import pickle from pathlib import Path import numpy as np import pandas as pd import streamlit as st import joblib import matplotlib.pyplot as plt from tensorflow.keras.models import load_model # ------------------------- # Page config # ------------------------- st.set_page_config( page_title="Airline Passengers LSTM Forecast", page_icon="✈️", layout="wide" ) st.title("✈️ Airline Passengers Forecast (LSTM)") st.caption("Time Series Forecasting using an LSTM model (train: up to 1958, test: 1959–1960).") # ------------------------- # Paths (robust: root or src) # ------------------------- BASE_DIR = Path(__file__).resolve().parent CANDIDATE_DIRS = [BASE_DIR, BASE_DIR / "src"] def find_file(filename: str) -> Path | None: for d in CANDIDATE_DIRS: p = d / filename if p.exists(): return p return None MODEL_PATH = find_file("lstm_airline_model.keras") SCALER_PATH = find_file("scaler.pkl") METRICS_PATH = find_file("metrics.pkl") DATA_URL = "https://raw.githubusercontent.com/amankharwal/Website-data/master/airline-passengers.csv" LOOKBACK = 12 # must match training # ------------------------- # Load artifacts # ------------------------- @st.cache_resource def load_artifacts(): if MODEL_PATH is None: raise FileNotFoundError("Missing model file: lstm_airline_model.keras") if SCALER_PATH is None: raise FileNotFoundError("Missing scaler file: scaler.pkl") model = load_model(MODEL_PATH, compile=False) scaler = joblib.load(SCALER_PATH) return model, scaler # ------------------------- # Load data # ------------------------- @st.cache_data def load_data(): df = pd.read_csv(DATA_URL) # Standardize df["Month"] = pd.to_datetime(df["Month"]) df = df.sort_values("Month").set_index("Month") df["Passengers"] = df["Passengers"].astype("float32") return df def create_sequences(data_2d: np.ndarray, lookback: int = 12): # data_2d shape: (n, 1) X = np.array([data_2d[i-lookback:i] for i in range(lookback, len(data_2d))]) y = data_2d[lookback:] return X, y def inverse_safe(scaler, arr_2d: np.ndarray): # scaler expects 2D return scaler.inverse_transform(arr_2d) def rolling_forecast(model, scaler, last_values: np.ndarray, steps: int, lookback: int = 12): """ last_values: unscaled original values, shape (lookback, 1) steps: how many future points to forecast """ # scale with the saved scaler window = scaler.transform(last_values.astype("float32")) preds_scaled = [] for _ in range(steps): x = window.reshape(1, lookback, 1) pred_scaled = model.predict(x, verbose=0) # shape (1,1) preds_scaled.append(pred_scaled[0, 0]) # shift window window = np.vstack([window[1:], pred_scaled]) preds_scaled = np.array(preds_scaled).reshape(-1, 1) preds = inverse_safe(scaler, preds_scaled) return preds # ------------------------- # Sidebar controls # ------------------------- st.sidebar.header("Settings") forecast_steps = st.sidebar.slider("Forecast months into the future", min_value=1, max_value=36, value=12, step=1) show_metrics = st.sidebar.checkbox("Show saved metrics (metrics.pkl)", value=True) show_test_plot = st.sidebar.checkbox("Show test forecast plot (1959–1960)", value=True) show_future_plot = st.sidebar.checkbox("Show future forecast plot", value=True) # ------------------------- # Main # ------------------------- try: model, scaler = load_artifacts() df = load_data() except Exception as e: st.error(str(e)) st.stop() col1, col2 = st.columns([1.2, 1]) with col1: st.subheader("Dataset preview") st.dataframe(df.head(10), use_container_width=True) with col2: st.subheader("Basic info") st.write(f"Rows: **{len(df)}**") st.write(f"Date range: **{df.index.min().date()} → {df.index.max().date()}**") st.write("Lookback window:", f"**{LOOKBACK} months**") if show_metrics and METRICS_PATH is not None: st.subheader("Saved metrics") try: m = joblib.load(METRICS_PATH) c1, c2, c3 = st.columns(3) c1.metric("RMSE", f"{m['rmse']:.2f}") c2.metric("MAE", f"{m['mae']:.2f}") c3.metric("MAPE (%)", f"{m['mape']:.2f}") except Exception as e: st.info(f"metrics.pkl found but could not be read. ({e})") st.divider() # Plot original series st.subheader("Time series") fig = plt.figure(figsize=(10, 4)) plt.plot(df.index, df["Passengers"]) plt.title("Monthly Airline Passengers (1949–1960)") plt.xlabel("Date") plt.ylabel("Passengers") plt.grid(alpha=0.3) st.pyplot(fig) # ------------------------- # Test forecast (1959–1960) # ------------------------- if show_test_plot: st.subheader("Test forecast (1959–1960)") # Split exactly like training train = df.loc[: "1958", ["Passengers"]].values.astype("float32") test = df.loc["1959":, ["Passengers"]].values.astype("float32") # Scale with loaded scaler (fitted on train during your training) train_scaled = scaler.transform(train) # NOTE: scaler already fitted test_scaled = scaler.transform(test) # Sequences x_test, y_test = create_sequences(test_scaled, LOOKBACK) # Predict y_pred_scaled = model.predict(x_test, verbose=0) y_pred = inverse_safe(scaler, y_pred_scaled) y_true = inverse_safe(scaler, y_test) test_dates = df.loc["1959":].index[LOOKBACK:] fig2 = plt.figure(figsize=(10, 4)) plt.plot(test_dates, y_true, label="Actual") plt.plot(test_dates, y_pred, label="Predicted") plt.title("LSTM Forecast vs Actual (Test)") plt.xlabel("Date") plt.ylabel("Passengers") plt.legend() plt.grid(alpha=0.3) st.pyplot(fig2) # ------------------------- # Future forecast # ------------------------- if show_future_plot: st.subheader("Future forecast") last_window = df[["Passengers"]].values.astype("float32")[-LOOKBACK:] future_preds = rolling_forecast(model, scaler, last_window, steps=forecast_steps, lookback=LOOKBACK) last_date = df.index.max() future_dates = pd.date_range(last_date + pd.offsets.MonthBegin(1), periods=forecast_steps, freq="MS") fig3 = plt.figure(figsize=(10, 4)) plt.plot(df.index, df["Passengers"].values, label="History") plt.plot(future_dates, future_preds, label="Forecast") plt.title("Future Forecast") plt.xlabel("Date") plt.ylabel("Passengers") plt.legend() plt.grid(alpha=0.3) st.pyplot(fig3) st.success("✅ App ready. Make sure model/scaler files are in the repo (root or src).")