Zenith-AI / backend /core /engine.py
Shads229's picture
Upload 14 files
f15f315 verified
import os, json, logging, time, base64, gc, asyncio, concurrent.futures
import cv2, numpy as np, torch
from pathlib import Path
from typing import List, Dict, Any, Optional, AsyncGenerator
from collections import Counter
from dataclasses import dataclass
from dotenv import load_dotenv
load_dotenv()
# Configuration
DEEPSEEK_API_URL = "https://ds2api-tau-woad.vercel.app/v1/chat/completions"
DEEPSEEK_API_KEY = "sk-ds2api-key-1-your-custom-key"
DEEPSEEK_MODEL = "deepseek-chat"
BASE_DIR = Path("video_analysis_pro")
OUTPUT_DIR, CACHE_DIR, REPORTS_DIR = BASE_DIR/"output", BASE_DIR/"cache", BASE_DIR/"reports"
for d in [BASE_DIR, OUTPUT_DIR, CACHE_DIR, REPORTS_DIR]: d.mkdir(parents=True, exist_ok=True)
logging.basicConfig(level=logging.INFO, format='%(asctime)s - %(levelname)s - %(message)s')
logger = logging.getLogger("ZenithEngine")
# Tools Availability
try:
from ultralytics import YOLO
YOLO_AVAILABLE = True
except ImportError:
YOLO_AVAILABLE = False
try:
from faster_whisper import WhisperModel
WHISPER_AVAILABLE = True
except ImportError:
WHISPER_AVAILABLE = False
@dataclass
class Frame:
path: Path
timestamp: float
metrics: Dict[str, float] = None
vision_content: str = ""
class DeepSeekClient:
def __init__(self):
self.api_url = DEEPSEEK_API_URL
self.api_key = DEEPSEEK_API_KEY
logger.info(f"✅ DeepSeek Client initialisé avec l'URL : {self.api_url}")
async def stream_content(self, model: str, messages: List[Dict[str, Any]], options: Dict[str, Any]) -> AsyncGenerator[Dict[str, Any], None]:
# Convertir les messages au format OpenAI compatible
formatted_messages = []
for msg in messages:
role = msg["role"]
content = msg.get("content", "")
# Si le contenu contient des images, on les convertit en format texte + images
if isinstance(content, list):
text_parts = []
image_parts = []
for part in content:
if part["type"] == "text":
text_parts.append(part["text"])
elif part["type"] == "image_url":
url = part["image_url"]["url"]
if url.startswith("data:"):
image_parts.append({"type": "image_url", "image_url": {"url": url}})
# DeepSeek supporte le format OpenAI vision
if image_parts:
formatted_messages.append({
"role": role,
"content": [{"type": "text", "text": " ".join(text_parts)}] + image_parts
})
else:
formatted_messages.append({"role": role, "content": " ".join(text_parts)})
else:
formatted_messages.append({"role": role, "content": content})
payload = {
"model": model,
"messages": formatted_messages,
"temperature": options.get("temperature", 0.7),
"stream": True
}
import httpx
async with httpx.AsyncClient(timeout=None) as client:
try:
async with client.stream(
"POST", self.api_url,
headers={
"Authorization": f"Bearer {self.api_key}",
"Content-Type": "application/json"
},
json=payload
) as response:
if response.status_code != 200:
error_text = await response.aread()
logger.error(f"❌ Erreur DeepSeek API (HTTP {response.status_code}): {error_text.decode()}")
yield {"error": f"Erreur API DeepSeek: {response.status_code}"}
return
async for line in response.aiter_lines():
if line.startswith("data: "):
data_str = line[6:]
if data_str.strip() == "[DONE]":
break
try:
data = json.loads(data_str)
# Format OpenAI streaming response
if "choices" in data and len(data["choices"]) > 0:
delta = data["choices"][0].get("delta", {})
content = delta.get("content", "")
if content:
# Convertir au format attendu par le frontend
yield {
"response": {
"candidates": [{
"content": {
"parts": [{"text": content}]
}
}]
}
}
except json.JSONDecodeError:
continue
except Exception as e:
logger.error(f"❌ Erreur lors du streaming DeepSeek : {str(e)}")
yield {"error": str(e)}
class VideoProcessor:
@staticmethod
def get_frame_metrics(frame: np.ndarray) -> dict:
try:
gray = cv2.cvtColor(frame, cv2.COLOR_BGR2GRAY)
hsv = cv2.cvtColor(frame, cv2.COLOR_BGR2HSV)
return {"brightness": float(np.mean(gray)), "contrast": float(np.std(gray)),
"saturation": float(np.mean(hsv[:, :, 1])), "sharpness": float(cv2.Laplacian(gray, cv2.CV_64F).var())}
except: return {"brightness": 0, "contrast": 0, "saturation": 0, "sharpness": 0}
def __init__(self, video_path: Path, output_dir: Path):
self.video_path, self.output_dir = video_path, output_dir
self.output_dir.mkdir(parents=True, exist_ok=True)
def extract_keyframes(self, max_frames: int = 30) -> List[Frame]:
"""
Extraction intelligente de keyframes avec échantillonnage adaptatif.
- Vidéos courtes (<2min) : 1 frame toutes les 3-4s
- Vidéos moyennes (2-10min) : 1 frame toutes les 10-15s
- Vidéos longues (>10min) : 1 frame toutes les 20-30s
"""
try:
from decord import VideoReader, cpu
vr = VideoReader(str(self.video_path), ctx=cpu(0))
total = len(vr)
fps = vr.get_avg_fps()
duration_seconds = total / fps
# Échantillonnage adaptatif basé sur la durée
if duration_seconds < 120: # < 2 minutes
target_interval = 3 # 1 frame toutes les 3 secondes
elif duration_seconds < 600: # 2-10 minutes
target_interval = 12 # 1 frame toutes les 12 secondes
else: # > 10 minutes
target_interval = 25 # 1 frame toutes les 25 secondes
# Calculer le nombre de frames optimal
optimal_frames = min(int(duration_seconds / target_interval), max_frames)
optimal_frames = max(optimal_frames, 10) # Minimum 10 frames
step = max(1, total // optimal_frames)
indices = range(0, total, step)[:optimal_frames]
frames_data = vr.get_batch(indices).asnumpy()
extracted = []
for i, idx in enumerate(indices):
img = cv2.cvtColor(frames_data[i], cv2.COLOR_RGB2BGR)
ts = idx / fps
p = self.output_dir / f"f_{idx}.jpg"
cv2.imwrite(str(p), img, [cv2.IMWRITE_JPEG_QUALITY, 70])
extracted.append(Frame(path=p, timestamp=ts, metrics=self.get_frame_metrics(img)))
logger.info(f"✅ Extraction adaptative : {len(extracted)} frames pour {duration_seconds:.1f}s de vidéo (1 frame/{target_interval}s)")
return extracted
except Exception as e:
logger.warning(f"Decord failed, fallback to CV2: {e}")
cap = cv2.VideoCapture(str(self.video_path))
fps = cap.get(cv2.CAP_PROP_FPS) or 30.0
total = int(cap.get(cv2.CAP_PROP_FRAME_COUNT)) or 1000
duration_seconds = total / fps
# Même logique adaptative pour le fallback CV2
if duration_seconds < 120:
target_interval = 3
elif duration_seconds < 600:
target_interval = 12
else:
target_interval = 25
optimal_frames = min(int(duration_seconds / target_interval), max_frames)
optimal_frames = max(optimal_frames, 10)
step = max(1, total // optimal_frames)
extracted = []
for idx in range(0, total, step):
if len(extracted) >= optimal_frames: break
cap.set(cv2.CAP_PROP_POS_FRAMES, idx)
ret, img = cap.read()
if ret:
ts = idx / fps
p = self.output_dir / f"f_{idx}.jpg"
cv2.imwrite(str(p), img, [cv2.IMWRITE_JPEG_QUALITY, 70])
extracted.append(Frame(path=p, timestamp=ts, metrics=self.get_frame_metrics(img)))
cap.release()
logger.info(f"✅ Extraction CV2 adaptative : {len(extracted)} frames pour {duration_seconds:.1f}s de vidéo")
return extracted
class AudioProcessor:
def __init__(self): self.model = None
def initialize(self):
if WHISPER_AVAILABLE and self.model is None:
try:
device = "cuda" if torch.cuda.is_available() else "cpu"
# Utiliser tiny au lieu de base pour plus de rapidité
self.model = WhisperModel("tiny", device=device, compute_type="int8")
except: pass
def transcribe(self, p: Path) -> str:
self.initialize()
if not self.model: return "Transcription indisponible"
try:
segments, info = self.model.transcribe(str(p), beam_size=5)
transcript = " ".join([s.text for s in segments])
return f"[Langue source détectée: {info.language.upper()}] {transcript}"
except: return "Erreur transcription"
class VideoDownloader:
@staticmethod
def download(url: str, output_dir: Path) -> Optional[Path]:
import yt_dlp
ydl_opts = {
'format': 'bestvideo[ext=mp4]+bestaudio[ext=m4a]/best[ext=mp4]/best',
'outtmpl': str(output_dir / 'downloaded_video.%(ext)s'),
'noplaylist': True, 'quiet': True, 'no_warnings': True, 'nocheckcertificate': True,
'user_agent': 'Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/120.0.0.0 Safari/537.36',
'referer': 'https://www.google.com/',
'http_headers': {'Accept': '*/*', 'Accept-Language': 'en-US,en;q=0.9'}
}
try:
with yt_dlp.YoutubeDL(ydl_opts) as ydl:
info = ydl.extract_info(url, download=True)
return Path(ydl.prepare_filename(info))
except: return None
class ZenithAnalyzer:
def __init__(self):
self.deepseek = DeepSeekClient()
self.audio_proc = AudioProcessor()
self.yolo = YOLO("yolov8n.pt") if YOLO_AVAILABLE else None
async def extract_frames_only(self, video_path: Path, session_id: str) -> List[str]:
session_dir = OUTPUT_DIR / f"session_{session_id}"
session_dir.mkdir(parents=True, exist_ok=True)
proc = VideoProcessor(video_path, session_dir)
frames = proc.extract_keyframes()
return [f"/output/session_{session_id}/{f.path.name}" for f in frames[:12]]
async def run_full_analysis(self, video_path: Path, session_id: str, custom_prompt: Optional[str] = None) -> AsyncGenerator[Dict[str, Any], None]:
session_dir = OUTPUT_DIR / f"session_{session_id}"
session_dir.mkdir(parents=True, exist_ok=True)
cache_file = session_dir / "analysis_cache.json"
# Optimisation : Ne pas ré-extraire si les frames existent déjà
existing_frames = list(session_dir.glob("f_*.jpg"))
if not existing_frames:
yield {"status": "sampling", "message": "Analyse des séquences..."}
proc = VideoProcessor(video_path, session_dir)
frames = proc.extract_keyframes()
else:
def get_idx(p):
try: return int(p.stem.split('_')[1])
except: return 0
existing_paths = sorted(existing_frames, key=get_idx)
frames = []
for p in existing_paths:
img = cv2.imread(str(p))
metrics = VideoProcessor.get_frame_metrics(img) if img is not None else {"brightness": 0, "contrast": 0, "saturation": 0, "sharpness": 0}
frames.append(Frame(path=p, timestamp=0.0, metrics=metrics))
yield {"status": "sampling", "message": "Récupération des séquences existantes..."}
# Envoyer les chemins des images au frontend
frame_urls = [f"/output/session_{session_id}/{f.path.name}" for f in frames[:12]]
yield {"status": "frames_ready", "frames": frame_urls, "message": "Séquences prêtes."}
# Vérifier si on a un cache pour l'audio et le visuel
cached_data = {}
if cache_file.exists():
try:
with open(cache_file, "r") as f:
cached_data = json.load(f)
logger.info(f"✅ Cache trouvé pour la session {session_id}")
except: pass
if "transcript" in cached_data and "vision_info" in cached_data:
transcript = cached_data["transcript"]
v_info = cached_data["vision_info"]
yield {"status": "fusion", "message": "Utilisation des données en cache..."}
else:
yield {"status": "audio", "message": "Traitement audio & visuel..."}
loop = asyncio.get_event_loop()
with concurrent.futures.ThreadPoolExecutor() as executor:
audio_task = loop.run_in_executor(executor, self.audio_proc.transcribe, video_path)
if self.yolo:
all_paths = [str(f.path) for f in frames]
batch_size = 20
for i in range(0, len(all_paths), batch_size):
batch = all_paths[i:i+batch_size]
results = await loop.run_in_executor(executor, lambda: self.yolo(batch, verbose=False, imgsz=256, stream=False))
for j, res in enumerate(results):
idx = i + j
objs = [res.names[int(b.cls[0])] for b in res.boxes if b.conf > 0.35]
ambiance = f"Ambiance: {'Sombre' if frames[idx].metrics['brightness'] < 50 else 'Lumineuse'}"
frames[idx].vision_content = f"{ambiance}, Objets: " + ", ".join([f"{v}x {k}" for k,v in Counter(objs).items()])
transcript = await audio_task
v_info = "\n".join([f"[{f.timestamp:.1f}s] {f.vision_content}" for f in frames[:40]])
# Sauvegarder dans le cache
try:
with open(cache_file, "w") as f:
json.dump({"transcript": transcript, "vision_info": v_info}, f)
except: pass
yield {"status": "fusion", "message": "Intelligence Artificielle en action..."}
# Utilisation du prompt personnalisé si fourni
base_instruction = custom_prompt if custom_prompt else "Résumer et continuer l'analyse du média"
prompt = f"""Tu es l'unité Zenith AI, un système d'analyse de données multimédias.
INSTRUCTION UTILISATEUR : {base_instruction}
DONNÉES D'ENTRÉE :
- TRANSCRIPTION : {transcript}
- DONNÉES VISUELLES : {v_info}
Produis un rapport TECHNIQUE, FACTUEL et STRUCTURÉ en Markdown."""
# Encodage parallèle des images - Sélection intelligente et équilibrée
# On prend des images réparties uniformément sur toute la durée
num_images_to_send = min(8, len(frames)) # Max 8 images pour l'IA
if len(frames) > 0:
step = max(1, len(frames) // num_images_to_send)
selected_frames = [frames[i] for i in range(0, len(frames), step)][:num_images_to_send]
else:
selected_frames = []
def encode_f(f):
img = cv2.imread(str(f.path))
# Redimensionner pour réduire la taille tout en gardant la qualité visuelle
img = cv2.resize(img, (800, 450), interpolation=cv2.INTER_AREA)
_, buf = cv2.imencode('.jpg', img, [cv2.IMWRITE_JPEG_QUALITY, 65])
return {"type": "image_url", "image_url": {"url": f"data:image/jpeg;base64,{base64.b64encode(buf).decode()}"}}
with concurrent.futures.ThreadPoolExecutor() as executor:
images = list(executor.map(encode_f, selected_frames))
messages = [{"role": "user", "content": [{"type": "text", "text": prompt}] + images}]
yield {"status": "generating", "message": "Génération du rapport par l'IA..."}
async for chunk in self.deepseek.stream_content(DEEPSEEK_MODEL, messages, {"temperature": 0.7}):
if "error" in chunk:
yield {"error": chunk["error"]}
break
resp = chunk.get("response", {})
candidates = resp.get("candidates", [])
if candidates:
for part in candidates[0].get("content", {}).get("parts", []):
text = part.get("text", "")
if text: yield {"status": "streaming", "text": text}
# Cleanup
gc.collect()
if torch.cuda.is_available(): torch.cuda.empty_cache()
yield {"status": "completed", "message": "Analyse terminée."}