#!/usr/bin/env python
# -*- encoding: utf-8 -*-
#
# This file is auto-generated by h2o-3/h2o-bindings/bin/gen_python.py
# Copyright 2016 H2O.ai; Apache License Version 2.0 (see LICENSE for details)
#
import h2o
from h2o.base import Keyed
from h2o.frame import H2OFrame
from h2o.expr import ExprNode
from h2o.expr import ASTId
from h2o.exceptions import H2OValueError
from h2o.estimators.estimator_base import H2OEstimator
from h2o.exceptions import H2OValueError
from h2o.frame import H2OFrame
from h2o.utils.typechecks import assert_is_type, Enum, numeric
[docs]class H2OModelSelectionEstimator(H2OEstimator):
"""
Model Selection
H2O ModelSelection is used to build the best model with one predictor, two predictors, ... up to max_predictor_number
specified in the algorithm parameters when mode=allsubsets. The best model is the one with the highest R2 value. When
mode=maxr, the model returned is no longer guaranteed to have the best R2 value.
"""
algo = "modelselection"
supervised_learning = True
def __init__(self,
model_id=None, # type: Optional[Union[None, str, H2OEstimator]]
training_frame=None, # type: Optional[Union[None, str, H2OFrame]]
validation_frame=None, # type: Optional[Union[None, str, H2OFrame]]
nfolds=0, # type: int
seed=-1, # type: int
fold_assignment="auto", # type: Literal["auto", "random", "modulo", "stratified"]
fold_column=None, # type: Optional[str]
response_column=None, # type: Optional[str]
ignored_columns=None, # type: Optional[List[str]]
ignore_const_cols=True, # type: bool
score_each_iteration=False, # type: bool
score_iteration_interval=0, # type: int
offset_column=None, # type: Optional[str]
weights_column=None, # type: Optional[str]
family="auto", # type: Literal["auto", "gaussian", "binomial", "fractionalbinomial", "quasibinomial", "poisson", "gamma", "tweedie", "negativebinomial"]
link="family_default", # type: Literal["family_default", "identity", "logit", "log", "inverse", "tweedie", "ologit"]
tweedie_variance_power=0.0, # type: float
tweedie_link_power=0.0, # type: float
theta=0.0, # type: float
solver="irlsm", # type: Literal["auto", "irlsm", "l_bfgs", "coordinate_descent_naive", "coordinate_descent", "gradient_descent_lh", "gradient_descent_sqerr"]
alpha=None, # type: Optional[List[float]]
lambda_=[0.0], # type: List[float]
lambda_search=False, # type: bool
early_stopping=False, # type: bool
nlambdas=0, # type: int
standardize=True, # type: bool
missing_values_handling="mean_imputation", # type: Literal["mean_imputation", "skip", "plug_values"]
plug_values=None, # type: Optional[Union[None, str, H2OFrame]]
compute_p_values=False, # type: bool
remove_collinear_columns=False, # type: bool
intercept=True, # type: bool
non_negative=False, # type: bool
max_iterations=0, # type: int
objective_epsilon=-1.0, # type: float
beta_epsilon=0.0001, # type: float
gradient_epsilon=-1.0, # type: float
startval=None, # type: Optional[List[float]]
prior=0.0, # type: float
cold_start=False, # type: bool
lambda_min_ratio=0.0, # type: float
beta_constraints=None, # type: Optional[Union[None, str, H2OFrame]]
max_active_predictors=-1, # type: int
obj_reg=-1.0, # type: float
stopping_rounds=0, # type: int
stopping_metric="auto", # type: Literal["auto", "deviance", "logloss", "mse", "rmse", "mae", "rmsle", "auc", "aucpr", "lift_top_group", "misclassification", "mean_per_class_error", "custom", "custom_increasing"]
stopping_tolerance=0.001, # type: float
balance_classes=False, # type: bool
class_sampling_factors=None, # type: Optional[List[float]]
max_after_balance_size=5.0, # type: float
max_confusion_matrix_size=20, # type: int
max_runtime_secs=0.0, # type: float
custom_metric_func=None, # type: Optional[str]
nparallelism=0, # type: int
max_predictor_number=1, # type: int
min_predictor_number=1, # type: int
mode="maxr", # type: Literal["allsubsets", "maxr", "maxrsweep", "backward"]
build_glm_model=False, # type: bool
p_values_threshold=0.0, # type: float
influence=None, # type: Optional[Literal["dfbetas"]]
multinode_mode=False, # type: bool
):
"""
:param model_id: Destination id for this model; auto-generated if not specified.
Defaults to ``None``.
:type model_id: Union[None, str, H2OEstimator], optional
:param training_frame: Id of the training data frame.
Defaults to ``None``.
:type training_frame: Union[None, str, H2OFrame], optional
:param validation_frame: Id of the validation data frame.
Defaults to ``None``.
:type validation_frame: Union[None, str, H2OFrame], optional
:param nfolds: Number of folds for K-fold cross-validation (0 to disable or >= 2).
Defaults to ``0``.
:type nfolds: int
:param seed: Seed for pseudo random number generator (if applicable)
Defaults to ``-1``.
:type seed: int
:param fold_assignment: Cross-validation fold assignment scheme, if fold_column is not specified. The
'Stratified' option will stratify the folds based on the response variable, for classification problems.
Defaults to ``"auto"``.
:type fold_assignment: Literal["auto", "random", "modulo", "stratified"]
:param fold_column: Column with cross-validation fold index assignment per observation.
Defaults to ``None``.
:type fold_column: str, optional
:param response_column: Response variable column.
Defaults to ``None``.
:type response_column: str, optional
:param ignored_columns: Names of columns to ignore for training.
Defaults to ``None``.
:type ignored_columns: List[str], optional
:param ignore_const_cols: Ignore constant columns.
Defaults to ``True``.
:type ignore_const_cols: bool
:param score_each_iteration: Whether to score during each iteration of model training.
Defaults to ``False``.
:type score_each_iteration: bool
:param score_iteration_interval: Perform scoring for every score_iteration_interval iterations
Defaults to ``0``.
:type score_iteration_interval: int
:param offset_column: Offset column. This will be added to the combination of columns before applying the link
function.
Defaults to ``None``.
:type offset_column: str, optional
:param weights_column: Column with observation weights. Giving some observation a weight of zero is equivalent
to excluding it from the dataset; giving an observation a relative weight of 2 is equivalent to repeating
that row twice. Negative weights are not allowed. Note: Weights are per-row observation weights and do
not increase the size of the data frame. This is typically the number of times a row is repeated, but
non-integer values are supported as well. During training, rows with higher weights matter more, due to
the larger loss function pre-factor. If you set weight = 0 for a row, the returned prediction frame at
that row is zero and this is incorrect. To get an accurate prediction, remove all rows with weight == 0.
Defaults to ``None``.
:type weights_column: str, optional
:param family: Family. For maxr/maxrsweep, only gaussian. For backward, ordinal and multinomial families are
not supported
Defaults to ``"auto"``.
:type family: Literal["auto", "gaussian", "binomial", "fractionalbinomial", "quasibinomial", "poisson", "gamma",
"tweedie", "negativebinomial"]
:param link: Link function.
Defaults to ``"family_default"``.
:type link: Literal["family_default", "identity", "logit", "log", "inverse", "tweedie", "ologit"]
:param tweedie_variance_power: Tweedie variance power
Defaults to ``0.0``.
:type tweedie_variance_power: float
:param tweedie_link_power: Tweedie link power
Defaults to ``0.0``.
:type tweedie_link_power: float
:param theta: Theta
Defaults to ``0.0``.
:type theta: float
:param solver: AUTO will set the solver based on given data and the other parameters. IRLSM is fast on on
problems with small number of predictors and for lambda-search with L1 penalty, L_BFGS scales better for
datasets with many columns.
Defaults to ``"irlsm"``.
:type solver: Literal["auto", "irlsm", "l_bfgs", "coordinate_descent_naive", "coordinate_descent",
"gradient_descent_lh", "gradient_descent_sqerr"]
:param alpha: Distribution of regularization between the L1 (Lasso) and L2 (Ridge) penalties. A value of 1 for
alpha represents Lasso regression, a value of 0 produces Ridge regression, and anything in between
specifies the amount of mixing between the two. Default value of alpha is 0 when SOLVER = 'L-BFGS'; 0.5
otherwise.
Defaults to ``None``.
:type alpha: List[float], optional
:param lambda_: Regularization strength
Defaults to ``[0.0]``.
:type lambda_: List[float]
:param lambda_search: Use lambda search starting at lambda max, given lambda is then interpreted as lambda min
Defaults to ``False``.
:type lambda_search: bool
:param early_stopping: Stop early when there is no more relative improvement on train or validation (if
provided)
Defaults to ``False``.
:type early_stopping: bool
:param nlambdas: Number of lambdas to be used in a search. Default indicates: If alpha is zero, with lambda
search set to True, the value of nlamdas is set to 30 (fewer lambdas are needed for ridge regression)
otherwise it is set to 100.
Defaults to ``0``.
:type nlambdas: int
:param standardize: Standardize numeric columns to have zero mean and unit variance
Defaults to ``True``.
:type standardize: bool
:param missing_values_handling: Handling of missing values. Either MeanImputation, Skip or PlugValues.
Defaults to ``"mean_imputation"``.
:type missing_values_handling: Literal["mean_imputation", "skip", "plug_values"]
:param plug_values: Plug Values (a single row frame containing values that will be used to impute missing values
of the training/validation frame, use with conjunction missing_values_handling = PlugValues)
Defaults to ``None``.
:type plug_values: Union[None, str, H2OFrame], optional
:param compute_p_values: Request p-values computation, p-values work only with IRLSM solver and no
regularization
Defaults to ``False``.
:type compute_p_values: bool
:param remove_collinear_columns: In case of linearly dependent columns, remove some of the dependent columns
Defaults to ``False``.
:type remove_collinear_columns: bool
:param intercept: Include constant term in the model
Defaults to ``True``.
:type intercept: bool
:param non_negative: Restrict coefficients (not intercept) to be non-negative
Defaults to ``False``.
:type non_negative: bool
:param max_iterations: Maximum number of iterations
Defaults to ``0``.
:type max_iterations: int
:param objective_epsilon: Converge if objective value changes less than this. Default (of -1.0) indicates: If
lambda_search is set to True the value of objective_epsilon is set to .0001. If the lambda_search is set
to False and lambda is equal to zero, the value of objective_epsilon is set to .000001, for any other
value of lambda the default value of objective_epsilon is set to .0001.
Defaults to ``-1.0``.
:type objective_epsilon: float
:param beta_epsilon: Converge if beta changes less (using L-infinity norm) than beta esilon, ONLY applies to
IRLSM solver
Defaults to ``0.0001``.
:type beta_epsilon: float
:param gradient_epsilon: Converge if objective changes less (using L-infinity norm) than this, ONLY applies to
L-BFGS solver. Default (of -1.0) indicates: If lambda_search is set to False and lambda is equal to zero,
the default value of gradient_epsilon is equal to .000001, otherwise the default value is .0001. If
lambda_search is set to True, the conditional values above are 1E-8 and 1E-6 respectively.
Defaults to ``-1.0``.
:type gradient_epsilon: float
:param startval: double array to initialize fixed and random coefficients for HGLM, coefficients for GLM.
Defaults to ``None``.
:type startval: List[float], optional
:param prior: Prior probability for y==1. To be used only for logistic regression iff the data has been sampled
and the mean of response does not reflect reality.
Defaults to ``0.0``.
:type prior: float
:param cold_start: Only applicable to multiple alpha/lambda values. If false, build the next model for next set
of alpha/lambda values starting from the values provided by current model. If true will start GLM model
from scratch.
Defaults to ``False``.
:type cold_start: bool
:param lambda_min_ratio: Minimum lambda used in lambda search, specified as a ratio of lambda_max (the smallest
lambda that drives all coefficients to zero). Default indicates: if the number of observations is greater
than the number of variables, then lambda_min_ratio is set to 0.0001; if the number of observations is
less than the number of variables, then lambda_min_ratio is set to 0.01.
Defaults to ``0.0``.
:type lambda_min_ratio: float
:param beta_constraints: Beta constraints
Defaults to ``None``.
:type beta_constraints: Union[None, str, H2OFrame], optional
:param max_active_predictors: Maximum number of active predictors during computation. Use as a stopping
criterion to prevent expensive model building with many predictors. Default indicates: If the IRLSM
solver is used, the value of max_active_predictors is set to 5000 otherwise it is set to 100000000.
Defaults to ``-1``.
:type max_active_predictors: int
:param obj_reg: Likelihood divider in objective value computation, default (of -1.0) will set it to 1/nobs
Defaults to ``-1.0``.
:type obj_reg: float
:param stopping_rounds: Early stopping based on convergence of stopping_metric. Stop if simple moving average of
length k of the stopping_metric does not improve for k:=stopping_rounds scoring events (0 to disable)
Defaults to ``0``.
:type stopping_rounds: int
:param stopping_metric: Metric to use for early stopping (AUTO: logloss for classification, deviance for
regression and anomaly_score for Isolation Forest). Note that custom and custom_increasing can only be
used in GBM and DRF with the Python client.
Defaults to ``"auto"``.
:type stopping_metric: Literal["auto", "deviance", "logloss", "mse", "rmse", "mae", "rmsle", "auc", "aucpr", "lift_top_group",
"misclassification", "mean_per_class_error", "custom", "custom_increasing"]
:param stopping_tolerance: Relative tolerance for metric-based stopping criterion (stop if relative improvement
is not at least this much)
Defaults to ``0.001``.
:type stopping_tolerance: float
:param balance_classes: Balance training data class counts via over/under-sampling (for imbalanced data).
Defaults to ``False``.
:type balance_classes: bool
:param class_sampling_factors: Desired over/under-sampling ratios per class (in lexicographic order). If not
specified, sampling factors will be automatically computed to obtain class balance during training.
Requires balance_classes.
Defaults to ``None``.
:type class_sampling_factors: List[float], optional
:param max_after_balance_size: Maximum relative size of the training data after balancing class counts (can be
less than 1.0). Requires balance_classes.
Defaults to ``5.0``.
:type max_after_balance_size: float
:param max_confusion_matrix_size: [Deprecated] Maximum size (# classes) for confusion matrices to be printed in
the Logs
Defaults to ``20``.
:type max_confusion_matrix_size: int
:param max_runtime_secs: Maximum allowed runtime in seconds for model training. Use 0 to disable.
Defaults to ``0.0``.
:type max_runtime_secs: float
:param custom_metric_func: Reference to custom evaluation function, format: `language:keyName=funcName`
Defaults to ``None``.
:type custom_metric_func: str, optional
:param nparallelism: number of models to build in parallel. Defaults to 0.0 which is adaptive to the system
capability
Defaults to ``0``.
:type nparallelism: int
:param max_predictor_number: Maximum number of predictors to be considered when building GLM models. Defaults
to 1.
Defaults to ``1``.
:type max_predictor_number: int
:param min_predictor_number: For mode = 'backward' only. Minimum number of predictors to be considered when
building GLM models starting with all predictors to be included. Defaults to 1.
Defaults to ``1``.
:type min_predictor_number: int
:param mode: Mode: Used to choose model selection algorithms to use. Options include 'allsubsets' for all
subsets, 'maxr' that uses sequential replacement and GLM to build all models, slow but works with cross-
validation, validation frames for more robust results, 'maxrsweep' that uses sequential replacement and
sweeping action, much faster than 'maxr', 'backward' for backward selection.
Defaults to ``"maxr"``.
:type mode: Literal["allsubsets", "maxr", "maxrsweep", "backward"]
:param build_glm_model: For maxrsweep mode only. If true, will return full blown GLM models with the desired
predictorsubsets. If false, only the predictor subsets, predictor coefficients are returned. This is
forspeeding up the model selection process. The users can choose to build the GLM models themselvesby
using the predictor subsets themselves. Defaults to false.
Defaults to ``False``.
:type build_glm_model: bool
:param p_values_threshold: For mode='backward' only. If specified, will stop the model building process when
all coefficientsp-values drop below this threshold
Defaults to ``0.0``.
:type p_values_threshold: float
:param influence: If set to dfbetas will calculate the difference in beta when a datarow is included and
excluded in the dataset.
Defaults to ``None``.
:type influence: Literal["dfbetas"], optional
:param multinode_mode: For maxrsweep only. If enabled, will attempt to perform sweeping action using multiple
nodes in the cluster. Defaults to false.
Defaults to ``False``.
:type multinode_mode: bool
"""
super(H2OModelSelectionEstimator, self).__init__()
self._parms = {}
self._id = self._parms['model_id'] = model_id
self.training_frame = training_frame
self.validation_frame = validation_frame
self.nfolds = nfolds
self.seed = seed
self.fold_assignment = fold_assignment
self.fold_column = fold_column
self.response_column = response_column
self.ignored_columns = ignored_columns
self.ignore_const_cols = ignore_const_cols
self.score_each_iteration = score_each_iteration
self.score_iteration_interval = score_iteration_interval
self.offset_column = offset_column
self.weights_column = weights_column
self.family = family
self.link = link
self.tweedie_variance_power = tweedie_variance_power
self.tweedie_link_power = tweedie_link_power
self.theta = theta
self.solver = solver
self.alpha = alpha
self.lambda_ = lambda_
self.lambda_search = lambda_search
self.early_stopping = early_stopping
self.nlambdas = nlambdas
self.standardize = standardize
self.missing_values_handling = missing_values_handling
self.plug_values = plug_values
self.compute_p_values = compute_p_values
self.remove_collinear_columns = remove_collinear_columns
self.intercept = intercept
self.non_negative = non_negative
self.max_iterations = max_iterations
self.objective_epsilon = objective_epsilon
self.beta_epsilon = beta_epsilon
self.gradient_epsilon = gradient_epsilon
self.startval = startval
self.prior = prior
self.cold_start = cold_start
self.lambda_min_ratio = lambda_min_ratio
self.beta_constraints = beta_constraints
self.max_active_predictors = max_active_predictors
self.obj_reg = obj_reg
self.stopping_rounds = stopping_rounds
self.stopping_metric = stopping_metric
self.stopping_tolerance = stopping_tolerance
self.balance_classes = balance_classes
self.class_sampling_factors = class_sampling_factors
self.max_after_balance_size = max_after_balance_size
self.max_confusion_matrix_size = max_confusion_matrix_size
self.max_runtime_secs = max_runtime_secs
self.custom_metric_func = custom_metric_func
self.nparallelism = nparallelism
self.max_predictor_number = max_predictor_number
self.min_predictor_number = min_predictor_number
self.mode = mode
self.build_glm_model = build_glm_model
self.p_values_threshold = p_values_threshold
self.influence = influence
self.multinode_mode = multinode_mode
@property
def training_frame(self):
"""
Id of the training data frame.
Type: ``Union[None, str, H2OFrame]``.
"""
return self._parms.get("training_frame")
@training_frame.setter
def training_frame(self, training_frame):
self._parms["training_frame"] = H2OFrame._validate(training_frame, 'training_frame')
@property
def validation_frame(self):
"""
Id of the validation data frame.
Type: ``Union[None, str, H2OFrame]``.
"""
return self._parms.get("validation_frame")
@validation_frame.setter
def validation_frame(self, validation_frame):
self._parms["validation_frame"] = H2OFrame._validate(validation_frame, 'validation_frame')
@property
def nfolds(self):
"""
Number of folds for K-fold cross-validation (0 to disable or >= 2).
Type: ``int``, defaults to ``0``.
"""
return self._parms.get("nfolds")
@nfolds.setter
def nfolds(self, nfolds):
assert_is_type(nfolds, None, int)
self._parms["nfolds"] = nfolds
@property
def seed(self):
"""
Seed for pseudo random number generator (if applicable)
Type: ``int``, defaults to ``-1``.
"""
return self._parms.get("seed")
@seed.setter
def seed(self, seed):
assert_is_type(seed, None, int)
self._parms["seed"] = seed
@property
def fold_assignment(self):
"""
Cross-validation fold assignment scheme, if fold_column is not specified. The 'Stratified' option will stratify
the folds based on the response variable, for classification problems.
Type: ``Literal["auto", "random", "modulo", "stratified"]``, defaults to ``"auto"``.
"""
return self._parms.get("fold_assignment")
@fold_assignment.setter
def fold_assignment(self, fold_assignment):
assert_is_type(fold_assignment, None, Enum("auto", "random", "modulo", "stratified"))
self._parms["fold_assignment"] = fold_assignment
@property
def fold_column(self):
"""
Column with cross-validation fold index assignment per observation.
Type: ``str``.
"""
return self._parms.get("fold_column")
@fold_column.setter
def fold_column(self, fold_column):
assert_is_type(fold_column, None, str)
self._parms["fold_column"] = fold_column
@property
def response_column(self):
"""
Response variable column.
Type: ``str``.
"""
return self._parms.get("response_column")
@response_column.setter
def response_column(self, response_column):
assert_is_type(response_column, None, str)
self._parms["response_column"] = response_column
@property
def ignored_columns(self):
"""
Names of columns to ignore for training.
Type: ``List[str]``.
"""
return self._parms.get("ignored_columns")
@ignored_columns.setter
def ignored_columns(self, ignored_columns):
assert_is_type(ignored_columns, None, [str])
self._parms["ignored_columns"] = ignored_columns
@property
def ignore_const_cols(self):
"""
Ignore constant columns.
Type: ``bool``, defaults to ``True``.
"""
return self._parms.get("ignore_const_cols")
@ignore_const_cols.setter
def ignore_const_cols(self, ignore_const_cols):
assert_is_type(ignore_const_cols, None, bool)
self._parms["ignore_const_cols"] = ignore_const_cols
@property
def score_each_iteration(self):
"""
Whether to score during each iteration of model training.
Type: ``bool``, defaults to ``False``.
"""
return self._parms.get("score_each_iteration")
@score_each_iteration.setter
def score_each_iteration(self, score_each_iteration):
assert_is_type(score_each_iteration, None, bool)
self._parms["score_each_iteration"] = score_each_iteration
@property
def score_iteration_interval(self):
"""
Perform scoring for every score_iteration_interval iterations
Type: ``int``, defaults to ``0``.
"""
return self._parms.get("score_iteration_interval")
@score_iteration_interval.setter
def score_iteration_interval(self, score_iteration_interval):
assert_is_type(score_iteration_interval, None, int)
self._parms["score_iteration_interval"] = score_iteration_interval
@property
def offset_column(self):
"""
Offset column. This will be added to the combination of columns before applying the link function.
Type: ``str``.
"""
return self._parms.get("offset_column")
@offset_column.setter
def offset_column(self, offset_column):
assert_is_type(offset_column, None, str)
self._parms["offset_column"] = offset_column
@property
def weights_column(self):
"""
Column with observation weights. Giving some observation a weight of zero is equivalent to excluding it from the
dataset; giving an observation a relative weight of 2 is equivalent to repeating that row twice. Negative
weights are not allowed. Note: Weights are per-row observation weights and do not increase the size of the data
frame. This is typically the number of times a row is repeated, but non-integer values are supported as well.
During training, rows with higher weights matter more, due to the larger loss function pre-factor. If you set
weight = 0 for a row, the returned prediction frame at that row is zero and this is incorrect. To get an
accurate prediction, remove all rows with weight == 0.
Type: ``str``.
"""
return self._parms.get("weights_column")
@weights_column.setter
def weights_column(self, weights_column):
assert_is_type(weights_column, None, str)
self._parms["weights_column"] = weights_column
@property
def family(self):
"""
Family. For maxr/maxrsweep, only gaussian. For backward, ordinal and multinomial families are not supported
Type: ``Literal["auto", "gaussian", "binomial", "fractionalbinomial", "quasibinomial", "poisson", "gamma",
"tweedie", "negativebinomial"]``, defaults to ``"auto"``.
"""
return self._parms.get("family")
@family.setter
def family(self, family):
assert_is_type(family, None, Enum("auto", "gaussian", "binomial", "fractionalbinomial", "quasibinomial", "poisson", "gamma", "tweedie", "negativebinomial"))
self._parms["family"] = family
@property
def link(self):
"""
Link function.
Type: ``Literal["family_default", "identity", "logit", "log", "inverse", "tweedie", "ologit"]``, defaults to
``"family_default"``.
"""
return self._parms.get("link")
@link.setter
def link(self, link):
assert_is_type(link, None, Enum("family_default", "identity", "logit", "log", "inverse", "tweedie", "ologit"))
self._parms["link"] = link
@property
def tweedie_variance_power(self):
"""
Tweedie variance power
Type: ``float``, defaults to ``0.0``.
"""
return self._parms.get("tweedie_variance_power")
@tweedie_variance_power.setter
def tweedie_variance_power(self, tweedie_variance_power):
assert_is_type(tweedie_variance_power, None, numeric)
self._parms["tweedie_variance_power"] = tweedie_variance_power
@property
def tweedie_link_power(self):
"""
Tweedie link power
Type: ``float``, defaults to ``0.0``.
"""
return self._parms.get("tweedie_link_power")
@tweedie_link_power.setter
def tweedie_link_power(self, tweedie_link_power):
assert_is_type(tweedie_link_power, None, numeric)
self._parms["tweedie_link_power"] = tweedie_link_power
@property
def theta(self):
"""
Theta
Type: ``float``, defaults to ``0.0``.
"""
return self._parms.get("theta")
@theta.setter
def theta(self, theta):
assert_is_type(theta, None, numeric)
self._parms["theta"] = theta
@property
def solver(self):
"""
AUTO will set the solver based on given data and the other parameters. IRLSM is fast on on problems with small
number of predictors and for lambda-search with L1 penalty, L_BFGS scales better for datasets with many columns.
Type: ``Literal["auto", "irlsm", "l_bfgs", "coordinate_descent_naive", "coordinate_descent",
"gradient_descent_lh", "gradient_descent_sqerr"]``, defaults to ``"irlsm"``.
"""
return self._parms.get("solver")
@solver.setter
def solver(self, solver):
assert_is_type(solver, None, Enum("auto", "irlsm", "l_bfgs", "coordinate_descent_naive", "coordinate_descent", "gradient_descent_lh", "gradient_descent_sqerr"))
self._parms["solver"] = solver
@property
def alpha(self):
"""
Distribution of regularization between the L1 (Lasso) and L2 (Ridge) penalties. A value of 1 for alpha
represents Lasso regression, a value of 0 produces Ridge regression, and anything in between specifies the
amount of mixing between the two. Default value of alpha is 0 when SOLVER = 'L-BFGS'; 0.5 otherwise.
Type: ``List[float]``.
"""
return self._parms.get("alpha")
@alpha.setter
def alpha(self, alpha):
# For `alpha` and `lambda` the server reports type float[], while in practice simple floats are also ok
assert_is_type(alpha, None, numeric, [numeric])
self._parms["alpha"] = alpha
@property
def lambda_(self):
"""
Regularization strength
Type: ``List[float]``, defaults to ``[0.0]``.
"""
return self._parms.get("lambda")
@lambda_.setter
def lambda_(self, lambda_):
assert_is_type(lambda_, None, numeric, [numeric])
self._parms["lambda"] = lambda_
@property
def lambda_search(self):
"""
Use lambda search starting at lambda max, given lambda is then interpreted as lambda min
Type: ``bool``, defaults to ``False``.
"""
return self._parms.get("lambda_search")
@lambda_search.setter
def lambda_search(self, lambda_search):
assert_is_type(lambda_search, None, bool)
self._parms["lambda_search"] = lambda_search
@property
def early_stopping(self):
"""
Stop early when there is no more relative improvement on train or validation (if provided)
Type: ``bool``, defaults to ``False``.
"""
return self._parms.get("early_stopping")
@early_stopping.setter
def early_stopping(self, early_stopping):
assert_is_type(early_stopping, None, bool)
self._parms["early_stopping"] = early_stopping
@property
def nlambdas(self):
"""
Number of lambdas to be used in a search. Default indicates: If alpha is zero, with lambda search set to True,
the value of nlamdas is set to 30 (fewer lambdas are needed for ridge regression) otherwise it is set to 100.
Type: ``int``, defaults to ``0``.
"""
return self._parms.get("nlambdas")
@nlambdas.setter
def nlambdas(self, nlambdas):
assert_is_type(nlambdas, None, int)
self._parms["nlambdas"] = nlambdas
@property
def standardize(self):
"""
Standardize numeric columns to have zero mean and unit variance
Type: ``bool``, defaults to ``True``.
"""
return self._parms.get("standardize")
@standardize.setter
def standardize(self, standardize):
assert_is_type(standardize, None, bool)
self._parms["standardize"] = standardize
@property
def missing_values_handling(self):
"""
Handling of missing values. Either MeanImputation, Skip or PlugValues.
Type: ``Literal["mean_imputation", "skip", "plug_values"]``, defaults to ``"mean_imputation"``.
"""
return self._parms.get("missing_values_handling")
@missing_values_handling.setter
def missing_values_handling(self, missing_values_handling):
assert_is_type(missing_values_handling, None, Enum("mean_imputation", "skip", "plug_values"))
self._parms["missing_values_handling"] = missing_values_handling
@property
def plug_values(self):
"""
Plug Values (a single row frame containing values that will be used to impute missing values of the
training/validation frame, use with conjunction missing_values_handling = PlugValues)
Type: ``Union[None, str, H2OFrame]``.
"""
return self._parms.get("plug_values")
@plug_values.setter
def plug_values(self, plug_values):
self._parms["plug_values"] = H2OFrame._validate(plug_values, 'plug_values')
@property
def compute_p_values(self):
"""
Request p-values computation, p-values work only with IRLSM solver and no regularization
Type: ``bool``, defaults to ``False``.
"""
return self._parms.get("compute_p_values")
@compute_p_values.setter
def compute_p_values(self, compute_p_values):
assert_is_type(compute_p_values, None, bool)
self._parms["compute_p_values"] = compute_p_values
@property
def remove_collinear_columns(self):
"""
In case of linearly dependent columns, remove some of the dependent columns
Type: ``bool``, defaults to ``False``.
"""
return self._parms.get("remove_collinear_columns")
@remove_collinear_columns.setter
def remove_collinear_columns(self, remove_collinear_columns):
assert_is_type(remove_collinear_columns, None, bool)
self._parms["remove_collinear_columns"] = remove_collinear_columns
@property
def intercept(self):
"""
Include constant term in the model
Type: ``bool``, defaults to ``True``.
"""
return self._parms.get("intercept")
@intercept.setter
def intercept(self, intercept):
assert_is_type(intercept, None, bool)
self._parms["intercept"] = intercept
@property
def non_negative(self):
"""
Restrict coefficients (not intercept) to be non-negative
Type: ``bool``, defaults to ``False``.
"""
return self._parms.get("non_negative")
@non_negative.setter
def non_negative(self, non_negative):
assert_is_type(non_negative, None, bool)
self._parms["non_negative"] = non_negative
@property
def max_iterations(self):
"""
Maximum number of iterations
Type: ``int``, defaults to ``0``.
"""
return self._parms.get("max_iterations")
@max_iterations.setter
def max_iterations(self, max_iterations):
assert_is_type(max_iterations, None, int)
self._parms["max_iterations"] = max_iterations
@property
def objective_epsilon(self):
"""
Converge if objective value changes less than this. Default (of -1.0) indicates: If lambda_search is set to
True the value of objective_epsilon is set to .0001. If the lambda_search is set to False and lambda is equal to
zero, the value of objective_epsilon is set to .000001, for any other value of lambda the default value of
objective_epsilon is set to .0001.
Type: ``float``, defaults to ``-1.0``.
"""
return self._parms.get("objective_epsilon")
@objective_epsilon.setter
def objective_epsilon(self, objective_epsilon):
assert_is_type(objective_epsilon, None, numeric)
self._parms["objective_epsilon"] = objective_epsilon
@property
def beta_epsilon(self):
"""
Converge if beta changes less (using L-infinity norm) than beta esilon, ONLY applies to IRLSM solver
Type: ``float``, defaults to ``0.0001``.
"""
return self._parms.get("beta_epsilon")
@beta_epsilon.setter
def beta_epsilon(self, beta_epsilon):
assert_is_type(beta_epsilon, None, numeric)
self._parms["beta_epsilon"] = beta_epsilon
@property
def gradient_epsilon(self):
"""
Converge if objective changes less (using L-infinity norm) than this, ONLY applies to L-BFGS solver. Default
(of -1.0) indicates: If lambda_search is set to False and lambda is equal to zero, the default value of
gradient_epsilon is equal to .000001, otherwise the default value is .0001. If lambda_search is set to True, the
conditional values above are 1E-8 and 1E-6 respectively.
Type: ``float``, defaults to ``-1.0``.
"""
return self._parms.get("gradient_epsilon")
@gradient_epsilon.setter
def gradient_epsilon(self, gradient_epsilon):
assert_is_type(gradient_epsilon, None, numeric)
self._parms["gradient_epsilon"] = gradient_epsilon
@property
def startval(self):
"""
double array to initialize fixed and random coefficients for HGLM, coefficients for GLM.
Type: ``List[float]``.
"""
return self._parms.get("startval")
@startval.setter
def startval(self, startval):
assert_is_type(startval, None, [numeric])
self._parms["startval"] = startval
@property
def prior(self):
"""
Prior probability for y==1. To be used only for logistic regression iff the data has been sampled and the mean
of response does not reflect reality.
Type: ``float``, defaults to ``0.0``.
"""
return self._parms.get("prior")
@prior.setter
def prior(self, prior):
assert_is_type(prior, None, numeric)
self._parms["prior"] = prior
@property
def cold_start(self):
"""
Only applicable to multiple alpha/lambda values. If false, build the next model for next set of alpha/lambda
values starting from the values provided by current model. If true will start GLM model from scratch.
Type: ``bool``, defaults to ``False``.
"""
return self._parms.get("cold_start")
@cold_start.setter
def cold_start(self, cold_start):
assert_is_type(cold_start, None, bool)
self._parms["cold_start"] = cold_start
@property
def lambda_min_ratio(self):
"""
Minimum lambda used in lambda search, specified as a ratio of lambda_max (the smallest lambda that drives all
coefficients to zero). Default indicates: if the number of observations is greater than the number of variables,
then lambda_min_ratio is set to 0.0001; if the number of observations is less than the number of variables, then
lambda_min_ratio is set to 0.01.
Type: ``float``, defaults to ``0.0``.
"""
return self._parms.get("lambda_min_ratio")
@lambda_min_ratio.setter
def lambda_min_ratio(self, lambda_min_ratio):
assert_is_type(lambda_min_ratio, None, numeric)
self._parms["lambda_min_ratio"] = lambda_min_ratio
@property
def beta_constraints(self):
"""
Beta constraints
Type: ``Union[None, str, H2OFrame]``.
"""
return self._parms.get("beta_constraints")
@beta_constraints.setter
def beta_constraints(self, beta_constraints):
self._parms["beta_constraints"] = H2OFrame._validate(beta_constraints, 'beta_constraints')
@property
def max_active_predictors(self):
"""
Maximum number of active predictors during computation. Use as a stopping criterion to prevent expensive model
building with many predictors. Default indicates: If the IRLSM solver is used, the value of
max_active_predictors is set to 5000 otherwise it is set to 100000000.
Type: ``int``, defaults to ``-1``.
"""
return self._parms.get("max_active_predictors")
@max_active_predictors.setter
def max_active_predictors(self, max_active_predictors):
assert_is_type(max_active_predictors, None, int)
self._parms["max_active_predictors"] = max_active_predictors
@property
def obj_reg(self):
"""
Likelihood divider in objective value computation, default (of -1.0) will set it to 1/nobs
Type: ``float``, defaults to ``-1.0``.
"""
return self._parms.get("obj_reg")
@obj_reg.setter
def obj_reg(self, obj_reg):
assert_is_type(obj_reg, None, numeric)
self._parms["obj_reg"] = obj_reg
@property
def stopping_rounds(self):
"""
Early stopping based on convergence of stopping_metric. Stop if simple moving average of length k of the
stopping_metric does not improve for k:=stopping_rounds scoring events (0 to disable)
Type: ``int``, defaults to ``0``.
"""
return self._parms.get("stopping_rounds")
@stopping_rounds.setter
def stopping_rounds(self, stopping_rounds):
assert_is_type(stopping_rounds, None, int)
self._parms["stopping_rounds"] = stopping_rounds
@property
def stopping_metric(self):
"""
Metric to use for early stopping (AUTO: logloss for classification, deviance for regression and anomaly_score
for Isolation Forest). Note that custom and custom_increasing can only be used in GBM and DRF with the Python
client.
Type: ``Literal["auto", "deviance", "logloss", "mse", "rmse", "mae", "rmsle", "auc", "aucpr", "lift_top_group",
"misclassification", "mean_per_class_error", "custom", "custom_increasing"]``, defaults to ``"auto"``.
"""
return self._parms.get("stopping_metric")
@stopping_metric.setter
def stopping_metric(self, stopping_metric):
assert_is_type(stopping_metric, None, Enum("auto", "deviance", "logloss", "mse", "rmse", "mae", "rmsle", "auc", "aucpr", "lift_top_group", "misclassification", "mean_per_class_error", "custom", "custom_increasing"))
self._parms["stopping_metric"] = stopping_metric
@property
def stopping_tolerance(self):
"""
Relative tolerance for metric-based stopping criterion (stop if relative improvement is not at least this much)
Type: ``float``, defaults to ``0.001``.
"""
return self._parms.get("stopping_tolerance")
@stopping_tolerance.setter
def stopping_tolerance(self, stopping_tolerance):
assert_is_type(stopping_tolerance, None, numeric)
self._parms["stopping_tolerance"] = stopping_tolerance
@property
def balance_classes(self):
"""
Balance training data class counts via over/under-sampling (for imbalanced data).
Type: ``bool``, defaults to ``False``.
"""
return self._parms.get("balance_classes")
@balance_classes.setter
def balance_classes(self, balance_classes):
assert_is_type(balance_classes, None, bool)
self._parms["balance_classes"] = balance_classes
@property
def class_sampling_factors(self):
"""
Desired over/under-sampling ratios per class (in lexicographic order). If not specified, sampling factors will
be automatically computed to obtain class balance during training. Requires balance_classes.
Type: ``List[float]``.
"""
return self._parms.get("class_sampling_factors")
@class_sampling_factors.setter
def class_sampling_factors(self, class_sampling_factors):
assert_is_type(class_sampling_factors, None, [float])
self._parms["class_sampling_factors"] = class_sampling_factors
@property
def max_after_balance_size(self):
"""
Maximum relative size of the training data after balancing class counts (can be less than 1.0). Requires
balance_classes.
Type: ``float``, defaults to ``5.0``.
"""
return self._parms.get("max_after_balance_size")
@max_after_balance_size.setter
def max_after_balance_size(self, max_after_balance_size):
assert_is_type(max_after_balance_size, None, float)
self._parms["max_after_balance_size"] = max_after_balance_size
@property
def max_confusion_matrix_size(self):
"""
[Deprecated] Maximum size (# classes) for confusion matrices to be printed in the Logs
Type: ``int``, defaults to ``20``.
"""
return self._parms.get("max_confusion_matrix_size")
@max_confusion_matrix_size.setter
def max_confusion_matrix_size(self, max_confusion_matrix_size):
assert_is_type(max_confusion_matrix_size, None, int)
self._parms["max_confusion_matrix_size"] = max_confusion_matrix_size
@property
def max_runtime_secs(self):
"""
Maximum allowed runtime in seconds for model training. Use 0 to disable.
Type: ``float``, defaults to ``0.0``.
"""
return self._parms.get("max_runtime_secs")
@max_runtime_secs.setter
def max_runtime_secs(self, max_runtime_secs):
assert_is_type(max_runtime_secs, None, numeric)
self._parms["max_runtime_secs"] = max_runtime_secs
@property
def custom_metric_func(self):
"""
Reference to custom evaluation function, format: `language:keyName=funcName`
Type: ``str``.
"""
return self._parms.get("custom_metric_func")
@custom_metric_func.setter
def custom_metric_func(self, custom_metric_func):
assert_is_type(custom_metric_func, None, str)
self._parms["custom_metric_func"] = custom_metric_func
@property
def nparallelism(self):
"""
number of models to build in parallel. Defaults to 0.0 which is adaptive to the system capability
Type: ``int``, defaults to ``0``.
"""
return self._parms.get("nparallelism")
@nparallelism.setter
def nparallelism(self, nparallelism):
assert_is_type(nparallelism, None, int)
self._parms["nparallelism"] = nparallelism
@property
def max_predictor_number(self):
"""
Maximum number of predictors to be considered when building GLM models. Defaults to 1.
Type: ``int``, defaults to ``1``.
"""
return self._parms.get("max_predictor_number")
@max_predictor_number.setter
def max_predictor_number(self, max_predictor_number):
assert_is_type(max_predictor_number, None, int)
self._parms["max_predictor_number"] = max_predictor_number
@property
def min_predictor_number(self):
"""
For mode = 'backward' only. Minimum number of predictors to be considered when building GLM models starting
with all predictors to be included. Defaults to 1.
Type: ``int``, defaults to ``1``.
"""
return self._parms.get("min_predictor_number")
@min_predictor_number.setter
def min_predictor_number(self, min_predictor_number):
assert_is_type(min_predictor_number, None, int)
self._parms["min_predictor_number"] = min_predictor_number
@property
def mode(self):
"""
Mode: Used to choose model selection algorithms to use. Options include 'allsubsets' for all subsets, 'maxr'
that uses sequential replacement and GLM to build all models, slow but works with cross-validation, validation
frames for more robust results, 'maxrsweep' that uses sequential replacement and sweeping action, much faster
than 'maxr', 'backward' for backward selection.
Type: ``Literal["allsubsets", "maxr", "maxrsweep", "backward"]``, defaults to ``"maxr"``.
:examples:
>>> import h2o
>>> from h2o.estimators import H2OModelSelectionEstimator
>>> h2o.init()
>>> prostate = h2o.import_file("http://s3.amazonaws.com/h2o-public-test-data/smalldata/logreg/prostate.csv")
>>> predictors = ["AGE", "RACE", "CAPSULE", "DCAPS", "PSA", "VOL", "DPROS"]
>>> response = "GLEASON"
>>> maxrModel = H2OModelSelectionEstimator(max_predictor_number=5,
... seed=12345,
... mode="maxr")
>>> maxrModel.train(x=predictors, y=response, training_frame=prostate)
>>> results = maxrModel.result()
>>> print(results)
"""
return self._parms.get("mode")
@mode.setter
def mode(self, mode):
assert_is_type(mode, None, Enum("allsubsets", "maxr", "maxrsweep", "backward"))
self._parms["mode"] = mode
@property
def build_glm_model(self):
"""
For maxrsweep mode only. If true, will return full blown GLM models with the desired predictorsubsets. If
false, only the predictor subsets, predictor coefficients are returned. This is forspeeding up the model
selection process. The users can choose to build the GLM models themselvesby using the predictor subsets
themselves. Defaults to false.
Type: ``bool``, defaults to ``False``.
:examples:
>>> import h2o
>>> from h2o.estimators import H2OModelSelectionEstimator
>>> h2o.init()
>>> prostate = h2o.import_file("http://s3.amazonaws.com/h2o-public-test-data/smalldata/logreg/prostate.csv")
>>> predictors = ["AGE", "RACE", "CAPSULE", "DCAPS", "PSA", "VOL", "DPROS"]
>>> response = "GLEASON"
>>> maxrModel = H2OModelSelectionEstimator(max_predictor_number=5,
... seed=12345,
... mode="maxrsweep",
... build_glm_model=True)
>>> maxrModel.train(x=predictors, y=response, training_frame=prostate)
>>> result = maxrModel.result()
>>> # get the GLM model with the best performance for a fixed predictor size:
>>> one_model = h2o.get_model(result["model_id"][1, 0])
>>> predict = one_model.predict(prostate)
>>> # print a version of the predict frame:
>>> print(predict)
"""
return self._parms.get("build_glm_model")
@build_glm_model.setter
def build_glm_model(self, build_glm_model):
assert_is_type(build_glm_model, None, bool)
self._parms["build_glm_model"] = build_glm_model
@property
def p_values_threshold(self):
"""
For mode='backward' only. If specified, will stop the model building process when all coefficientsp-values drop
below this threshold
Type: ``float``, defaults to ``0.0``.
:examples:
>>> import h2o
>>> from h2o.estimators import H2OModelSelectionEstimator
>>> h2o.init()
>>> prostate = h2o.import_file("http://s3.amazonaws.com/h2o-public-test-data/smalldata/logreg/prostate.csv")
>>> predictors = ["AGE", "RACE", "CAPSULE", DCAPS", "PSA", "VOL", "DPROS"]
>>> response = "GLEASON"
>>> backwardModel = H2OModelSelectionEstimator(min_predictor_number=2,
... seed=12345,
... mode="backward",
... p_values_threshold=0.001)
>>> backwardModel.train(x=predictors, y=response, training_frame=prostate)
>>> result = backwardModel.result()
>>> print(result)
"""
return self._parms.get("p_values_threshold")
@p_values_threshold.setter
def p_values_threshold(self, p_values_threshold):
assert_is_type(p_values_threshold, None, numeric)
self._parms["p_values_threshold"] = p_values_threshold
@property
def influence(self):
"""
If set to dfbetas will calculate the difference in beta when a datarow is included and excluded in the dataset.
Type: ``Literal["dfbetas"]``.
:examples:
>>> import h2o
>>> from h2o.estimators import H2OModelSelectionEstimator
>>> h2o.init()
>>> prostate = h2o.import_file("http://s3.amazonaws.com/h2o-public-test-data/smalldata/logreg/prostate.csv")
>>> predictors = ["AGE", "RACE", "CAPSULE", "DCAPS", "PSA", "VOL", "DPROS"]
>>> response = "GLEASON"
>>> maxrModel = H2OModelSelectionEstimator(max_predictor_number=5,
... seed=12345,
... mode="maxr",
... influence="dfbetas")
>>> maxrModel.train(x=predictors, y=response, training_frame=prostate)
>>> glm_rid = maxrModel.get_regression_influence_diagnostics()
>>> print(glm_rid)
"""
return self._parms.get("influence")
@influence.setter
def influence(self, influence):
assert_is_type(influence, None, Enum("dfbetas"))
self._parms["influence"] = influence
@property
def multinode_mode(self):
"""
For maxrsweep only. If enabled, will attempt to perform sweeping action using multiple nodes in the cluster.
Defaults to false.
Type: ``bool``, defaults to ``False``.
"""
return self._parms.get("multinode_mode")
@multinode_mode.setter
def multinode_mode(self, multinode_mode):
assert_is_type(multinode_mode, None, bool)
self._parms["multinode_mode"] = multinode_mode
[docs] def get_regression_influence_diagnostics(self, predictor_size=None):
"""
Get the regression influence diagnostics frames for all models with different number of predictors. If a
predictor size is specified, only one frame is returned for that predictor size.
:param predictor_size: predictor subset size, will return regression influence diagnostics frame of that size
:return: list of H2OFrames or just one frame that contains predictors, response and DFBETA_ predictors
"""
if self.actual_params["mode"] == "maxrsweep" and not(self.actual_params["build_glm_model"]):
raise H2OValueError("get_regression_influence_diagnostics can only be called if glm models are built."
" For mode == 'maxrsweep', build_glm_model should be set to True.")
model_ids = self._model_json["output"]["best_model_ids"]
num_models = len(model_ids)
if self.actual_params["influence"] == "dfbetas":
if predictor_size is None or len(predictor_size) == 0:
frame_list = [None]*num_models
for index in range(0, num_models):
one_model = h2o.get_model(model_ids[index]['name'])
frame_list[index] = h2o.get_frame(one_model._model_json["output"]["regression_influence_diagnostics"]["name"])
return frame_list
else:
max_pred_numbers = len(self._model_json["output"]["best_predictors_subset"][num_models-1])
if predictor_size <= 0 or predictor_size > max_pred_numbers:
raise H2OValueError("predictor_size must be between 1 and the maximum number of predictors in"
" the model {0}".format(max_pred_numbers))
if mode == 'backward':
offset = max_pred_numbers - predictor_size
one_model = h2o.get_model(model_ids[num_models-1-offset]['name'])
else:
one_model = h2o.get_model(model_ids[predictor_size-1]['name'])
return h2o.get_frame(one_model._model_json["output"]["regression_influence_diagnostics"]["name"])
else:
raise H2OValueError("influence must be set to dfbetas in order to enable regression influence "
"diagnostics generation.")
[docs] def coef_norm(self, predictor_size=None):
"""
Get the normalized coefficients for all models built with different number of predictors.
:param predictor_size: predictor subset size, will only return model coefficients of that subset size.
:return: list of Python Dicts of coefficients for all models built with different predictor numbers
:examples:
>>> import h2o
>>> from h2o.estimators import H2OModelSelectionEstimator
>>> h2o.init()
>>> prostate = h2o.import_file("http://s3.amazonaws.com/h2o-public-test-data/smalldata/logreg/prostate.csv")
>>> predictors = ["AGE", "RACE", "CAPSULE", "DCAPS", "PSA", "VOL", "DPROS"]
>>> response = "GLEASON"
>>> maxrModel = H2OModelSelectionEstimator(max_predictor_number=5,
... seed=12345,
... mode="maxr")
>>> maxrModel.train(x=predictors, y=response, training_frame=prostate)
>>> coeff_norm = maxrModel.coef_norm()
>>> print(coeff_norm)
>>> coeff_norm_3 = maxrModel.coef_norm(predictor_size=3) # print coefficient norm with 3 predictors
>>> print(coeff_norm_3)
"""
model_ids = self._model_json["output"]["best_model_ids"]
if not(self.actual_params["build_glm_model"]) and self.actual_params["mode"]=="maxrsweep":
coef_names = self._model_json["output"]["coefficient_names"]
coef_values = self._model_json["output"]["coefficient_values_normalized"]
num_models = len(coef_names)
if predictor_size is None or len(predictor_size) == 0:
coefs = [None]*num_models
for index in range(0, num_models):
coef_name = coef_names[index]
coef_val = coef_values[index]
coefs[index] = dict(zip(coef_name, coef_val))
return coefs
else:
if predictor_size > num_models:
raise H2OValueError("predictor_size (predictor subset size) cannot exceed the total number of predictors used.")
if predictor_size <= 0:
raise H2OValueError("predictor_size (predictor subset size) must be between 0 and the total number of predictors used.")
coef_name = coef_names[predictor_size-1]
coef_val = coef_values[predictor_size-1]
return dict(zip(coef_name, coef_val))
else:
model_numbers = len(model_ids)
mode = self.get_params()['mode']
if predictor_size is None:
coefs = [None]*model_numbers
for index in range(0, model_numbers):
one_model = h2o.get_model(model_ids[index]['name'])
tbl = one_model._model_json["output"]["coefficients_table"]
if tbl is not None:
coefs[index] = {name: coef for name, coef in zip(tbl["names"], tbl["standardized_coefficients"])}
return coefs
max_pred_numbers = len(self._model_json["output"]["best_predictors_subset"][model_numbers-1])
if predictor_size > max_pred_numbers:
raise H2OValueError("predictor_size (predictor subset size) cannot exceed the total number of predictors used.")
if predictor_size == 0:
raise H2OValueError("predictor_size (predictor subset size) must be between 0 and the total number of predictors used.")
if mode=='backward':
offset = max_pred_numbers - predictor_size
one_model = h2o.get_model(model_ids[model_numbers-1-offset]['name'])
else:
one_model = h2o.get_model(model_ids[predictor_size-1]['name'])
tbl = one_model._model_json["output"]["coefficients_table"]
if tbl is not None:
return {name: coef for name, coef in zip(tbl["names"], tbl["standardized_coefficients"])}
[docs] def coef(self, predictor_size=None):
"""
Get the coefficients for all models built with different number of predictors.
:param predictor_size: predictor subset size, will only return model coefficients of that subset size.
:return: list of Python Dicts of coefficients for all models built with different predictor numbers
:examples:
>>> import h2o
>>> from h2o.estimators import H2OModelSelectionEstimator
>>> h2o.init()
>>> prostate = h2o.import_file("http://s3.amazonaws.com/h2o-public-test-data/smalldata/logreg/prostate.csv")
>>> predictors = ["AGE", "RACE", "CAPSULE", "DCAPS", "PSA", "VOL", "DPROS"]
>>> response = "GLEASON"
>>> maxrModel = H2OModelSelectionEstimator(max_predictor_number=5,
... seed=12345,
... mode="maxr")
>>> maxrModel.train(x=predictors, y=response, training_frame=prostate)
>>> coeff = maxrModel.coef()
>>> print(coeff)
>>> coeff_3 = maxrModel.coef(predictor_size=3)
>>> print(coeff_3)
"""
if not self.actual_params["build_glm_model"] and self.actual_params["mode"]=="maxrsweep":
coef_names = self._model_json["output"]["coefficient_names"]
coef_values = self._model_json["output"]["coefficient_values"]
num_models = len(coef_names)
if predictor_size is None:
coefs = [None]*num_models
for index in range(0, num_models):
coef_name = coef_names[index]
coef_val = coef_values[index]
coefs[index] = dict(zip(coef_name, coef_val))
return coefs
else:
if predictor_size > num_models:
raise H2OValueError("predictor_size (predictor subset size) cannot exceed the total number of predictors used.")
if predictor_size <= 0:
raise H2OValueError("predictor_size (predictor subset size) must be between 0 and the total number of predictors used.")
coef_name = coef_names[predictor_size-1]
coef_val = coef_values[predictor_size-1]
return dict(zip(coef_name, coef_val))
else:
model_ids = self._model_json["output"]["best_model_ids"]
if model_ids is None:
return None
else:
model_numbers = len(model_ids)
mode = self.get_params()['mode']
if predictor_size is None:
coefs = [None]*model_numbers
for index in range(0, model_numbers):
one_model = h2o.get_model(model_ids[index]['name'])
tbl = one_model._model_json["output"]["coefficients_table"]
if tbl is not None:
coefs[index] = dict(zip(tbl["names"], tbl["coefficients"]))
return coefs
max_pred_numbers = len(self._model_json["output"]["best_predictors_subset"][model_numbers-1])
if predictor_size > max_pred_numbers:
raise H2OValueError("predictor_size (predictor subset size) cannot exceed the total number of predictors used.")
if predictor_size == 0:
raise H2OValueError("predictor_size (predictor subset size) must be between 0 and the total number of predictors used.")
if mode=='backward':
offset = max_pred_numbers - predictor_size
one_model = h2o.get_model(model_ids[model_numbers-1-offset]['name'])
else:
one_model = h2o.get_model(model_ids[predictor_size-1]['name'])
tbl = one_model._model_json["output"]["coefficients_table"]
if tbl is not None:
return dict(zip(tbl["names"], tbl["coefficients"]))
[docs] def result(self):
"""
Get result frame that contains information about the model building process like for modelselection and anovaglm.
:return: the H2OFrame that contains information about the model building process like for modelselection and anovaglm.
"""
return H2OFrame._expr(expr=ExprNode("result", ASTId(self.key)))._frame(fill_cache=True)
[docs] def get_best_R2_values(self):
"""
Get list of best R2 values of models with 1 predictor, 2 predictors, ..., max_predictor_number of predictors
:return: a list of best r2 values
"""
return self._model_json["output"]["best_r2_values"]
[docs] def get_predictors_added_per_step(self):
"""
Get list of predictors added at each step of the model building process
:return: a list of predictors added at each step
"""
if not(self.get_params()["mode"] == "backward"):
return self._model_json["output"]["predictors_added_per_step"]
else:
print("backward mode does not have list predictor_added_per_step")
[docs] def get_predictors_removed_per_step(self):
"""
Get list of predictors removed at each step of the model building process
:return: a list of predictors removed at each step
"""
return self._model_json["output"]["predictors_removed_per_step"]
[docs] def get_best_model_predictors(self):
"""
Get list of best models with 1 predictor, 2 predictors, ..., max_predictor_number of predictors that have the
highest r2 values
:return: a list of best predictors subset
"""
return self._model_json["output"]["best_predictors_subset"]