app/model.py

import torch
import torch.nn as nn
import timm
from pathlib import Path
import logging
import os

logging.basicConfig(level=logging.INFO)
logger = logging.getLogger(__name__)


class EfficientNetDeepFakeDetector(nn.Module):
    """Frame-level EfficientNet-B0 with temporal mean-pooling."""

    FEAT_DIM = 1280

    def __init__(self, dropout: float = 0.4):
        super().__init__()

        # Backbone
        backbone = timm.create_model(
            'efficientnet_b0',
            pretrained=False,
            num_classes=0,
            global_pool='avg'
        )

        # Freeze BatchNorm layers
        for m in backbone.modules():
            if isinstance(m, (nn.BatchNorm2d, nn.SyncBatchNorm)):
                m.eval()
                for p in m.parameters():
                    p.requires_grad = False

        self.backbone = backbone

        # Classifier head
        self.head = nn.Sequential(
            nn.LayerNorm(self.FEAT_DIM),
            nn.Dropout(dropout),
            nn.Linear(self.FEAT_DIM, 256),
            nn.GELU(),
            nn.Dropout(dropout * 0.5),
            nn.Linear(256, 1)
        )

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        B, T, C, H, W = x.shape
        x = x.view(B * T, C, H, W)
        feat = self.backbone(x)
        feat = feat.view(B, T, self.FEAT_DIM)
        feat = feat.mean(dim=1)
        logit = self.head(feat).squeeze(-1)
        return logit


class DeepFakeModel:
    def __init__(self, model_path: str, device: str = "cpu"):
        self.device = torch.device(device)
        self.model = EfficientNetDeepFakeDetector(dropout=0.4).to(self.device)
        self._load_model(model_path)
        self.model.eval()
        logger.info(f"Model loaded on {self.device}")

    def _load_model(self, model_path: str):
        checkpoint = torch.load(model_path, map_location=self.device, weights_only=False)
        self.model.load_state_dict(checkpoint['model_state_dict'])
        logger.info(f"Loaded checkpoint from epoch {checkpoint.get('epoch', 'unknown')}")

    @torch.no_grad()
    def predict(self, video_tensor: torch.Tensor, threshold: float = 0.4) -> dict:
        """
        Predict if video is real or fake.
        
        Args:
            video_tensor: Tensor of shape (T, 3, H, W) or (1, T, 3, H, W)
            threshold: Decision threshold (default: 0.4 from notebook testing)
        
        Returns:
            dict with prediction, confidence, and probabilities
        """
        if video_tensor.dim() == 4:
            video_tensor = video_tensor.unsqueeze(0)

        video_tensor = video_tensor.to(self.device)
        logit = self.model(video_tensor)
        prob = torch.sigmoid(logit).item()
        
        # prob = P(REAL), because training used label 1=REAL, 0=FAKE
        prediction = "REAL" if prob >= threshold else "FAKE"
        confidence = prob if prediction == "REAL" else 1 - prob

        return {
            "prediction": prediction,
            "confidence": round(confidence, 4),
            "probability_real": round(prob, 4),
            "probability_fake": round(1 - prob, 4),
            "threshold": threshold
        }