train.py

"""
AI-Generated Image Detector
Architecture: SwinV2 backbone + SRM High-Pass Filter + DCT Frequency Analysis + FFT Spectral Analysis
Dataset: OwensLab/CommunityForensics-Small (556K images, 4803 generators)
Based on: AIDE paper (arxiv:2406.19435) + CommunityForensics (CVPR 2025)
"""
import os
import io
import math
import random
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
from PIL import Image
from torch.utils.data import Dataset, DataLoader
from transformers import (
    AutoImageProcessor,
    Swinv2Model,
    TrainingArguments,
    Trainer,
    DefaultDataCollator,
)
from torchvision.transforms import (
    Compose, Normalize, Resize, CenterCrop, RandomResizedCrop,
    RandomHorizontalFlip, ToTensor, ColorJitter,
)
import evaluate
import datasets as hf_datasets
import trackio


# ============================================================
# 1. SRM HIGH-PASS FILTER BANK (30 fixed kernels, no gradient)
# ============================================================

def get_srm_kernels():
    """Generate 30 SRM high-pass filter kernels (5x5)."""
    f1 = np.array([[0,0,0,0,0],[0,0,0,0,0],[0,0,-1,1,0],[0,0,0,0,0],[0,0,0,0,0]], dtype=np.float32)
    f2 = np.array([[0,0,0,0,0],[0,0,0,0,0],[0,0,-1,0,0],[0,0,1,0,0],[0,0,0,0,0]], dtype=np.float32)
    f3 = np.array([[0,0,0,0,0],[0,0,0,0,0],[0,1,-2,1,0],[0,0,0,0,0],[0,0,0,0,0]], dtype=np.float32)
    f4 = np.array([[0,0,0,0,0],[0,0,1,0,0],[0,0,-2,0,0],[0,0,1,0,0],[0,0,0,0,0]], dtype=np.float32)
    f5 = np.array([[0,0,0,0,0],[0,1,0,0,0],[0,0,-2,0,0],[0,0,0,1,0],[0,0,0,0,0]], dtype=np.float32)
    f6 = np.array([[0,0,0,0,0],[0,0,0,1,0],[0,0,-2,0,0],[0,1,0,0,0],[0,0,0,0,0]], dtype=np.float32)
    f7 = np.array([[0,0,0,0,0],[0,0,0,0,0],[0,-1,3,-3,1],[0,0,0,0,0],[0,0,0,0,0]], dtype=np.float32)
    f8 = np.array([[0,0,0,0,0],[0,0,-1,0,0],[0,0,3,0,0],[0,0,-3,0,0],[0,0,1,0,0]], dtype=np.float32)
    f9 = np.array([[0,0,0,0,0],[0,0,1,0,0],[0,1,-4,1,0],[0,0,1,0,0],[0,0,0,0,0]], dtype=np.float32)
    f10 = np.array([[0,0,0,0,0],[0,1,1,1,0],[0,1,-8,1,0],[0,1,1,1,0],[0,0,0,0,0]], dtype=np.float32) / 3.0
    f11 = np.array([[0,0,0,0,0],[0,-1,2,-1,0],[0,2,-4,2,0],[0,-1,2,-1,0],[0,0,0,0,0]], dtype=np.float32)
    f12 = np.array([[0,0,0,0,0],[0,0,0,0,0],[-1,2,-2,2,-1],[0,0,0,0,0],[0,0,0,0,0]], dtype=np.float32) / 2.0
    f13 = np.array([[0,0,-1,0,0],[0,0,2,0,0],[0,0,-2,0,0],[0,0,2,0,0],[0,0,-1,0,0]], dtype=np.float32) / 2.0
    f14 = np.array([[0,0,-1,0,0],[0,0,2,0,0],[-1,2,-4,2,-1],[0,0,2,0,0],[0,0,-1,0,0]], dtype=np.float32) / 4.0
    f15 = np.array([[-1,2,-2,2,-1],[2,-6,8,-6,2],[-2,8,-12,8,-2],[2,-6,8,-6,2],[-1,2,-2,2,-1]], dtype=np.float32) / 12.0
    spam_h = np.array([[0,0,0,0,0],[0,0,0,0,0],[0,-1,2,-1,0],[0,0,0,0,0],[0,0,0,0,0]], dtype=np.float32)
    spam_v = np.array([[0,0,0,0,0],[0,0,-1,0,0],[0,0,2,0,0],[0,0,-1,0,0],[0,0,0,0,0]], dtype=np.float32)
    spam_d1 = np.array([[0,0,0,0,0],[0,-1,0,0,0],[0,0,2,0,0],[0,0,0,-1,0],[0,0,0,0,0]], dtype=np.float32)
    spam_d2 = np.array([[0,0,0,0,0],[0,0,0,-1,0],[0,0,2,0,0],[0,-1,0,0,0],[0,0,0,0,0]], dtype=np.float32)
    spam3_h = np.array([[0,0,0,0,0],[0,0,0,0,0],[1,-3,3,-1,0],[0,0,0,0,0],[0,0,0,0,0]], dtype=np.float32)
    spam3_v = np.array([[0,0,1,0,0],[0,0,-3,0,0],[0,0,3,0,0],[0,0,-1,0,0],[0,0,0,0,0]], dtype=np.float32)
    sq5_1 = np.array([[0,0,0,0,0],[0,0,1,0,0],[0,1,-4,1,0],[0,0,1,0,0],[0,0,0,0,0]], dtype=np.float32) / 2.0
    sq5_2 = np.array([[0,0,0,0,0],[0,1,0,1,0],[0,0,-4,0,0],[0,1,0,1,0],[0,0,0,0,0]], dtype=np.float32) / 2.0
    cross1 = np.array([[0,0,-1,0,0],[0,0,2,0,0],[0,0,-2,0,0],[0,0,2,0,0],[0,0,-1,0,0]], dtype=np.float32) / 2.0
    cross2 = np.array([[0,0,0,0,0],[0,0,0,0,0],[-1,2,-2,2,-1],[0,0,0,0,0],[0,0,0,0,0]], dtype=np.float32) / 2.0
    edge_d1 = np.array([[-1,0,0,0,0],[0,2,0,0,0],[0,0,-2,0,0],[0,0,0,2,0],[0,0,0,0,-1]], dtype=np.float32) / 2.0
    edge_d2 = np.array([[0,0,0,0,-1],[0,0,0,2,0],[0,0,-2,0,0],[0,2,0,0,0],[-1,0,0,0,0]], dtype=np.float32) / 2.0
    gabor_h = np.array([[0,0,0,0,0],[1,-1,0,-1,1],[0,0,0,0,0],[-1,1,0,1,-1],[0,0,0,0,0]], dtype=np.float32) / 4.0
    gabor_v = np.array([[0,1,0,-1,0],[0,-1,0,1,0],[0,0,0,0,0],[0,-1,0,1,0],[0,1,0,-1,0]], dtype=np.float32) / 4.0
    
    all_filters = [f1,f2,f3,f4,f5,f6,f7,f8,f9,f10,f11,f12,f13,f14,f15,
                   spam_h,spam_v,spam_d1,spam_d2,spam3_h,spam3_v,
                   sq5_1,sq5_2,cross1,cross2,edge_d1,edge_d2,gabor_h,gabor_v,f15]
    return all_filters[:30]


class SRMFilterBank(nn.Module):
    """SRM High-Pass Filter Bank - 30 fixed forensic filters."""
    def __init__(self):
        super().__init__()
        kernels = get_srm_kernels()
        weight = torch.stack([torch.tensor(k) for k in kernels]).unsqueeze(1)
        weight = weight.repeat(1, 3, 1, 1) / 3.0
        self.register_buffer('weight', weight)
    
    def forward(self, x):
        return F.conv2d(x, self.weight, padding=2)


# ============================================================
# 2. DCT FREQUENCY ANALYSIS MODULE
# ============================================================

class DCTFrequencyAnalyzer(nn.Module):
    """Extracts frequency-domain features using 2D DCT on image patches."""
    def __init__(self, patch_size=32, num_freq_bands=8):
        super().__init__()
        self.patch_size = patch_size
        self.num_freq_bands = num_freq_bands
        N = patch_size
        dct_mat = torch.zeros(N, N)
        for k in range(N):
            for n in range(N):
                if k == 0:
                    dct_mat[k, n] = math.sqrt(1.0 / N)
                else:
                    dct_mat[k, n] = math.sqrt(2.0 / N) * math.cos(math.pi * (2*n + 1) * k / (2*N))
        self.register_buffer('dct_mat', dct_mat)
        self.register_buffer('dct_mat_t', dct_mat.t())
    
    def dct2d(self, x):
        return torch.matmul(torch.matmul(self.dct_mat, x), self.dct_mat_t)
    
    def forward(self, x):
        B, C, H, W = x.shape
        ps = self.patch_size
        gray = 0.299 * x[:, 0] + 0.587 * x[:, 1] + 0.114 * x[:, 2]
        h_patches = H // ps
        w_patches = W // ps
        gray = gray[:, :h_patches*ps, :w_patches*ps]
        patches = gray.unfold(1, ps, ps).unfold(2, ps, ps)
        B_p, hp, wp = patches.shape[:3]
        patches = patches.reshape(B_p * hp * wp, ps, ps)
        dct_patches = self.dct2d(patches)
        dct_patches = dct_patches.reshape(B_p, hp * wp, ps, ps)
        
        features = []
        freq_y = torch.arange(ps, device=x.device).float()
        freq_x = torch.arange(ps, device=x.device).float()
        fy, fx = torch.meshgrid(freq_y, freq_x, indexing='ij')
        freq_dist = torch.sqrt(fy**2 + fx**2)
        max_freq = math.sqrt(2) * ps
        
        for band in range(self.num_freq_bands):
            lo = band * max_freq / self.num_freq_bands
            hi = (band + 1) * max_freq / self.num_freq_bands
            mask = ((freq_dist >= lo) & (freq_dist < hi)).float()
            band_energy = (dct_patches ** 2 * mask.unsqueeze(0).unsqueeze(0)).sum(dim=(-2, -1))
            features.append(band_energy.mean(dim=1, keepdim=True))
            features.append(band_energy.std(dim=1, keepdim=True))
        
        total_energy = (dct_patches ** 2).sum(dim=(-2, -1))
        weighted_freq = (dct_patches ** 2 * freq_dist.unsqueeze(0).unsqueeze(0)).sum(dim=(-2, -1))
        spectral_centroid = weighted_freq / (total_energy + 1e-8)
        features.append(spectral_centroid.mean(dim=1, keepdim=True))
        features.append(spectral_centroid.std(dim=1, keepdim=True))
        
        mid = ps // 2
        low_mask = (freq_dist < mid).float()
        high_mask = (freq_dist >= mid).float()
        low_energy = (dct_patches ** 2 * low_mask).sum(dim=(-2, -1))
        high_energy = (dct_patches ** 2 * high_mask).sum(dim=(-2, -1))
        hl_ratio = high_energy / (low_energy + 1e-8)
        features.append(hl_ratio.mean(dim=1, keepdim=True))
        features.append(hl_ratio.std(dim=1, keepdim=True))
        
        dc_values = dct_patches[:, :, 0, 0]
        features.append(dc_values.mean(dim=1, keepdim=True))
        features.append(dc_values.std(dim=1, keepdim=True))
        
        return torch.cat(features, dim=1)


# ============================================================
# 3. FFT SPECTRAL ANALYSIS MODULE
# ============================================================

class FFTSpectralAnalyzer(nn.Module):
    """Azimuthally averaged power spectrum + 1/f deviation analysis."""
    def __init__(self, num_bins=32):
        super().__init__()
        self.num_bins = num_bins
    
    def forward(self, x):
        B, C, H, W = x.shape
        gray = 0.299 * x[:, 0] + 0.587 * x[:, 1] + 0.114 * x[:, 2]
        hann_y = torch.hann_window(H, device=x.device)
        hann_x = torch.hann_window(W, device=x.device)
        window = hann_y.unsqueeze(1) * hann_x.unsqueeze(0)
        gray = gray * window.unsqueeze(0)
        
        fft = torch.fft.fft2(gray)
        fft_shift = torch.fft.fftshift(fft)
        power = torch.abs(fft_shift) ** 2
        
        cy, cx = H // 2, W // 2
        y = torch.arange(H, device=x.device).float() - cy
        xx = torch.arange(W, device=x.device).float() - cx
        yy, xx = torch.meshgrid(y, xx, indexing='ij')
        radius = torch.sqrt(yy**2 + xx**2)
        
        max_radius = min(cy, cx)
        bin_width = max_radius / self.num_bins
        
        features = []
        for i in range(self.num_bins):
            r_lo = i * bin_width
            r_hi = (i + 1) * bin_width
            mask = ((radius >= r_lo) & (radius < r_hi)).float()
            count = mask.sum() + 1e-8
            bin_power = (power * mask.unsqueeze(0)).sum(dim=(-2, -1)) / count
            features.append(bin_power.unsqueeze(1))
        
        radial_spectrum = torch.cat(features, dim=1)
        log_spectrum = torch.log1p(radial_spectrum)
        
        log_freq = torch.log1p(torch.arange(self.num_bins, device=x.device).float() + 1)
        log_freq = log_freq.unsqueeze(0).expand(B, -1)
        
        xm = log_freq - log_freq.mean(dim=1, keepdim=True)
        ym = log_spectrum - log_spectrum.mean(dim=1, keepdim=True)
        slope = (xm * ym).sum(dim=1, keepdim=True) / ((xm**2).sum(dim=1, keepdim=True) + 1e-8)
        intercept = log_spectrum.mean(dim=1, keepdim=True) - slope * log_freq.mean(dim=1, keepdim=True)
        
        predicted = slope * log_freq + intercept
        residuals = log_spectrum - predicted
        residual_std = residuals.std(dim=1, keepdim=True)
        residual_max = residuals.max(dim=1, keepdim=True)[0]
        
        return torch.cat([log_spectrum, slope, intercept, residual_std, residual_max], dim=1)


# ============================================================
# 4. FREQUENCY-AWARE DETECTOR MODEL
# ============================================================

class FrequencyAwareDetector(nn.Module):
    """Multi-branch AI image detector: SwinV2 + SRM + DCT + FFT"""
    def __init__(self, backbone_name="microsoft/swinv2-tiny-patch4-window8-256",
                 num_labels=2, dct_patch_size=32, num_freq_bands=8, fft_bins=32):
        super().__init__()
        self.num_labels = num_labels
        self.supports_gradient_checkpointing = True
        
        self.backbone = Swinv2Model.from_pretrained(backbone_name)
        backbone_dim = self.backbone.config.hidden_size
        
        self.srm = SRMFilterBank()
        self.srm_encoder = nn.Sequential(
            nn.Conv2d(30, 64, kernel_size=3, stride=2, padding=1),
            nn.BatchNorm2d(64), nn.GELU(),
            nn.Conv2d(64, 128, kernel_size=3, stride=2, padding=1),
            nn.BatchNorm2d(128), nn.GELU(),
            nn.Conv2d(128, 256, kernel_size=3, stride=2, padding=1),
            nn.BatchNorm2d(256), nn.GELU(),
            nn.AdaptiveAvgPool2d((1, 1)), nn.Flatten(),
        )
        srm_dim = 256
        
        self.dct_analyzer = DCTFrequencyAnalyzer(patch_size=dct_patch_size, num_freq_bands=num_freq_bands)
        dct_dim = num_freq_bands * 2 + 6
        
        self.fft_analyzer = FFTSpectralAnalyzer(num_bins=fft_bins)
        fft_dim = fft_bins + 4
        
        freq_total_dim = srm_dim + dct_dim + fft_dim
        self.freq_proj = nn.Sequential(
            nn.Linear(freq_total_dim, 256), nn.GELU(), nn.Dropout(0.3),
            nn.Linear(256, 128),
        )
        
        self.classifier = nn.Sequential(
            nn.Linear(backbone_dim + 128, 512), nn.GELU(), nn.Dropout(0.3),
            nn.Linear(512, 128), nn.GELU(), nn.Dropout(0.1),
            nn.Linear(128, num_labels),
        )
        
        self.loss_fn = nn.CrossEntropyLoss(label_smoothing=0.1)
    
    def gradient_checkpointing_enable(self, gradient_checkpointing_kwargs=None):
        self.backbone.gradient_checkpointing_enable(gradient_checkpointing_kwargs=gradient_checkpointing_kwargs)
    
    def gradient_checkpointing_disable(self):
        self.backbone.gradient_checkpointing_disable()
    
    def forward(self, pixel_values, labels=None, **kwargs):
        backbone_out = self.backbone(pixel_values=pixel_values)
        semantic_feats = backbone_out.pooler_output
        
        srm_maps = self.srm(pixel_values)
        srm_feats = self.srm_encoder(srm_maps)
        
        dct_feats = self.dct_analyzer(pixel_values)
        fft_feats = self.fft_analyzer(pixel_values)
        
        freq_feats = torch.cat([srm_feats, dct_feats, fft_feats], dim=1)
        freq_proj = self.freq_proj(freq_feats)
        
        fused = torch.cat([semantic_feats, freq_proj], dim=1)
        logits = self.classifier(fused)
        
        loss = None
        if labels is not None:
            loss = self.loss_fn(logits, labels)
        
        return {"loss": loss, "logits": logits}


# ============================================================
# 5. DATASET & TRANSFORMS
# ============================================================

class CommunityForensicsDataset(Dataset):
    """Wraps OwensLab/CommunityForensics-Small for PyTorch training."""
    def __init__(self, hf_dataset, transform=None, is_train=True):
        self.dataset = hf_dataset
        self.transform = transform
        self.is_train = is_train
    
    def __len__(self):
        return len(self.dataset)
    
    def __getitem__(self, idx):
        item = self.dataset[idx]
        img_data = item['image_data']
        if isinstance(img_data, str):
            import base64
            img_data = base64.b64decode(img_data)
        elif isinstance(img_data, list):
            img_data = bytes(img_data)
        
        img = Image.open(io.BytesIO(img_data)).convert('RGB')
        
        if self.is_train:
            img = self._social_media_augment(img)
        
        if self.transform:
            pixel_values = self.transform(img)
        else:
            pixel_values = ToTensor()(img)
        
        return {'pixel_values': pixel_values, 'labels': item['label']}
    
    def _social_media_augment(self, img):
        """Simulate social media compression/resize artifacts for robustness."""
        if random.random() < 0.10:
            quality = random.randint(30, 95)
            buffer = io.BytesIO()
            img.save(buffer, format='JPEG', quality=quality)
            buffer.seek(0)
            img = Image.open(buffer).convert('RGB')
        
        if random.random() < 0.10:
            from PIL import ImageFilter
            radius = random.uniform(0.1, 2.0)
            img = img.filter(ImageFilter.GaussianBlur(radius=radius))
        
        if random.random() < 0.05:
            w, h = img.size
            scale = random.uniform(0.5, 0.9)
            small = img.resize((int(w*scale), int(h*scale)), Image.BILINEAR)
            img = small.resize((w, h), Image.BILINEAR)
        
        return img


# ============================================================
# 6. CUSTOM TRAINER
# ============================================================

class FreqDetectorTrainer(Trainer):
    def compute_loss(self, model, inputs, return_outputs=False, num_items_in_batch=None):
        labels = inputs.pop("labels")
        outputs = model(pixel_values=inputs["pixel_values"], labels=labels)
        loss = outputs["loss"]
        return (loss, outputs) if return_outputs else loss


# ============================================================
# 7. MAIN TRAINING SCRIPT
# ============================================================

def main():
    import argparse
    parser = argparse.ArgumentParser()
    parser.add_argument("--max_train_samples", type=int, default=None)
    parser.add_argument("--max_eval_samples", type=int, default=None)
    parser.add_argument("--num_train_epochs", type=int, default=5)
    parser.add_argument("--per_device_train_batch_size", type=int, default=16)
    parser.add_argument("--gradient_accumulation_steps", type=int, default=4)
    parser.add_argument("--learning_rate", type=float, default=2e-5)
    parser.add_argument("--output_dir", type=str, default="ai-image-detector")
    parser.add_argument("--hub_model_id", type=str, default="Reju983/ai-generated-image-detector")
    parser.add_argument("--image_size", type=int, default=256)
    parser.add_argument("--test_mode", action="store_true")
    args = parser.parse_args()
    
    print("=" * 60)
    print("AI-Generated Image Detector Training")
    print("Architecture: SwinV2 + SRM + DCT + FFT")
    print(f"Dataset: OwensLab/CommunityForensics-Small")
    print("=" * 60)
    
    trackio.init(project="ai-image-detector", name="swinv2-srm-dct-fft")
    
    print("\n[1/5] Loading dataset...")
    if args.test_mode:
        ds = hf_datasets.load_dataset(
            "OwensLab/CommunityForensics-Small", split="train[:200]", trust_remote_code=True,
        )
        ds = ds.train_test_split(test_size=0.5, seed=42)
        train_ds = ds["train"]
        eval_ds = ds["test"]
    else:
        full_ds = hf_datasets.load_dataset(
            "OwensLab/CommunityForensics-Small", split="train", trust_remote_code=True,
        )
        split = full_ds.train_test_split(test_size=0.05, seed=42, stratify_by_column="label")
        train_ds = split["train"]
        eval_ds = split["test"]
    
    if args.max_train_samples:
        train_ds = train_ds.select(range(min(args.max_train_samples, len(train_ds))))
    if args.max_eval_samples:
        eval_ds = eval_ds.select(range(min(args.max_eval_samples, len(eval_ds))))
    
    print(f"  Train: {len(train_ds)}, Eval: {len(eval_ds)}")
    
    img_size = args.image_size
    train_transform = Compose([
        RandomResizedCrop((img_size, img_size), scale=(0.8, 1.0)),
        RandomHorizontalFlip(),
        ColorJitter(brightness=0.1, contrast=0.1, saturation=0.1),
        ToTensor(),
        Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
    ])
    eval_transform = Compose([
        Resize((img_size + 32, img_size + 32)),
        CenterCrop((img_size, img_size)),
        ToTensor(),
        Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
    ])
    
    train_dataset = CommunityForensicsDataset(train_ds, transform=train_transform, is_train=True)
    eval_dataset = CommunityForensicsDataset(eval_ds, transform=eval_transform, is_train=False)
    
    print("\n[3/5] Building model...")
    model = FrequencyAwareDetector(
        backbone_name="microsoft/swinv2-tiny-patch4-window8-256",
        num_labels=2, dct_patch_size=32, num_freq_bands=8, fft_bins=32,
    )
    total_params = sum(p.numel() for p in model.parameters())
    print(f"  Parameters: {total_params:,}")
    
    accuracy_metric = evaluate.load("accuracy")
    def compute_metrics(eval_pred):
        predictions, labels = eval_pred
        if isinstance(predictions, dict): predictions = predictions["logits"]
        predictions = np.argmax(predictions, axis=1)
        acc = accuracy_metric.compute(predictions=predictions, references=labels)
        real_mask = labels == 0
        fake_mask = labels == 1
        real_acc = (predictions[real_mask] == labels[real_mask]).mean() if real_mask.sum() > 0 else 0
        fake_acc = (predictions[fake_mask] == labels[fake_mask]).mean() if fake_mask.sum() > 0 else 0
        return {"accuracy": acc["accuracy"], "real_accuracy": float(real_acc), "fake_accuracy": float(fake_acc)}
    
    def data_collator(features):
        pixel_values = torch.stack([f["pixel_values"] for f in features])
        labels = torch.tensor([f["labels"] for f in features], dtype=torch.long)
        return {"pixel_values": pixel_values, "labels": labels}
    
    training_args = TrainingArguments(
        output_dir=args.output_dir,
        remove_unused_columns=False,
        eval_strategy="epoch", save_strategy="epoch",
        learning_rate=args.learning_rate,
        per_device_train_batch_size=args.per_device_train_batch_size,
        gradient_accumulation_steps=args.gradient_accumulation_steps,
        per_device_eval_batch_size=args.per_device_train_batch_size,
        num_train_epochs=args.num_train_epochs,
        warmup_ratio=0.1, weight_decay=0.01, bf16=True,
        logging_steps=25, logging_strategy="steps", logging_first_step=True,
        disable_tqdm=True,
        load_best_model_at_end=True, metric_for_best_model="accuracy", greater_is_better=True,
        push_to_hub=True, hub_model_id=args.hub_model_id, hub_strategy="end",
        save_total_limit=2, dataloader_num_workers=4,
        gradient_checkpointing=True,
        report_to="trackio", run_name="swinv2-srm-dct-fft",
        label_names=["labels"],
    )
    
    trainer = FreqDetectorTrainer(
        model=model, args=training_args, data_collator=data_collator,
        train_dataset=train_dataset, eval_dataset=eval_dataset,
        compute_metrics=compute_metrics,
    )
    
    print("\n[5/5] Training...")
    trainer.train()
    
    metrics = trainer.evaluate()
    for k, v in metrics.items(): print(f"  {k}: {v}")
    
    save_dir = os.path.join(args.output_dir, "final_model")
    os.makedirs(save_dir, exist_ok=True)
    torch.save(model.state_dict(), os.path.join(save_dir, "model_state_dict.pt"))
    
    import json
    config = {
        "architecture": "FrequencyAwareDetector",
        "backbone_name": "microsoft/swinv2-tiny-patch4-window8-256",
        "num_labels": 2, "dct_patch_size": 32, "num_freq_bands": 8, "fft_bins": 32,
        "id2label": {"0": "real", "1": "ai_generated"},
        "label2id": {"real": 0, "ai_generated": 1},
    }
    with open(os.path.join(save_dir, "config.json"), "w") as f:
        json.dump(config, f, indent=2)
    
    trainer.push_to_hub(
        commit_message="AI-Generated Image Detector: SwinV2 + SRM + DCT + FFT",
        tags=["image-classification", "ai-image-detection", "deepfake-detection",
              "frequency-analysis", "swinv2", "srm", "dct", "fft"],
    )
    
    from huggingface_hub import HfApi
    api = HfApi()
    api.upload_file(
        path_or_fileobj=os.path.join(save_dir, "model_state_dict.pt"),
        path_in_repo="model_state_dict.pt", repo_id=args.hub_model_id,
    )
    api.upload_file(
        path_or_fileobj=os.path.join(save_dir, "config.json"),
        path_in_repo="detector_config.json", repo_id=args.hub_model_id,
    )
    
    print(f"\nDone! Model at: https://huggingface.co/{args.hub_model_id}")

if __name__ == "__main__":
    main()