Frequently Asked Questions

What is difference between parent and child iterations/algorithm?

AutoTab operates based upon parent and child optimization iterations The parent iteration is responsible for preprocessing step optimization and model optimization. During each parent iteration, when the preprocessing and model is selected/suggested by the algorithm for this iteration, the child optimization loops starts. The job of child optimization loop is to optimize hyperparameters of the selected/suggested model. The user can specify any algorithm from following algorithms for parent and child optimization algorithms.

bayes

random

grid

bayes_rf

tpe

atpe

cmaes

what splitting scheme is used

By default it is supposed that the data is split into 3 sets i.e. training, validation and test sets. validation data is only used during pipeline optimization inside .fit method while the test data is only used after optimization. If you have only two sets i.e. training and validation, set fit_on_all_train_data to False during post_fit

Is the pipeline optimized for test data or validation data?

for validation data

I don’t want to optimize preprocessing step

If you dont want any preprocessing steps, keep inputs_to_transform and outputs_to_transform arguments equal to None or an empty list. In this way transformations will not be optimized for both inputs and targets. As shown in below example,

>>> from autotab import OptimizePipeline
>>> pl = OptimizePipeline(
...     inputs_to_transform=[],
...     outputs_to_transform=[],
...     )
>>> results = pl.fit(data=data)

I don’t want to optimize hyperprameters of the models

If you dont want to optimize hyperparameters of the models, the child iterations needs to be set to zero. As shown in below example,

>>> from autotab import OptimizePipeline
>>> pl = OptimizePipeline(child_iterations=0)
>>> results = pl.fit(data=data)

I don’t want to optimize model selection

If you dont want to optimize model selection, keep models argument equals to None or an empty list. As shown in below example,

>>> from autotab import OptimizePipeline
>>> pl = OptimizePipeline(models=[])
>>> results = pl.fit(data=data)

I want to optimize pipeline for only one model

You can set models parameter to the desired model. In this way, pipeline will be optimized by using only one model. For example, in the following code, only AdaBoostRegressor will be used in pipeline optimization.

>>> from autotab import OptimizePipeline
>>> pl = OptimizePipeline(
>>> models=["AdaBoostRegressor"])
>>> results = pl.fit(data=data)

I want to optimize pipeline for only selected models

List the desired models in models as a list. In this way, pipeline will be optimized for the selected models.

>>> from autotab import OptimizePipeline
>>> pl = OptimizePipeline(
>>> models=[
...     "GradientBoostingRegressor",
...    "HistGradientBoostingRegressor",
...    "DecisionTreeRegressor",
...    "CatBoostRegressor",
...    "ExtraTreeRegressor",
...    "ExtraTreesRegressor",
...    ])
>>> results = pl.fit(data=data)

Can I use different optimization algorithms for parent and child iterations

Different optimization algorithms can be set by parent_algorithm and child_algorithm.

>>> from autotab import OptimizePipeline
>>> pl = OptimizePipeline(
...        parent_algorithm="bayes",
...        child_algorithm="bayes"
...    )
>>> results = pl.fit(data=data)

How to monitor more than one metrics

The metrics you want to monitor can be given to monitor as a list. In this example, two metrics NSE and $R^2$ are being monitored.

>>> from autotab import OptimizePipeline
>>> pl = OptimizePipeline(monitor=['r2', 'nse'])
>>> results = pl.fit(data=data)

How to find best/optimized pipeline

There are two functions to get best pipeline after optimization. They are get_best_pipeline_by_metric which returns optimized pipeline according to given metric. On the other hand, get_best_pipeline_by_model gives us best pipeline according to given model.

>>> from autotab import OptimizePipeline
>>> pl = OptimizePipeline()
>>> results = pl.fit(data=data)
>>> pl.get_best_pipeline_by_metric(metric_name='nse')
>>> pl.get_best_pipeline_by_model(model_name='RandomForest_regressor')

Find best pipeline with respect to a specific (performance) metric

get_best_pipeline_by_metric function can be used to get best pipeline with respect to a specific (performance) metric.

>>> from autotab import OptimizePipeline
>>> pl = OptimizePipeline()
>>> results = pl.fit(data=data)
>>> pl.get_best_pipeline_by_metric(metric_name='nse')

Find best pipeline with respect to a particular model

get_best_pipeline_by_model returns the best pipeline with respect to a particular model and performance metric. The metric must be recorded i.e. must be given as monitor argument.

>>> from autotab import OptimizePipeline
>>> pl = OptimizePipeline()
>>> results = pl.fit(data=data)
>>> pl.get_best_pipeline_by_model(model_name='RandomForest_regressor')

Change search space of a particular model

update_model_space updates or changes the search space of an already existing model.

>>> pl = OptimizePipeline(...)
>>> rf_space = {'max_depth': [5,10, 15, 20],
>>>          'n_models': [5,10, 15, 20]}
>>> pl.update_model_space({"RandomForestRegressor": rf_space})

consider only selected transformations

Selected transformations can be given to input_transformations and output_transformations. In this way, the given transformations will be used for preprocessing steps.

>>> from autotab import OptimizePipeline
>>> pl = OptimizePipeline(
...                    input_transformations=['minmax', 'log', 'zscore'],
...                    output_transformations=['quantile', 'box-cox', 'yeo-johnson']
...                       )
>>> results = pl.fit(data=data)

do not optimize transformations for input data

If you dont want to optimize transformations for input data, keep inputs_to_transform argument equal to None or an empty list. In this way transformations will not be optimized for input data.

>>> from autotab import OptimizePipeline
>>> pl = OptimizePipeline(inputs_to_transform=[])
>>> results = pl.fit(data=data)

change number of optimization iterations of a specific model

Number of optimization iterations for a particular model can be changed by using change_child_iteration function after initializing the OptimizePipeline class. For example we may want to change the child hpo iterations for one or more models. We may want to run only 10 iterations for LinearRegression but 40 iterations for XGBRegressor. In such a case we can use this function to modify child hpo iterations for one or more models. The iterations for all the remaining models will remain same as defined by the user at the start.

>>> from autotab import OptimizePipeline
>>> pl = OptimizePipeline(...)
>>> pl.change_child_iteration({"XGBRegressor": 10})
#If we want to change iterations for more than one models
>>> pl.change_child_iteration(({"XGBRegressor": 30,
>>>                             "RandomForestRegressor": 20}))

where are all the results stored

The results are stored in folder named results in the current working directory. The exact path of stored results can be checked by printing model.path.

>>> from autotab import OptimizePipeline
>>> pl = OptimizePipeline(...)
>>> print(pl.path)

what if optimization stops in the middle

If optimization stops in the middle due to an error, remaining results can be saved and analyzed by using these commands.

>>> from autotab import OptimizePipeline
>>> pl = OptimizePipeline(...)
>>> pl.fit(data=data)
.. # if above command stops in the middle due to an error
>>> pl.save_results()
>>> pl.post_fit(data=data)

what is `config.json` file

config.json is a simply plain text file that stores information about pipeline such as parameters, pipeline configuration. The pipeline can be built again by using from_config_file method as shown below.

>>> from autotab import OptimizePipeline
>>> config_path = "path/to/config.json"
>>> new_pipeline = OptimizePipeline.from_config_file(config_path)

How to include results from previous runs

The path to iterations.json from previous pipeline results has to be given to fit function in order to include results from previous runs.

>>> from autotab import OptimizePipeline
>>> pl = OptimizePipeline(...)
>>> fpath = "path/to/previous/iterations.json"
>>> results = pl.fit(data=data, previous_results=fpath)

What versions of underlying libraries do this package depends

Currently AutoTab is strongly coupled with a ML python framework AI4Water, whose version should be 1.2 or greater. Another dependency is h5py which does not have any specific version requirement.

how to use cross validation during pipeline optimization

By default the pipeline is evaluated on the validation data according to eval_metric. However, you can choose to perform cross validation on child or parent or on both iterations. To perform cross validation at parent iterations set cv_parent_hpo to True. Similarly to perform cross validation at child iteration, set cv_child_hpo to True. You must pass the cross_validator argument as well to determine what kind of cross validation to be performed. Consider the following example

>>> from autotab import OptimizePipeline
>>> pl = OptimizePipeline(
...
...           cv_parent_hpo=True,
...           cross_validator={"KFold": {"n_splits": 5}},
...    )

Instead of KFold, we also choose LeaveOneOut, or ShuffleSplit or TimeSeriesSplit.

how to change search space for batch_size and learning rate

The learning_rate and batch_size search space is only active for deep learning models i.e. when the category is “DL”. The default search space for learning rate is Real(low=1e-5, high=0.05, num_samples=10, name="lr") while for batch_size, the default search space is [8, 16, 32, 64]. We can change the default search space by making use of change_batch_size_space and change_lr_space methods after class initialization. For example we can achieve a different batch_size search space as below

>>> from autotab import OptimizePipeline
>>> pl = OptimizePipeline(
...         ...
...         category="DL
...           )
... pl.change_batch_size_space([32, 64, 128, 256, 512])