Machine learning tools to make a Data Scientist's work more efficient
Project description
verstack 3.2.2 Documentation
Machine learning tools to make a Data Scientist's work efficient
veratack package contains the following tools:
-
Stacker automated stacking ensemble configuration/train/features creation in train/test sets
-
FeatureSelector automated feature selection class based on recursive feature elimination
-
DateParser automated date columns finder and parser
-
LGBMTuner automated lightgbm models tuner with optuna
-
NaNImputer impute all missing values in a pandas dataframe using advanced machine learning with 1 line of code
-
Multicore execute any function in concurrency using all the available cpu cores
-
ThreshTuner tune threshold for binary classification predictions
-
stratified_continuous_split create train/test splits stratified on the continuous variable
-
categoric_encoders encode categoric variable by numeric labels
- Factorizer encode categoric variable by numeric labels
- OneHotEncoder represent categoric variable as a set of binary variables
- FrequencyEncoder encode categoric variable by class frequencies
- MeanTargetEncoder encode categoric variable by mean of the target variable
- WeightOfEvidenceEncoder encode categoric variable as a weight of evidence of a binary target variable
-
timer convenient timer decorator to quickly measure and display time of any function execution
-
Printer a convenient function to set up and execute print statements based on the 'global' verbosity setting within large projects
Note
Getting verstack
$ pip install verstack
$ pip install --upgrade verstack
:::
Stacker
Fully automated highly configurable stacking ensemble creation class. Can create single or multiple layers of stacked features. Applicable for train/test set features creation. Any number of layers and models within layers can be added to Stacker instance (models in layers must contain fit / predict / predict_proba ([if classification]{.title-ref}) methods for the Stacker to properly create features using these models).
Additional metafeatures can be created from stacked features if metafeats parameter is set to True.
Subsequent (>1) layers can be trained either on predictions from one previous layer / or predictions from one previous layer and meta features / or predictions from all previous layers / or predictions from all previous layers and meta features [subject to stacking_feats_depth parameter configuration]{.title-ref}; original X features can also be used for training the subsequent layers [subjuect to include_X parameter configuration]{.title-ref}.
Stacker includes auto mode which will create two layers of stacked features with layer 1 consisting of 14 diverse models and layer 2 consisting of a linear and boosed model
Models' RandomizedSearchCV
hyperparameters tuning is enabled if
gridsearch_iteration parameter is > 0 [subject to model being
supported by built in parameters optimization function]{.title-ref}.
Stacked feats creation on the train set is perfromed by train/predict operations on 4 folds. Each stacked feature in the test set is created by predicting with 4 models [fitted on train set]{.title-ref} and averaging predictions. When averaging for regression tasks - mean of predicted values is computed; for binary - mean of positive class probabilities is computed; for multiclass - the most commonly predicted class from the 4 predictions is selected.
... the output of fit_transfrom() / transform() methods will return the dataframe with original features and stacked features.
auto mode models
layer_1: 14 models
- LGBM(max_depth = 12)
- XGB(max_depth = 10, n_jobs = -1)
- GradientBoosting(max_depth = 7)
- kerasModel(num_layers = 3)
- kerasModel(num_layers = 2)
- kerasModel(num_layers = 1)
- ExtraTree(max_depth = 12)
- RandomForest(max_depth = 7)
- Linear/LogisticRegression()
- KNeighbors(n_neighbors=15)
- KNeighbors(n_neighbors=10)
- SVR(kernel = 'rbf')
- DecisionTree(max_depth = 15)
- DecisionTree(max_depth = 8)
layer_2: two models
- LGBM(max_depth = 3)
- Ridge()
Initialize Stacker
from verstack import Stacker
# initialize with default parameters
stacker = Stacker(objective = 'regression')
# initialize with selected parameters
stacker = Stacker(objective = 'regression',
auto = True,
auto_num_layers = 2,
metafeats = True,
epochs = 500,
gridsearch_iterations = 20,
stacking_feats_depth = 1,
include_X = False,
verbose = True)
Parameters
parameters
metafeats
,gridsearch_iterations
,stacking_feats_depth
,include_X
can be configured independently for any layer in the follwoing manner: E.g. If need to optimize the models' hyperparameters only in layer_2:: -
stacker = Stacker('regression', gridsearch_iterations = 0)
-stacker.add_layer([model_1, model_2, model_3])
-X_transformed = stacker.fit_transform(X, y)
-stacker.add_layer([model_4, model_5])
-stacker.gridsearch_iterations = 20
-X_transformed = stacker.fit_transform(X_transformed, y)
-
objective
[default=None]Training objective. Can take values: 'regression', 'binary', 'multiclass'
-
auto
[default=False]Enable/disable automatic configuration of 1 or 2 layers of models to create stacked features. If True will automatically populate the self.layers with 1 or 2 lists of preconfigured diverse models.
-
auto_num_layers
[default=2]Number of automatically generated layers. Can take values 1 and 2
-
metafeats
[default=True]Additional statistical meta features creation from the stacked predictions:
: - pairwise differences between the stacked predictions are created for all pairs (recursively) - mean and std for all the stacked features in a layer are created as two extra meta feats
-
epochs
[default=200]Number of neural networks epochs. Applicable for the three automatically configured neural networks in the auto mode
-
gridsearch_iterations
[default=10]Number of hyperparameters optimization iterations. If set to 0, hyperparameters will not be optimized. If > 0, hyperparameters in all layers will be optimized. E.g. Supported models for optimization:
- lightgbm.sklearn.LGBMRegressor / lightgbm.sklearn.LGBMClassifier
- xgboost.sklearn.XGBRegressor / xgboost.sklearn.XGBClassifier
- sklearn.ensemble.GradientBoostingRegressor / sklearn.ensemble.GradientBoostingClassifier
- sklearn.tree._classes.ExtraTreeRegressor / sklearn.tree._classes.ExtraTreeClassifier
- sklearn.ensemble._forest.RandomForestRegressor / sklearn.ensemble._forest.RandomForestClassifier
- sklearn.linear_model._logistic.LogisticRegression
- sklearn.linear_model._ridge.Ridge
- sklearn.neighbors._regression.KNeighborsRegressor / sklearn.neighbors._classification.KNeighborsClassifier
- sklearn.svm._classes.SVR / sklearn.svm._classes.SVC
- sklearn.tree._classes.DecisionTreeRegressor / sklearn.tree._classes.DecisionTreeClassifier
-
stacking_feats_depth
[default=1]Defines the features used by subsequent (>1) layers to train the stacking models. Can take values between 1 and 4 where:
: - 1 = use predictions from one previous layer - 2 = use predictions from one previous layer and meta features - 3 = use predictions from all previous layers - 4 = use predictions from all previous layers and meta features
-
include_X
[default=False]Flag to use original X features for subsequent layer training
-
verbose
[default=True]Print progress outputs or silent
Methods
-
add_layer([model_1, model_2(), model_3])
Add layer with models to Stacker instance.
Parameters
-
models_list
[list]List containing initiated models instances. Each model must contain fit() / predict() / predict_proba() ([if classification]{.title-ref}) methods
returns
: None
-
-
fit_transform(X, y)
Train/predict/append to X the stacking features from models defined in self.layers
Parameters
-
X
[pd.DataFrame]train features
-
y
[pd.Series]train labels
returns
: pd.DataFrame train featues with appended stacking features
-
-
transform(X)
Create stacking features on the test set from models saved in self.trained_models
Parameters
-
X
[pd.DataFrame]test features
returns
: pd.DataFrame test featues with appended stacking features
-
Attributes
-
layers
Dictionary with 'layer_n' as key and list of models in layer as value
-
trained_models
Dictionary with 'layer_n' as key and dictionary with stacked feature name as key and list of 4 [trained on different folds]{.title-ref} models instances for predicting on test set
Examples
Using Stacker in auto mode
from verstack import Stacker
stacker = Stacker(objective = 'multiclass', auto = True)
X_with_stacked_feats = stacker.fit_transform(X, y)
Add two custom layers, for training subsequent (>1) layers use not only the predictions of the previous layer, but also metafeats in the previous layer and X original features Then add one more layer and disable hyperparameters optimization for this layer
# initialize Stacker
stacker = Stacker(objective = 'multiclass',
auto = False,
stacking_feats_depth = 2,
include_X = True)
# add layers
stacker.add_layer([model_1, model_2, model_3])
stacker.add_layer([model_4, model_5])
# add stacking features to train/test
X_with_stacked_feats = stacker.fit_transform(X, y)
test_with_stacked_feats = stacker.transform(test)
# add extra layer
stacker.add_layer([model_6, model_7])
# change the gridsearch_iteration setting
stacker.gridsearch_iterations = 0
# pass the transformed dataset if need to call .fit_transform() after adding extra layers to the fitted instance of Stacker
X_with_stacked_feats = stacker.fit_transform(X_with_stacked_feats, y)
test_with_stacked_feats = stacker.transform(test_with_stacked_feats)
FeatureSelector
Automated feature selector based on recursive feature elimination. FeatureSelector has built-in & configured models (linear/logistic regression & RandomForest) and employs logic to recursively eliminate features with one of these models taking advantage of sklearn.feature_selection.RFECV. Different modes preform feature selection in different modes: - one of the built-in models - any other model, which should be passed by user at init - auto mode: a competition between feature selection independently by linear model and RandomForest is evaluated by a third model (LGBM by default, can be configured by user) to select a subset which yields higher accuracy
Additional arguments allow to: - reduce input data size for running experiments - in auto mode, allows to automatically select features from a model with smaller accuracy if number of selected features for this model is smaller and percent difference between accuracy is within the allowed_score_gap parameter
Initialize FeatureSelector
from verstack import FeatureSelector
# initialize with default parameters
FS = FeatureSelector(objective = 'regression')
# initialize with custom model
from lightgbm import LGBMRegressor
model_for_feature_selection = LGBMRegressor()
FS = FeatureSelector(objective = 'regression',
custom_model = model_for_feature_selection)
# initialize with selected parameters
stacker = Stacker(objective = 'regression',
auto = True,
subset_size_mb = 50,
allowed_score_gap = 0.05,
verbose = True)
Parameters
-
objective
[default='regression']Training objective. Can take values: 'regression' and any other string which will be interpreted as 'classification'
-
auto
[default=False]Enable/disable automatic feature selection comparison between linear model and RandomForest. FeatureSelector will select two independent sets of features by LR/RF and score with a third model (LGBM by default). Features that yield a higher accuracy are returned
-
allowed_score_gap
[default=0.0](If
auto
==True) If a user requires a smaller set of features and can compromise a controlled value of accuracy, theallowed_score_gap
parameter can take values between 0.0 and 1.0 to control the allowance for potential model lower validation score if model has a smaller number of selected features. E.g.allowed_score_gap
= 0.05 will allow to return selected features from one of the two models if: - its accuracy is up to 5% worse than the competing model - it has selected a smaller number of features
-
final_scoring_model
[default=None](If
auto
==True) Pass model instance to compare scores between features selected by linear model and by RandomForest model. The default value is None, in this case lightgbm model is used -
default_model_linear
[default=False]Flag to deploy linear model or RandomForest model for feature selection
-
custom_model
[default=None]Pass model instance to be used for feature selection instead of built-in linear/RandomForest models
-
subset_size_mb
[default=20]Value to reduce data dimensionality (row-wise) for running feature selection experiments
-
verbose
[default=True]Verbosity setting
Methods
-
fit_transform(X, y, kwargs)
Apply feature selection on features and target
Parameters
-
X
[pd.DataFrame]Train features
-
y
[pd.Series]Train labels
-
kwargs
[keyword arguments]Arguments for sklearn.feature_selection.RFECV
returns
: pd.DataFrame selected features
-
-
transform(X)
Apply trained FeatureSelector instance to transform another dataset by subsetting it to the selected features
Parameters
-
X
[pd.DataFrame]Features
returns
: pd.DataFrame selected features
-
Attributes
-
layers
Dictionary with 'layer_n' as key and list of models in layer as value
-
trained_models
Dictionary with 'layer_n' as key and dictionary with stacked feature name as key and list of 4 [trained on different folds]{.title-ref} models instances for predicting on test set
Examples
Using FeatureSelector in auto mode
from verstack import FeatureSelector
FS = FeatureSelector(objective = 'regression', auto = True)
selected_feats = FS.fit_transform(X, y)
Use built-in RandomForest model for feature selection
FS = FeatureSelector(objective = 'regression', default_model_linear=False)
selected_feats = FS.fit_transform(X, y)
Pass custom model for feature selection
from lightgbm import LGBMRegressor
model = LGBMRegressor()
FS = FeatureSelector(objective = 'regression', custom_model=model)
selected_feats = FS.fit_transform(X, y)}
DateParser
Fully automated DateParser tool that takes as input a pandas.DataFrame and returns a pandas.DataFrame with parsed datetime features. Holidays flags and names are created as features subject to user passing the country argument (E.g. country = 'US'). Holiday features extraction are based on utilizing the [holidays]{.title-ref} package. Datetime columns will be found automatically, transformed to pd.Timestamp format, new columns with the follwing features (if applicable to the specific datetime format) will be created: - year - month - day (monthday) - quarter - week - weekday - dayofyear - hour - minute - second - part_of_day - timediff (if two datetime columns are found) - is_holiday (if country argument is passed) - holiday_name (if country argument is passed) - is_payday (if payday argument is passed) - days_from_epoch (1970/01/01)
... same set of features will be created (with column name prefix) for each of the datetime columns DateParser detects.
Supported datetime formats
- '28-OCT-90',
- '28-OCT-1990',
- '10/28/90',
- '10/28/1990',
- '28.10.90',
- '28.10.1990',
- '90/10/28',
- '1990/10/28',
- '4 Q 90',
- '4 Q 1990',
- 'OCT 90',
- 'OCT 1990',
- '43 WK 90',
- '43 WK 1990',
- '01:02',
- '02:34',
- '02:34.75',
- '20-JUN-1990 08:03',
- '20-JUN-1990 08:03:00',
- '1990-06-20 08:03',
- '1990-06-20 08:03:00.0'
Initialize DateParser
from verstack import DateParser
# initialize with default parameters
parser = DateParser()
# initialize with selected parameters
parser = DateParser(country = 'US',
state = 'CA',
payday = [1, 15])
Parameters
-
country
[default=None]Country name or abreviation. For a full list of supported countries call parser.list_supported_countries()
-
state
[default=None]State abreviation. Correct state abreviations are available at https://pypi.org/project/holidays/
-
prov
[default=None]Province abreviation. Correct province abreviations are available at https://pypi.org/project/holidays/
-
payday
[default=None]List of paydays applicable in a specific country. E.g. [1, 15]
-
verbose
[default=True]Enable or desable console prints
Methods
-
fit_transform(df)
Fully automatic search of datetime columns and features extraction. Apart from all the conventional datetime features will automatically parse holidays / paydays if specified and init. Saves the found datetime columns names and feature extraction pipelines for the transform() method.
Parameters
-
df
[pd.DataFrame]Data with raw features
returns
: pd.DataFrame with new features
-
-
transform(df)
Parse identical set of features from a new dataset. Usually applied to test set transformation. E.g. if test set datetime columns include a short timeframe so that quarter feature is constant and thus should not be created, the dataset will still be populated by this feature in order to preserve the identical columns names and order between train/test sets. Think machine learning.
Parameters
-
df
[pd.DataFrame]Data with raw features (test/valid set)
returns
: pd.DataFrame with new features
-
-
parse_holidays(datetime_col_series, country, state, province, holiday_names)
Create series with holidays names or flags for a defined country based on series of datetime-like strings.
-
datetime_col_series
[pd.Series]Series of datetime-like strings in line with supported_formats
-
country
[str]Country name or abreviation. For a full list of supported countries call parser.list_supported_countries()
-
state
[str, default = None]State abreviation. Correct state abreviations are available at https://pypi.org/project/holidays/
-
prov
[str, default = None]Province abreviation. Correct province abreviations are available at https://pypi.org/project/holidays/
-
holiday_names
[bool, default = False]Flag to return holidays as a binary feature or string holidays names
returns
: pd.Series with holidays binary flags or holidays string names
-
-
get_holidays_calendar(country, years, state = None, prov = None)
Get data on the holidays in a given country (optinally in a certain state/province) for a given year(s).
-
country
[str]Country name or abreviation. For a full list of supported countries call parser.list_supported_countries()
-
state
[str, default = None]State abreviation. Correct state abreviations are available at https://pypi.org/project/holidays/
-
prov
[str, default = None]Province abreviation. Correct province abreviations are available at https://pypi.org/project/holidays/
returns
: dictionary with holidays dates and names
-
-
list_supported_countries()
Print a list of supported countries and abreviations.
Attributes
-
datetime_cols
List of found datetime columns names. Available after fit_transform()
-
created_datetime_cols
List of created datetime features. Available after fit_transform()
-
supported formats
List of supported datetime formats
Examples
Using DateParser with all default parameters
parser = DateParser()
train_with_parsed_dt_feats = parser.fit_transform(train)
test_with_parsed_dt_feats = parser.transform(test)
DateParser with holidays/paydays
parser = DateParser(country = 'US', payday = [1, 15])
train_with_parsed_dt_feats = parser.fit_transform(train)
test_with_parsed_dt_feats = parser.transform(test)
LGBMTuner
Fully automated lightgbm model hyperparameter tuning class with optuna under the hood. LGBMTuner selects optimal hyperparameters based on executed trials (configurable), optimizes n_estimators and fits the final model to the whole train set. Feature importances are available in numeric format, as a static plot, and as an interactive plot (html). Optimization history and parameters importance in static and interactive formats are alse accesable by built in methods.
Logic
The only required user inputs are the X (features), y (labels) and evaluation metric name, LGBMTuner will handle the rest
- lgbm model type (regression/classification) is inferred from the labels and evaluation metric (passed by user)
- optimization metric may be different from the evaluation metric (passed by user). LGBMTuner at hyperparameters search stage imploys the error reduction strategy, thus:
- most regression task type metrics are supported for optimization, if not, MSE is selected for optimization
- for classification task types hyperparameters are tuned by optimizing log_loss, n_estimators are tuned with evaluation_metric
- early stopping is engaged at each stage of LGBMTuner optimizations
- for every trial (iteration) a random train_test_split is performed (stratified for classification)
- lgbm model initial parameters!=defaults and are inferred from the data stats and built in logic
- optimization parameters and their search space are inferred from the data stats and built in logic
- LGBMTuner class instance (after optimization) can be used for making predictions with conventional syntaxis (predict/predict_proba)
- verbosity is controlled and by default outputs only the necessary optimization process/results information
Initialize LGBMTuner
from verstack import LGBMTuner
# initialize with default parameters
tuner = LGBMTuner('rmse')
# initialize with selected parameters
tuner = LGBMTuner(metric = 'rmse',
trials = 200,
refit = False,
verbosity = 0,
visualization = False,
seed = 999)
Parameters
-
metric
[default=None]Evaluation metric for hyperparameters optimization. LGBMTuner supports the following metrics (note the syntax)
: ['mae', 'mse', 'rmse', 'rmsle', 'mape', 'smape', 'rmspe', 'r2', 'auc', 'gini', 'log_loss', 'accuracy', 'balanced_accuracy', 'precision', 'precision_weighted', 'precision_macro', 'recall', 'recall_weighted', 'recall_macro', 'f1', 'f1_weighted', 'f1_macro', 'lift']
-
trials
[default=100]Number of trials to run
-
refit
[default=True]Fit the model with optimized hyperparameters on the whole train set (required for feature_importances, plot_importances() and prediction methods)
-
verbosity
[default=1]Console verbosity level: 0 - no output except for optuna CRITICAL errors and builtin exceptions; (1-5) based on optuna.logging options. The default is 1
-
visualization
[default=True]Automatically output feature_importance & optimization plots into the console after tuning. Plots are also available on demand by corresponding methods
-
seed
[default=42]Random state parameter
-
eval_results_callback
[default=None]Callback function to be applied on the eval_results dictionary that is being populated with evaluation metric score upon completion of each training trial
Methods
-
fit(X, y)
Execute LGBM model hyperparameters tuning
Parameters
-
X
[pd.DataFrame]Train features
-
y
[pd.Series]Train labels
-
-
optimize_n_estimators(X, y, params, verbose_eval = 100)
Optimize n_estimators for lgb model.
Parameters
-
X
[np.array]Train features
-
y
[np.array]Train labels
-
params
[dict]parameters to use for training the model with early stopping
-
verbose_eval
[int]evaluation output at each
verbose_eval
iteratio n
returns
: (best_iteration, best_score)
-
-
fit_optimized(X, y)
Train model with tuned params on whole train data
-
X
[np.array]Train features
-
y
[np.array]
-
-
predict(test, threshold = 0.5)
Predict by optimized model on new data
-
test
[pd.DataFrame]Test features
-
threshold
[default=0.5]Classification threshold (applicable for binary classification)
returns
: array of int
-
-
predict_proba(test)
Predict probabilities by optimized model on new data
-
test
[pd.DataFrame]Test features
returns
: array of float
-
-
plot_importances(n_features = 15, figsize = (10,6), interactive = False, display = True, plotly_fig_update_layout_kwargs = {})
Plot feature importance
-
n_features
[default=15]Number of important features to plot
-
figsize
[default=(10,6)]plot size
-
interactive
[default=False]Create & display with the default browser the interactive html plot or (if browser disply is unavailable) save to current wd.
-
display
[default=True]Display plot in browser. If False, plot will be saved in cwd.
-
plotly_fig_update_layout_kwargs
[default={}]kwargs for plotly.fig.update_layout() function. The default is empty dict and default_plotly_fig_update_layout_kwargs configured inside the plot_importances() will be used.
-
-
plot_optimization_history(interactive = False)
Plot optimization function improvement history
-
interactive
[default=False]Create & display with the default browser the interactive html plot or (if browser disply is unavailable) save to current wd.
-
display
[default=True]Display plot in browser. If False, plot will be saved in cwd.
-
-
plot_param_importances(interactive = False)
Plot params importance plot
-
interactive
[default=False]Create & display with the default browser the interactive html plot or (if browser disply is unavailable) save to current wd.
-
display
[default=True]Display plot in browser. If False, plot will be saved in cwd.
-
-
plot_intermediate_values(interactive = False, legend = False)
Plot optimization trials history. Shows successful and terminated trials. If trials > 50 it is better to study the interactive version
-
interactive
[default=False]Create & display with the default browser the interactive html plot or (if browser disply is unavailable) save to current wd.
-
legend
[default=False]Plot legen on a static plot
-
display
[default=True]Display plot in browser. If False, plot will be saved in cwd.
-
Attributes
-
metric
Evaluation metric defined by user at LGBMTuner init
-
refit
Setting for refitting the optimized model on whole train dataset
-
verbosity
Verbosity level settings
-
visualization
Automatic plots output after optimization setting
-
seed
Random state value
-
fitted_model
Trained LGBM booster model with optimized parameters
-
feature_importances
Feature importance values
-
study
optuna.study.study.Study object after hyperparameters tuning
-
init_params
initial LGBM model parameters
-
best_params
learned optimized parameters
-
eval_results
dictionary with evaluation results per each of non-pruned trials measured by a function derived from the
metric
argument
Examples
Using LGBMTuner with all default parameters
imputer = LGBMTuner('auc')
tuner.fit(X, y)
tuner.feature_importances
tuner.plot_importances()
tuner.plot_intermediate_values()
tuner.plot_optimization_history()
tuner.plot_param_importances()
tuner.best_params
tuner.predict(test)
LGBMTuner with custom settings
imputer = LGBMTuner(metric = 'auc', trials = 300, verbosity = 3, visualization = False)
tuner.fit(X, y)
tuner.plot_importances(legend = True)
tuner.plot_intermediate_values(interactive = True)
tuner.predict(test, threshold = 0.3)
NaNImputer
Impute all missing values in a pandas dataframe by xgboost models in multiprocessing mode using a single line of code.
Logic
With NaNImputer you can fill missing values in numeric, binary and categoric columns in your pandas dataframe using advanced XGBRegressor/XGBClassifier models with just 1 line of code. Regardless of the data types in your dataframe (string/bool/numeric):
- all of the columns will be checked for missing values
- transformed into numeric formats
- split into subsets with and without missing values
- applicalbe models will be selected and configured for each of the columns with NaNs
- models will be trained in multiprocessing mode utilizing all the available cores and threads of your cpu (this saves a lot of time)
- NaNs will be predicted and placed into corresponding indixes
- columns with all NaNs will be droped
- columns containing NaNs and known values as a single constant
- data will be reverse-transformed into original format
The module is highly configurable with default argumets set for the highest performance and verbosity
The only limitation is:
- NaNs in pure text columns are not imputed. By default they are filled with 'Missing_data' value. Configurable. If disabled - will return these columns with missing values untouched
Initialize NaNImputer
from verstack import NaNImputer
# initialize with default parameters
imputer = NaNImputer()
# initialize with selected parameters
imputer = NaNImputer(conservative = False,
n_feats = 10,
nan_cols = None,
fix_string_nans = True,
multiprocessing_load = 3,
verbose = True,
fill_nans_in_pure_text = True,
drop_empty_cols = True,
drop_nan_cols_with_constant = True)
Parameters
-
conservative
[default=False]Model complexity level used to impute missing values. If
True
: model will be set to less complex and much faster. -
n_feats
[default=10]Number of corellated independent features to be used forcorresponding column (with NaN) model training and imputation.
-
nan_cols
[default=None]List of columns to impute missing values in. If None: all the columns with missing values will be used.
-
fix_string_nans
[default=True]Find possible missing values in numeric columns that had been (mistakenly) encoded as strings, E.g. 'Missing'/'NaN'/'No data' and replace them with np.nan for further imputation.
-
multiprocessing_load
[default=3]- Levels of parallel multiprocessing compute
- 1 = single core
- 2 = half of all available cores
- 3 = all available cores
- Levels of parallel multiprocessing compute
-
verbose
[default=True]Print the imputation progress.
-
fill_nans_in_pure_text
[default=True]Fill the missing values in text fields by string 'Missing_data'.Applicable for text fields (not categoric).
-
drop_empty_cols
[default=True]Drop columns with all NaNs.
-
drop_nan_cols_with_constant
[default=True]Drop columns containing NaNs and known values as a single constant.
-
feature_selection
[default="correlation"]- Define algorithm to select most important feats for each column imputation. Quick option: "correlation" is based on selecting n_feats with the highest binary correlation with each column for NaNs imputation. Less quick but more precise: "feature_importance" is based on extracting feature_importances from an xgboost model.
Methods
-
impute(data)
Execute NaNs imputation columnwise in a pd.DataFrame
Parameters
-
data
pd.DataFramedataframe with missing values in a single/multiple columns
-
Examples
Using NaNImputer with all default parameters
imputer = NaNImputer()
df_imputed = imputer.impute(df)
Say you would like to impute missing values in a list of specific columns, use 20 most important features for each of these columns imputation and deploy a half of the available cpu cores
imputer = NaNImputer(nan_cols = ['col1', 'col2'], n_feats = 20, multiprocessing_load = 2)
df_imputed = imputer.impute(df)
Multicore
Execute any function in concurrency using all the available cpu cores.
Logic
Multicore module is built on top of concurrent.futures package. Passed iterables are divided into chunks according to the number of workers and passed into separate processes.
Results are extracted from finished processes and combined into a single/multiple output as per the defined function output requirements.
Multiple outputs are returned as a nested list.
Initialize Multicore
from verstack import Multicore
# initialize with default parameters
multicore = Multicore()
# initialize with selected parameters
multicore = Multicore(workers = 6,
multiple_iterables = True)
Parameters
-
workers
int or bool [default=False]Number of workers if passed by user. If
False
: all available cpu cores will be used. -
multiple_iterables
bool [default=False]If function needs to iterate over multiple iterables, set to
True
.Multiple iterables must be passed as a list (see examples below).
-
verbose
bool [default=True]Enable function execution progress print to the console
Methods
-
execute(func, iterable)
Execute passed function and iterable(s) in concurrency.
Parameters
-
func
functionfunction to execute in parallel
-
iterable
list/pd.Series/pd.DataFrame/dictionarydata to iterate over
-
Examples
Use Multicore with all default parameters
multicore = Multicore()
result = multicore.execute(function, iterable_list)
If you want to use a limited number of cpu cores and need to iterate over two objects:
multicore = Multicore(workers = 2, multiple_iterables = True)
result = multicore.execute(function, [iterable_dataframe, iterable_list])
ThreshTuner
Find the best threshold to split your predictions in a binary classification task. Most applicable for imbalance target cases. In addition to thresholds & loss_func scores, the predicted_ratio (predicted fraction of 1) will be calculated and saved for every threshold. This will help the identify the appropriate threshold not only based on the score, but also based on the resulting distribution of 0 and 1 in the predictions.
Logic
Default behavior (only pass the labels and predictions):
: - Calculate the labels balance (fraction_of_1 in labels) - Define the min_threshold as fraction_of_1 * 0.8 - Define the max_threshold as fraction_of_1 * 1.2 but not greater than 1 - Define the n_thresholds = 200 - Create 200 threshold options uniformly distributed between min_threshold & max_threshold - Deploy the balanced_accuracy_score as loss_func - Peform loss function calculation and save results in class instance placeholders
Customization options
: - Change the n_thresholds to the desired value - Change the min_threshold & max_threshold to the desired values - Pass the loss_func of choice, e.g. sklearn.metrics.f1_score
This will result in user defined granulation of thresholds to test
Initialize ThreshTuner
from verstack import ThreshTuner
# initialize with default parameters
thresh = ThreshTuner()
# initialize with selected parameters
thresh = ThreshTuner(n_thresholds = 500,
min_threshold = 0.3,
max_threshold = 0.7)
Parameters
-
n_thresholds
int [default=200]Number of thresholds to test. If not set by user: 200 thresholds will be tested.
-
min_threshold
float or int [default=None]Minimum threshold value. If not set by user: will be inferred from labels balance based on fraction_of_1
-
max_threshold
float or int [default=None]Maximum threshold value. If not set by user: will be inferred from labels balance based on fraction_of_1
Methods
-
fit(labels, pred, loss_func)
Calculate loss_func results for labels & preds for the defined/default thresholds. Print the threshold(s) with the best loss_func scores
Parameters
-
labels
array/list/series [default=balanced_accuracy_score]y_true labels represented as 0 or 1
-
pred
array/list/seriespredicted probabilities of 1
-
loss_func
functionloss function for scoring the predictions, e.g. sklearn.metrics.f1_score
-
-
result()
Display a dataframe with thresholds/loss_func_scores/fraction_of_1 for for all the the defined/default thresholds
-
best_score()
Display a dataframe with thresholds/loss_func_scores/fraction_of_1 for the best loss_func_score
-
best_predict_ratio()
Display a dataframe with thresholds/loss_func_scores/fraction_of_1 for the (predicted) fraction_of_1 which is closest to the (actual) labels_fraction_of_1
Examples
Use ThreshTuner with all default parameters
thresh = ThreshTuner()
thres.fit(labels, pred)
Customized ThreshTuner application
from sklearn.metrics import f1_score
thresh = ThreshTuner(n_thresholds = 500, min_threshold = 0.2, max_threshold = 0.6)
thresh.fit(labels, pred, f1_score)
Access the results after .fit()
thresh = ThreshTuner()
thres.fit(labels, pred)
# return pd.DataFrame with all the results
thresh.result
# return pd.DataFrame with the best loss_func score
thresh.best_score()
thresh.best_score()['threshold']
# return pd.DataFrame with the best predicted fraction_of_1
thresh.best_predict_ratio()
# return the actual labels fraction_of_1
thresh.labels_fractio_of_1
stratified_continuous_split
Create stratified splits based on either continuous or categoric target variable.
: - For continuous target variable verstack uses binning and categoric split based on bins - For categoric target enhanced sklearn.model_selection.train_test_split is used: in case there are not enough categories for the split, the minority classes will be combined with nearest neighbors.
Can accept only pandas.DataFrame/pandas.Series as data input.
verstack.stratified_continuous_split.scsplit(*args,
stratify,
test_size = 0.3,
train_size = 0.7,
continuous = True,
random_state = None)
Parameters
-
X,y,data
data input for the split in pandas.DataFrame/pandas.Series format.
-
stratify
target variable for the split in pandas/eries format.
-
test_size
[default=0.3]test split ratio.
-
train_size
[default=0.7]train split ratio.
-
continuous
[default=True]stratification target definition. If True, verstack will perform the stratification on the continuous target variable, if False, sklearn.model_selection.train_test_split will be performed with verstack enhancements.
-
random_state
[default=5]random state value.
Examples
from verstack.stratified_continuous_split import scsplit
train, test = scsplit(data, stratify = data['continuous_column_name'])
X_train, X_val, y_train, y_val = scsplit(X, y, stratify = y,
test_size = 0.3, random_state = 5)
categoric_encoders
::: {.note} ::: {.title} Note :::
All the categoric encoders are conveniently integrated to work with pandas.DataFrame. Modules receive pd.DataFrame and kwargs as inputs and return pd.DataFrame with encoded column. All the necessary attributes for further transform/inverse_transform are saved in instance objects and can be seralized (e.g. pickle) for latter application. :::
Factorizer
Encode categoric column by numeric labels.
Logic
Assign numeric labels starting with 0 to all unique variable's categories.
Missing values can be encoded by an integer value (defaults to -1) / float / string or can be left untransformed.
When transform () - unseen categories will be be represented as NaN.
Initialize Factorizer
from verstack import Factorizer
# initialize with default parameters
factorizer = Factorizer()
# initialize with changing the NaN encoding value
factorizer = Factorizer(na_sentinel = np.nan) #-999/0.33333/'No data')
Attributes
-
na_sentinel
Defined (at init) missing values encoding value.
-
colname
Defined (at fit_transform()) column that had been transformed.
-
pattern
Defined (at fit_transform()) encoding map.
Parameters
-
na_sentinel
[default=-1]Missing values encoding value. Can take int/float/str/np.nan values.
Methods
-
fit_transform(df, colname)
Fit Factorizer to data and return transformed data.
Parameters
-
df
pd.DataFramedf containing the colname to transform.
-
colname
strColumn name in df to be transformed.
-
-
transform(df)
Apply the fitted Factorizer to new data and return transformed data. Unseen categories will be represented by NaN.
Parameters
-
df
pd.DataFrameData containing the colname to transform.
-
-
inverse_transform(df)
Inverse transform data that had been encoded by Factorizer. Data must contain colname that was passed at fit_transform().
Parameters
-
df
pd.DataFrameData containing the colname to transform.
-
Examples
Use with default na_sentinel:
factorizer = Factorizer()
train_encoded = factorizer.fit_transform(train, 'colname') # will encode NaN values by -1
test_encoded = factorizer.transform(test)
train_reversed_to_original = factorizer.inverse_transform(train_encoded)
test_reversed_to_original = factorizer.inverse_transform(test_encoded)
Keep missing values untransformed:
factorizer = Factorizer(na_sentinel = np.nan)
train_encoded = factorizer.fit_transform(train)
OneHotEncoder
Encode categoric column by a set of binary columns.
Logic
Categoric 'column':['a','b','c'] will be represented by three binary columns 'a', 'b', 'c'. Original categoric 'column' is droped.
Missing values can be represented by a separate column or omited.
When transform() - unseen categories will not be represented by new columns, missing categories will be represented by empty (all zeros) columns.
Initialize OneHotEncoder
from verstack import OneHotEncoder
ohe = OneHotEncoder()
train_encoded = ohe.fit_transform(train, 'colname') # will create a separate column for NaN values (if any)
test_encoded = ohe.transform(test)
train_reversed_to_original = ohe.inverse_transform(train_encoded)
test_reversed_to_original = ohe.inverse_transform(test_encoded)
Attributes
-
na_sentinel
Defined (at init) missing values encoding value.
-
colname
Defined (at fit_transform()) column that had been transformed.
-
categories
Defined (at fit_transform()) unique class categories which will be represented by binary columns.
Parameters
-
na_sentinel
[default=True]If True: create separate class column for NaN values.
Methods
-
fit_transform(df, colname, prefix)
Fit OneHotEncoder to data and return transformed data.
Parameters
-
df
pd.DataFramedf containing the colname to transform.
-
colname
strColumn name in df to be transformed.
-
prefix
str/int/float/bool/None, optionalString to append DataFrame column names. The default is None.
-
-
transform(df)
Apply the fitted OneHotEncoder to new data and return transformed data. Unseen categories will not be represented by new columns, missing categories will be represented by empty (all zeros) columns.
Parameters
-
df
pd.DataFrameData containing the colname to transform.
-
-
inverse_transform(df)
Inverse transform data that had been encoded by OneHotEncoder. Data must contain one-hot-encoded columns that was created at fit_transform().
Parameters
-
df
pd.DataFrameData containing the colname to transform.
-
Examples
ohe = OneHotEncoder()
train_encoded = ohe.fit_transform(train, 'colname', prefix = 'colname')
test_encoded = ohe.transform(test)
train_reversed_to_original = ohe.inverse_transform(train_encoded)
test_reversed_to_original = ohe.inverse_transform(test_encoded)
FrequencyEncoder
Encoder to represent categoric variable classes' frequency across the dataset.
Logic
Original column ['a', 'a', 'a', 'b', 'b', 'c', 'c', 'c', 'c', np.nan]
Encoded column [0.3, 0.3, 0.3, 0.2, 0.2, 0.4, 0.4, 0.4, 0.4, 0.1] # np.nan]
When transform() - unseen categories will be represented by the most common (highest) frequency.
Can handle missing values - encode NaN by NaN frequency or leave NaN values untransformed. Resulting frequencies are normalized as a percentage.
Initialize FrequencyEncoder
from verstack import FrequencyEncoder
fe = FrequencyEncoder()
train_encoded = fe.fit_transform(train, 'colname')
test_encoded = fe.transform(test)
train_reversed_to_original = fe.inverse_transform(train_encoded)
test_reversed_to_original = fe.inverse_transform(test_encoded)
Attributes
-
na_sentinel
Defined (at init) missing values encoding value.
-
colname
Defined (at fit_transform()) column that had been transformed.
-
pattern
Defined (at fit_transform()) encoding map.
Parameters
na_sentinel
[default=True]- If True: Encode NaN values by their frequency. If False return np.nan in the encoded column.
Methods
-
fit_transform(df, colname)
Fit FrequencyEncoder to data and return transformed data.
Parameters
-
df
pd.DataFramedf containing the colname to transform.
-
colname
strColumn name in df to be transformed.
-
-
transform(df)
Apply the fitted FrequencyEncoder to new data and return transformed data. Unseen categories will be represented as NaN.
Parameters
-
df
pd.DataFrameData containing the colname to transform.
-
-
inverse_transform(df)
Inverse transform data that had been encoded by FrequencyEncoder. Data must contain colname that was passed at fit_transform().
Parameters
-
df
pd.DataFrameData containing the colname to transform.
-
Examples
frequency_encoder = FrequencyEncoder()
train_encoded = frequency_encoder.fit_transform(train, 'colname')
test_encoded = frequency_encoder.transform(test)
train_reversed_to_original = frequency_encoder.inverse_transform(train_encoded)
test_reversed_to_original = frequency_encoder.inverse_transform(test_encoded)
MeanTargetEncoder
Encode train cat cols by mean target value for category.
Logic
To avoid target leakage train set encoding is performed by breaking data into 5 folds & encoding categories of each fold with their respective target mean values calculated on the other 4 folds. This will introduce minor noize to train data encoding (at fit_transform()) as a normalization technique. Test set (transform()) is encoded without normalization.
When transform() - unseen categories will be represented by the global target mean.
Can handle missing values - encode NaN by global mean or leave NaN values untransformed.
Initialize MeanTargetEncoder
from verstack import MeanTargetEncoder
mean_target_encoder = MeanTargetEncoder(save_inverse_transform = True)
train_encoded = mean_target_encoder.fit_transform(train, 'colname', 'targetname')
test_encoded = mean_target_encoder.transform(test)
train_reversed_to_original = mean_target_encoder.inverse_transform(train_encoded)
test_reversed_to_original = mean_target_encoder.inverse_transform(test_encoded)
Attributes
-
na_sentinel
Defined (at init) missing values encoding value.
-
colname
Defined (at fit_transform()) column that had been transformed.
-
pattern
Defined (at fit_transform()) encoding map.
-
save_inverse_transform
Defined (at init) flag for saving the pattern for inverse transform.
Parameters
-
na_sentinel
[default=True]If True: Encode NaN values by target global mean. If False return np.nan in the encoded column.
-
save_inverse_transform
[default=False]If True: Saves mean target values for each category at each encoding fold. Enable if need to inverse_transform the encoded data. Defaults to False because for large datasets saved pattern can significantly increase instance object size.
Methods
-
fit_transform(df, colname, targetname)
Fit MeanTargetEncoder to data and return transformed data.
Parameters
-
df
pd.DataFramedf containing the colname to transform.
-
colname
strColumn name in df to be transformed.
-
targetname
strTarget column name in df for extracting the mean values for each colname category.
-
-
transform(df)
Apply the fitted MeanTargetEncoder to new data and return transformed data. Unseen categories will be encoded by the global target mean.
Parameters
-
df
pd.DataFrameData containing the colname to transform.
-
-
inverse_transform(df)
Inverse transform data that had been encoded by MeanTargetEncoder. Data must contain colname that was passed at fit_transform().
Parameters
-
df
pd.DataFrameData containing the colname to transform.
-
Examples
mean_target_encoder = MeanTargetEncoder(save_inverse_transform = True)
train_encoded = mean_target_encoder.fit_transform(train, 'colname', 'targetname')
test_encoded = mean_target_encoder.transform(test)
train_reversed_to_original = mean_target_encoder.inverse_transform(train_encoded)
test_reversed_to_original = mean_target_encoder.inverse_transform(test_encoded)
WeightOfEvidenceEncoder
Encoder to represent categoric variables by Weight of Evidence in regards to the binary target variable.
Logic
Built on top of sclearn package category_encoders.woe.WOEEncoder.
If encoded value is negative - it represents a category that is more heavily enclided to the negative target class (0). Positive encoding result represents inclination to the positive target class (1).
When fit_transform() is used on a train set, variable is encoded with adding minor noize to reduce the risk of overfitting.
Can handle missing values - encode NaN by zero WoE or leave NaN untransformed.
Initialize WeightOfEvidenceEncoder
from verstack import WeightOfEvidenceEncoder
WOE = WeightOfEvidenceEncoder()
train_encoded = WOE.fit_transform(train, 'colname', 'targetname')
test_encoded = WOE.transform(test)
train_reversed_to_original = WOE.inverse_transform(train_encoded)
test_reversed_to_original = WOE.inverse_transform(test_encoded)
Attributes
-
na_sentinel
Defined (at init) missing values encoding value.
-
colname
Defined (at fit_transform()) column that had been transformed.
-
params
Defined (at init) category_encoders.woe.WOEEncoder parameters
Parameters
-
na_sentinel
[default=True]If True: Encode NaN values by zero WoE. If False return np.nan in the encoded column.
-
kwargs
category_encoders.woe.WOEEncoder parameters. Following parameters are set by default:
'randomized':True
,'random_state':42
,'handle_missing':'return_nan'
<- inferred from na_sentinel setting.
Methods
-
fit_transform(df, colname, targetname)
Fit WeightOfEvidenceEncoder to data and return transformed data.
Parameters
-
df
pd.DataFramedf containing the colname to transform.
-
colname
strColumn name in df to be transformed.
-
targetname
strTarget column name in df for calculating WoE for each colname category.
-
-
transform(df)
Apply the fitted WeightOfEvidenceEncoder to new data and return transformed data. Unseen categories' WoE is set to 0.
Parameters
-
df
pd.DataFrameData containing the colname to transform.
-
-
inverse_transform(df)
Inverse transform data that had been encoded by WeightOfEvidenceEncoder. Data must contain colname that was passed at fit_transform().
Parameters
-
df
pd.DataFrameData containing the colname to transform.
-
Examples
WOE = WeightOfEvidenceEncoder()
train_encoded = WOE.fit_transform(train, 'colname', 'targetname')
test_encoded = WOE.transform(test)
train_reversed_to_original = WOE.inverse_transform(train_encoded)
test_reversed_to_original = WOE.inverse_transform(test_encoded)
timer
Timer decorator to measure any function execution time and create elapsed time output: hours/minues/seconds will be calculated and returned conveniently.
verstack.tools.timer
Examples
timer is a decorator function: it must placed above the function (that needs to be timed) definition
from verstack.tools import timer
@timer
def func(a,b):
print(f'Result is: {a + b}')
func(2,3)
>>>Result is: 5
>>>Time elapsed for func execution: 0.0002 seconds
Printer
Class to execute print statements subject to verbose argument and order
of printed message. Includes errors stack trace if order == 'error'.
Add print statements to your program with different level of indentation
for different messages and have them printed subject on the global
verbosity setting in your program. A convenient way to set up verbosity
for large projects without having to define all the print statements
with if verbose == True
. Just pass the verbose argument to the Printer
class instance at initialisation, devine all the print messages with
Printer.print() instaed of builtin print() and if verbose==True
the
messages will be printed, else only the messages with order=='error'
will be printed. Also includes the force_print argument, which will
print the selected messages even if verbose==False
. Applicable for
non-error important messages that need to be printed.
from verstack.tools import Printer
Examples
Abstract example
from verstack.tools import Printer
def long_program_with_multiple_modules(verbose):
printer = Printer(verbose=verbose)
printer.print('Program header', order = 0)
printer.print('Module/major step/epoch name', order = 1)
printer.print('Function inside module name', order = 2)
printer.print('func first order result 1', order = 3)
printer.print('func first order result 2', order = 3)
printer.print('func second order result 1', order = 4)
printer.print('func second order result 2', order = 4)
printer.print('func third order result 1', order = 5)
printer.print('func third order result 2', order = 5)
printer.print(breakline = '=')
printer.print('message with breakline below', order = 1, breakline='.')
try:
5/0
except:
printer.print('5/0 division not executed', order='error')
long_program_with_multiple_modules(verbose=True)
>>> ---------------------------------------------------------------------------
>>> Program header
>>> ---------------------------------------------------------------------------
>>>
>>> * Module/major step/epoch name
>>>
>>> - Function inside module name
>>> . func first order result 1
>>> . func first order result 2
>>> .. func second order result 1
>>> .. func second order result 2
>>> ... func third order result 1
>>> ... func third order result 2
>>> ===========================================================================
>>>
>>> * message with breakline below
>>> ...........................................................................
>>> Traceback (most recent call last):
>>> File "<ipython-input-37-f1aa2de68f72>", line 18, in long_program_with_multiple_modules
>>> 5/0
>>> ZeroDivisionError: division by zero
>>>
>>> ! 5/0 division not executed
Applied example
from verstack.tools import Printer
# define a function/program/code
def do_something(a, b, c, verbose):
printer = Printer(verbose=verbose)
printer.print('Executing do_something() function', order = 0)
printer.print('Running addition operations', order = 1)
printer.print('adding a+b and b+c', order = 2)
result_1 = a + b
result_2 = b + c
printer.print(f'a + b result is {result_1}', order = 3)
printer.print(f'b + c result is {result_2}', order = 3)
printer.print('Trying to make an error', order = 1)
try:
a / b
except ZeroDivisionError:
printer.print('Argument b can not be zero', order = 'error')
do_something(1,0,5, verbose = False)
>>> Traceback (most recent call last):
>>> File "<ipython-input-17-bb8dafd4f34d>", line 9, in do_something
>>> a / b
>>> ZeroDivisionError: division by zero
>>> ! Argument b can not be zero
# only error message gets printed
do_something(1,0,5, verbose = True)
>>> ---------------------------------------------------------------------------
>>> Executing do_something() function
>>> ---------------------------------------------------------------------------
>>>
>>> * Running addition operations
>>>
>>> - adding a+b and b+c
>>> . a + b result is 4
>>> . b + c result is 8
>>>
>>> * Trying to make an error
>>> Traceback (most recent call last):
>>> File "<ipython-input-38-050165db3ba2>", line 13, in do_something
>>> a / b
>>> ZeroDivisionError: division by zero
>>>
>>> ! Argument b can not be zero
Links
Project details
Release history Release notifications | RSS feed
Download files
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.