app.py

"""
DeepShield AI — Full-Stack FastAPI Backend (SupCon Version)
Serves the frontend UI + deepfake detection API from one HF Space.
98.3% Accuracy — Supervised Contrastive Learning Model
"""

import os
import sys
import uuid
import shutil
import logging
import tempfile
from pathlib import Path
from functools import lru_cache

import cv2
import torch
import torch.nn as nn
import torch.nn.functional as F
import numpy as np
from PIL import Image, ImageFile
from facenet_pytorch import MTCNN
from fastapi import FastAPI, File, UploadFile, HTTPException
from fastapi.middleware.cors import CORSMiddleware
from fastapi.responses import JSONResponse, FileResponse
from fastapi.staticfiles import StaticFiles
import torchvision.transforms as T

ImageFile.LOAD_TRUNCATED_IMAGES = True
logging.basicConfig(level=logging.INFO, format="%(asctime)s %(levelname)s %(message)s")
logger = logging.getLogger(__name__)

# ─────────────────────────────────────────────
# Model Definition (Self-Contained SupCon Architecture)
# ─────────────────────────────────────────────

class DINOv2Extractor(nn.Module):
    def __init__(self, variant: str = "dinov2_vitb14"):
        super().__init__()
        logger.info(f"Loading {variant} from torch.hub...")
        self.backbone = torch.hub.load(
            "facebookresearch/dinov2", variant, pretrained=True
        )
        self.feature_dim = 768
        for p in self.backbone.parameters():
            p.requires_grad = False

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        return self.backbone(x)

class MLPClassifier(nn.Module):
    def __init__(self, input_dim: int, num_classes: int = 2, dropout: float = 0.4):
        super().__init__()
        self.net = nn.Sequential(
            nn.Linear(input_dim, 512),
            nn.BatchNorm1d(512),
            nn.GELU(),
            nn.Dropout(dropout),
            nn.Linear(512, 256),
            nn.BatchNorm1d(256),
            nn.GELU(),
            nn.Dropout(dropout * 0.75),
            nn.Linear(256, num_classes),
        )

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        return self.net(x)

class SupConDeepfakeClassifier(nn.Module):
    """
    Supervised Contrastive Version of the DINOv2 Deepfake Detector.
    Matches the architecture used in scripts3.
    """
    def __init__(self, dual_input: bool = True, proj_dim: int = 128):
        super().__init__()
        self.dual_input = dual_input
        self.extractor = DINOv2Extractor()
        
        feat_dim = 768
        classifier_input = feat_dim * 2 if dual_input else feat_dim
        
        # Projection Head for SupCon (needed for weight loading, even if not used in inference)
        self.head = nn.Sequential(
            nn.Linear(classifier_input, classifier_input),
            nn.BatchNorm1d(classifier_input),
            nn.ReLU(inplace=True),
            nn.Linear(classifier_input, proj_dim)
        )
        
        self.classifier = MLPClassifier(classifier_input)

    def forward(self, full_image: torch.Tensor, face_crop: torch.Tensor = None):
        full_feat = self.extractor(full_image)
        if self.dual_input:
            face_feat = self.extractor(face_crop if face_crop is not None else full_image)
            features = torch.cat([full_feat, face_feat], dim=1)
        else:
            features = full_feat
            
        logits = self.classifier(features)
        # We don't need 'proj' for inference
        return logits

# ─────────────────────────────────────────────
# App Setup
# ─────────────────────────────────────────────

app = FastAPI(
    title="DeepShield AI",
    description="DINO-G50 deepfake detector — SupCon SOTA version",
    version="3.0.0",
)

app.add_middleware(
    CORSMiddleware,
    allow_origins=["*"],
    allow_credentials=True,
    allow_methods=["*"],
    allow_headers=["*"],
)

DEVICE = torch.device("cuda" if torch.cuda.is_available() else "cpu")
CHECKPOINT_PATH = Path("best_model.pth")
MAX_FRAMES = 20
MAX_FILE_MB = 30
MAX_DURATION_SEC = 60

# MTCNN face detector
try:
    MTCNN_DETECTOR = MTCNN(
        image_size=224,
        margin=40,
        keep_all=False,
        post_process=False,
        device='cpu'
    )
    logger.info("MTCNN face detector initialized.")
except Exception as e:
    MTCNN_DETECTOR = None
    logger.warning(f"MTCNN init failed: {e}")

TRANSFORM = T.Compose([
    T.Resize((224, 224)),
    T.CenterCrop(224),
    T.ToTensor(),
    T.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
])

def detect_face_crop(img: Image.Image) -> Image.Image:
    if MTCNN_DETECTOR is None:
        return None
    try:
        boxes, probs = MTCNN_DETECTOR.detect(img)
        if boxes is None or len(boxes) == 0:
            return None
        
        best_idx = np.argmax(probs)
        if probs[best_idx] < 0.9:
            return None
            
        box = boxes[best_idx]
        w, h = img.size
        x1, y1, x2, y2 = [int(b) for b in box]
        margin = 40
        x1, y1 = max(0, x1-margin), max(0, y1-margin)
        x2, y2 = min(w, x2+margin), min(h, y2+margin)
        
        face = img.crop((x1, y1, x2, y2))
        return face.resize((224, 224), Image.LANCZOS)
    except Exception:
        pass
    return None

@lru_cache(maxsize=1)
def load_model() -> SupConDeepfakeClassifier:
    if not CHECKPOINT_PATH.exists():
        fallback = Path("models3/checkpoints/best_model.pth")
        if fallback.exists():
            shutil.copy(fallback, CHECKPOINT_PATH)
        else:
            raise RuntimeError("best_model.pth not found. Please upload the model from models3/.")

    logger.info(f"Loading SupCon checkpoint on {DEVICE}...")
    ckpt = torch.load(CHECKPOINT_PATH, map_location=DEVICE)
    state = ckpt.get("model_state_dict", ckpt)

    # Auto-detect dual input from weights
    mlp_w = state.get("classifier.net.0.weight", None)
    dual = (mlp_w.shape[1] == 1536) if mlp_w is not None else True

    model = SupConDeepfakeClassifier(dual_input=dual).to(DEVICE)
    model.load_state_dict(state, strict=False)
    model.eval()
    logger.info(f"SupCon Model ready. dual_input={dual}, device={DEVICE}")
    return model

def extract_frames(video_path: str, output_dir: str, num_frames: int = MAX_FRAMES) -> list:
    cap = cv2.VideoCapture(video_path)
    if not cap.isOpened():
        raise ValueError("Cannot open video file.")
    total_frames = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
    if total_frames <= 0: total_frames = 300
    step = max(1, total_frames // num_frames)
    target_indices = set(range(0, total_frames, step))
    saved_paths = []
    frame_idx = 0
    while len(saved_paths) < num_frames:
        ret, frame = cap.read()
        if not ret: break
        if frame_idx in target_indices:
            rgb = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
            path = os.path.join(output_dir, f"frame_{len(saved_paths):04d}.jpg")
            Image.fromarray(rgb).save(path, quality=90)
            saved_paths.append(path)
        frame_idx += 1
    cap.release()
    return saved_paths

def run_inference(model: SupConDeepfakeClassifier, frame_paths: list) -> dict:
    fake_probs = []
    with torch.no_grad():
        for fpath in frame_paths:
            try:
                img = Image.open(fpath).convert("RGB")
                t_img = TRANSFORM(img).unsqueeze(0).to(DEVICE)
                t_face = t_img
                if model.dual_input:
                    face_crop = detect_face_crop(img)
                    if face_crop is not None:
                        t_face = TRANSFORM(face_crop).unsqueeze(0).to(DEVICE)

                logits = model(t_img, t_face if model.dual_input else None)
                prob = torch.softmax(logits, dim=1)[0, 1].item()
                fake_probs.append(prob)
            except Exception as e:
                logger.warning(f"Error on {fpath}: {e}")

    if not fake_probs: raise ValueError("No frames processed.")
    
    # Matching test_real.py simple mean logic for consistency
    video_fake_prob = float(np.mean(fake_probs))
    is_fake = video_fake_prob > 0.5
    avg_real = 1.0 - video_fake_prob

    return {
        "verdict": "FAKE" if is_fake else "REAL",
        "fake_probability": round(video_fake_prob * 100, 1),
        "real_probability": round(avg_real * 100, 1),
        "frame_count": len(fake_probs),
        "confidence": round(max(video_fake_prob, avg_real) * 100, 1),
        "per_frame_scores": [round(p * 100, 1) for p in fake_probs],
    }

@app.on_event("startup")
async def startup_event():
    try:
        load_model()
    except Exception as e:
        logger.error(f"Startup model load failed: {e}")

@app.get("/health")
def health_check():
    return {
        "status": "ok",
        "model": "DINO-G50 SupCon Detector",
        "model_loaded": CHECKPOINT_PATH.exists(),
    }

@app.post("/predict")
async def predict(file: UploadFile = File(...)):
    allowed_exts = {".mp4", ".mov", ".avi", ".mkv", ".jpg", ".jpeg", ".png", ".webp"}
    ext = Path(file.filename).suffix.lower() if file.filename else ""
    if ext not in allowed_exts:
        raise HTTPException(400, f"Unsupported file type '{ext}'.")

    content = await file.read()
    size_mb = len(content) / (1024 * 1024)
    if size_mb > MAX_FILE_MB:
        raise HTTPException(413, f"File too large ({size_mb:.1f} MB). Max: {MAX_FILE_MB} MB.")

    job_id = str(uuid.uuid4())[:8]
    temp_dir = Path(tempfile.gettempdir()) / f"deepshield_{job_id}"
    frames_dir = temp_dir / "frames"
    frames_dir.mkdir(parents=True, exist_ok=True)
    file_path = temp_dir / f"input{ext}"

    try:
        with open(file_path, "wb") as f:
            f.write(content)
        del content
        model = load_model()
        
        if ext in {".mp4", ".mov", ".avi", ".mkv"}:
            frame_paths = extract_frames(str(file_path), str(frames_dir))
        else:
            img_path = frames_dir / f"frame_0000{ext}"
            shutil.copy(file_path, img_path)
            frame_paths = [str(img_path)]

        if not frame_paths: raise HTTPException(422, "Failed to extract frames.")
        
        result = run_inference(model, frame_paths)
        result.update({"filename": file.filename, "file_size_mb": round(size_mb, 2)})
        return JSONResponse(content=result)
    except Exception as e:
        logger.error(f"Error: {e}", exc_info=True)
        raise HTTPException(500, str(e))
    finally:
        shutil.rmtree(temp_dir, ignore_errors=True)

app.mount("/", StaticFiles(directory="static", html=True), name="static")