src/streamlit_app.py

import pandas as pd
import streamlit as st
import joblib
from pathlib import Path

st.set_page_config(page_title='Star System Classification (LightGBM)', page_icon='🪐', layout='centered')

BASE_DIR = Path(__file__).resolve().parent

MODEL_PATH = BASE_DIR / 'lightgbm_model.pkl'
FEATURES_PATH = BASE_DIR / 'featurer.pkl'  # you saved it with this name
PLANET_ENCODER_PATH = BASE_DIR / 'planet_encoder.pkl'
STAR_ENCODER_PATH = BASE_DIR / 'star_encoder.pkl'

# --- Fixed mapping you used in training ---
ACTIVITY_MAP = {'Low': 0, 'Medium': 1, 'High': 2}

# Optional: label names (edit if your competition uses different names)
LABEL_NAMES = {
    0: 'Habitable',
    1: 'Young',
    2: 'Old',
    3: 'Exotic'
}

@st.cache_resource
def load_artifacts():
    missing = [p.name for p in [MODEL_PATH, FEATURES_PATH, PLANET_ENCODER_PATH, STAR_ENCODER_PATH] if not p.exists()]
    if missing:
        raise FileNotFoundError(
            'Missing files in repo root: ' + ', '.join(missing) +
            '\n\nMake sure these files are in the same folder as app.py:\n'
            '- lightgbm_model.pkl\n- featurer.pkl\n- planet_encoder.pkl\n- star_encoder.pkl'
        )

    model = joblib.load(MODEL_PATH)
    features = joblib.load(FEATURES_PATH)
    le_planet = joblib.load(PLANET_ENCODER_PATH)
    le_star = joblib.load(STAR_ENCODER_PATH)
    return model, features, le_planet, le_star

def safe_transform(le, value: str, col_name: str) -> int:
    """Transform a single category value with a saved LabelEncoder.
    If unseen value appears, show a helpful error."""
    try:
        return int(le.transform([value])[0])
    except Exception:
        known = list(getattr(le, 'classes_', []))
        st.error(f'Unknown category for {col_name}: {value}. Known values: {known}')
        st.stop()

model, FEATURES, le_planet, le_star = load_artifacts()

st.title('🪐 Star System Classification (LightGBM)')
st.write('Predict the star system type using 10 astrophysical measurements (multiclass).')

with st.expander('ℹ️ Required files in this folder', expanded=False):
    st.code(
        'app.py\n'
        'lightgbm_model.pkl\n'
        'featurer.pkl\n'
        'planet_encoder.pkl\n'
        'star_encoder.pkl\n'
        'requirements.txt'
    )

st.subheader('Enter feature values')

# --- Inputs ---
# Numeric
star_size = st.number_input('star_size', min_value=0.0, value=1.0, step=0.01)
star_brightness = st.number_input('star_brightness', min_value=0.0, value=1.2, step=0.01)
distance_from_earth = st.number_input('distance_from_earth', min_value=0.0, value=90.0, step=1.0)
star_mass = st.number_input('star_mass', min_value=0.0, value=1.3, step=0.01)
metallicity = st.number_input('metallicity', value=0.02, step=0.001, format='%.4f')

# Discrete numeric / encoded-like
galaxy_region = st.selectbox('galaxy_region', options=[0, 1, 2], index=1)
galaxy_type = st.selectbox('galaxy_type', options=[0, 1, 2], index=0)

# Categorical (original strings)
star_spectral_class = st.selectbox(
    'star_spectral_class',
    options=list(le_star.classes_),
    index=0
)

planet_configuration = st.selectbox(
    'planet_configuration',
    options=list(le_planet.classes_),
    index=0
)

stellar_activity_class = st.selectbox(
    'stellar_activity_class',
    options=['Low', 'Medium', 'High'],
    index=0
)

# --- Build row in the ORIGINAL feature space ---
row = {
    'star_size': float(star_size),
    'star_brightness': float(star_brightness),
    'galaxy_region': int(galaxy_region),
    'distance_from_earth': float(distance_from_earth),
    'galaxy_type': int(galaxy_type),
    'star_spectral_class': star_spectral_class,
    'planet_configuration': planet_configuration,
    'stellar_activity_class': stellar_activity_class,
    'star_mass': float(star_mass),
    'metallicity': float(metallicity),
}

# --- Apply same preprocessing as training ---
# Mapping for activity (ordinal)
row['stellar_activity_class'] = ACTIVITY_MAP[row['stellar_activity_class']]

# LabelEncoders for the other two categorical columns
row['planet_configuration'] = safe_transform(le_planet, planet_configuration, 'planet_configuration')
row['star_spectral_class'] = safe_transform(le_star, star_spectral_class, 'star_spectral_class')

# Make DataFrame and enforce correct column order
X_input = pd.DataFrame([row])

# Ensure all expected feature columns exist
missing_cols = [c for c in FEATURES if c not in X_input.columns]
extra_cols = [c for c in X_input.columns if c not in FEATURES]
if missing_cols:
    st.error(f'Missing columns for model: {missing_cols}')
    st.stop()
if extra_cols:
    # Not an error, but we will drop extras to be safe
    X_input = X_input.drop(columns=extra_cols)

X_input = X_input[FEATURES]

st.divider()

col1, col2 = st.columns(2)

with col1:
    if st.button('🔮 Predict', use_container_width=True):
        pred = model.predict(X_input)[0]
        pred_int = int(pred)
        label = LABEL_NAMES.get(pred_int, str(pred_int))
        st.success(f'Prediction: **{label}** (class {pred_int})')

with col2:
    if st.button('📊 Predict probabilities', use_container_width=True):
        if hasattr(model, 'predict_proba'):
            proba = model.predict_proba(X_input)[0]
            proba_df = pd.DataFrame({'class': list(range(len(proba))), 'probability': proba}).sort_values('probability', ascending=False)
            proba_df['label'] = proba_df['class'].map(LABEL_NAMES).fillna(proba_df['class'].astype(str))
            st.dataframe(proba_df[['label', 'class', 'probability']], use_container_width=True)
        else:
            st.warning('This model does not support predict_proba().')

st.caption('Tip: If predictions look wrong, ensure the same encoders and feature order are used as during training.')