I have the following prediction code:
\n\n\n\n\n\n\n\n\n\n\n os\n gc\n joblib\n numpy np\n pandas pd\n polars pl\n pathlib Path\n traceback\n datetime datetime\n\n\n\n\n\nARTIFACTS_DIRECTORY = Path()\n\n\n ():\n \n ()\n\n\nLOADED_ARTIFACTS = \n\n\n\n\nlog()\n scipy.spatial.transform Rotation R\n\n\n ():\n acc_values = acc_data[[, , ]].values.astype(np.float64)\n quat_values = rot_data[[, , , ]].values.astype(np.float64)\n linear_accel = np.zeros_like(acc_values)\n gravity_world = np.array([, , ])\n :\n valid_mask = ~(np.isnan(quat_values).(axis=) | (np.(quat_values).(axis=) < ))\n np.(valid_mask):\n rotations = R.from_quat(quat_values[valid_mask])\n gravity_sensor_frames = rotations.apply(gravity_world, inverse=)\n linear_accel[valid_mask] = acc_values[valid_mask] - gravity_sensor_frames\n linear_accel[~valid_mask] = acc_values[~valid_mask]\n Exception:\n linear_accel = acc_values\n linear_accel\n\n ():\n quat_values = rot_data[[, , , ]].values.astype(np.float64)\n angular_vel = np.zeros((quat_values.shape[], ))\n quat_values.shape[] < : angular_vel\n :\n valid_current = ~np.isnan(quat_values[:-]).(axis=)\n valid_next = ~np.isnan(quat_values[:]).(axis=)\n valid_pairs = valid_current & valid_next\n np.(valid_pairs):\n rot_current = R.from_quat(quat_values[:-][valid_pairs])\n rot_next = R.from_quat(quat_values[:][valid_pairs])\n delta_rots = rot_current.inv() * rot_next\n angular_vel[:-][valid_pairs] = delta_rots.as_rotvec() / time_delta\n Exception:\n \n angular_vel\n\nlog()\n\n\n\n\n\n () -> :\n \n LOADED_ARTIFACTS\n log()\n\n :\n \n LOADED_ARTIFACTS :\n log()\n tensorflow tf\n\n \n (): tf.keras.backend.(x, axis=)\n (): tf.squeeze(x, axis=-)\n (): tf.expand_dims(x, axis=-)\n custom_objects = {: time_sum, : squeeze_last_axis, : expand_last_axis}\n\n \n scaler = joblib.load(ARTIFACTS_DIRECTORY / )\n label_encoder = joblib.load(ARTIFACTS_DIRECTORY / )\n feature_cols = np.load(ARTIFACTS_DIRECTORY / , allow_pickle=)\n pad_length = (np.load(ARTIFACTS_DIRECTORY / , allow_pickle=))\n\n \n models = []\n model_files = ((ARTIFACTS_DIRECTORY.glob()))\n model_files:\n FileNotFoundError()\n\n model_path model_files:\n log()\n model = tf.keras.models.load_model(model_path, custom_objects=custom_objects)\n models.append(model)\n\n LOADED_ARTIFACTS = {\n : scaler,\n : label_encoder,\n : feature_cols,\n : pad_length,\n : models,\n }\n log()\n :\n log()\n\n \n log()\n df = sequence.to_pandas()\n linear_accel = remove_gravity_from_acc(df, df)\n df[], df[], df[] = linear_accel[:, ], linear_accel[:, ], linear_accel[:, ]\n df[] = np.sqrt(np.(linear_accel**, axis=))\n df[] = df[].diff().fillna()\n angular_vel = calculate_angular_velocity_from_quat(df)\n df[], df[], df[] = angular_vel[:, ], angular_vel[:, ], angular_vel[:, ]\n\n \n log()\n feature_cols = LOADED_ARTIFACTS[]\n \n col feature_cols:\n col df.columns:\n df[col] = \n features = df[feature_cols].ffill().bfill().fillna().values.astype(np.float32)\n\n \n pad_length = LOADED_ARTIFACTS[]\n padded = np.zeros((pad_length, features.shape[]), dtype=np.float32)\n seq_len = ((features), pad_length)\n padded[:seq_len] = features[:seq_len]\n\n scaled = LOADED_ARTIFACTS[].transform(padded)\n X_input = scaled.reshape(, pad_length, (feature_cols))\n\n \n log()\n all_preds = []\n model LOADED_ARTIFACTS[]:\n pred_probs = model.predict(X_input, verbose=)\n all_preds.append(pred_probs[])\n\n \n avg_pred_proba = np.mean(all_preds, axis=)\n predicted_class_int = np.argmax(avg_pred_proba)\n prediction = LOADED_ARTIFACTS[].classes_[predicted_class_int]\n log()\n log()\n prediction\n\n Exception e:\n \n log()\n \n \n \n\n\n\n\n kaggle_evaluation.cmi_inference_server\n\nlog()\ninference_server = kaggle_evaluation.cmi_inference_server.CMIInferenceServer(predict)\n\n\n\n os.getenv():\n log()\n inference_server.serve()\n:\n log()\n \n local_test_data = \n local_test_demographics = \n os.path.exists(local_test_data):\n inference_server.run_local_gateway(\n data_paths=(local_test_data, local_test_demographics)\n )\n :\n log()\n\nlog()\nI'm getting the following error:
\n[::] Defining helper functions feature engineering.\n[::] Global helper functions are defined.\n[::] Initializing Kaggle inference server...\n[::] Starting server local test mode.\n[::] --- PREDICTION START ---\n[::] First run: Loading models artifacts...\n[::] -> Loading model: model_fold_0.h5\n[::] -> Loading model: model_fold_1.h5\n[::] -> Loading model: model_fold_2.h5\n[::] -> Loading model: model_fold_3.h5\n[::] Artifacts models loaded cached.\n[::] Performing feature engineering...\n[::] Preparing data matrix model...\n[::] Running ensemble prediction...\n[::] Prediction complete. Predicted gesture: \n[::] --- PREDICTION END ---\n\n---------------------------------------------------------------------------\nGatewayRuntimeError Traceback (most recent call last)\n/tmp/ipykernel_36/py <cell line: >()\n local_test_demographics = \n os.path.exists(local_test_data):\n--> inference_server.run_local_gateway(\n data_paths=(local_test_data, local_test_demographics)\n )\n\n/kaggle//cmi-detect-behavior--sensor-data/kaggle_evaluation/core/templates.py run_local_gateway(, data_paths, file_share_dir, *args, **kwargs)\n .gateway.run()\n Exception err:\n--> err \n :\n .server.stop()\n\n/kaggle//cmi-detect-behavior--sensor-data/kaggle_evaluation/core/templates.py run_local_gateway(, data_paths, file_share_dir, *args, **kwargs)\n :\n .gateway = ._get_gateway_for_test(data_paths, file_share_dir, *args, **kwargs)\n--> .gateway.run()\n Exception err:\n err \n\n/kaggle//cmi-detect-behavior--sensor-data/kaggle_evaluation/core/templates.py run()\n error:\n \n--> error\n \n \n\n/kaggle//cmi-detect-behavior--sensor-data/kaggle_evaluation/core/templates.py run()\n :\n .unpack_data_paths()\n---> predictions, row_ids = .get_all_predictions()\n .write_submission(predictions, row_ids)\n kaggle_evaluation.core.base_gateway.GatewayRuntimeError gre:\n\n/kaggle//cmi-detect-behavior--sensor-data/kaggle_evaluation/core/templates.py get_all_predictions()\n all_row_ids = []\n data_batch, row_ids .generate_data_batches():\n---> predictions = .predict(*data_batch)\n .validate_prediction_batch(predictions, row_ids)\n all_predictions.append(predictions)\n\n/kaggle//cmi-detect-behavior--sensor-data/kaggle_evaluation/core/templates.py predict(, *args, **kwargs)\n .client.send(, *args, **kwargs)\n Exception e:\n---> .handle_server_error(e, )\n \n () -> :\n\n/kaggle//cmi-detect-behavior--sensor-data/kaggle_evaluation/core/base_gateway.py handle_server_error(, exception, endpoint)\n message_match = re.search(, exception_str, re.IGNORECASE)\n message = message_match.group() message_match exception_str\n--> GatewayRuntimeError(GatewayRuntimeErrorType.SERVER_RAISED_EXCEPTION, message) \n (exception, grpc._channel._InactiveRpcError):\n GatewayRuntimeError(GatewayRuntimeErrorType.SERVER_CONNECTION_FAILED, exception_str) \n\nGatewayRuntimeError: (<GatewayRuntimeErrorType.SERVER_RAISED_EXCEPTION: >, )\nCan you help me on fixing this so that I can submit?
", "messages": [ { "id": "3253921", "postDate": "07/25/2025 13:52:41", "content": "I have the following prediction code:
\n\n\n\n\n\n\n\n\n\n\n os\n gc\n joblib\n numpy np\n pandas pd\n polars pl\n pathlib Path\n traceback\n datetime datetime\n\n\n\n\n\nARTIFACTS_DIRECTORY = Path()\n\n\n ():\n \n ()\n\n\nLOADED_ARTIFACTS = \n\n\n\n\nlog()\n scipy.spatial.transform Rotation R\n\n\n ():\n acc_values = acc_data[[, , ]].values.astype(np.float64)\n quat_values = rot_data[[, , , ]].values.astype(np.float64)\n linear_accel = np.zeros_like(acc_values)\n gravity_world = np.array([, , ])\n :\n valid_mask = ~(np.isnan(quat_values).(axis=) | (np.(quat_values).(axis=) < ))\n np.(valid_mask):\n rotations = R.from_quat(quat_values[valid_mask])\n gravity_sensor_frames = rotations.apply(gravity_world, inverse=)\n linear_accel[valid_mask] = acc_values[valid_mask] - gravity_sensor_frames\n linear_accel[~valid_mask] = acc_values[~valid_mask]\n Exception:\n linear_accel = acc_values\n linear_accel\n\n ():\n quat_values = rot_data[[, , , ]].values.astype(np.float64)\n angular_vel = np.zeros((quat_values.shape[], ))\n quat_values.shape[] < : angular_vel\n :\n valid_current = ~np.isnan(quat_values[:-]).(axis=)\n valid_next = ~np.isnan(quat_values[:]).(axis=)\n valid_pairs = valid_current & valid_next\n np.(valid_pairs):\n rot_current = R.from_quat(quat_values[:-][valid_pairs])\n rot_next = R.from_quat(quat_values[:][valid_pairs])\n delta_rots = rot_current.inv() * rot_next\n angular_vel[:-][valid_pairs] = delta_rots.as_rotvec() / time_delta\n Exception:\n \n angular_vel\n\nlog()\n\n\n\n\n\n () -> :\n \n LOADED_ARTIFACTS\n log()\n\n :\n \n LOADED_ARTIFACTS :\n log()\n tensorflow tf\n\n \n (): tf.keras.backend.(x, axis=)\n (): tf.squeeze(x, axis=-)\n (): tf.expand_dims(x, axis=-)\n custom_objects = {: time_sum, : squeeze_last_axis, : expand_last_axis}\n\n \n scaler = joblib.load(ARTIFACTS_DIRECTORY / )\n label_encoder = joblib.load(ARTIFACTS_DIRECTORY / )\n feature_cols = np.load(ARTIFACTS_DIRECTORY / , allow_pickle=)\n pad_length = (np.load(ARTIFACTS_DIRECTORY / , allow_pickle=))\n\n \n models = []\n model_files = ((ARTIFACTS_DIRECTORY.glob()))\n model_files:\n FileNotFoundError()\n\n model_path model_files:\n log()\n model = tf.keras.models.load_model(model_path, custom_objects=custom_objects)\n models.append(model)\n\n LOADED_ARTIFACTS = {\n : scaler,\n : label_encoder,\n : feature_cols,\n : pad_length,\n : models,\n }\n log()\n :\n log()\n\n \n log()\n df = sequence.to_pandas()\n linear_accel = remove_gravity_from_acc(df, df)\n df[], df[], df[] = linear_accel[:, ], linear_accel[:, ], linear_accel[:, ]\n df[] = np.sqrt(np.(linear_accel**, axis=))\n df[] = df[].diff().fillna()\n angular_vel = calculate_angular_velocity_from_quat(df)\n df[], df[], df[] = angular_vel[:, ], angular_vel[:, ], angular_vel[:, ]\n\n \n log()\n feature_cols = LOADED_ARTIFACTS[]\n \n col feature_cols:\n col df.columns:\n df[col] = \n features = df[feature_cols].ffill().bfill().fillna().values.astype(np.float32)\n\n \n pad_length = LOADED_ARTIFACTS[]\n padded = np.zeros((pad_length, features.shape[]), dtype=np.float32)\n seq_len = ((features), pad_length)\n padded[:seq_len] = features[:seq_len]\n\n scaled = LOADED_ARTIFACTS[].transform(padded)\n X_input = scaled.reshape(, pad_length, (feature_cols))\n\n \n log()\n all_preds = []\n model LOADED_ARTIFACTS[]:\n pred_probs = model.predict(X_input, verbose=)\n all_preds.append(pred_probs[])\n\n \n avg_pred_proba = np.mean(all_preds, axis=)\n predicted_class_int = np.argmax(avg_pred_proba)\n prediction = LOADED_ARTIFACTS[].classes_[predicted_class_int]\n log()\n log()\n prediction\n\n Exception e:\n \n log()\n \n \n \n\n\n\n\n kaggle_evaluation.cmi_inference_server\n\nlog()\ninference_server = kaggle_evaluation.cmi_inference_server.CMIInferenceServer(predict)\n\n\n\n os.getenv():\n log()\n inference_server.serve()\n:\n log()\n \n local_test_data = \n local_test_demographics = \n os.path.exists(local_test_data):\n inference_server.run_local_gateway(\n data_paths=(local_test_data, local_test_demographics)\n )\n :\n log()\n\nlog()\nI'm getting the following error:
\n[::] Defining helper functions feature engineering.\n[::] Global helper functions are defined.\n[::] Initializing Kaggle inference server...\n[::] Starting server local test mode.\n[::] --- PREDICTION START ---\n[::] First run: Loading models artifacts...\n[::] -> Loading model: model_fold_0.h5\n[::] -> Loading model: model_fold_1.h5\n[::] -> Loading model: model_fold_2.h5\n[::] -> Loading model: model_fold_3.h5\n[::] Artifacts models loaded cached.\n[::] Performing feature engineering...\n[::] Preparing data matrix model...\n[::] Running ensemble prediction...\n[::] Prediction complete. Predicted gesture: \n[::] --- PREDICTION END ---\n\n---------------------------------------------------------------------------\nGatewayRuntimeError Traceback (most recent call last)\n/tmp/ipykernel_36/py <cell line: >()\n local_test_demographics = \n os.path.exists(local_test_data):\n--> inference_server.run_local_gateway(\n data_paths=(local_test_data, local_test_demographics)\n )\n\n/kaggle//cmi-detect-behavior--sensor-data/kaggle_evaluation/core/templates.py run_local_gateway(, data_paths, file_share_dir, *args, **kwargs)\n .gateway.run()\n Exception err:\n--> err \n :\n .server.stop()\n\n/kaggle//cmi-detect-behavior--sensor-data/kaggle_evaluation/core/templates.py run_local_gateway(, data_paths, file_share_dir, *args, **kwargs)\n :\n .gateway = ._get_gateway_for_test(data_paths, file_share_dir, *args, **kwargs)\n--> .gateway.run()\n Exception err:\n err \n\n/kaggle//cmi-detect-behavior--sensor-data/kaggle_evaluation/core/templates.py run()\n error:\n \n--> error\n \n \n\n/kaggle//cmi-detect-behavior--sensor-data/kaggle_evaluation/core/templates.py run()\n :\n .unpack_data_paths()\n---> predictions, row_ids = .get_all_predictions()\n .write_submission(predictions, row_ids)\n kaggle_evaluation.core.base_gateway.GatewayRuntimeError gre:\n\n/kaggle//cmi-detect-behavior--sensor-data/kaggle_evaluation/core/templates.py get_all_predictions()\n all_row_ids = []\n data_batch, row_ids .generate_data_batches():\n---> predictions = .predict(*data_batch)\n .validate_prediction_batch(predictions, row_ids)\n all_predictions.append(predictions)\n\n/kaggle//cmi-detect-behavior--sensor-data/kaggle_evaluation/core/templates.py predict(, *args, **kwargs)\n .client.send(, *args, **kwargs)\n Exception e:\n---> .handle_server_error(e, )\n \n () -> :\n\n/kaggle//cmi-detect-behavior--sensor-data/kaggle_evaluation/core/base_gateway.py handle_server_error(, exception, endpoint)\n message_match = re.search(, exception_str, re.IGNORECASE)\n message = message_match.group() message_match exception_str\n--> GatewayRuntimeError(GatewayRuntimeErrorType.SERVER_RAISED_EXCEPTION, message) \n (exception, grpc._channel._InactiveRpcError):\n GatewayRuntimeError(GatewayRuntimeErrorType.SERVER_CONNECTION_FAILED, exception_str) \n\nGatewayRuntimeError: (<GatewayRuntimeErrorType.SERVER_RAISED_EXCEPTION: >, )\nCan you help me on fixing this so that I can submit?
", "rawMarkdown": "I have the following prediction code:\n\n```py\n# =================================================================================\n#\n# Gesture Recognition - Kaggle Submission Script\n#\n# This script is designed for inference. It loads pre-trained models and\n# all necessary artifacts to predict gestures from sensor data sequences.\n# It is optimized for the Kaggle environment, loading artifacts only once.\n#\n# =================================================================================\n\nimport os\nimport gc\nimport joblib\nimport numpy as np\nimport pandas as pd\nimport polars as pl\nfrom pathlib import Path\nimport traceback\nfrom datetime import datetime\n\n# --- Configuration ---\n# !!! IMPORTANT !!!\n# Change this path to point to the Kaggle dataset containing your trained models and artifacts.\n# For example: \"/kaggle/input/my-gesture-model-artifacts/\"\nARTIFACTS_DIRECTORY = Path(\"/kaggle/input/dosamasters-sub1/output/\")\n\n# --- Global Logging & Artifact Cache ---\ndef log(message: str):\n \"\"\"A simple, timestamped print function for clear logging during inference.\"\"\"\n print(f\"[{datetime.now().strftime('%H:%M:%S')}] {message}\")\n\n# This global variable will cache the loaded artifacts to prevent reloading on every call.\nLOADED_ARTIFACTS = None\n\n# =================================================================================\n# STEP 1: All necessary helper functions from training are defined globally.\n# =================================================================================\nlog(\"Defining global helper functions for feature engineering.\")\nfrom scipy.spatial.transform import Rotation as R\n\n# The following functions are identical to the ones used during training to ensure consistency.\ndef remove_gravity_from_acc(acc_data, rot_data):\n acc_values = acc_data[['acc_x', 'acc_y', 'acc_z']].values.astype(np.float64)\n quat_values = rot_data[['rot_x', 'rot_y', 'rot_z', 'rot_w']].values.astype(np.float64)\n linear_accel = np.zeros_like(acc_values)\n gravity_world = np.array([0, 0, 9.81])\n try:\n valid_mask = ~(np.isnan(quat_values).any(axis=1) | (np.abs(quat_values).sum(axis=1) < 1e-8))\n if np.any(valid_mask):\n rotations = R.from_quat(quat_values[valid_mask])\n gravity_sensor_frames = rotations.apply(gravity_world, inverse=True)\n linear_accel[valid_mask] = acc_values[valid_mask] - gravity_sensor_frames\n linear_accel[~valid_mask] = acc_values[~valid_mask]\n except Exception:\n linear_accel = acc_values\n return linear_accel\n\ndef calculate_angular_velocity_from_quat(rot_data, time_delta=1/200):\n quat_values = rot_data[['rot_x', 'rot_y', 'rot_z', 'rot_w']].values.astype(np.float64)\n angular_vel = np.zeros((quat_values.shape[0], 3))\n if quat_values.shape[0] < 2: return angular_vel\n try:\n valid_current = ~np.isnan(quat_values[:-1]).any(axis=1)\n valid_next = ~np.isnan(quat_values[1:]).any(axis=1)\n valid_pairs = valid_current & valid_next\n if np.any(valid_pairs):\n rot_current = R.from_quat(quat_values[:-1][valid_pairs])\n rot_next = R.from_quat(quat_values[1:][valid_pairs])\n delta_rots = rot_current.inv() * rot_next\n angular_vel[:-1][valid_pairs] = delta_rots.as_rotvec() / time_delta\n except Exception:\n pass\n return angular_vel\n\nlog(\"Global helper functions are defined.\")\n\n\n# =================================================================================\n# STEP 2: The Official `predict` Function for the Kaggle API\n# =================================================================================\ndef predict(sequence: pl.DataFrame, demographics: pl.DataFrame) -> str:\n \"\"\"\n This function is called by the Kaggle server for each test sequence.\n \"\"\"\n global LOADED_ARTIFACTS\n log(\"--- PREDICTION START ---\")\n\n try:\n # --- Stage A: Load Models and Artifacts (ONCE) ---\n if LOADED_ARTIFACTS is None:\n log(\"First run: Loading models and artifacts...\")\n import tensorflow as tf\n\n # Define custom objects needed to load the model\n def time_sum(x): return tf.keras.backend.sum(x, axis=1)\n def squeeze_last_axis(x): return tf.squeeze(x, axis=-1)\n def expand_last_axis(x): return tf.expand_dims(x, axis=-1)\n custom_objects = {'time_sum': time_sum, 'squeeze_last_axis': squeeze_last_axis, 'expand_last_axis': expand_last_axis}\n\n # Load preprocessing tools and metadata\n scaler = joblib.load(ARTIFACTS_DIRECTORY / \"scaler.pkl\")\n label_encoder = joblib.load(ARTIFACTS_DIRECTORY / \"label_encoder.pkl\")\n feature_cols = np.load(ARTIFACTS_DIRECTORY / \"feature_cols.npy\", allow_pickle=True)\n pad_length = int(np.load(ARTIFACTS_DIRECTORY / \"sequence_maxlen.npy\", allow_pickle=True))\n\n # Load all trained model folds\n models = []\n model_files = sorted(list(ARTIFACTS_DIRECTORY.glob(\"model_fold_*.h5\")))\n if not model_files:\n raise FileNotFoundError(f\"No model files (.h5) found in {ARTIFACTS_DIRECTORY}\")\n\n for model_path in model_files:\n log(f\" -> Loading model: {model_path.name}\")\n model = tf.keras.models.load_model(model_path, custom_objects=custom_objects)\n models.append(model)\n\n LOADED_ARTIFACTS = {\n 'scaler': scaler,\n 'label_encoder': label_encoder,\n 'feature_cols': feature_cols,\n 'pad_length': pad_length,\n 'models': models,\n }\n log(f\"Artifacts and {len(models)} models loaded and cached.\")\n else:\n log(\"Artifacts already in memory.\")\n\n # --- Stage B: Feature Engineering ---\n log(\"Performing feature engineering...\")\n df = sequence.to_pandas()\n linear_accel = remove_gravity_from_acc(df, df)\n df[\"linear_acc_x\"], df[\"linear_acc_y\"], df[\"linear_acc_z\"] = linear_accel[:, 0], linear_accel[:, 1], linear_accel[:, 2]\n df[\"linear_acc_mag\"] = np.sqrt(np.sum(linear_accel**2, axis=1))\n df[\"linear_acc_mag_jerk\"] = df[\"linear_acc_mag\"].diff().fillna(0)\n angular_vel = calculate_angular_velocity_from_quat(df)\n df[\"angular_vel_x\"], df[\"angular_vel_y\"], df[\"angular_vel_z\"] = angular_vel[:, 0], angular_vel[:, 1], angular_vel[:, 2]\n\n # --- Stage C: Data Preparation ---\n log(\"Preparing data matrix for model...\")\n feature_cols = LOADED_ARTIFACTS['feature_cols']\n # Ensure all columns from training are present, filling missing ones with 0\n for col in feature_cols:\n if col not in df.columns:\n df[col] = 0\n features = df[feature_cols].ffill().bfill().fillna(0).values.astype(np.float32)\n\n # Pad and scale the sequence exactly as in training\n pad_length = LOADED_ARTIFACTS['pad_length']\n padded = np.zeros((pad_length, features.shape[1]), dtype=np.float32)\n seq_len = min(len(features), pad_length)\n padded[:seq_len] = features[:seq_len]\n\n scaled = LOADED_ARTIFACTS['scaler'].transform(padded)\n X_input = scaled.reshape(1, pad_length, len(feature_cols))\n\n # --- Stage D: Ensemble Prediction ---\n log(\"Running ensemble prediction...\")\n all_preds = []\n for model in LOADED_ARTIFACTS['models']:\n pred_probs = model.predict(X_input, verbose=0)\n all_preds.append(pred_probs[0])\n\n # Average the predictions from all models\n avg_pred_proba = np.mean(all_preds, axis=0)\n predicted_class_int = np.argmax(avg_pred_proba)\n prediction = LOADED_ARTIFACTS['label_encoder'].classes_[predicted_class_int]\n log(f\"Prediction complete. Predicted gesture: '{prediction}'\")\n log(\"--- PREDICTION END ---\\n\")\n return prediction\n\n except Exception as e:\n # If any error occurs, log it and return a default value as required by the competition.\n log(f\"FATAL ERROR in predict function: {e}\\n{traceback.format_exc()}\")\n # Returning a default valid gesture is often better than failing.\n # You can change 'unknown' to the most frequent class if needed.\n return 'unknown'\n\n# =================================================================================\n# STEP 3: Kaggle Server Initialization\n# =================================================================================\nimport kaggle_evaluation.cmi_inference_server\n\nlog(\"Initializing Kaggle inference server...\")\ninference_server = kaggle_evaluation.cmi_inference_server.CMIInferenceServer(predict)\n\n# The server is started only when the KAGGLE_IS_COMPETITION_RERUN environment variable is set.\n# For local testing, it runs a local gateway.\nif os.getenv('KAGGLE_IS_COMPETITION_RERUN'):\n log(\"Starting server in competition mode.\")\n inference_server.serve()\nelse:\n log(\"Starting server in local test mode.\")\n # In local mode, you need to provide paths to local test data.\n local_test_data = '/kaggle/input/cmi-detect-behavior-with-sensor-data/test.csv'\n local_test_demographics = '/kaggle/input/cmi-detect-behavior-with-sensor-data/test_demographics.csv'\n if os.path.exists(local_test_data):\n inference_server.run_local_gateway(\n data_paths=(local_test_data, local_test_demographics)\n )\n else:\n log(\"Local test files not found. Submission script finished without local test.\")\n\nlog(\"Submission script execution finished.\")\n```\n\n\nI'm getting the following error:\n\n```\n[13:49:22] Defining global helper functions for feature engineering.\n[13:49:22] Global helper functions are defined.\n[13:49:22] Initializing Kaggle inference server...\n[13:49:22] Starting server in local test mode.\n[13:49:22] --- PREDICTION START ---\n[13:49:22] First run: Loading models and artifacts...\n[13:49:22] -> Loading model: model_fold_0.h5\n[13:49:23] -> Loading model: model_fold_1.h5\n[13:49:24] -> Loading model: model_fold_2.h5\n[13:49:24] -> Loading model: model_fold_3.h5\n[13:49:25] Artifacts and 4 models loaded and cached.\n[13:49:25] Performing feature engineering...\n[13:49:25] Preparing data matrix for model...\n[13:49:25] Running ensemble prediction...\n[13:49:31] Prediction complete. Predicted gesture: 'Forehead - scratch'\n[13:49:31] --- PREDICTION END ---\n\n---------------------------------------------------------------------------\nGatewayRuntimeError Traceback (most recent call last)\n/tmp/ipykernel_36/1179446127.py inModify kaggle_evaluation folder of this competition and create a new dataset or model using it. Then you can edit the code and import this folder as module and replace kaggle_evaluation.cmi_inference_server for debug.
@rithulkamesh hey were you able to resolve this issue?
", "rawMarkdown": "rithulkamesh hey were you able to resolve this issue?", "votes": null }, { "id": "3271758", "postDate": "08/19/2025 15:37:47", "content": "I had the same issue; check the result type of your model; if it's numpy.str_ you should convert it to str.
\nFor me it worked.
Modify kaggle_evaluation folder of this competition and create a new dataset or model using it. Then you can edit the code and import this folder as module and replace kaggle_evaluation.cmi_inference_server for debug.
@rithulkamesh hey were you able to resolve this issue?
", "votes": null, "replies": [] }, { "id": 3271758, "author_name": "obiaf88", "author_url": "", "post_date": "08/19/2025 15:37:47", "content": "I had the same issue; check the result type of your model; if it's numpy.str_ you should convert it to str.
\nFor me it worked.