app.py

import streamlit as st
import numpy as np
import torch
import shap
import matplotlib.pyplot as plt
import joblib
import pandas as pd
# Load scalers and model
@st.cache_resource
def load_resources():
    scaler_X = joblib.load("scaler_X.joblib")
    scaler_y = joblib.load("scaler_y.joblib")
    
    model = torch.jit.load("scripted_model.pt")
    model.eval()
    
    return scaler_X, scaler_y, model

# Create a wrapper function for SHAP
def model_wrapper(X):
    with torch.no_grad():
        X_tensor = torch.tensor(X, dtype=torch.float32)
        output = model(X_tensor).numpy()
    return scaler_y.inverse_transform(output)

# Streamlit app
st.title("Resilience Modulus Predictor")

# Load resources
scaler_X, scaler_y, model = load_resources()

# Define feature names and default values
feature_names = [
    "25", "19", "12.5", "9.5", "4.75", "2.36", "1.18", "0.6", "0.3", "0.15", "0.075", "CA", "FA", "type"
]
default_values = [100, 100, 81.593, 68.395, 49.318, 29.283, 17.261, 14.257, 6.041, 3.000, 2.115, 0.600, 0.350, 1.0]

# Input features
st.sidebar.header("Input Features")
input_features = {}
for feature, default_value in zip(feature_names, default_values):
    if feature == "type":
        type_option = st.sidebar.selectbox(f"Enter {feature}", options=["1 - Limestone", "2 - Basalt"], index=0)
        input_features[feature] = 1.0 if type_option == "1 - Limestone" else 2.0
    else:
        input_features[feature] = st.sidebar.number_input(f"Enter {feature}", value=default_value)

# Create input array
input_array = np.array([input_features[feature] for feature in feature_names]).reshape(1, -1)
input_scaled = scaler_X.transform(input_array)

# Make prediction
with torch.no_grad():
    prediction = model(torch.tensor(input_scaled, dtype=torch.float32)).numpy()
    prediction_unscaled = scaler_y.inverse_transform(prediction)

st.write(f"Predicted Resilience Modulus: {prediction_unscaled[0][0]:.2f} MPa")

# SHAP explanation
if st.button("Explain Prediction"):
    # Generate some random background data for SHAP
    background_data = np.random.randn(100, 14)  # 100 samples, 14 features
    background_data_scaled = scaler_X.transform(background_data)
    
    explainer = shap.KernelExplainer(model_wrapper, background_data_scaled)
    shap_values = explainer.shap_values(input_scaled)
    
    shap_values_single = shap_values[0].flatten()
    expected_value = explainer.expected_value[0]
    
    feature_values = [f"{x:.1f}" for x in input_array[0]]
    
    explanation = shap.Explanation(
        values=shap_values_single,
        base_values=expected_value,
        data=feature_values,
        feature_names=feature_names
    )
    
    fig, ax = plt.subplots(figsize=(10, 6))
    shap.plots.waterfall(explanation, show=False)
    st.pyplot(fig)
    
    st.write(f"Base value (unscaled): {([[expected_value]])[0][0]:.2f} MPa")
    st.write(f"Output value (unscaled): {prediction_unscaled[0][0]:.2f} MPa")
    
    st.write("\nFeature contributions (unscaled):")
    feature_contributions = pd.DataFrame({
        'Contribution': shap_values_single
    }, index=feature_names)
    feature_contributions['Contribution'] = feature_contributions['Contribution'].round(4)
    st.table(feature_contributions)