Delete main.py

main.py

@@ -1,160 +0,0 @@
-import os
-import cv2
-import torch
-import shutil
-from fastapi import FastAPI, UploadFile, File
-from fastapi.middleware.cors import CORSMiddleware
-from transformers import VideoMAEImageProcessor, VideoMAEForVideoClassification
-from facenet_pytorch import MTCNN
-from PIL import Image
-
-app = FastAPI()
-
-# Enable CORS for the React frontend
-app.add_middleware(
-    CORSMiddleware,
-    allow_origins=["*"], 
-    allow_credentials=True,
-    allow_methods=["*"],
-    allow_headers=["*"],
-)
-
-# Initialize Models globally so they load once on startup
-print("Loading MTCNN Face Detector...")
-mtcnn = MTCNN(keep_all=False, select_largest=True, post_process=False)
-
-print("Loading Hugging Face Temporal Deepfake Detector (VideoMAE)...")
-model_name = "Ammar2k/videomae-base-finetuned-deepfake-subset"
-processor = VideoMAEImageProcessor.from_pretrained(model_name)
-model = VideoMAEForVideoClassification.from_pretrained(model_name)
-
-# Ensure temp directory exists
-os.makedirs("temp", exist_ok=True)
-
-def extract_faces_sequence(video_path, sequence_length=16):
-    """Extracts a sequence of continuous frames and tracks the face temporally."""
-    cap = cv2.VideoCapture(video_path)
-    frames_count = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
-    
-    if frames_count == 0:
-        return []
-        
-    # Try to get frames from the middle of the video
-    start_frame = max(0, (frames_count // 2) - (sequence_length // 2))
-    cap.set(cv2.CAP_PROP_POS_FRAMES, start_frame)
-    
-    faces = []
-    last_box = None
-    
-    for _ in range(sequence_length):
-        ret, frame = cap.read()
-        if not ret:
-            break
-            
-        # Convert BGR to RGB for MTCNN and PIL
-        frame_rgb = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
-        pil_img = Image.fromarray(frame_rgb)
-        
-        # Detect and crop face
-        boxes, _ = mtcnn.detect(pil_img)
-        
-        # Bounding box smoothing: fallback to last known box if detection fails on a frame
-        if boxes is not None and len(boxes) > 0:
-            box = boxes[0]
-            last_box = box
-        elif last_box is not None:
-            box = last_box
-        else:
-            continue # Skip if no face found yet
-            
-        # Add 30% padding around the face for better context
-        w = box[2] - box[0]
-        h = box[3] - box[1]
-        pad_w = int(w * 0.3)
-        pad_h = int(h * 0.3)
-        
-        x1 = max(0, int(box[0]) - pad_w)
-        y1 = max(0, int(box[1]) - pad_h)
-        x2 = min(pil_img.width, int(box[2]) + pad_w)
-        y2 = min(pil_img.height, int(box[3]) + pad_h)
-        
-        if x2 > x1 and y2 > y1:
-            face_crop = pil_img.crop((x1, y1, x2, y2))
-            faces.append(face_crop)
-            
-    cap.release()
-    
-    # VideoMAE requires exactly `sequence_length` frames. Pad by duplicating last frame if short.
-    if len(faces) == 0:
-        return []
-        
-    while len(faces) < sequence_length:
-        faces.append(faces[-1])
-        
-    return faces
-
-@app.post("/api/analyze")
-async def analyze_video(file: UploadFile = File(...)):
-    print(f"Received file: {file.filename}")
-    
-    temp_video_path = os.path.join("temp", file.filename)
-    with open(temp_video_path, "wb") as buffer:
-        shutil.copyfileobj(file.file, buffer)
-        
-    try:
-        print("Extracting facial sequence from video...")
-        # Extract 16 consecutive frames
-        faces = extract_faces_sequence(temp_video_path, sequence_length=16)
-        
-        if not faces:
-            return {
-                "isFake": False, 
-                "confidence": 0, 
-                "explanation": "Could not detect a clear face in the video sequence.",
-                "details": []
-            }
-            
-        print(f"Extracted {len(faces)} frame sequence. Running temporal inference...")
-        
-        # Prepare for VideoMAE 3D model
-        inputs = processor(list(faces), return_tensors="pt")
-        
-        # Run inference
-        with torch.no_grad():
-            outputs = model(**inputs)
-            
-        # Video classification models return logits for the whole sequence
-        probabilities = torch.nn.functional.softmax(outputs.logits[0], dim=-1)
-        
-        # Get highest probability label
-        predicted_class_idx = probabilities.argmax(-1).item()
-        label = model.config.id2label[predicted_class_idx].lower()
-        
-        is_fake = 'fake' in label
-        confidence = round(probabilities[predicted_class_idx].item() * 100, 2)
-            
-        explanation = "Our 3D Temporal AI analyzed facial motion and detected unnatural movement, micro-expressions, or spatial inconsistencies typical of deepfakes." if is_fake else "Our 3D Temporal AI analyzed the facial movement and found natural micro-expressions and consistent temporal flow."
-        
-        print(f"Result: isFake={is_fake}, confidence={confidence}%")
-        
-        return {
-            "isFake": is_fake,
-            "confidence": confidence,
-            "explanation": explanation,
-            "details": [
-                {"title": "Temporal Analysis", "desc": "Analyzed a continuous 16-frame clip using a VideoMAE 3D Transformer."},
-                {"title": "Motion Tracking", "desc": "Tracked facial landmarks across time to detect jitter, blending artifacts, and lip-sync inconsistencies."}
-            ]
-        }
-        
-    except Exception as e:
-        print(f"Error analyzing video: {e}")
-        return {"error": str(e)}
-        
-    finally:
-        if os.path.exists(temp_video_path):
-            os.remove(temp_video_path)
-
-@app.get("/")
-def health_check():
-    return {"status": "3D Temporal Backend is running!"}