app.py

mport os
import torch
import io
from fastapi import FastAPI, File, UploadFile
from transformers import AutoProcessor, AutoModelForCausalLM
from ultralytics import YOLO
from PIL import Image
import uvicorn

# --- 1. إعداد التطبيق والموديلات ---
app = FastAPI(title="YOLO + GIT Large: Final Visual Description API")

device = "cuda" if torch.cuda.is_available() else "cpu"
MY_MODEL_PATH = 'best.pt'

print(f"🔄 جاري التحميل على جهاز: {device}...")

# تحميل YOLO الخاص بكِ
try:
    detection_model = YOLO(MY_MODEL_PATH)
    print("✅ YOLO Model: Loaded successfully")
except Exception as e:
    print(f"⚠️ YOLO Warning: Using default yolov8n.pt - {e}")
    detection_model = YOLO("yolov8n.pt")

# تحميل موديل الوصف GIT-Large
model_name = "microsoft/git-large"
processor = AutoProcessor.from_pretrained(model_name)
caption_model = AutoModelForCausalLM.from_pretrained(model_name).to(device)
print(f"✅ Caption Model: {model_name} Loaded")

@app.get("/")
def home():
    return {"status": "Online", "instruction": "Use /docs to test the /analyze endpoint"}

# --- 2. وظيفة المعالجة والتحليل ---

@app.post("/analyze")
async def analyze_image(file: UploadFile = File(...)):
    # قراءة الصورة
    data = await file.read()
    original_image = Image.open(io.BytesIO(data)).convert("RGB")

    # كشف الأجسام باستخدام YOLO
    results = detection_model(original_image, conf=0.25)
    integrated_results = []

    for r in results:
        for i, box in enumerate(r.boxes):
            label = r.names[int(box.cls)]
            coords = box.xyxy[0].tolist() 

            # قص العنصر مع هامش (Padding) 20 بكسل لرؤية الشكل واللون بوضوح
            pad = 20
            left = max(0, coords[0] - pad)
            top = max(0, coords[1] - pad)
            right = min(original_image.width, coords[2] + pad)
            bottom = min(original_image.height, coords[3] + pad)
            
            cropped_img = original_image.crop((left, top, right, bottom))

            # --- استراتيجية الوصف الحر (بدون برومبت نصي مقيد) ---
            # نترك الموديل يحلل الصورة بصرياً فقط
            inputs = processor(images=cropped_img, return_tensors="pt").to(device)
            
            generated_ids = caption_model.generate(
                pixel_values=inputs.pixel_values,
                max_length=60,       # طول كافٍ لوصف اللون والشكل
                min_length=12,       # إجبار الموديل على التفصيل وعدم الاختصار
                num_beams=5,         # جودة عالية في اختيار الكلمات
                repetition_penalty=1.5,
                early_stopping=True
            )
            
            # فك التشفير للوصف الناتج
            description = processor.batch_decode(generated_ids, skip_special_tokens=True)[0]

            integrated_results.append({
                "object_id": i + 1,
                "label": label,
                "confidence": f"{float(box.conf[0]):.2f}",
                "visual_description": f"Detected {label}: {description.strip()}"
            })

    # في حال لم يتم كشف أي شيء
    if not integrated_results:
        inputs = processor(images=original_image, return_tensors="pt").to(device)
        out = caption_model.generate(pixel_values=inputs.pixel_values, max_length=50)
        general_desc = processor.batch_decode(out, skip_special_tokens=True)[0]
        return {
            "message": "No specific objects detected by YOLO.",
            "general_scene_description": general_desc
        }

    return {
        "detected_count": len(integrated_results),
        "results": integrated_results
    }

# --- 3. تشغيل السيرفر ---
if __name__ == "__main__":
    uvicorn.run(app, host="0.0.0.0", port=7860)