    print(f"  {'='*65}")


if __name__ == "__main__":
    train()