How to incorporate pipeline in bayesian_optuna_search?

[andreale28]

Hi there, recently I have tried skforecast version 0.5.x with _bayesian_optuna_search. In order to make sklearn SVR work properly, I have to add StandardScaler and intend to use Pipeline.

However, after several trials, I cannot implement the pipeline in _bayesian_optuna_search. Are there any solutions?

[JoaquinAmatRodrigo]

Hi @andreale28
Could you add a example of code and the output error?

Thanks

[andreale28]

hi @JoaquinAmatRodrigo , this is my code:

lags_grid = [i for i in range(5, 20, 1)]
steps = 18
pipeline = Pipeline(steps=[
	('scaler', preprocessor),
	('regression', SVR(gamma='scale'))
])
svr_forecaster = ForecasterAutoreg(
		regressor=pipeline,
		lags=5
)
def search_space_svr(trial):
	space = {

		'kernel' : trial.suggest_categorical('kernel', ['linear', 'poly', 'rbf']),
		'degree' : trial.suggest_int('degree', 2, 5),
		'C'      : trial.suggest_float('C', 1e-2, 1e+2, log=True),
		'epsilon': trial.suggest_float('epsilon', 1e-3, 1e-1, log=True),
	}

	return space


results_svr, results_opt_dict_svr = _bayesian_search_optuna(
		forecaster=svr_forecaster,
		y=data_vn.loc [:end_val, 'new_confirmed'],
		exog=data_vn.loc [:end_val, exog_var],
		search_space=search_space_svr,
		lags_grid=lags_grid,
		steps=steps,
		refit=False,
		metric='mean_absolute_percentage_error',
		n_trials=10,
		initial_train_size=len(df_confirmed_train),
		fixed_train_size=False,
		return_best=True,
		verbose=False,
		kwargs_create_study={'direction': 'minimize'},
		kwargs_study_optimize={'gc_after_trial': True}
)
results_svr

The error return is
ValueError: Invalid parameter kernel for estimator Pipeline(steps=[('scaler', StandardScaler()), ('regression', SVR())]). Check the list of available parameters with estimator.get_params().keys().

[JoaquinAmatRodrigo]

Hi @andreale28,
The model's name precedes the parameters' name when performing grid search over a sklearn pipeline.
https://joaquinamatrodrigo.github.io/skforecast/0.4.3/notebooks/sklearn-pipeline.html#grid-search

I think the problem may be solved if you use:

`
def search_space_svr(trial):
space = {

	'regression__kernel' : trial.suggest_categorical('kernel', ['linear', 'poly', 'rbf']),
	'regression__degree' : trial.suggest_int('degree', 2, 5),
	'regression__C'      : trial.suggest_float('C', 1e-2, 1e+2, log=True),
	'regression__epsilon': trial.suggest_float('epsilon', 1e-3, 1e-1, log=True),
}

return space

`

[andreale28]

Hi @andreale28, The model's name precedes the parameters' name when performing grid search over a sklearn pipeline. https://joaquinamatrodrigo.github.io/skforecast/0.4.3/notebooks/sklearn-pipeline.html#grid-search

I think the problem may be solved if you use:

` def search_space_svr(trial): space = {

	'regression__kernel' : trial.suggest_categorical('kernel', ['linear', 'poly', 'rbf']),
	'regression__degree' : trial.suggest_int('degree', 2, 5),
	'regression__C'      : trial.suggest_float('C', 1e-2, 1e+2, log=True),
	'regression__epsilon': trial.suggest_float('epsilon', 1e-3, 1e-1, log=True),
}

return space

`

I will try and report it later, thanks for help