# /snake-ai-api/api/model_service.py import os import io from pathlib import Path import numpy as np import pandas as pd from PIL import Image import onnxruntime as ort from scipy.special import softmax import google.generativeai as genai from dotenv import load_dotenv # Load environment variables # ===================== 0. LOAD ENV ===================== load_dotenv() # loads .env file in root automatically GEMINI_API_KEY = os.getenv("GEMINI_API_KEY") if GEMINI_API_KEY: genai.configure(api_key=GEMINI_API_KEY) else: print("WARNING: GEMINI_API_KEY not found. Image generation will fail.") # ===================== 1. PATHS & MODEL ===================== BASE_DIR = Path(__file__).resolve().parent ONNX_MODEL_PATH = BASE_DIR / "ViT_PreProcessing-ops11-preprocessing-int-dynam_graph.onnx" # Load ONNX model globally try: session = ort.InferenceSession(str(ONNX_MODEL_PATH), providers=["CPUExecutionProvider"]) except FileNotFoundError: raise SystemExit(f"ONNX model not found at {ONNX_MODEL_PATH}. Deployment cannot proceed.") # ===================== 2. IMAGE PREPROCESSING ===================== def preprocess_image(image: Image.Image) -> np.ndarray: image = image.convert("RGB").resize((224, 224)) arr = np.array(image, dtype=np.float32) / 255.0 arr = np.transpose(arr, (2, 0, 1)) return np.expand_dims(arr, axis=0) # ===================== 3. PREDICTION FUNCTION ===================== def predict_snake(image: Image.Image) -> dict: try: # Lazy-load CSVs to save memory metadata = pd.read_csv(BASE_DIR / "train_metadata.csv") venom_data = pd.read_csv(BASE_DIR / "venomstatus_with_antivenom.csv") country_metadata = pd.read_csv(BASE_DIR / "min-train_metadata.csv") class_names = metadata['binomial'].unique() input_tensor = preprocess_image(image) input_name = session.get_inputs()[0].name output_name = session.get_outputs()[0].name logits = session.run([output_name], {input_name: input_tensor})[0][0] probabilities = softmax(logits) predicted_class_id = int(np.argmax(probabilities)) species_name = class_names[predicted_class_id] confidence = float(np.max(probabilities)) species_rows = venom_data[venom_data['species'] == species_name] species_row = species_rows.iloc[0].to_dict() if not species_rows.empty else {} country_list = country_metadata[country_metadata['binomial'] == species_name]['country'].unique() countries = [c for c in country_list if c.lower() not in ('unknown', 'unidentified', 'n/a')] country_output = ", ".join(countries) if countries else "Distribution data unavailable" return { "species_name": species_name, "venom_status": int(species_row.get('venom_status', -1)), "antivenom_name": species_row.get('antivenom Name', 'N/A'), "manufacturer": species_row.get('manufacturer', 'N/A'), "confidence": round(confidence, 4), "countries": country_output } except Exception as e: return {"error": "Prediction failed", "detail": str(e)} # ===================== 4. GEMINI IMAGE GENERATION ===================== def generate_image_from_text(prompt: str) -> tuple[bytes, Image.Image]: if not GEMINI_API_KEY: raise ConnectionError("Gemini API key not configured.") full_prompt = f"Highly detailed, scientifically accurate photo of a snake: {prompt}" # Use the official image generation model result = genai.models.get("imagen-3.0-generate-002").generate_content( prompt=full_prompt, config=dict( number_of_images=1, output_mime_type="image/png", aspect_ratio="1:1" ) ) if not result.generated_images: raise RuntimeError("Gemini API returned no image.") image_bytes = result.generated_images[0].image.image_bytes gen_image_pil = Image.open(io.BytesIO(image_bytes)).convert("RGB") gen_image_pil = gen_image_pil.resize((224, 224)) return image_bytes, gen_image_pil