Hyperparameter tuning overview

In machine learning, hyperparameter tuning identifies a set of optimal hyperparameters for a learning algorithm. A hyperparameter is a model argument whose value is set before the learning process begins. By contrast, the values of other parameters such as coefficients of a linear model are learned.

Hyperparameter tuning lets you spend less time manually iterating hyperparameters and more time focusing on exploring insights from data.

You can specify hyperparameter tuning options for the following model types:

For these types of models, hyperparameter tuning is enabled when you specify a value for the NUM_TRIALS option in the CREATE MODEL statement.

To try running hyperparameter tuning on a linear regression model, see Use the BigQuery ML hyperparameter tuning to improve model performance.

The following models also support hyperparameter tuning but don't allow you to specify particular values:

For information about the supported SQL statements and functions for each model type, see End-to-end user journey for each model.

Locations

For information about which locations support hyperparameter tuning, see BigQuery ML locations.

Set hyperparameters

To tune a hyperparameter, you must specify a range of values for that hyperparameter that the model can use for a set of trials. You can do this by using one of the following keywords when setting the hyperparameter in the CREATE MODEL statement, instead of providing a single value:

  • HPARAM_RANGE: A two-element ARRAY(FLOAT64) value that defines the minimum and maximum bounds of the search space of continuous values for a hyperparameter. Use this option to specify a range of values for a hyperparameter, for example LEARN_RATE = HPARAM_RANGE(0.0001, 1.0).

  • HPARAM_CANDIDATES: A ARRAY(STRUCT) value that specifies the set of discrete values for the hyperparameter. Use this option to specify a set of values for a hyperparameter, for example OPTIMIZER = HPARAM_CANDIDATES(['ADAGRAD', 'SGD', 'FTRL']).

Hyperparameters and objectives

The following table lists the supported hyperparameters and objectives for each model type that supports hyperparameter tuning:

Model type Hyperparameter objectives Hyperparameter Valid range Default range Scale type
LINEAR_REG MEAN_ABSOLUTE_ERROR

MEAN_SQUARED_ERROR

MEAN_SQUARED_LOG_ERROR

MEDIAN_ABSOLUTE_ERROR

R2_SCORE (default)

EXPLAINED_VARIANCE 		L1_REG

L2_REG 		(0, ∞]

(0, ∞] 		(0, 10]

(0, 10] 		LOG

LOG
LOGISTIC_REG PRECISION

RECALL

ACCURACY

F1_SCORE

LOG_LOSS

ROC_AUC (default) 		L1_REG

L2_REG 		(0, ∞]

(0, ∞] 		(0, 10]

(0, 10] 		LOG

LOG
KMEANS DAVIES_BOULDIN_INDEX NUM_CLUSTERS [2, 100] [2, 10] LINEAR
MATRIX_
FACTORIZATION (explicit) 		MEAN_SQUARED_ERROR NUM_FACTORS

L2_REG 		[2, 200]

(0, ∞) 		[2, 20]

(0, 10] 		LINEAR

LOG
MATRIX_
FACTORIZATION (implicit) 		MEAN_AVERAGE_PRECISION (default)

MEAN_SQUARED_ERROR

NORMALIZED_DISCOUNTED_CUMULATIVE_GAIN

AVERAGE_RANK 		NUM_FACTORS

L2_REG

WALS_ALPHA 		[2, 200]

(0, ∞)

[0, ∞) 		[2, 20]

(0, 10]

[0, 100] 		LINEAR

LOG

LINEAR
AUTOENCODER MEAN_ABSOLUTE_ERROR

MEAN_SQUARED_ERROR (default)

MEAN_SQUARED_LOG_ERROR 		LEARN_RATE

BATCH_SIZE

L1_REG

L2_REG

L1_REG_ACTIVATION

DROPOUT

HIDDEN_UNITS


OPTIMIZER



ACTIVATION_FN 		[0, 1]

(0, ∞)

(0, ∞)

(0, ∞)

(0, ∞)


[0, 1)

Array of [1, ∞)

{ADAM, ADAGRAD, FTRL, RMSPROP, SGD}

{RELU, RELU6, CRELU, ELU, SELU, SIGMOID, TANH} 		[0, 1]

[16, 1024]

(0, 10]

(0, 10]

(0, 10]


[0, 0.8]

N/A

{ADAM, ADAGRAD, FTRL, RMSPROP, SGD}

N/A 		LOG

LOG

LOG

LOG

LOG


LINEAR

N/A

N/A



N/A
DNN_CLASSIFIER PRECISION

RECALL

ACCURACY

F1_SCORE

LOG_LOSS

ROC_AUC (default) 		BATCH_SIZE

DROPOUT

HIDDEN_UNITS

LEARN_RATE

OPTIMIZER



L1_REG

L2_REG

ACTIVATION_FN 		(0, ∞)

[0, 1)

Array of [1, ∞)

[0, 1]

{ADAM, ADAGRAD, FTRL, RMSPROP, SGD}

(0, ∞)

(0, ∞)

{RELU, RELU6, CRELU, ELU, SELU, SIGMOID, TANH} 		[16, 1024]

[0, 0.8]

N/A

[0, 1]

{ADAM, ADAGRAD, FTRL, RMSPROP, SGD}

(0, 10]

(0, 10]

N/A 		LOG

LINEAR

N/A

LINEAR

N/A



LOG

LOG

N/A
DNN_REGRESSOR MEAN_ABSOLUTE_ERROR

MEAN_SQUARED_ERROR

MEAN_SQUARED_LOG_ERROR

MEDIAN_ABSOLUTE_ERROR

R2_SCORE (default)

EXPLAINED_VARIANCE
DNN_LINEAR_
COMBINED_
CLASSIFIER 		PRECISION

RECALL

ACCURACY

F1_SCORE

LOG_LOSS

ROC_AUC (default) 		BATCH_SIZE

DROPOUT

HIDDEN_UNITS

L1_REG

L2_REG

ACTIVATION_FN 		(0, ∞)

[0, 1)

Array of [1, ∞)

(0, ∞)

(0, ∞)

{RELU, RELU6, CRELU, ELU, SELU, SIGMOID, TANH} 		[16, 1024]

[0, 0.8]

N/A

(0, 10]

(0, 10]

N/A 		LOG

LINEAR

N/A

LOG

LOG

N/A
DNN_LINEAR_
COMBINED_
REGRESSOR 		MEAN_ABSOLUTE_ERROR

MEAN_SQUARED_ERROR

MEAN_SQUARED_LOG_ERROR

MEDIAN_ABSOLUTE_ERROR

R2_SCORE (default)

EXPLAINED_VARIANCE
BOOSTED_TREE_
CLASSIFIER 		PRECISION

RECALL

ACCURACY

F1_SCORE

LOG_LOSS

ROC_AUC (default) 		LEARN_RATE

L1_REG

L2_REG

DROPOUT

MAX_TREE_DEPTHMAX_TREE_DEPTH

SUBSAMPLE

MIN_SPLIT_LOSS

NUM_PARALLEL_TREE

MIN_TREE_CHILD_WEIGHT

COLSAMPLE_BYTREE

COLSAMPLE_BYLEVEL

COLSAMPLE_BYNODE

BOOSTER_TYPE

DART_NORMALIZE_TYPE

TREE_METHOD 		[0, ∞)

(0, ∞)

(0, ∞)

[0, 1]

[1, 20]



(0, 1]

[0, ∞)

[1, ∞)


[0, ∞)


[0, 1]


[0, 1]


[0, 1]


{GBTREE, DART}

{TREE, FOREST}

{AUTO, EXACT, APPROX, HIST} 		[0, 1]

(0, 10]

(0, 10]

N/A

[1, 10]



(0, 1]

N/A

N/A


N/A


N/A


N/A


N/A


N/A

N/A

N/A 		LINEAR

LOG

LOG

LINEAR

LINEAR



LINEAR

LINEAR

LINEAR


LINEAR


LINEAR


LINEAR


LINEAR


N/A

N/A

N/A
BOOSTED_TREE_
REGRESSOR





 MEAN_ABSOLUTE_ERROR

MEAN_SQUARED_ERROR

MEAN_SQUARED_LOG_ERROR

MEDIAN_ABSOLUTE_ERROR

R2_SCORE (default)

EXPLAINED_VARIANCE
RANDOM_FOREST_
CLASSIFIER 		PRECISION

RECALL

ACCURACY

F1_SCORE

LOG_LOSS

ROC_AUC (default) 		L1_REG

L2_REG

MAX_TREE_DEPTH

SUBSAMPLE

MIN_SPLIT_LOSS

NUM_PARALLEL_TREE

MIN_TREE_CHILD_WEIGHT

COLSAMPLE_BYTREE

COLSAMPLE_BYLEVEL

COLSAMPLE_BYNODE

TREE_METHOD 		(0, ∞)

(0, ∞)

[1, 20]

(0, 1)

[0, ∞)

[2, ∞)


[0, ∞)


[0, 1]


[0, 1]


[0, 1]

{AUTO, EXACT, APPROX, HIST} 		(0, 10]

(0, 10]

[1, 20]

(0, 1)

N/A

[2, 200]


N/A


N/A


N/A


N/A


N/A 		LOG

LOG

LINEAR

LINEAR

LINEAR

LINEAR


LINEAR


LINEAR


LINEAR


LINEAR


N/A
RANDOM_FOREST_
REGRESSOR





 MEAN_ABSOLUTE_ERROR

MEAN_SQUARED_ERROR

MEAN_SQUARED_LOG_ERROR

MEDIAN_ABSOLUTE_ERROR

R2_SCORE (default)

EXPLAINED_VARIANCE

Most LOG scale hyperparameters use the open lower boundary of 0. You can still set 0 as the lower boundary by using the HPARAM_RANGE keyword to set the hyperparameter range. For example, in a boosted tree classifier model, you could set the range for the L1_REG hyperparameter as L1_REG = HPARAM_RANGE(0, 5). A value of 0 gets converted to 1e-14.

Conditional hyperparameters are supported. For example, in a boosted tree regressor model, you can only tune the DART_NORMALIZE_TYPE hyperparameter when the value of the BOOSTER_TYPE hyperparameter is DART. In this case, you specify both search spaces and the conditions are handled automatically, as shown in the following example:

BOOSTER_TYPE = HPARAM_CANDIDATES(['DART', 'GBTREE'])
DART_NORMALIZE_TYPE = HPARAM_CANDIDATES(['TREE', 'FOREST'])

Search starting point

If you don't specify a search space for a hyperparameter by using HPARAM_RANGE or HPARAM_CANDIDATES, the search starts from the default value of that hyperparameter, as documented in the CREATE MODEL topic for that model type. For example, if you are running hyperparameter tuning for a boosted tree model, and you don't specify a value for the L1_REG hyperparameter, then the search starts from 0, the default value.

If you specify a search space for a hyperparameter by using HPARAM_RANGE or HPARAM_CANDIDATES, the search starting points depends on whether the specified search space includes the default value for that hyperparameter, as documented in the CREATE MODEL topic for that model type:

  • If the specified range contains the default value, that's where the search starts. For example, if you are running hyperparameter tuning for an implicit matrix factorization model, and you specify the value [20, 30, 40, 50] for the WALS_ALPHA hyperparameter, then the search starts at 40, the default value.
  • If the specified range doesn't contain the default value, the search starts from the point in the specified range that is closest to the default value. For example,if you specify the value [10, 20, 30] for the WALS_ALPHA hyperparameter, then the search starts from 30, which is the closest value to the default value of 40.

Data split

When you specify a value for the NUM_TRIALS option, the service identifies that you are doing hyperparameter tuning and automatically performs a 3-way split on input data to divide it into training, evaluation, and test sets. By default, the input data is randomized and then split 80% for training, 10% for evaluation, and 10% for testing.

The training and evaluation sets are used in each trial training, the same as in models that don't use hyperparameter tuning. The trial hyperparameter suggestions are calculated based on the model evaluation metrics for that model type. At the end of each trial training, the test set is used to test the trial and record its metrics in the model. This ensures the objectivity of the final reporting evaluation metrics by using data that has not yet been analyzed by the model. Evaluation data is used to calculate the intermediate metrics for hyperparameter suggestion, while the test data is used to calculate the final, objective model metrics.

If you want to use only a training set, specify NO_SPLIT for the DATA_SPLIT_METHOD option of the CREATE MODEL statement.

If you want to use only training and evaluation sets, specify 0 for the DATA_SPLIT_TEST_FRACTION option of the CREATE MODEL statement. When the test set is empty, the evaluation set is used as the test set for the final evaluation metrics reporting.

The metrics from models that are generated from a normal training job and those from a hyperparameter tuning training job are only comparable when the data split fractions are equal. For example, the following models are comparable:

  • Non-hyperparameter tuning: DATA_SPLIT_METHOD='RANDOM', DATA_SPLIT_EVAL_FRACTION=0.2
  • Hyperparameter tuning: DATA_SPLIT_METHOD='RANDOM', DATA_SPLIT_EVAL_FRACTION=0.2, DATA_SPLIT_TEST_FRACTION=0

Performance

Model performance when using hyperparameter tuning is typically no worse than model performance when using the default search space and not using hyperparameter tuning. A model that uses the default search space and doesn't use hyperparameter tuning always uses the default hyperparameters in the first trial.

To confirm the model performance improvements provided by hyperparameter tuning, compare the optimal trial for the hyperparameter tuning model to the first trial for the non-hyperparameter tuning model.

Transfer learning

Transfer learning is enabled by default when you set the HPARAM_TUNING_ALGORITHM option in the CREATE MODEL statement to VIZIER_DEFAULT. The hyperparameter tuning for a model benefits by learning from previously tuned models if it meets the following requirements:

  • It has the same model type as previously tuned models.
  • It resides in the same project as previously tuned models.
  • It use the same hyperparameter search space OR a subset of the hyperparameter search space of previously tuned models. A subset uses the same hyperparameter names and types, but doesn't have to have the same ranges. For example, (a:[0, 10]) is considered as a subset of (a:[-1, 1], b:[0, 1]).

Transfer learning doesn't require that the input data be the same.

Transfer learning helps solve the cold start problem where the system performs random exploration during the first trial batch. Transfer learning provides the system with some initial knowledge about the hyperparameters and their objectives. To continuously improve the model quality, always train a new hyperparameter tuning model with the same or a subset of hyperparameters.

Transfer learning helps hyperparameter tuning converge faster, instead of helping submodels to converge.

Error handling

Hyperparameter tuning handles errors in the following ways:

  • Cancellation: If a training job is cancelled while running, then all successful trials remain usable.

  • Invalid input: If the user input is invalid, then the service returns a user error.

  • Invalid hyperparameters: If the hyperparameters are invalid for a trial, then the trial is skipped and marked as INFEASIBLE in the output from the ML.TRIAL_INFO function.

  • Trial internal error: If more than 10% of the NUM_TRIALS value fail due to INTERNAL_ERROR, then the training job stops and returns a user error.

  • If less than 10% of the NUM_TRIALS value fail due to INTERNAL_ERROR, the training continues with the failed trials marked as FAILED in the output from the ML.TRIAL_INFO function.

Model serving functions

You can use output models from hyperparameter tuning with a number of existing model serving functions. To use these functions, follow these rules:

  • When the function takes input data, only the result from one trial is returned. By default this is the optimal trial, but you can also choose a particular trial by specifying the TRIAL_ID as an argument for the given function. You can get the TRIAL_ID from the output of the ML.TRIAL_INFO function. The following functions are supported:

  • When the function doesn't take input data, all trial results are returned, and the first output column is TRIAL_ID. The following functions are supported:

The output from ML.FEATURE_INFO doesn't change, because all trials share the same input data.

Evaluation metrics from ML.EVALUATE and ML.TRIAL_INFO can be different because of the way input data is split. By default, ML.EVALUATE runs against the test data, while ML.TRIAL_INFO runs against the evaluation data. For more information, see Data split.

Unsupported functions

The ML.TRAINING_INFO function returns information for each iteration, and iteration results aren't saved in hyperparameter tuning models. Trial results are saved instead. You can use the ML.TRIAL_INFO function to get information about trial results.

Model export

You can export models created with hyperparameter tuning to Cloud Storage locations using the EXPORT MODEL statement. You can export the default optimal trial or any specified trial.

Pricing

The cost of hyperparameter tuning training is the sum of the cost of all executed trials. The pricing of a trial is consistent with the existing BigQuery ML pricing model.

FAQ

How many trials do I need to tune a model?

We recommend using at least 10 trials for one hyperparameter, so the total number of trials should be at least 10 * num_hyperparameters. If you are using the default search space, refer to the Hyperparameters column in the Hyperparameters and objectives table for the number of hyperparameters tuned by default for a given model type.

What if I don't see performance improvements by using hyperparameter tuning?

Make sure you follow the guidance in this document to get a fair comparison. If you still don't see performance improvements, it might mean the default hyperparameters already work well for you. You might want to focus on feature engineering or try other model types before trying another round of hyperparameter tuning.

What if I want to continue tuning a model?

Train a new hyperparameter tuning model with the same search space. The built-in transfer learning helps to continue tuning based on your previously tuned models.

Do I need to retrain the model with all data and the optimal hyperparameters?

It depends on the following factors:

  • K-means models already use all data as the training data, so there's no need to retrain the model.

  • For matrix factorization models, you can retrain the model with the selected hyperparameters and all input data for better coverage of users and items.

  • For all other model types, retraining is usually unnecessary. The service already keeps 80% of the input data for training during the default random data split. You can still retrain the model with more training data and the selected hyperparameters if your dataset is small, but leaving little evaluation data for early stop might worsen overfitting.

What's next

To try running hyperparameter tuning, see Use the BigQuery ML hyperparameter tuning to improve model performance.