Complected Project

Author: Sayed-Hossein-Hosseini

RhythmAttention_Hybrid_CNN_Transformer_Architecture_for_Arrhythmia_Classification.ipynb

@@ -259,7 +259,7 @@
         "    return signal * factor\n",
         "\n",
         "def apply_time_shift(signal, shift_limit=10):\n",
-        "    \"\"\"سیگنال را به چپ یا راست شیفت می‌دهد تا مدل به مکان دقیق موج حساس نباشد.\"\"\"\n",
+        "    \"\"\"Shifts the signal left or right so that the model is not sensitive to the exact location of the wave.\"\"\"\n",
         "    shift = np.random.randint(-shift_limit, shift_limit)\n",
         "    return np.roll(signal, shift, axis=0)"
       ]
@@ -304,17 +304,16 @@
         "        \n",
         "        if augment:\n",
         "            for i in range(batch_size):\n",
-        "                # اعمال آگمنتیشن سنگین روی کلاس‌های اقلیت (1, 2, 3, 4)\n",
-        "                # در بخش augment ژنراتور:\n",
+        "                # Apply heavy augmentation to minority classes (1, 2, 3, 4)                \n",
         "                if y_batch[i] != 0:\n",
-        "                    if np.random.rand() > 0.3: # فقط ۷۰ درصد مواقع آگمنت کن، ۳۰ درصد نسخه اصلی را بده\n",
+        "                    if np.random.rand() > 0.3: # Only augment 70% of the time, give the original 30% of the time.\n",
         "                        X_batch[i] = apply_time_shift(apply_amplitude_scaling(add_gaussian_noise(X_batch[i])))\n",
         "                else:\n",
-        "                    # نویز بسیار سبک برای کلاس اکثریت (0)\n",
+        "                    # Very light noise for the majority class (0)\n",
         "                    if np.random.rand() > 0.7:\n",
         "                        X_batch[i] = add_gaussian_noise(X_batch[i], noise_level=0.002)\n",
         "        \n",
-        "        # تبدیل به One-Hot\n",
+        "        # Become One-Hot\n",
         "        y_batch_onehot = tf.keras.utils.to_categorical(y_batch, num_classes=5)\n",
         "        \n",
         "        yield X_batch, y_batch_onehot\n",
@@ -348,17 +347,17 @@
         "def focal_loss(gamma=2.0, alpha=0.25):\n",
         "    def focal_loss_fixed(y_true, y_pred):\n",
         "        y_true = tf.cast(y_true, tf.float32)\n",
-        "        # استفاده از epsilon برای جلوگیری از log(0)\n",
+        "        # Using epsilon to avoid log(0)\n",
         "        epsilon = K.epsilon()\n",
         "        y_pred = K.clip(y_pred, epsilon, 1.0 - epsilon)\n",
         "        \n",
-        "        # محاسبه Cross Entropy به روش پایدار\n",
+        "        # Calculating Cross Entropy in a Stable Way\n",
         "        cross_entropy = -y_true * K.log(y_pred)\n",
         "        \n",
-        "        # وزن‌دهی Focal\n",
+        "        # Focal weighting\n",
         "        loss = alpha * K.pow(1 - y_pred, gamma) * cross_entropy\n",
         "        \n",
-        "        # جمع روی کلاس‌ها و میانگین روی بچ\n",
+        "        # Sum on classes and average on batch\n",
         "        return tf.reduce_mean(tf.reduce_sum(loss, axis=-1))\n",
         "    return focal_loss_fixed"
       ]
@@ -620,16 +619,16 @@
       ],
       "source": [
         "def build_rhythm_attention_model(input_shape=(187, 1), num_classes=5):\n",
-        "    # تنظیم مقدار Regularization\n",
+        "    # Setting the Regularization value\n",
         "    reg = l2(0.001)\n",
         "    \n",
         "    inputs = layers.Input(shape=input_shape)\n",
         "\n",
         "    # --- BLOCK 1: Robust CNN Feature Extractor ---\n",
-        "    # لایه اول با فیلتر بزرگ برای دیدن ساختار کلی موج\n",
+        "    # First layer with large filter to see the overall structure of the wave\n",
         "    x = layers.Conv1D(64, kernel_size=7, padding='same', activation='relu', kernel_regularizer=reg)(inputs)\n",
         "    x = layers.BatchNormalization()(x)\n",
-        "    x = layers.SpatialDropout1D(0.1)(x) # حذف Feature Map های نویزی\n",
+        "    x = layers.SpatialDropout1D(0.1)(x) # Removing noisy feature maps\n",
         "    \n",
         "    # Residual Block 1\n",
         "    shortcut = x\n",
@@ -642,20 +641,19 @@
         "    x = layers.SpatialDropout1D(0.1)(x)\n",
         "\n",
         "    # Residual Block 2\n",
-        "    # تطبیق ابعاد برای Shortcut\n",
         "    shortcut = layers.Conv1D(128, kernel_size=1, kernel_regularizer=reg)(x)\n",
         "    x = layers.Conv1D(128, kernel_size=3, padding='same', activation='relu', kernel_regularizer=reg)(x)\n",
         "    x = layers.BatchNormalization()(x)\n",
         "    x = layers.MaxPool1D(pool_size=2)(x) \n",
         "    shortcut = layers.MaxPool1D(pool_size=2)(shortcut)\n",
         "    x = layers.Add()([x, shortcut]) \n",
-        "    x = layers.SpatialDropout1D(0.2)(x) # دراپ‌اوت قوی‌تر در لایه‌های عمیق\n",
+        "    x = layers.SpatialDropout1D(0.2)(x) # Stronger dropout in deep layers\n",
         "\n",
         "    # --- BLOCK 2: Positional Encoding + Transformer ---\n",
-        "    # اضافه کردن اطلاعات زمانی\n",
+        "    # Add time information\n",
         "    x = PositionalEmbedding(sequence_length=x.shape[1], embed_dim=x.shape[2])(x)\n",
         "\n",
-        "    # Multi-Head Attention (کاهش هدها به 4 برای تعمیم‌پذیری بهتر)\n",
+        "    # Multi-Head Attention \n",
         "    attention_out = layers.MultiHeadAttention(\n",
         "        num_heads=4, key_dim=128, dropout=0.2\n",
         "    )(x, x)\n",
@@ -668,7 +666,7 @@
         "    x = layers.Concatenate()([avg_pool, max_pool])\n",
         "    \n",
         "    x = layers.Dense(128, activation='relu', kernel_regularizer=reg)(x)\n",
-        "    x = layers.Dropout(0.5)(x) # افزایش Dropout به 0.5 برای مقابله با Overfitting\n",
+        "    x = layers.Dropout(0.5)(x) \n",
         "    x = layers.Dense(64, activation='relu', kernel_regularizer=reg)(x)\n",
         "    \n",
         "    outputs = layers.Dense(num_classes, activation='softmax')(x)\n",
@@ -694,10 +692,9 @@
       "metadata": {},
       "outputs": [],
       "source": [
-        "# تنظیمات پیشنهادی برای کامپایل\n",
         "model.compile(\n",
-        "    optimizer=tf.keras.optimizers.Adam(learning_rate=0.0001), # سرعت یادگیری را کمتر کنید\n",
-        "    loss=focal_loss(gamma=2.0, alpha=1.0), # آلفا ۱ چون ژنراتور وزن‌دهی شده است\n",
+        "    optimizer=tf.keras.optimizers.Adam(learning_rate=0.0001), \n",
+        "    loss=focal_loss(gamma=2.0, alpha=1.0), \n",
         "    metrics=['accuracy']\n",
         ")"
       ]
@@ -1476,6 +1473,13 @@
         "print(\"\\nClassification Report:\\n\")\n",
         "print(classification_report(y_true, y_pred, target_names=['N', 'S', 'V', 'F', 'Q']))"
       ]
+    },
+    {
+      "cell_type": "markdown",
+      "metadata": {},
+      "source": [
+        "## **THE END**"
+      ]
     }
   ],
   "metadata": {