How to retrieve trained Hyperparameters in Facebook Prophet (Python)?

Question

I want to retrieve the hyperparameters after training a time-series model using Facebook Prophet, so that I can use the learned values in a second model.

Here's what I tried.

    from prophet import Prophet

    # Train the first DataFrame
    m1 = Prophet() # initialized with default parameters
    m1.fit(df1)
    future = m1.make_future_dataframe(periods=365)
    forecast = m1.predict(df1)

    m2 = Prophet(
        daily_seasonality = m1.daily_seasonality,
        weekly_seasonality = m1.weekly_seasonality,
        yearly_seasonality = m1.yearly_seasonality,
        seasonality_prior_scale = m1.seasonality_prior_scale,
        changepoint_prior_scale = m1.changepoint_prior_scale,
        uncertainty_samples = m1.uncertainty_samples,
        seasonality_mode = m1.seasonality_mode,
        interval_width = m1.interval_width,
        n_changepoints = m1.n_changepoints,
    ) # initialized with trained parameters from m1

    # Train the second DataFrame
    m2.fit(df2)

However, when I try debugging the hyparameter values, it looks like m2 is initialized with the same values m1 was initialized with.

How do I get the trained parameters from m1?

Answer 1

You can use MLFlow for the purpose.

import mlflow
from prophet import Prophet, serialize
def extract_params(pr_model):
    return {attr: getattr(pr_model, attr) for attr in serialize.SIMPLE_ATTRIBUTES}

mlflow.log_params(params)    
m1_params = extract_params(m1)

Afterward it should be simple to set the m2 hyper parameters with the m1 ones.