Update app.py

app.py

@@ -1,10 +1,8 @@
-import os
-os.environ["TF_USE_LEGACY_KERAS"] = "1"  # tell TF to use tf-keras (legacy)
-
 import streamlit as st
 import tensorflow as tf
 import numpy as np
 from PIL import Image
+import os
 
 # ================== SETTINGS ==================
 IMG_SIZE = (170, 170)  # same as target_size in your ImageDataGenerator
@@ -16,22 +14,22 @@ CLASS_NAMES = [
     "Leaf Blight",
 ]
 
-MODEL_FILENAME = "grape_disease_model.h5"  # make sure this file is in the same folder
+TFLITE_MODEL_FILENAME = "grape_disease_model.tflite"  # new TFLite model
 # =============================================
 
 
-
 @st.cache_resource
-def load_model():
+def load_interpreter():
     """
-    Load the Keras model once and cache it.
+    Load the TFLite model once and cache the interpreter.
     """
-    if not os.path.exists(MODEL_FILENAME):
-        st.error(f"Model file '{MODEL_FILENAME}' not found in this directory.")
+    if not os.path.exists(TFLITE_MODEL_FILENAME):
+        st.error(f"Model file '{TFLITE_MODEL_FILENAME}' not found in this directory.")
         st.stop()
 
-    model = tf.keras.models.load_model(MODEL_FILENAME)
-    return model
+    interpreter = tf.lite.Interpreter(model_path=TFLITE_MODEL_FILENAME)
+    interpreter.allocate_tensors()
+    return interpreter
 
 
 def preprocess_image(img: Image.Image) -> np.ndarray:
@@ -54,16 +52,23 @@ def preprocess_image(img: Image.Image) -> np.ndarray:
 
 def predict_image(img: Image.Image):
     """
-    Returns (predicted_label, confidence, all_probabilities)
-    for multi-class classification.
+    Run inference using the TFLite interpreter.
+    Returns (predicted_label, confidence, all_probabilities).
     """
-    # ⬇️ load model only when needed
-    model = load_model()
+    interpreter = load_interpreter()
+
+    input_details = interpreter.get_input_details()
+    output_details = interpreter.get_output_details()
+
+    x = preprocess_image(img).astype(np.float32)
+
+    # feed input
+    interpreter.set_tensor(input_details[0]['index'], x)
+    interpreter.invoke()
 
-    x = preprocess_image(img)
-    preds = model.predict(x)
+    # get output
+    preds = interpreter.get_tensor(output_details[0]['index'])
 
-    # expected: shape (1, 4) with softmax
     probs = preds[0]
     idx = int(np.argmax(probs))
     label = CLASS_NAMES[idx]