Medium "How to actually forecast COVID-19" embed

eval.py

from sir_models.utils import eval_on_select_dates_and_k_days_ahead
from sir_models.utils import smape
from sklearn.metrics import mean_absolute_error

K = 30
last_day = train_subset.date.iloc[-1] - pd.to_timedelta(K, unit='D')
eval_dates = pd.date_range(start='2020-06-01', end=last_day)[::20]

def eval_hidden_moscow(train_df, t, train_t, eval_t):
    weights = {
        'I': 0.25,
        'R': 0.25,
        'D': 0.5,
    }
    model = SEIRHidden()
    fitter = HiddenCurveFitter(
        new_deaths_col='deaths_per_day_ma7',
        new_cases_col='infected_per_day_ma7',
        new_recoveries_col='recovered_per_day_ma7',
        weights=weights,
        max_iters=1000,
        save_params_every=500)
    fitter.fit(model, train_df)

    train_initial_conditions = model.get_initial_conditions(train_df)
    train_states, history = model.predict(train_t, train_initial_conditions, history=False)

    test_initial_conds = [compartment[-1] for compartment in train_states]
    test_states, history = model.predict(eval_t, test_initial_conds, history=False)
        
    return model, fitter, test_states
    
(models, fitters, 
model_predictions, 
train_dfs, test_dfs) = eval_on_select_dates_and_k_days_ahead(train_subset,
                                        eval_func=eval_hidden_moscow,
                                        eval_dates=eval_dates,
                                        k=K)
model_pred_D = [pred[7] for pred in model_predictions]
true_D = [tdf.total_dead.values for tdf in test_dfs]
baseline_pred_D = [[tdf.iloc[-1].total_dead]*K for tdf in train_dfs]

overall_errors_model = [mean_absolute_error(true, pred) for true, pred in zip(true_D, model_pred_D)]
overall_errors_baseline = [mean_absolute_error(true, pred) for true, pred in zip(true_D, baseline_pred_D)]

print('Mean overall error baseline', np.mean(overall_errors_baseline).round(3))
print('Mean overall error model', np.mean(overall_errors_model).round(3))
overall_smape_model = [smape(true, pred) for true, pred in zip(true_D, model_pred_D)]
np.median(overall_smape_model)