{"metadata":{"kernelspec":{"language":"python","display_name":"Python 3","name":"python3"},"language_info":{"name":"python","version":"3.11.13","mimetype":"text/x-python","codemirror_mode":{"name":"ipython","version":3},"pygments_lexer":"ipython3","nbconvert_exporter":"python","file_extension":".py"},"kaggle":{"accelerator":"nvidiaTeslaT4","dataSources":[{"sourceId":59093,"databundleVersionId":7469972,"sourceType":"competition"}],"dockerImageVersionId":31090,"isInternetEnabled":true,"language":"python","sourceType":"notebook","isGpuEnabled":true}},"nbformat_minor":4,"nbformat":4,"cells":[{"cell_type":"code","source":"# This Python 3 environment comes with many helpful analytics libraries installed\n# It is defined by the kaggle/python Docker image: https://github.com/kaggle/docker-python\n# For example, here's several helpful packages to load\n\nimport numpy as np # linear algebra\nimport pandas as pd # data processing, CSV file I/O (e.g. pd.read_csv)\n!pip uninstall -y shap tensorflow keras\n!pip install tensorflow==2.12.0 keras==2.12.0 shap==0.44.1 --quiet\n\n# Input data files are available in the read-only \"../input/\" directory\n# For example, running this (by clicking run or pressing Shift+Enter) will list all files under the input directory\n\n#import os\n#for dirname, _, filenames in os.walk('/kaggle/input/hms-harmful-brain-activity-classification'):\n#    for filename in filenames:\n#        print(os.path.join(dirname, filename))\n\n# You can write up to 20GB to the current directory (/kaggle/working/) that gets preserved as output when you create a version using \"Save & Run All\" \n# You can also write temporary files to /kaggle/temp/, but they won't be saved outside of the current session","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-08-04T15:59:22.339855Z","iopub.execute_input":"2025-08-04T15:59:22.340422Z","iopub.status.idle":"2025-08-04T16:00:19.519756Z","shell.execute_reply.started":"2025-08-04T15:59:22.340391Z","shell.execute_reply":"2025-08-04T16:00:19.519032Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"import numpy as np\nimport pandas as pd\nimport matplotlib.pyplot as plt\nimport os\nimport tensorflow as tf\nfrom tensorflow import keras\nfrom tensorflow.keras import layers\nfrom sklearn.model_selection import train_test_split\nfrom sklearn.preprocessing import StandardScaler, LabelEncoder\nfrom sklearn.metrics import accuracy_score, classification_report, confusion_matrix\nfrom tensorflow.keras.utils import to_categorical\nimport pyarrow.parquet as pq\nimport io\nimport scipy.signal\nfrom PIL import Image\n\n# Define dataset directories\ndata_csv = \"/kaggle/input/hms-harmful-brain-activity-classification/train.csv\"\nparquet_dir = \"/kaggle/input/hms-harmful-brain-activity-classification/train_spectrograms\"\n\n# Load metadata\ndf = pd.read_csv(data_csv)\n\n# Filter rows that have corresponding parquet files\nexisting_files = set(os.listdir(parquet_dir))\ndf = df[df[\"spectrogram_id\"].apply(lambda x: f\"{x}.parquet\" in existing_files)]\n\n# Limit to first 5000 samples\ndf = df.head(5000)\n\nif df.empty:\n    raise ValueError(\"No matching parquet files found. Check your directory and filenames.\")\n\n# Encode labels\ny = df[\"expert_consensus\"].dropna()\ndf = df.loc[y.index]\nle = LabelEncoder()\ny_encoded = le.fit_transform(y)\ny_categorical = to_categorical(y_encoded)\n\n# Metadata features\nX_meta = df[['seizure_vote', 'lpd_vote', 'gpd_vote', 'lrda_vote', 'grda_vote', 'other_vote']].values\nscaler = StandardScaler()\nX_meta = scaler.fit_transform(X_meta)\n\n# Helper to convert spectrogram parquet to image array\ndef load_spectrogram_as_image(spectrogram_id, size=(224, 224)):\n    file_path = os.path.join(parquet_dir, f\"{spectrogram_id}.parquet\")\n    table = pq.read_table(file_path)\n    data = table.to_pandas().values.T  # Transpose to get (freq, time)\n    data = np.log1p(data)\n    data = (data - np.min(data)) / (np.max(data) - np.min(data))\n    img = Image.fromarray(np.uint8(data * 255)).convert(\"RGB\").resize(size)\n    return np.array(img) / 255.0\n\n# Load spectrogram images\nX_images = np.stack([load_spectrogram_as_image(sid) for sid in df[\"spectrogram_id\"]])\n\n# Split dataset\nX_img_train, X_img_temp, X_meta_train, X_meta_temp, y_train, y_temp = train_test_split(\n    X_images, X_meta, y_categorical, test_size=0.3, random_state=42, stratify=y_categorical)\nX_img_val, X_img_test, X_meta_val, X_meta_test, y_val, y_test = train_test_split(\n    X_img_temp, X_meta_temp, y_temp, test_size=0.5, random_state=42, stratify=y_temp)\n\n# Define RetNet-inspired image model \nfrom tensorflow.keras.applications import ResNet50\n\nimage_input = keras.Input(shape=(224, 224, 3), name=\"image_input\")\nx = layers.Conv2D(32, (3, 3), activation='relu', padding='same')(image_input)\nx = layers.MaxPooling2D((2, 2))(x)\nx = layers.Conv2D(64, (3, 3), activation='relu', padding='same')(x)\nx = layers.MaxPooling2D((2, 2))(x)\nx = layers.Conv2D(128, (3, 3), activation='relu', padding='same')(x)\nx = layers.GlobalAveragePooling2D()(x)\nx = layers.Dense(128, activation='relu')(x)\nimage_features = layers.Dropout(0.3)(x)\n\n# Metadata model\nmeta_input = keras.Input(shape=(X_meta.shape[1],), name=\"meta_input\")\ny_meta = layers.Dense(32, activation='relu')(meta_input)\nmeta_features = layers.Dense(16, activation='relu')(y_meta)\n\n# Merge models\nmerged = layers.concatenate([image_features, meta_features])\nz = layers.Dense(64, activation='relu')(merged)\nout = layers.Dense(y_categorical.shape[1], activation='softmax')(z)\n\nmodel = keras.Model(inputs=[image_input, meta_input], outputs=out)\nmodel.compile(optimizer=keras.optimizers.Adam(learning_rate=1e-4),\n              loss='categorical_crossentropy', metrics=['accuracy'])\n\n# Train model\nmodel.fit(\n    x=[X_img_train, X_meta_train],\n    y=y_train,\n    epochs=10,\n    batch_size=32,\n    validation_data=([X_img_val, X_meta_val], y_val)\n)\n\n# Evaluate model\ny_pred = model.predict([X_img_test, X_meta_test])\ny_pred_classes = np.argmax(y_pred, axis=1)\ny_true_classes = np.argmax(y_test, axis=1)\n\nprint(\"Test Accuracy:\", accuracy_score(y_true_classes, y_pred_classes))\nprint(\"Test Report:\\n\", classification_report(y_true_classes, y_pred_classes))\n\n# Confusion Matrix\nplt.figure(figsize=(6, 4))\nplt.imshow(confusion_matrix(y_true_classes, y_pred_classes), cmap='Blues', interpolation='nearest')\nplt.colorbar()\nplt.title(\"Confusion Matrix - Test Set\")\nplt.xlabel(\"Predicted\")\nplt.ylabel(\"True\")\nplt.show()","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-08-04T16:02:32.776060Z","iopub.execute_input":"2025-08-04T16:02:32.776358Z","iopub.status.idle":"2025-08-04T16:36:38.893663Z","shell.execute_reply.started":"2025-08-04T16:02:32.776328Z","shell.execute_reply":"2025-08-04T16:36:38.893064Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"!pip install lime --quiet\n\nfrom lime import lime_image\nfrom skimage.segmentation import mark_boundaries\nimport random\n\n# Select a test image index\nidx = random.randint(0, len(X_img_test) - 1)\ntest_img = X_img_test[idx]\ntrue_label = np.argmax(y_test[idx])\nprint(f\"Explaining Test Sample Index: {idx}, True Class: {true_label}\")\n\n# LIME expects batch input to model, so define a wrapper for image input only\ndef predict_fn(images):\n    images = tf.image.resize(images, (224, 224))  # Ensure size\n    metas = np.tile(X_meta_test[idx], (images.shape[0], 1))  # Repeat metadata\n    return model.predict([images, metas])\n\n# Create LIME image explainer\nexplainer = lime_image.LimeImageExplainer()\n\n# Run explanation\nexplanation = explainer.explain_instance(\n    image=test_img,\n    classifier_fn=predict_fn,\n    top_labels=1,\n    hide_color=0,\n    num_samples=1000\n)\n\n# Get image + mask\nfrom skimage.color import label2rgb\n\ntemp, mask = explanation.get_image_and_mask(\n    label=explanation.top_labels[0],\n    positive_only=True,\n    num_features=10,\n    hide_rest=False\n)\n\n# Plot explanation\nplt.figure(figsize=(6, 6))\nplt.title(f\"LIME Explanation for Class: {explanation.top_labels[0]}\")\nplt.imshow(mark_boundaries(temp, mask))\nplt.axis('off')\nplt.show()\n","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-08-04T16:42:41.479188Z","iopub.execute_input":"2025-08-04T16:42:41.479470Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"import shap\nimport numpy as np\nimport matplotlib.pyplot as plt\n\n# Step 1: Build image-only submodel manually (reuse CNN layers)\nimage_input = tf.keras.Input(shape=(224, 224, 3), name=\"image_input\")\nx = model.get_layer(\"conv2d_3\")(image_input)\nx = model.get_layer(\"max_pooling2d_2\")(x)\nx = model.get_layer(\"conv2d_4\")(x)\nx = model.get_layer(\"max_pooling2d_3\")(x)\nx = model.get_layer(\"conv2d_5\")(x)\nx = model.get_layer(\"global_average_pooling2d_1\")(x)\nx = model.get_layer(\"dense_5\")(x)\nx = model.get_layer(\"dropout_1\")(x)\n\n# Image feature extractor model\nimage_only_model = tf.keras.Model(inputs=image_input, outputs=x)\n\n# Step 2: Prepare background + test image\nbackground = X_img_train[np.random.choice(X_img_train.shape[0], 50, replace=False)]\ntest_image = X_img_test[0:1]\n\n# Step 3: SHAP GradientExplainer\nexplainer = shap.GradientExplainer(image_only_model, background)\n\n# Step 4: Compute SHAP values\nshap_values = explainer.shap_values(test_image)\n\n# Step 5: Visualize SHAP explanation\nshap.image_plot(shap_values, test_image)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-08-04T15:50:35.483714Z","iopub.execute_input":"2025-08-04T15:50:35.484562Z","iopub.status.idle":"2025-08-04T15:50:35.865513Z","shell.execute_reply.started":"2025-08-04T15:50:35.484514Z","shell.execute_reply":"2025-08-04T15:50:35.864569Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"","metadata":{"trusted":true},"outputs":[],"execution_count":null}]}