Create app.py

app.py

@@ -0,0 +1,90 @@
+import pickle
+import cv2
+import mediapipe as mp
+import numpy as np
+import tensorflow as tf
+from collections import deque
+import gradio as gr
+
+# Constants (fallback defaults)
+SEQUENCE_LENGTH = 30
+FEATURE_LENGTH = 168
+
+# Load model & preprocessing
+def load_model_and_preprocessing():
+    global scaler, label_encoder, model, labels_dict
+    with open('lstm_preprocessing.pickle', 'rb') as f:
+        preproc = pickle.load(f)
+    scaler = preproc['scaler']
+    label_encoder = preproc['label_encoder']
+    timesteps = preproc['timesteps']
+    n_features = preproc['n_features']
+
+    global SEQUENCE_LENGTH, FEATURE_LENGTH
+    SEQUENCE_LENGTH = timesteps
+    FEATURE_LENGTH = n_features
+
+    model = tf.keras.models.load_model('lstm_model.h5')
+
+    labels_dict = {0:'salam',1:'good morning',2:'thanks',3:'ana',4:'anta',5:'ante',6:'hua',
+                   7:'hea',8:'antm',9:'hm',10:'name',11:'how r u',12:'thanks god',13:'happy',
+                   14:'sad',15:'angry',16:'good',17:'bad',18:'tired',19:'sick',20:'see',21:'say',
+                   22:'talk',23:'walk',24:'went',25:'came',26:'home',27:'eat',28:'slept',29:'university',
+                   30:'today',31:'tmrw',32:'sunday',33:'tuesday',34:'thursday',35:'friday',36:'week',
+                   37:'month',38:'year',39:'when',40:'I know',41:'thinking',42:'forgetten',43:'love',
+                   44:'I want',45:'helps',46:'not allowed',47:'agree',48:'together',49:'different'}
+
+load_model_and_preprocessing()
+
+# MediaPipe setup
+mp_hands = mp.solutions.hands
+hands = mp_hands.Hands(static_image_mode=True, max_num_hands=2)
+
+def extract_hand_features(image):
+    data_aux = np.zeros(FEATURE_LENGTH, dtype=np.float32)
+    image_rgb = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
+    results = hands.process(image_rgb)
+    if not results.multi_hand_landmarks:
+        return None
+
+    for hand_idx, hand_landmarks in enumerate(results.multi_hand_landmarks):
+        if hand_idx >= 2: break
+        min_x = min([lm.x for lm in hand_landmarks.landmark])
+        min_y = min([lm.y for lm in hand_landmarks.landmark])
+        base_idx = hand_idx * 84
+        for i, lm in enumerate(hand_landmarks.landmark):
+            data_aux[base_idx + i * 2] = lm.x - min_x
+            data_aux[base_idx + i * 2 + 1] = lm.y - min_y
+
+    return data_aux
+
+sequence_buffer = deque(maxlen=SEQUENCE_LENGTH)
+
+def predict_sign(image):
+    image = np.array(image)
+    features = extract_hand_features(image)
+    if features is None:
+        return "No hand detected"
+    sequence_buffer.append(features)
+
+    if len(sequence_buffer) < SEQUENCE_LENGTH:
+        return f"Waiting for sequence: {len(sequence_buffer)}/{SEQUENCE_LENGTH}"
+
+    sequence_data = np.array(list(sequence_buffer)).reshape(1, SEQUENCE_LENGTH, FEATURE_LENGTH)
+    scaled = scaler.transform(sequence_data.reshape(-1, FEATURE_LENGTH)).reshape(1, SEQUENCE_LENGTH, FEATURE_LENGTH)
+    prediction_scores = model.predict(scaled, verbose=0)[0]
+    predicted_idx = np.argmax(prediction_scores)
+    confidence = prediction_scores[predicted_idx]
+    label = labels_dict.get(predicted_idx, "Unknown")
+
+    return f"{label} ({confidence:.2f})"
+
+# Gradio interface
+interface = gr.Interface(fn=predict_sign,
+                         inputs=gr.Image(source="webcam", tool="editor", type="numpy"),
+                         outputs="text",
+                         title="Sign Language Recognition",
+                         description="Perform a sign and capture it. Wait for prediction...")
+
+# ✅ LAUNCH THE APP
+interface.launch()