src/streamlit_app.py

import json
import joblib
import numpy as np
import pandas as pd
import streamlit as st
from pathlib import Path

# -------------------------
# Page config
# -------------------------
st.set_page_config(
    page_title='Sales Forecast (LightGBM)',
    page_icon='📈',
    layout='centered'
)

st.title('📈 Sales Forecast (LightGBM)')
st.write('Predict **num_sold** using a trained LightGBM model + saved encoders and preprocessing.')

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

MODEL_PATH = BASE_DIR / 'model_lgbm.pkl'
FEATURES_PATH = BASE_DIR / 'feature_names.pkl'
ENCODERS_PATH = BASE_DIR / 'encoders.pkl'
FILLMAP_PATH = BASE_DIR / 'fill_map.pkl'
META_PATH = BASE_DIR / 'meta.json'


@st.cache_resource
def load_assets():
    if not MODEL_PATH.exists():
        raise FileNotFoundError(f'Missing {MODEL_PATH.name} (put it next to app.py).')
    if not FEATURES_PATH.exists():
        raise FileNotFoundError(f'Missing {FEATURES_PATH.name} (put it next to app.py).')
    if not ENCODERS_PATH.exists():
        raise FileNotFoundError(f'Missing {ENCODERS_PATH.name} (put it next to app.py).')
    if not FILLMAP_PATH.exists():
        raise FileNotFoundError(f'Missing {FILLMAP_PATH.name} (put it next to app.py).')

    model = joblib.load(MODEL_PATH)
    features = joblib.load(FEATURES_PATH)
    encoders = joblib.load(ENCODERS_PATH)
    fill_map = joblib.load(FILLMAP_PATH)

    meta = None
    if META_PATH.exists():
        with open(META_PATH, 'r') as f:
            meta = json.load(f)

    return model, features, encoders, fill_map, meta


model, FEATURES, encoders, fill_map, meta = load_assets()

with st.expander('ℹ️ Model info'):
    if meta:
        st.write(meta)
    else:
        st.write('No meta.json found.')


# -------------------------
# Helpers
# -------------------------
def make_date_features(date_value: pd.Timestamp) -> dict:
    # date_value is a Timestamp
    year = int(date_value.year)
    month = int(date_value.month)
    week = int(date_value.isocalendar().week)
    dayofweek = int(date_value.dayofweek)  # Monday=0
    is_weekend = int(dayofweek >= 5)
    dayofyear = int(date_value.dayofyear)

    return {
        'year': year,
        'month': month,
        'week': week,
        'dayofweek': dayofweek,
        'is_weekend': is_weekend,
        'dayofyear': dayofyear
    }


def safe_encode(col_name: str, value: str) -> int:
    # If unseen label appears, fall back to the most frequent label (index 0) or safe default.
    le = encoders.get(col_name)
    if le is None:
        return 0

    classes = set(le.classes_.astype(str))
    v = str(value)

    if v in classes:
        return int(le.transform([v])[0])

    # fallback: use first known class
    return int(le.transform([str(le.classes_[0])])[0])


# -------------------------
# UI Inputs
# -------------------------
st.subheader('🧾 Input')

date_in = st.date_input('Date', value=pd.to_datetime('2019-01-01'))
country_in = st.text_input('Country', value='Finland')
store_in = st.text_input('Store', value='KaggleMart')
product_in = st.text_input('Product', value='Kaggle Mug')

st.markdown('---')
st.subheader('⏳ Lag features')

use_manual_lags = st.checkbox('Enter lag values manually (recommended if you know them)', value=False)

default_lag_364 = float(fill_map.get('lag_364', 0.0))
default_lag_365 = float(fill_map.get('lag_365', 0.0))
default_lag_371 = float(fill_map.get('lag_371', 0.0))

if use_manual_lags:
    lag_364 = st.number_input('lag_364', value=default_lag_364)
    lag_365 = st.number_input('lag_365', value=default_lag_365)
    lag_371 = st.number_input('lag_371', value=default_lag_371)
else:
    st.write('Using default lag values (from training medians):')
    st.write({
        'lag_364': default_lag_364,
        'lag_365': default_lag_365,
        'lag_371': default_lag_371
    })
    lag_364, lag_365, lag_371 = default_lag_364, default_lag_365, default_lag_371


# -------------------------
# Predict
# -------------------------
if st.button('Predict'):
    date_ts = pd.to_datetime(date_in)

    d_feats = make_date_features(date_ts)

    row = {}
    row.update(d_feats)

    row['lag_364'] = float(lag_364)
    row['lag_365'] = float(lag_365)
    row['lag_371'] = float(lag_371)

    row['country'] = safe_encode('country', country_in)
    row['store'] = safe_encode('store', store_in)
    row['product'] = safe_encode('product', product_in)

    X = pd.DataFrame([row])

    # Ensure all FEATURES exist and order is correct
    for c in FEATURES:
        if c not in X.columns:
            # numeric fallback from fill_map
            X[c] = fill_map.get(c, 0.0)

    X = X[FEATURES].copy()

    # Fill numeric NaNs just in case
    for c in X.columns:
        if X[c].isna().any():
            X[c] = X[c].fillna(fill_map.get(c, 0.0))

    pred = model.predict(X)[0]

    st.success(f'✅ Predicted num_sold: **{pred:.2f}**')
    st.caption('Note: Lag features heavily influence the forecast. If you can compute real lags, the prediction will be more accurate.')

    with st.expander('Show model input vector'):
        st.dataframe(X)