Unable to extract LR coefficients

[jwretham]

Hi, I'm using NP as part of a forecasting tool I'm building.

I'd really like to be able to access the linear regression coefficient estimates for my additional regressors.

In Prophet I'm able to do this using the regressor_coefficients() function, however I can't find anything in the NP documentation about how to do this.

Any help/suggestions appreciated

Thanks

Joe

[weberpals]

Hi Joe,

I was able to get the coefficients like this. Not sure if you are only interested in future regressors, but I also added lagged and event just in case:

def get_all_regressor_coefficients(model):
    # Access the TimeNet model 
    time_net = model.model
    coefficients = {}

    # Future Regressors
    if time_net.config_regressors is not None and time_net.config_regressors.regressors is not None:
        for name, config in time_net.config_regressors.regressors.items():
            regressor_component = time_net.future_regressors
            if config.mode == 'additive':
                coefficients[name] = regressor_component.regressor_params['additive'].data.cpu().numpy()
            elif config.mode == 'multiplicative':
                coefficients[name] = regressor_component.regressor_params['multiplicative'].data.cpu().numpy()
    
    # Lagged Regressors
    if time_net.config_lagged_regressors is not None:
        for name, config in time_net.config_lagged_regressors.items():
            layer_weights = []
            for layer in time_net.covar_net:
                if isinstance(layer, nn.Linear):
                    layer_weights.append(layer.weight.data.cpu().numpy())
            if layer_weights:
                coefficients[name] = np.array(layer_weights)

    
    # Event Regressors
    if time_net.config_events is not None:
        for event, event_config in time_net.config_events.items():
            if event_config.mode == 'additive':
                coefficients[event] = time_net.event_params['additive'].data.cpu().numpy()
            elif event_config.mode == 'multiplicative':
                coefficients[event] = time_net.event_params['multiplicative'].data.cpu().numpy()

    return coefficients

# Get the regressor coefficients
coefficients = get_all_regressor_coefficients(m)
print("Coefficients:", temp_coefficients)

I can also create a PR to make the coefficients more easily accessible. @ourownstory would be great to know what you think Oskar.

Hope this helps!

Constantin

[jwretham]

Hi Constantin - so sorry, I absolutely missed that you'd responded on this. Thanks for taking the time to put this together - let me give it a try and I'll get back to you.

Joe