Piyush23890/Sign_Language_Decoder

1"""2train_dynamic_model.py3======================4Train an LSTM sequence classifier on dynamic ISL gesture recordings.5 6Expects  dynamic_dataset/<action>/*.npy  with shape (30, 126).7Saves    dynamic_sign_model.h58 9Usage10-----11    python train_dynamic_model.py12"""13 14import os15import numpy as np16from sklearn.model_selection import train_test_split17 18# ── Config ──────────────────────────────────────────────────────────────────────19DATASET_PATH    = "dynamic_dataset"20ACTIONS         = ["hello", "thank_you"]    # must match collection labels21SEQUENCE_LENGTH = 3022FEATURES        = 12623EPOCHS          = 3024BATCH_SIZE      = 1625MODEL_PATH      = "dynamic_sign_model.h5"26 27# ── TF import (isolated so the mock in app.py doesn't interfere) ────────────────28import tensorflow as tf29from tensorflow.keras.models import Sequential30from tensorflow.keras.layers import LSTM, Dense, Dropout31from tensorflow.keras.utils import to_categorical32from tensorflow.keras.callbacks import EarlyStopping, ModelCheckpoint33 34print("=" * 50)35print("SignBridge — Dynamic Model Trainer")36print(f"TF version : {tf.__version__}")37print("=" * 50)38 39# ── 1. Load .npy sequences ──────────────────────────────────────────────────────40X, y = [], []41for label_idx, action in enumerate(ACTIONS):42    folder = os.path.join(DATASET_PATH, action)43    if not os.path.isdir(folder):44        print(f"  [WARN] Missing folder: {folder}")45        continue46    files = [f for f in os.listdir(folder) if f.endswith(".npy")]47    print(f"  [{action}]  {len(files)} sequences")48    for fname in files:49        arr = np.load(os.path.join(folder, fname))50        if arr.shape == (SEQUENCE_LENGTH, FEATURES):51            X.append(arr)52            y.append(label_idx)53        else:54            print(f"    [skip] bad shape {arr.shape}: {fname}")55 56X = np.array(X, dtype=np.float32)57y = to_categorical(y, num_classes=len(ACTIONS))58print(f"\nDataset: {X.shape[0]} sequences × {SEQUENCE_LENGTH} frames × {FEATURES} features")59 60# ── 2. Split ─────────────────────────────────────────────────────────────────────61X_train, X_test, y_train, y_test = train_test_split(62    X, y, test_size=0.20, random_state=4263)64print(f"Train: {len(X_train)}  |  Test: {len(X_test)}")65 66# ── 3. Model ──────────────────────────────────────────────────────────────────────67model = Sequential([68    LSTM(64, return_sequences=True,69         input_shape=(SEQUENCE_LENGTH, FEATURES)),70    Dropout(0.30),71    LSTM(64),72    Dense(32, activation="relu"),73    Dense(len(ACTIONS), activation="softmax"),74], name="dynamic_isl")75 76model.compile(77    optimizer="adam",78    loss="categorical_crossentropy",79    metrics=["accuracy"],80)81model.summary()82 83# ── 4. Train ─────────────────────────────────────────────────────────────────────84callbacks = [85    EarlyStopping(monitor="val_accuracy", patience=5,86                  restore_best_weights=True, verbose=1),87    ModelCheckpoint(MODEL_PATH, monitor="val_accuracy",88                    save_best_only=True, verbose=1),89]90 91print(f"\nTraining for up to {EPOCHS} epochs …")92history = model.fit(93    X_train, y_train,94    epochs=EPOCHS,95    batch_size=BATCH_SIZE,96    validation_data=(X_test, y_test),97    callbacks=callbacks,98    verbose=1,99)100 101# ── 5. Final evaluation ───────────────────────────────────────────────────────────102loss, acc = model.evaluate(X_test, y_test, verbose=0)103print(f"\nFinal Test Accuracy : {acc * 100:.2f}%")104print(f"Model saved         : {MODEL_PATH}")105