app/services/sign_recognition_service.py

import numpy as np
import os

try:
    import tensorflow as tf
    tflite = tf.lite
except ImportError:
    try:
        import tflite_runtime.interpreter as tflite
    except ImportError:
        raise ImportError("Neither tensorflow nor tflite_runtime is installed.")

class GestureRecognizer:
    def __init__(self, model_path=None, label_path=None, threshold=0.8):
        self.threshold = threshold
        self.sequence_length = 30
        self.sequence_buffer = []
        
        # Determine paths relative to this file
        base_dir = os.path.dirname(os.path.dirname(os.path.abspath(__file__))) # app/
        if model_path is None:
            model_path = os.path.join(base_dir, "models", "lstm_model.tflite")
        if label_path is None:
            label_path = os.path.join(base_dir, "models", "labels.txt")

        # Load Labels
        self.labels = []
        try:
            with open(label_path, "r") as f:
                self.labels = [line.strip() for line in f.readlines()]
        except FileNotFoundError:
            print(f"Warning: {label_path} not found.")
            
        # Load Model
        try:
            # We use the Interpreter class from whichever module we imported
            if hasattr(tflite, 'Interpreter'):
                self.interpreter = tflite.Interpreter(model_path=model_path)
            else:
                # Some versions of tflite_runtime have it at the top level
                self.interpreter = tflite(model_path=model_path)
            
            self.interpreter.allocate_tensors()
            self.input_details = self.interpreter.get_input_details()
            self.output_details = self.interpreter.get_output_details()
        except Exception as e:
            print(f"Error loading LSTM model: {e}")
            raise

    def process_landmarks(self, landmarks):
        lm_np = np.array(landmarks)
        wrist = lm_np[:3]
        relative_lm = lm_np - np.tile(wrist, 21)
        
        self.sequence_buffer.append(relative_lm)
        if len(self.sequence_buffer) > self.sequence_length:
            self.sequence_buffer.pop(0)
            
        if len(self.sequence_buffer) == self.sequence_length:
            return self._predict()
        return None

    def _predict(self):
        input_data = np.array([self.sequence_buffer], dtype=np.float32)
        try:
            self.interpreter.set_tensor(self.input_details[0]['index'], input_data)
            self.interpreter.invoke()
            output_data = self.interpreter.get_tensor(self.output_details[0]['index'])
            prediction = np.squeeze(output_data)
            max_index = np.argmax(prediction)
            confidence = prediction[max_index]
            
            if confidence > self.threshold:
                return {
                    "text": self.labels[max_index] if self.labels else str(max_index),
                    "confidence": float(confidence)
                }
        except Exception as e:
            print(f"Inference error: {e}")
        return None

    def clear(self):
        self.sequence_buffer = []