# ダミーデータを作成
import pandas as pd
df = pd.DataFrame()
df["instagram_cost"] = np.random.normal(1000, 100, 365)
df["instagram_imp"] = df["instagram_cost"] * 10
df["instagram_install"] = df["instagram_cost"] / 1.0 +  np.random.normal(100, 10, 365)
df["x_cost"] = np.random.normal(2000, 200, 365)
df["x_imp"] = df["x_cost"] * 10
df["x_install"] = df["x_cost"] / 2.0 +  np.random.normal(100, 10, 365)
df["google_cost"] = np.random.normal(3000, 300, 365)
df["google_imp"] = df["google_cost"] * 10
df["google_install"] = df["google_cost"] / 3.0 + np.random.normal(100, 10, 365)

# 仮に真の成果がGoogleとXとinstagramが1:1:1だったとする
df["target"] = df["google_install"]  + df["x_install"]  + df["instagram_install"] 

from lightweight_mmm import preprocessing, lightweight_mmm, plot, optimize_media
import jax.numpy as jnp

# impression等
media_scaled      = preprocessing.CustomScaler(divide_operation=jnp.mean).fit_transform(df[["instagram_imp", "x_imp", "google_imp"]].values)
# 目的変数
target_scaled     = preprocessing.CustomScaler(divide_operation=jnp.mean).fit_transform(df[["target"]].values)
# 投資したコスト等
costs_scaled      = preprocessing.CustomScaler(divide_operation=jnp.mean).fit_transform(df[["instagram_cost", "x_cost", "google_cost"]].mean(axis=0).values)
mmm = lightweight_mmm.LightweightMMM(model_name="hill_adstock")
mmm.fit(
        media       = media_scaled,
        media_prior = costs_scaled,
        target      = target_scaled,
        number_warmup=500, # 1000
        number_samples=500, # 1000
        number_chains=1,
        degrees_seasonality=1,
        weekday_seasonality=True,
        seasonality_frequency=365,
        seed=1)

effect_hat, roi_hat = mmm.get_posterior_metrics()
rf = pd.DataFrame()
rf["media_name"] = columns
rf["effect_hat"] = media_effect_hat.mean(axis=0)
rf["roi_hat"] = roi_hat.mean(axis=0)
rf.sort_values(by=["effect_hat"], ascending=False)