inference_script.py

import pandas as pd
import joblib
from huggingface_hub import hf_hub_download
from sklearn.impute import SimpleImputer
import numpy as np

# Define the Hugging Face repository ID and filenames
REPO_ID = "DP1110/mlp-accessibility-model"
MODEL_FILENAME = 'mlp_regressor_model.joblib'
IMPUTER_FILENAME = 'simple_imputer.joblib'

# Define the feature columns, matching the training data order
FEATURE_COLUMNS = ['% ASF (Euclidean)', '% Built-Up Area', '% ASF (Network)', '% ASF from Bus Stops ', '% ASF from Bus Stops', '% ASF (Network) ']

# Download the model and imputer from Hugging Face Hub
try:
    model_path = hf_hub_download(repo_id=REPO_ID, filename=MODEL_FILENAME)
    imputer_path = hf_hub_download(repo_id=REPO_ID, filename=IMPUTER_FILENAME)
except Exception as e:
    print('Error downloading files from Hugging Face Hub:', e)
    model_path = None
    imputer_path = None

# Load the model and imputer
loaded_mlp_model = None
loaded_imputer = None

if model_path:
    loaded_mlp_model = joblib.load(model_path)
    print('MLP model loaded from', model_path)

if imputer_path:
    loaded_imputer = joblib.load(imputer_path)
    print('Imputer loaded from', imputer_path)

def predict_accessibility_score(new_data_df: pd.DataFrame) -> pd.Series:
    """
    Predicts the overall accessibility score for new, raw input data.

    Args:
        new_data_df (pd.DataFrame): A DataFrame containing new data with the same
                                    feature columns as the training data, before imputation.

    Returns:
        pd.Series: Predicted overall accessibility scores.
    """
    if loaded_mlp_model is None or loaded_imputer is None:
        raise RuntimeError('Model or imputer not loaded. Cannot make predictions.')

    # Ensure the order of columns matches the training data
    # Handle cases where new_data_df might have different columns or order
    missing_cols = set(FEATURE_COLUMNS) - set(new_data_df.columns)
    for c in missing_cols:
        new_data_df[c] = np.nan  # Or appropriate default value

    # Reorder columns to match the training features
    new_data_df = new_data_df[FEATURE_COLUMNS]

    # Apply the loaded imputer to handle missing values in new data
    new_data_imputed = loaded_imputer.transform(new_data_df)
    new_data_imputed_df = pd.DataFrame(new_data_imputed, columns=FEATURE_COLUMNS)

    # Make predictions using the loaded MLP model
    predictions = loaded_mlp_model.predict(new_data_imputed_df)

    return pd.Series(predictions, name='Predicted_Overall_Accessibility_Score')

if __name__ == '__main__':
    print("\n--- Demonstrating prediction with sample data ---")

    # Create a sample DataFrame for new raw data.
    # This should mimic the structure of the features used for training.
    sample_data_dict = {}
    for i, col_name in enumerate(FEATURE_COLUMNS):
        # Assign arbitrary values for demonstration
        sample_data_dict[col_name] = [0.5 + (i * 0.005) % 0.1]  # Varying slightly for demonstration

    new_sample_data = pd.DataFrame(sample_data_dict)

    # Make predictions using the defined function
    try:
        predictions = predict_accessibility_score(new_sample_data)

        # Display the new sample data and the predictions
        print("\n--- New Sample Data for Prediction ---")
        print(new_sample_data)
        print("\n--- Predicted Overall Accessibility Score ---")
        print(predictions)
    except Exception as e:
        print('Error during prediction:', e)