app.py

import os, json, calendar
from datetime import date, timedelta

# Keras 3 поверх TF
os.environ["KERAS_BACKEND"] = "tensorflow"
import numpy as np
import pandas as pd
import joblib
import gradio as gr
import keras  # standalone Keras 3

# ====== ПУТИ К АРТЕФАКТАМ ======
MODEL_DIR = "neiro1"  # в Space лежит папка neiro1 рядом с app.py
MODEL_PATH = os.path.join(MODEL_DIR, "model_v6.keras")
PP_PATH    = os.path.join(MODEL_DIR, "preprocess_v6.joblib")
META_PATH  = os.path.join(MODEL_DIR, "meta_v6.json")

# ====== ЗАГРУЗКА ======
model = keras.models.load_model(MODEL_PATH, compile=False)
ct    = joblib.load(PP_PATH)
with open(META_PATH, "r", encoding="utf-8") as f:
    META = json.load(f)

# ====== ПАРАМЕТРЫ И МЕТАДАННЫЕ ======
ROUTE = META.get("route", "SFO_MIA")
ORIGIN, DEST = ROUTE.split("_")
AIRLINES = META.get("airlines", [])
M_BLEND  = {int(k): float(v) for k, v in META.get("blend_month", {}).items()}
A_BLEND  = {str(k): float(v) for k, v in META.get("blend_airline", {}).items()}
AM_MEANS = {k: float(v) for k, v in META.get("am_means", {}).items()}
AM_COUNTS= {k: int(v)   for k, v in META.get("am_counts", {}).items()}
W_MONTH, W_AIR, W_AM = META.get("w_month", 0.0), META.get("w_air", 0.0), META.get("w_am", 0.0)
FLOOR_FRAC = float(META.get("floor_frac", 0.25))
EXOG       = META.get("exog_params", {})
K_AIRLINE  = float(META.get("calibration_k_airline", 1.0))
K_ROUTE    = float(META.get("calibration_k_route", 1.0))
WHITELIST  = META.get("airline_whitelist", AIRLINES)
SHARE      = META.get("airline_share_overall", {})
CAP_TARGET = float(META.get("capacity_target_summer", 1000.0))

# ====== ХЕЛПЕРЫ ======
def nth_weekday(year, month, weekday, n):
    d = date(year, month, 1)
    add = (weekday - d.weekday()) % 7
    return d + timedelta(days=add) + timedelta(weeks=n-1)

def last_weekday(year, month, weekday):
    d = date(year, month, calendar.monthrange(year, month)[1])
    sub = (d.weekday() - weekday) % 7
    return d - timedelta(days=sub)

def daterange(d0, d1):
    cur, out = d0, []
    while cur <= d1:
        out.append(cur)
        cur += timedelta(days=1)
    return out

def holiday_windows_for_year(y):
    mem  = set(daterange(last_weekday(y,5,0)-timedelta(days=3), last_weekday(y,5,0)))
    july = set(daterange(date(y,6,30), date(y,7,7)))
    labor= set(daterange(nth_weekday(y,9,0,1)-timedelta(days=3), nth_weekday(y,9,0,1)))
    tg   = set(daterange(nth_weekday(y,11,3,4)-timedelta(days=2), nth_weekday(y,11,3,4)+timedelta(days=3)))
    xny  = set(daterange(date(y,12,20), date(y,12,31))) | set(daterange(date(y+1,1,1), date(y+1,1,5)))
    return {"memorial":mem, "july4":july, "labor":labor, "thanksgiving":tg, "xmas_newyear":xny}

def exog_multiplier(ts: pd.Timestamp) -> float:
    y = int(ts.year); m = int(ts.month); dow = int(ts.weekday())
    P = {k: float(EXOG.get(k, v)) for k, v in dict(
        SUMMER_MULT=1.60, FRI_MULT=1.10, SAT_MULT=1.15, SUN_MULT=1.10,
        SPRING_BREAK_MULT=1.12, MEMORIAL_MULT=1.35, JULY4_MULT=1.30,
        LABOR_MULT=1.25, THANKSGIVING_MULT=1.60, XMAS_NEWYEAR_MULT=1.70
    ).items()}
    mult = 1.0
    if m in (6,7,8): mult *= P["SUMMER_MULT"]
    if pd.Timestamp(year=y, month=3, day=10) <= ts <= pd.Timestamp(year=y, month=4, day=10):
        mult *= P["SPRING_BREAK_MULT"]
    if dow == 4: mult *= P["FRI_MULT"]
    if dow == 5: mult *= P["SAT_MULT"]
    if dow == 6: mult *= P["SUN_MULT"]
    hw = holiday_windows_for_year(y); hw_prev = holiday_windows_for_year(y-1)
    if ts.date() in hw["memorial"]:     mult *= P["MEMORIAL_MULT"]
    if ts.date() in hw["july4"]:        mult *= P["JULY4_MULT"]
    if ts.date() in hw["labor"]:        mult *= P["LABOR_MULT"]
    if ts.date() in hw["thanksgiving"]: mult *= P["THANKSGIVING_MULT"]
    if ts.date() in hw["xmas_newyear"] or ts.date() in hw_prev["xmas_newyear"]:
        mult *= P["XMAS_NEWYEAR_MULT"]
    return float(mult)

def _blend(child_mean, child_n, prior_mean, prior_w):
    child_n = float(child_n)
    return float((child_n*child_mean + prior_w*prior_mean) / (child_n + prior_w))

def baseline_pair(a, m):
    gmean = np.mean(list(A_BLEND.values())) if A_BLEND else 0.0
    a0 = A_BLEND.get(str(a), gmean)
    m0 = M_BLEND.get(int(m), gmean)
    prior = 0.5*a0 + 0.5*m0
    key = f"{a}_{int(m)}"
    if key in AM_MEANS:
        mean = AM_MEANS[key]; cnt = AM_COUNTS.get(key, 0)
        return _blend(mean, cnt, prior, W_AM)
    return float(prior)

def make_inputs(airline: str, date_str: str):
    d = pd.to_datetime(date_str)
    dow, month = int(d.weekday()), int(d.month)
    weekofyear = int(d.isocalendar().week)
    is_weekend = int(dow in [5,6])
    sin_dow, cos_dow = np.sin(2*np.pi*dow/7), np.cos(2*np.pi*dow/7)
    sin_mon, cos_mon = np.sin(2*np.pi*(month-1)/12), np.cos(2*np.pi*(month-1)/12)
    exog = exog_multiplier(d)
    X_df = pd.DataFrame([{
        "airline": airline,
        "sin_dow": sin_dow, "cos_dow": cos_dow,
        "sin_month": sin_mon, "cos_month": cos_mon,
        "weekofyear": weekofyear, "is_weekend": is_weekend,
        "exog_mult": exog,
    }])
    base = baseline_pair(airline, month) * exog
    return ct.transform(X_df), np.array([[base]], dtype=float), dict(
        dow=dow, month=month, weekofyear=weekofyear, is_weekend=bool(is_weekend),
        exog=exog, baseline=float(base)
    )

def predict_one(airline, date_str):
    X, B, info = make_inputs(airline, date_str)
    y_raw = float(model.predict([X, B], verbose=0)[0,0])
    y_cal = y_raw * K_AIRLINE
    # safety floor: минимум 25% от baseline (или META['floor_frac'])
    floor_val = float(B[0,0]) * max(FLOOR_FRAC, 0.25)
    y_cal = max(y_cal, floor_val)
    return y_cal, info, y_raw

# ====== UI ======
with gr.Blocks(title="Прогноз пассажиров (v6.2)") as demo:
    gr.Markdown("### Прогноз числа пассажиров (учтены сезонность/праздники)")
    with gr.Row():
        gr.Textbox(label="Маршрут", value=ROUTE.replace('_', '→'), interactive=False)
        dd_airline = gr.Dropdown(
            choices=(WHITELIST or AIRLINES or ["American airlines","United"]),
            value=(WHITELIST[0] if WHITELIST else (AIRLINES[0] if AIRLINES else "American airlines")),
            label="Авиакомпания",
            allow_custom_value=False
        )
        tb_date = gr.Textbox(label="Дата (YYYY-MM-DD)", value=str(pd.Timestamp.today().date()))
    sum_all = gr.Checkbox(label="Суммарно по всем авиакомпаниям", value=True)

    def predict_ui(airline, date_str, sum_all):
        # простая валидация даты
        try:
            _ = pd.to_datetime(date_str)
        except Exception:
            return 0, "Некорректная дата, формат YYYY-MM-DD."

        if sum_all:
            total_raw, lines = 0.0, []
            for a in (WHITELIST or AIRLINES or [airline]):
                X, B, info = make_inputs(a, date_str)
                y_raw = float(model.predict([X, B], verbose=0)[0,0])
                total_raw += y_raw
                lines.append(f"- {a}: base×M={info['baseline']:.1f} (M={info['exog']:.2f}) → raw {y_raw:.1f}")
            total = total_raw * K_ROUTE
            msg = (
                f"Маршрут: **{ORIGIN}→{DEST}**, дата: **{date_str}**  \n"
                f"Суммарный прогноз: **{round(total)}** "
                f"(raw_sum={total_raw:.1f}, k_route={K_ROUTE:.2f}; целевой летний≈{int(CAP_TARGET)})\n"
                + "\n".join(lines)
            )
            return round(total), msg

        y, info, y_raw = predict_one(airline, date_str)
        warn = ""
        if SHARE and SHARE.get(airline, 0.0) < 0.05:
            warn = (
                "\n\n_Примечание: доля этой авиакомпании на маршруте <5%, "
                "прогноз может быть занижен; попробуйте «Суммарно по всем»._"
            )
        msg = (
            f"Маршрут: **{ORIGIN}→{DEST}**, авиакомпания: **{airline}**, дата: **{date_str}**  \n"
            f"M(date)={info['exog']:.2f}; baseline×M={info['baseline']:.1f}  \n"
            f"Прогноз: **{round(y)}** (raw={y_raw:.1f}, k_airline={K_AIRLINE:.2f}){warn}"
        )
        return round(y), msg

    btn = gr.Button("Рассчитать")
    out_pred = gr.Number(label="Прогноз пассажиров", interactive=False)
    out_expl = gr.Markdown()

    btn.click(predict_ui, [dd_airline, tb_date, sum_all], [out_pred, out_expl])

    # очередь без спорных параметров (кросс-версийно)
    try:
        demo.queue(max_size=32)
    except TypeError:
        demo.queue()

# HF Spaces сам запускает; локально — так:
if __name__ == "__main__":
    demo.launch(server_name="0.0.0.0", server_port=7860)