Blame for code/train_univfd.py

#!/usr/bin/env python3
"""

Training script for UnivFD-style binary classification (Real vs Fake).



Uses a frozen CLIP ViT-L/14 backbone with a trainable FC layer,

following the UniversalFakeDetect architecture but with the same

data sampling strategy as train.py.

"""
import argparse
import json
import os
import random
import sys
from dataclasses import dataclass
from pathlib import Path
from typing import Dict, List, Tuple

import yaml
from PIL import Image

import torch
import torch.nn as nn
import torch.optim as optim
from torch.utils.data import Dataset, DataLoader
import torchvision.transforms as T
from accelerate import Accelerator
from tqdm import tqdm

try:
    import clip
except ImportError:
    raise ImportError("Please install clip: pip install git+https://github.com/openai/CLIP.git")


# CLIP normalization statistics
CLIP_MEAN = [0.48145466, 0.4578275, 0.40821073]
CLIP_STD = [0.26862954, 0.26130258, 0.27577711]

ALLOWED_EXTS = {".jpg", ".jpeg", ".png", ".webp"}


@dataclass
class Sample:
    image_path: Path
    class_index: int


class ImageClsDataset(Dataset):
    """Dataset for loading images organized by model ID folders."""
    REAL_MODEL_ID = 0  # Reserved model ID for real images

    @classmethod
    def from_scratch(cls, ai_root, model_mapping, transform: T.Compose, real_root: str = None):
        samples = dict()

        # Load real images first (model_id = 0)
        if real_root and os.path.isdir(real_root):
            real_files = []
            for file in os.listdir(real_root):
                ext = Path(file).suffix.lower()
                if ext in ALLOWED_EXTS:
                    real_files.append((os.path.join(real_root, file), cls.REAL_MODEL_ID))
            if real_files:
                samples[cls.REAL_MODEL_ID] = real_files
                print(f"Loaded {len(real_files)} real images")

        # Load AI-generated images
        subfolders = os.listdir(ai_root)
        for subfolder in subfolders:
            if '.' in subfolder:
                continue
            try:
                idx = int(subfolder)
            except ValueError:
                continue
            if idx not in model_mapping.values():
                print(subfolder, 'is not listed as a model.')
            else:
                samples[idx] = []
                files = os.listdir(os.path.join(ai_root, subfolder))
                for file in files:
                    ext = Path(file).suffix.lower()
                    if ext in ALLOWED_EXTS:
                        samples[idx].append((os.path.join(ai_root, subfolder, file), idx))
        return cls(samples, model_mapping, transform)

    def __init__(self, samples, model_mapping, transform):
        super().__init__()
        self.samples_dict = samples  # Keep dict for subset operations
        self.model_mapping = model_mapping
        self.transform = transform

        # Create mapping from model_id to class index (0, 1, 2, ...)
        self.model_id_to_class = {mid: i for i, mid in enumerate(sorted(samples.keys()))}
        self.class_to_model_id = {i: mid for mid, i in self.model_id_to_class.items()}

        # Track which class index corresponds to real images (None if no real images)
        self.real_class_idx = self.model_id_to_class.get(self.REAL_MODEL_ID, None)

        # Flatten samples into a list for indexing: [(image_path, class_index), ...]
        self.samples = []
        for model_id, items in samples.items():
            class_idx = self.model_id_to_class[model_id]
            for img_path, _ in items:
                self.samples.append((img_path, class_idx))

    @property
    def num_classes(self):
        return len(self.samples_dict)

    def __len__(self) -> int:
        return len(self.samples)

    def __getitem__(self, idx: int):
        img_path, label = self.samples[idx]
        try:
            img = Image.open(img_path).convert("RGB")
        except Exception as exc:
            raise RuntimeError(f"Unreadable benchmark image: {img_path}") from exc
        img = self.transform(img)
        return img, label

    def crop_by_num_samples(self, limitation, min_threshold=None, real_limit=None):
        """Subsample to limitation samples per model."""
        if min_threshold is None:
            min_threshold = limitation
        subset = dict()
        for key, v in self.samples_dict.items():
            if key == self.REAL_MODEL_ID:
                limit = real_limit if real_limit is not None else limitation
                if len(v) >= min_threshold:
                    subset[key] = random.sample(v, k=min(limit, len(v)))
            else:
                if len(v) >= min_threshold:
                    subset[key] = random.sample(v, k=min(limitation, len(v)))
        return ImageClsDataset(subset, self.model_mapping, self.transform)

    def subsample_from(self, n_samples: int, real_ratio: float = None):
        """Create a smaller dataset by taking first n_samples from each AI model."""
        subset = dict()
        num_ai_models = len([k for k in self.samples_dict.keys() if k != self.REAL_MODEL_ID])

        for key, v in self.samples_dict.items():
            if key == self.REAL_MODEL_ID and real_ratio is not None:
                real_count = int(real_ratio * n_samples * num_ai_models)
                subset[key] = v[:real_count]
            else:
                subset[key] = v[:n_samples]
        return ImageClsDataset(subset, self.model_mapping, self.transform)

    def safe_split(self, ratio=0.9, real_test_count: int = None):
        """Split dataset into train and test sets."""
        train, val = dict(), dict()
        for k, v in self.samples_dict.items():
            v_copy = list(v)
            random.shuffle(v_copy)
            if k == self.REAL_MODEL_ID and real_test_count is not None:
                val[k] = v_copy[:real_test_count]
                train[k] = v_copy[real_test_count:]
            else:
                train[k] = v_copy[:int(len(v_copy) * ratio)]
                val[k] = v_copy[int(len(v_copy) * ratio):]
        return ImageClsDataset(train, self.model_mapping, self.transform), ImageClsDataset(val, self.model_mapping, self.transform)

    def get_split_config(self) -> Dict:
        """Export the current dataset configuration for reproducibility."""
        config = {}
        for model_id, items in self.samples_dict.items():
            config[str(model_id)] = [img_path for img_path, _ in items]
        return config

    @classmethod
    def from_split_config(cls, config: Dict, model_mapping, transform: T.Compose):
        """Create dataset from a saved split configuration."""
        samples = {}
        for model_id_str, paths in config.items():
            model_id = int(model_id_str)
            samples[model_id] = [(path, model_id) for path in paths]
        return cls(samples, model_mapping, transform)


class BinaryDatasetWrapper(Dataset):
    """Wrapper that converts multi-class labels to binary (real=0, fake=1)."""

    def __init__(self, base_dataset: ImageClsDataset):
        self.base_dataset = base_dataset
        self.real_class_idx = base_dataset.real_class_idx

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

    def __getitem__(self, idx):
        img, class_idx = self.base_dataset[idx]
        # Binary label: 0 for real, 1 for fake (AI-generated)
        if self.real_class_idx is not None:
            binary_label = 0.0 if class_idx == self.real_class_idx else 1.0
        else:
            # If no real images, all are fake
            binary_label = 1.0
        return img, binary_label, class_idx  # Also return original class for per-model analysis

    @property
    def samples_dict(self):
        return self.base_dataset.samples_dict


class UnivFDModel(nn.Module):
    """

    UnivFD-style model: Frozen CLIP backbone + trainable FC layer.



    Architecture:

        Input -> CLIP ViT-L/14 (frozen) -> FC(768, 1) -> Logit

    """

    def __init__(self, init_gain: float = 0.02):
        super().__init__()

        # Load CLIP model (will be moved to device later by accelerate)
        self.clip_model, _ = clip.load("ViT-L/14", device="cpu")
        self.clip_model.eval()

        # Freeze CLIP backbone
        for param in self.clip_model.parameters():
            param.requires_grad = False

        # Trainable FC layer
        self.fc = nn.Linear(768, 1)
        nn.init.normal_(self.fc.weight, 0.0, init_gain)
        nn.init.zeros_(self.fc.bias)

    def forward(self, x):
        # Extract features with frozen backbone
        with torch.no_grad():
            features = self.clip_model.encode_image(x)
        # Apply trainable FC layer
        return self.fc(features.float())


def save_split_config(train_config: Dict, test_config: Dict, output_path: Path):
    """Save train/test split configuration to JSON file."""
    split_data = {
        "train": train_config,
        "test": test_config,
    }
    with open(output_path, "w", encoding="utf-8") as f:
        json.dump(split_data, f, indent=2)
    print(f"Saved split config to: {output_path}")


def load_split_config(split_path: Path) -> Tuple[Dict, Dict]:
    """Load train/test split configuration from JSON file."""
    with open(split_path, "r", encoding="utf-8") as f:
        split_data = json.load(f)
    return split_data["train"], split_data["test"]


def make_clip_transforms(img_size: int = 224) -> Tuple[T.Compose, T.Compose]:
    """Create CLIP-normalized transforms for training and testing."""
    train_tf = T.Compose([
        T.RandomResizedCrop(img_size, scale=(0.6, 1.0)),
        T.RandomHorizontalFlip(),
        T.ToTensor(),
        T.Normalize(CLIP_MEAN, CLIP_STD),
    ])
    test_tf = T.Compose([
        T.Resize(int(img_size * 1.15)),
        T.CenterCrop(img_size),
        T.ToTensor(),
        T.Normalize(CLIP_MEAN, CLIP_STD),
    ])
    return train_tf, test_tf


def train_one_epoch(

    model: nn.Module,

    loader: DataLoader,

    criterion: nn.Module,

    optimizer: optim.Optimizer,

    accelerator: Accelerator,

    device: torch.device,

    epoch: int = 0,

    total_epochs: int = 0

) -> Tuple[float, float]:
    """Train for one epoch with BCE loss."""
    model.train()
    if hasattr(model, "clip_model"):
        model.clip_model.eval()
    running_loss = 0.0
    correct = 0
    total = 0

    pbar = tqdm(loader, desc=f"Epoch {epoch}/{total_epochs}", disable=not accelerator.is_local_main_process)
    for images, labels, _ in pbar:  # labels are binary floats, _ is original class
        images = images.to(device, non_blocking=True)
        labels = labels.to(device, non_blocking=True)

        optimizer.zero_grad(set_to_none=True)
        outputs = model(images).squeeze(-1)  # [B, 1] -> [B]
        loss = criterion(outputs, labels)
        accelerator.backward(loss)
        optimizer.step()

        running_loss += loss.item() * images.size(0)
        preds = (torch.sigmoid(outputs) > 0.5).float()
        correct += (preds == labels).sum().item()
        total += labels.size(0)

        pbar.set_postfix(loss=f"{loss.item():.4f}", acc=f"{correct/total:.4f}")

    # Reduce metrics across processes
    loss_tensor = torch.tensor([running_loss], device=device)
    correct_tensor = torch.tensor([correct], device=device)
    total_tensor = torch.tensor([total], device=device)
    loss_sum = accelerator.reduce(loss_tensor, reduction="sum").item()
    correct_sum = accelerator.reduce(correct_tensor, reduction="sum").item()
    total_sum = accelerator.reduce(total_tensor, reduction="sum").item()

    avg_loss = loss_sum / max(1, total_sum)
    acc = correct_sum / max(1, total_sum)
    return avg_loss, acc


def compute_f1_and_balanced_acc(real_correct: int, real_total: int, fake_correct: int, fake_total: int) -> Tuple[float, float, float]:
    """

    Compute balanced accuracy and F1 scores for both real and fake classes.



    Returns:

        balanced_acc: (real_acc + fake_acc) / 2

        f1_real: F1 score for the real class

        f1_fake: F1 score for the fake class

    """
    # Compute per-class accuracy
    real_acc = real_correct / max(1, real_total)
    fake_acc = fake_correct / max(1, fake_total)

    # Balanced accuracy
    balanced_acc = (real_acc + fake_acc) / 2

    # For real class (label=0):
    # TP = real_correct (predicted real, actually real)
    # FN = real_total - real_correct (predicted fake, actually real)
    # FP = fake_total - fake_correct (predicted real, actually fake)
    tp_real = real_correct
    fn_real = real_total - real_correct
    fp_real = fake_total - fake_correct

    precision_real = tp_real / max(1, tp_real + fp_real)
    recall_real = tp_real / max(1, tp_real + fn_real)
    f1_real = 2 * precision_real * recall_real / max(1e-8, precision_real + recall_real)

    # For fake class (label=1):
    tp_fake = fake_correct
    fn_fake = fake_total - fake_correct
    fp_fake = real_total - real_correct

    precision_fake = tp_fake / max(1, tp_fake + fp_fake)
    recall_fake = tp_fake / max(1, tp_fake + fn_fake)
    f1_fake = 2 * precision_fake * recall_fake / max(1e-8, precision_fake + recall_fake)

    return balanced_acc, f1_real, f1_fake


def evaluate(

    model: nn.Module,

    loader: DataLoader,

    criterion: nn.Module,

    accelerator: Accelerator,

    device: torch.device,

    num_models: int = None,

    class_to_model_id: Dict = None,

    real_class_idx: int = None

) -> Dict:
    """

    Evaluate model with binary classification metrics.



    Returns:

        Dict with keys: 'loss', 'accuracy', 'real_acc', 'fake_acc', 'balanced_acc',

                       'f1_real', 'f1_fake', 'per_model_detection_rate'

    """
    model.eval()
    running_loss = 0.0
    correct = 0
    total = 0

    # Track per-class metrics
    real_correct = 0
    real_total = 0
    fake_correct = 0
    fake_total = 0

    # Track per-model detection rates (for AI models only)
    if num_models is not None and class_to_model_id is not None:
        per_model_correct = {mid: 0 for mid in class_to_model_id.values() if mid != ImageClsDataset.REAL_MODEL_ID}
        per_model_total = {mid: 0 for mid in class_to_model_id.values() if mid != ImageClsDataset.REAL_MODEL_ID}
    else:
        per_model_correct = None
        per_model_total = None

    with torch.no_grad():
        for images, labels, orig_classes in loader:
            images = images.to(device, non_blocking=True)
            labels = labels.to(device, non_blocking=True)
            orig_classes = orig_classes.to(device, non_blocking=True)

            outputs = model(images).squeeze(-1)
            loss = criterion(outputs, labels)

            running_loss += loss.item() * images.size(0)
            preds = (torch.sigmoid(outputs) > 0.5).float()
            correct += (preds == labels).sum().item()
            total += labels.size(0)

            # Per-class accuracy
            real_mask = (labels == 0)
            fake_mask = (labels == 1)

            real_correct += ((preds == labels) & real_mask).sum().item()
            real_total += real_mask.sum().item()
            fake_correct += ((preds == labels) & fake_mask).sum().item()
            fake_total += fake_mask.sum().item()

            # Per-model detection rate (for fake images only)
            if per_model_correct is not None and class_to_model_id is not None:
                for i in range(len(orig_classes)):
                    orig_class = orig_classes[i].item()
                    model_id = class_to_model_id.get(orig_class)
                    if model_id is not None and model_id != ImageClsDataset.REAL_MODEL_ID:
                        per_model_total[model_id] += 1
                        if preds[i] == labels[i]:
                            per_model_correct[model_id] += 1

    # Reduce metrics across processes
    loss_tensor = torch.tensor([running_loss], device=device)
    correct_tensor = torch.tensor([correct], device=device)
    total_tensor = torch.tensor([total], device=device)
    real_correct_tensor = torch.tensor([real_correct], device=device)
    real_total_tensor = torch.tensor([real_total], device=device)
    fake_correct_tensor = torch.tensor([fake_correct], device=device)
    fake_total_tensor = torch.tensor([fake_total], device=device)

    loss_sum = accelerator.reduce(loss_tensor, reduction="sum").item()
    correct_sum = accelerator.reduce(correct_tensor, reduction="sum").item()
    total_sum = accelerator.reduce(total_tensor, reduction="sum").item()
    real_correct_sum = accelerator.reduce(real_correct_tensor, reduction="sum").item()
    real_total_sum = accelerator.reduce(real_total_tensor, reduction="sum").item()
    fake_correct_sum = accelerator.reduce(fake_correct_tensor, reduction="sum").item()
    fake_total_sum = accelerator.reduce(fake_total_tensor, reduction="sum").item()

    avg_loss = loss_sum / max(1, total_sum)
    accuracy = correct_sum / max(1, total_sum)
    real_acc = real_correct_sum / max(1, real_total_sum) if real_total_sum > 0 else None
    fake_acc = fake_correct_sum / max(1, fake_total_sum) if fake_total_sum > 0 else None

    # Compute balanced accuracy and F1 scores
    balanced_acc, f1_real, f1_fake = None, None, None
    if real_total_sum > 0 and fake_total_sum > 0:
        balanced_acc, f1_real, f1_fake = compute_f1_and_balanced_acc(
            int(real_correct_sum), int(real_total_sum),
            int(fake_correct_sum), int(fake_total_sum)
        )

    # Compute per-model detection rates
    per_model_detection_rate = None
    if per_model_correct is not None:
        per_model_detection_rate = {}
        for model_id in per_model_correct:
            if per_model_total[model_id] > 0:
                per_model_detection_rate[model_id] = per_model_correct[model_id] / per_model_total[model_id]

    return {
        'loss': avg_loss,
        'accuracy': accuracy,
        'real_acc': real_acc,
        'fake_acc': fake_acc,
        'balanced_acc': balanced_acc,
        'f1_real': f1_real,
        'f1_fake': f1_fake,
        'per_model_detection_rate': per_model_detection_rate,
    }


def save_outputs(output_dir: Path, model: nn.Module, is_best: bool = False, is_final: bool = False):
    """Save model checkpoints."""
    output_dir.mkdir(parents=True, exist_ok=True)

    if is_best:
        # Save only the FC layer weights (like UnivFD)
        fc_path = output_dir / "fc_weights_best.pth"
        torch.save(model.fc.state_dict(), fc_path)
        print(f"Saved best FC weights to: {fc_path}")

    if is_final:
        fc_path = output_dir / "fc_weights_final.pth"
        torch.save(model.fc.state_dict(), fc_path)
        print(f"Saved final FC weights to: {fc_path}")


def validation_epoch_interval(num_epochs: int, num_validations: int | None) -> int:
    """Return the epoch interval that keeps validation passes bounded."""
    if num_validations is None or num_validations <= 0:
        return 1
    return max(1, num_epochs // num_validations)


def should_validate_epoch(

    epoch: int,

    num_epochs: int,

    num_validations: int | None,

    no_val_between_epochs: bool,

) -> bool:
    if no_val_between_epochs:
        return epoch == num_epochs
    interval = validation_epoch_interval(num_epochs, num_validations)
    return epoch == num_epochs or epoch % interval == 0


def train_with_sample_size(

    train_ds: ImageClsDataset,

    test_ds: ImageClsDataset,

    sample_size: int,

    num_epochs: int,

    args,

    accelerator: Accelerator,

    output_dir: Path,

):
    """Train a model with a specific sample size setting."""
    accelerator.print(f"\n{'='*60}")
    accelerator.print(f"Training with sample_size={sample_size}, epochs={num_epochs}")
    accelerator.print(f"{'='*60}")

    # Create subsampled training set
    train_tf, test_tf = make_clip_transforms(args.img_size)
    n_samples_per_ai = int(sample_size * args.train_ratio)
    sub_train_ds = train_ds.subsample_from(n_samples_per_ai, real_ratio=args.real_ratio)

    sub_train_ds.transform = train_tf
    test_ds.transform = test_tf

    # Wrap with binary labels
    binary_train_ds = BinaryDatasetWrapper(sub_train_ds)
    binary_test_ds = BinaryDatasetWrapper(test_ds)

    # Count samples
    real_model_id = ImageClsDataset.REAL_MODEL_ID
    real_train_count = len(sub_train_ds.samples_dict.get(real_model_id, []))
    ai_train_count = sum(len(v) for k, v in sub_train_ds.samples_dict.items() if k != real_model_id)
    accelerator.print(f"Train samples: {len(binary_train_ds)} (Real: {real_train_count}, AI: {ai_train_count})")
    accelerator.print(f"Test samples: {len(binary_test_ds)}")

    pin_memory = torch.cuda.is_available()
    train_loader = DataLoader(
        binary_train_ds, batch_size=args.batch_size, shuffle=True, num_workers=args.num_workers, pin_memory=pin_memory
    )
    test_loader = DataLoader(
        binary_test_ds, batch_size=args.batch_size, shuffle=False, num_workers=args.num_workers, pin_memory=pin_memory
    )

    device = accelerator.device
    model = UnivFDModel(init_gain=args.init_gain)
    criterion = nn.BCEWithLogitsLoss()

    # Only optimize FC layer parameters
    optimizer = optim.AdamW(model.fc.parameters(), lr=args.lr, weight_decay=args.weight_decay)

    model, optimizer, train_loader, test_loader = accelerator.prepare(model, optimizer, train_loader, test_loader)

    best_acc = -1.0
    results = []

    # Get class to model_id mapping for per-model analysis
    class_to_model_id = sub_train_ds.class_to_model_id
    num_models = sub_train_ds.num_classes
    real_class_idx = sub_train_ds.real_class_idx

    best_per_model_detection = None

    for epoch in range(1, num_epochs + 1):
        train_loss, train_acc = train_one_epoch(
            model, train_loader, criterion, optimizer, accelerator, device, epoch, num_epochs
        )

        validate_now = should_validate_epoch(
            epoch, num_epochs, args.num_validations, args.no_val_between_epochs
        )
        if not validate_now:
            accelerator.print(f"Epoch {epoch:02d}/{num_epochs} | Train loss {train_loss:.4f} acc {train_acc:.4f}")
            results.append({
                'epoch': epoch,
                'validated': False,
                'train_loss': train_loss,
                'train_acc': train_acc,
                'test_loss': None,
                'test_acc': None,
                'real_acc': None,
                'fake_acc': None,
                'balanced_acc': None,
                'f1_real': None,
                'f1_fake': None,
                'per_model_detection_rate': None,
            })
            continue

        metrics = evaluate(
            model, test_loader, criterion, accelerator, device,
            num_models=num_models, class_to_model_id=class_to_model_id, real_class_idx=real_class_idx
        )

        test_loss = metrics['loss']
        test_acc = metrics['accuracy']
        real_acc = metrics['real_acc']
        fake_acc = metrics['fake_acc']
        balanced_acc = metrics['balanced_acc']
        f1_real = metrics['f1_real']
        f1_fake = metrics['f1_fake']
        per_model_detection = metrics['per_model_detection_rate']

        # Build output string
        output_str = (
            f"Epoch {epoch:02d}/{num_epochs} | Train loss {train_loss:.4f} acc {train_acc:.4f} | "
            f"Test loss {test_loss:.4f} acc {test_acc:.4f}"
        )
        if real_acc is not None:
            output_str += f" real_acc {real_acc:.4f}"
        if fake_acc is not None:
            output_str += f" fake_acc {fake_acc:.4f}"
        if balanced_acc is not None:
            output_str += f" bal_acc {balanced_acc:.4f}"
        if f1_real is not None:
            output_str += f" f1_real {f1_real:.4f}"
        accelerator.print(output_str)

        is_best = test_acc > best_acc
        if is_best:
            best_acc = test_acc
            best_per_model_detection = per_model_detection

        # Save best model (only when validation is performed)

        results.append({
            'epoch': epoch,
            'validated': True,
            'train_loss': train_loss,
            'train_acc': train_acc,
            'test_loss': test_loss,
            'test_acc': test_acc,
            'real_acc': real_acc,
            'fake_acc': fake_acc,
            'balanced_acc': balanced_acc,
            'f1_real': f1_real,
            'f1_fake': f1_fake,
            'per_model_detection_rate': per_model_detection,
        })

    # Save final model
    accelerator.wait_for_everyone()
    if accelerator.is_main_process:
        unwrapped = accelerator.unwrap_model(model)
        save_outputs(output_dir, unwrapped, is_final=True)

    # Save results summary
    # Get final epoch metrics for summary
    final_metrics = results[-1] if results else {}
    if accelerator.is_main_process:
        results_path = output_dir / "results.json"
        with open(results_path, "w") as f:
            json.dump({
                'sample_size': sample_size,
                'num_epochs': num_epochs,
                'num_validations': args.num_validations,
                'validation_interval': validation_epoch_interval(num_epochs, args.num_validations),
                'best_accuracy': best_acc,
                'final_accuracy': final_metrics.get('test_acc'),
                'final_balanced_acc': final_metrics.get('balanced_acc'),
                'final_f1_real': final_metrics.get('f1_real'),
                'final_f1_fake': final_metrics.get('f1_fake'),
                'final_real_acc': final_metrics.get('real_acc'),
                'final_fake_acc': final_metrics.get('fake_acc'),
                'best_per_model_detection_rate': best_per_model_detection,
                'class_to_model_id': {str(k): v for k, v in class_to_model_id.items()},
                'epochs': results,
            }, f, indent=2)
        accelerator.print(f"Saved results to: {results_path}")

    return final_metrics['test_acc']


def parse_args():
    parser = argparse.ArgumentParser(description="Train UnivFD-style binary classifier (Real vs Fake).")
    parser.add_argument("--mapping", type=str, default="metadata/model_mapping.yaml",
                       help="Path to model_mapping.yaml")
    parser.add_argument("--images-root", type=str, default="data_local/images/ai",
                       help="Root directory containing {id} subfolders with images")
    parser.add_argument("--real-root", type=str, default="data_local/images/real",
                       help="Root directory containing real images")
    parser.add_argument("--output-dir", type=str, default="./outputs_univfd",
                       help="Directory to save model and results")
    parser.add_argument("--batch-size", type=int, default=64)
    parser.add_argument("--lr", type=float, default=1e-4, help="Learning rate (UnivFD default: 1e-4)")
    parser.add_argument("--weight-decay", type=float, default=0.0, help="Weight decay for AdamW")
    parser.add_argument("--init-gain", type=float, default=0.02, help="FC layer weight initialization std")
    parser.add_argument("--num-workers", type=int, default=4)
    parser.add_argument("--img-size", type=int, default=224)
    parser.add_argument("--train-ratio", type=float, default=1.0, help="Train split ratio per class")
    parser.add_argument("--sample-sizes", type=str, default="900,90,45,18",
                       help="Comma-separated sample sizes to train on (descending order)")
    parser.add_argument("--total-samples-per-model", type=int, default=9000,
                       help="Total samples per model across all epochs")
    parser.add_argument("--real-ratio", type=float, default=0.1,
                       help="Ratio of real images to total AI images in training")
    parser.add_argument("--real-test-count", type=int, default=100,
                       help="Fixed number of real images in test set")
    parser.add_argument("--no-val-between-epochs", action="store_true",
                       help="Skip validation between epochs, only evaluate at the end")
    parser.add_argument("--num-validations", type=int, default=None,
                       help="Target validation passes per sample size. Default validates every epoch.")
    parser.add_argument("--split-file", type=str, required=True,
                       help="Path to split config JSON. If provided, reuse this split")
    parser.add_argument("--seed", type=int, default=42)
    parser.add_argument("--mixed-precision", type=str, default="no", choices=["no", "fp16", "bf16"],
                       help="Mixed precision mode for accelerate")
    return parser.parse_args()


def main():
    args = parse_args()
    accelerator = Accelerator(mixed_precision=args.mixed_precision)

    random.seed(args.seed)
    torch.manual_seed(args.seed)

    images_root = Path(args.images_root)
    output_dir = Path(args.output_dir)

    # Parse sample sizes
    sample_sizes = [int(s.strip()) for s in args.sample_sizes.split(",")]
    sample_sizes = sorted(sample_sizes, reverse=True)
    max_sample_size = sample_sizes[0]

    accelerator.print("=" * 60)
    accelerator.print("UnivFD-style Binary Classification Training")
    accelerator.print("=" * 60)
    accelerator.print(f"Model: CLIP ViT-L/14 (frozen) + FC layer (trainable)")
    accelerator.print(f"Task: Binary classification (Real vs Fake)")
    accelerator.print(f"Sample sizes to train: {sample_sizes}")
    accelerator.print(f"Total samples per model target: {args.total_samples_per_model}")
    accelerator.print(f"Real image ratio: {args.real_ratio}")
    accelerator.print(f"Real test count: {args.real_test_count}")

    accelerator.print("\nLoading mapping...")
    with open(args.mapping, "r", encoding="utf-8") as f:
        model_mapping = yaml.safe_load(f)

    accelerator.print("Preparing base dataset...")
    train_tf, test_tf = make_clip_transforms(args.img_size)

    if not Path(args.split_file).is_file():
        raise FileNotFoundError(args.split_file)
    with open(args.split_file, encoding="utf-8") as split_handle:
        release_split = json.load(split_handle)
    if release_split.get("protocol", {}).get("seed") != args.seed:
        raise ValueError("Build a release split for this seed before training")
    if args.split_file:
        accelerator.print(f"Loading split config from: {args.split_file}")
        train_config, test_config = load_split_config(Path(args.split_file))
        base_train_ds = ImageClsDataset.from_split_config(train_config, model_mapping, train_tf)
        base_test_ds = ImageClsDataset.from_split_config(test_config, model_mapping, test_tf)
        accelerator.print(f"Loaded {len(base_train_ds)} train samples, {len(base_test_ds)} test samples")
    else:
        dataset = ImageClsDataset.from_scratch(images_root, model_mapping, train_tf, real_root=args.real_root)

        num_ai_models = len([k for k in dataset.samples_dict.keys() if k != ImageClsDataset.REAL_MODEL_ID])
        max_real_train = int(args.real_ratio * (max_sample_size * args.train_ratio) * num_ai_models)
        real_limit = max_real_train + args.real_test_count
        accelerator.print(f"Real images needed: {real_limit} (train: {max_real_train}, test: {args.real_test_count})")

        subset = dataset.crop_by_num_samples(max_sample_size, min_threshold=1000, real_limit=real_limit)
        base_train_ds, base_test_ds = subset.safe_split(args.train_ratio, real_test_count=args.real_test_count)
        base_test_ds.transform = test_tf

        output_dir.mkdir(parents=True, exist_ok=True)
        split_path = output_dir / "split_config.json"
        if accelerator.is_main_process:
            save_split_config(base_train_ds.get_split_config(), base_test_ds.get_split_config(), split_path)

    accelerator.print(f"Base dataset: {base_train_ds.num_classes} classes, {len(base_train_ds)} train, {len(base_test_ds)} test")

    real_class_idx = base_train_ds.real_class_idx
    if real_class_idx is not None:
        accelerator.print(f"Real images included as class {real_class_idx}")
    else:
        accelerator.print("No real images - only fake detection accuracy will be computed")

    # Train for each sample size
    all_results = {}
    for sample_size in sample_sizes:
        num_epochs = args.total_samples_per_model // sample_size
        accelerator.print(f"\nSample size {sample_size}: {num_epochs} epochs")

        exp_output_dir = output_dir / f"exp_{sample_size}"

        final_acc = train_with_sample_size(
            base_train_ds,
            base_test_ds,
            sample_size,
            num_epochs,
            args,
            accelerator,
            exp_output_dir,
        )

        all_results[sample_size] = {
            'num_epochs': num_epochs,
            'final_accuracy': final_acc,
        }

    # Save summary
    if accelerator.is_main_process:
        summary_path = output_dir / "summary.json"
        with open(summary_path, "w") as f:
            json.dump(all_results, f, indent=2)
        accelerator.print(f"\nSaved summary to: {summary_path}")

    accelerator.print("\nAll training completed!")
    accelerator.print("\nResults summary:")
    for sample_size, result in all_results.items():
        accelerator.print(f"  Sample size {sample_size}: {result['num_epochs']} epochs, final acc {result['final_accuracy']:.4f}")


if __name__ == "__main__":
    try:
        main()
    except Exception as e:
        print(f"Error: {e}", file=sys.stderr)
        sys.exit(1)