app.py

"""
TUNILip+ — HuggingFace Spaces
Nouveau endpoint /extract-features-frames :
  Reçoit 16 frames JPEG base64 déjà croppées sur la bouche (par MediaPipe côté browser)
  → VideoMAE frozen → mean-pool spatial → (8, 768)
  → Identique au pipeline d'entraînement Kaggle
"""

from fastapi import FastAPI, UploadFile, File, HTTPException, Form
from fastapi.middleware.cors import CORSMiddleware
from fastapi.responses import JSONResponse
from typing import List
import numpy as np
import cv2
import torch
import tempfile
import os
import base64
import logging
from contextlib import asynccontextmanager

logging.basicConfig(level=logging.INFO)
logger = logging.getLogger("tunilip")

vmae_processor = None
vmae_model     = None
DEVICE         = None
VMAE_MODEL_ID  = "MCG-NJU/videomae-base"
NUM_FRAMES     = 16


@asynccontextmanager
async def lifespan(app: FastAPI):
    global vmae_processor, vmae_model, DEVICE
    logger.info(f"⏳ Chargement {VMAE_MODEL_ID} …")
    try:
        from transformers import VideoMAEModel, VideoMAEImageProcessor
        DEVICE = torch.device("cuda" if torch.cuda.is_available() else "cpu")
        logger.info(f"   Device : {DEVICE}")
        vmae_processor = VideoMAEImageProcessor.from_pretrained(VMAE_MODEL_ID)
        vmae_model     = VideoMAEModel.from_pretrained(VMAE_MODEL_ID)
        vmae_model.eval()
        vmae_model = vmae_model.to(DEVICE)
        for p in vmae_model.parameters():
            p.requires_grad = False
        n = sum(p.numel() for p in vmae_model.parameters())
        logger.info(f"✅ VideoMAE chargé — {n:,} params (GELÉS) sur {DEVICE}")
    except Exception as e:
        logger.error(f"❌ Erreur chargement VideoMAE : {e}")
    yield
    logger.info("Shutdown")


app = FastAPI(title="TUNILip+ Feature Extractor", lifespan=lifespan)

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


# ── Helper : base64 → numpy (H, W, 3) uint8 ──────────────────
def b64_to_frame(b64str: str) -> np.ndarray:
    img_bytes = base64.b64decode(b64str)
    arr = np.frombuffer(img_bytes, dtype=np.uint8)
    img = cv2.imdecode(arr, cv2.IMREAD_COLOR)
    if img is None:
        raise ValueError("Impossible de décoder une frame JPEG")
    img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
    # S'assurer que c'est bien 224×224
    if img.shape[:2] != (224, 224):
        img = cv2.resize(img, (224, 224))
    return img   # uint8 RGB


@torch.no_grad()
def run_videomae(frames_np: List[np.ndarray]) -> np.ndarray:
    """
    frames_np : liste de 16 arrays uint8 (224, 224, 3) RGB
    Retourne  : np.ndarray float32 (8, 768)
    Identique à extract_videomae_features() du notebook Kaggle.
    """
    if vmae_model is None or vmae_processor is None:
        raise RuntimeError("VideoMAE non chargé")

    # vmae_processor attend une liste de arrays uint8 (H, W, 3)
    inputs = vmae_processor(frames_np, return_tensors="pt")
    inputs = {k: v.to(DEVICE) for k, v in inputs.items()}

    out    = vmae_model(**inputs)
    hidden = out.last_hidden_state.squeeze(0).cpu().numpy()  # (1568, 768)

    # VideoMAE-base : 8 temporal × 196 spatial = 1568
    T_temp, T_spat = 8, 196
    hidden = hidden[:T_temp * T_spat].reshape(T_temp, T_spat, 768)
    hidden = hidden.mean(axis=1)   # mean-pool spatial → (8, 768)

    logger.info(f"Features stats — mean:{hidden.mean():.4f} std:{hidden.std():.4f}")
    return hidden.astype(np.float32)


# ══════════════════════════════════════════════════════════════
# ROUTES
# ══════════════════════════════════════════════════════════════

@app.get("/health")
def health():
    return {
        "status": "ok",
        "model_ready": vmae_model is not None,
        "device": str(DEVICE) if DEVICE else "unknown",
        "model_id": VMAE_MODEL_ID,
    }


@app.post("/extract-features-frames")
async def extract_features_frames(frames_json: str = Form(...)):
    """
    Reçoit : FormData { frames_json: "['<base64>', ...]" }
    Retourne: { "features": [[...], ...], "shape": [8, 768] }
    Utilise FormData (multipart) pour éviter le preflight CORS de HuggingFace.
    """
    if vmae_model is None:
        raise HTTPException(status_code=503, detail="VideoMAE non chargé")

    try:
        import json as _json
        frames_list = _json.loads(frames_json)
    except Exception:
        raise HTTPException(status_code=422, detail="frames_json invalide")

    n = len(frames_list)
    if n == 0:
        raise HTTPException(status_code=422, detail="Aucune frame reçue")

    # Padding ou troncature à NUM_FRAMES
    frames_b64 = frames_list[:NUM_FRAMES]
    while len(frames_b64) < NUM_FRAMES:
        frames_b64.append(frames_b64[-1])

    try:
        frames_np = [b64_to_frame(f) for f in frames_b64]
    except Exception as e:
        raise HTTPException(status_code=422, detail=f"Erreur décodage frames: {e}")

    try:
        features = run_videomae(frames_np)
    except Exception as e:
        logger.error(f"Erreur VideoMAE : {e}", exc_info=True)
        raise HTTPException(status_code=500, detail=str(e))

    return JSONResponse({
        "features": features.tolist(),
        "shape": list(features.shape),
        "model_id": VMAE_MODEL_ID,
        "frames_received": n,
    })


# Ancien endpoint gardé pour compatibilité (envoie vidéo brute)
@app.post("/extract-features")
async def extract_features(video: UploadFile = File(...)):
    """Endpoint legacy — envoie vidéo brute (sans crop bouche)."""
    suffix = os.path.splitext(video.filename or "video.mp4")[-1] or ".mp4"
    with tempfile.NamedTemporaryFile(delete=False, suffix=suffix) as tmp:
        content = await video.read()
        tmp.write(content)
        tmp_path = tmp.name
    try:
        cap = cv2.VideoCapture(tmp_path)
        total = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
        if total == 0:
            cap.release()
            raise HTTPException(status_code=422, detail="Vidéo vide")
        indices = np.linspace(0, total - 1, NUM_FRAMES, dtype=int)
        frames_np = []
        for idx in indices:
            cap.set(cv2.CAP_PROP_POS_FRAMES, int(idx))
            ret, frame = cap.read()
            if ret:
                frame = cv2.resize(frame, (224, 224))
                frame = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
                frames_np.append(frame)
        cap.release()
        while len(frames_np) < NUM_FRAMES:
            frames_np.append(np.zeros((224, 224, 3), dtype=np.uint8))

        features = run_videomae(frames_np)
        return JSONResponse({
            "features": features.tolist(),
            "shape": list(features.shape),
            "model_id": VMAE_MODEL_ID,
        })
    except HTTPException:
        raise
    except Exception as e:
        raise HTTPException(status_code=500, detail=str(e))
    finally:
        os.unlink(tmp_path)


@app.get("/")
def root():
    return {
        "service": "TUNILip+ VideoMAE Feature Extractor",
        "endpoints": {
            "POST /extract-features-frames": "Frames base64 croppées → (8,768) — RECOMMANDÉ",
            "POST /extract-features":        "Vidéo brute → (8,768) — legacy",
            "GET  /health":                  "Statut",
        }
    }


if __name__ == "__main__":
    import uvicorn
    uvicorn.run(app, host="0.0.0.0", port=7860)