LightAutoML documentation

LightAutoML is open-source Python library aimed at automated machine learning. It is designed to be lightweight and efficient for various tasks with tabular, text data. LightAutoML provides easy-to-use pipeline creation, that enables:

• Automatic hyperparameter tuning, data processing.

• Automatic typing, feature selection.

• Automatic time utilization.

• Automatic report creation.

• Graphical profiling system.

• Easy-to-use modular scheme to create your own pipelines.

lightautoml.automl

The main module, which includes the AutoML class, blenders and ready-made presets.

AutoML

Class for compile full pipeline of AutoML task.

Presets

Presets for end-to-end model training for special tasks.

base.AutoMLPreset	Basic class for automl preset.
whitebox_presets.WhiteBoxPreset	Preset for AutoWoE - logistic regression over binned features (scorecard).

Blenders

Blender	Basic class for blending.
BestModelSelector	Select best single model from level.
MeanBlender	Simple average level predictions.
WeightedBlender	Weighted Blender based on coord descent, optimize task metric directly.

lightautoml.addons

Extensions of core functionality.

Utilization

TimeUtilization

Class that helps to utilize given time to AutoMLPreset.

lightautoml.dataset

Provides an internal interface for working with data.

Dataset Interfaces

base.LAMLColumn	Basic class for pair - column, role.
base.LAMLDataset	Basic class to create dataset.
np_pd_dataset.NumpyDataset	Dataset that contains info in np.ndarray format.
np_pd_dataset.PandasDataset	Dataset that contains pd.DataFrame features and pd.Series targets.
np_pd_dataset.CSRSparseDataset	Dataset that contains sparse features and np.ndarray targets.

Roles

Role contains information about the column, which determines how it is processed.

ColumnRole	Abstract class for column role.
NumericRole	Numeric role.
CategoryRole	Category role.
TextRole	Text role.
DatetimeRole	Datetime role.
TargetRole	Target role.
GroupRole	Group role.
DropRole	Drop role.
WeightsRole	Weights role.
FoldsRole	Folds role.
PathRole	Path role.

Utils

Utilities for working with the structure of a dataset.

roles_parser	Parser of roles.
get_common_concat	Get concatenation function for datasets of different types.
numpy_and_pandas_concat	Concat of numpy and pandas dataset.
concatenate	Dataset concatenation function.

lightautoml.image

Provides an internal interface for working with image features.

Image Feature Extractors

Image feature extractors based on color histograms and CNN embeddings.

CreateImageFeatures	Class for parallel histogram computation.
EffNetImageEmbedder	Class to compute EfficientNet embeddings.

PyTorch Image Datasets

ImageDataset	Image Dataset Class.
DeepImageEmbedder	Transformer for image embeddings.

Utils

pil_loader

Load image from pathes.

lightautoml.ml_algo

Models used for machine learning pipelines.

Base Classes

MLAlgo	Abstract class for machine learning algorithm.
TabularMLAlgo	Machine learning algorithms that accepts numpy arrays as input.

Linear Models

LinearLBFGS	LBFGS L2 regression based on torch.
LinearL1CD	Coordinate descent based on sklearn implementation.

Boosted Trees

BoostLGBM	Gradient boosting on decision trees from LightGBM library.
BoostCB	Gradient boosting on decision trees from catboost library.

WhiteBox

WbMLAlgo

WhiteBox - scorecard model.

lightautoml.ml_algo.tuning

Bunch of classes for hyperparameters tuning.

Base Classes

ParamsTuner	Base abstract class for hyperparameters tuners.
DefaultTuner	Default realization of ParamsTuner - just take algo's defaults.

Tuning with Optuna

OptunaTuner

Wrapper for optuna tuner.

lightautoml.pipelines

Pipelines for solving different tasks.

Utils

map_pipeline_names	Pipelines create name in the way 'prefix__feature_name'.
get_columns_by_role	Search for columns with specific role and attributes when building pipeline.

lightautoml.pipelines.selection

Feature selection module for ML pipelines.

Base Classes

ImportanceEstimator	Abstract class, that estimates feature importances.
SelectionPipeline	Abstract class, performing feature selection.

Importance Based Selectors

ModelBasedImportanceEstimator	Base class for performing feature selection using model feature importances.
ImportanceCutoffSelector	Selector based on importance threshold.
NpPermutationImportanceEstimator	Permutation importance based estimator.
NpIterativeFeatureSelector	Select features sequentially using chunks to find the best combination of chunks.

Other Selectors

HighCorrRemoval

Selector to remove highly correlated features.

lightautoml.pipelines.features

Pipelines for features generation.

Base Classes

FeaturesPipeline	Abstract class.
EmptyFeaturePipeline	Dummy feature pipeline - .fit_transform and transform do nothing.
TabularDataFeatures	Helper class contains basic features transformations for tabular data.

Feature Pipelines for Boosting Models

LGBSimpleFeatures	Creates simple pipeline for tree based models.
LGBAdvancedPipeline	Create advanced pipeline for trees based models.

Feature Pipelines for Linear Models

LinearFeatures

Creates pipeline for linear models and nnets.

Feature Pipelines for WhiteBox

WBFeatures

Simple WhiteBox pipeline.

Image Feature Pipelines

ImageDataFeatures	Class contains basic features transformations for image data.
ImageSimpleFeatures	Class contains simple color histogram features for image data.
ImageAutoFeatures	Class contains efficient-net embeddings features for image data.

Text Feature Pipelines

NLPDataFeatures	Class contains basic features transformations for text data.
TextAutoFeatures	Class contains embedding features for text data.
NLPTFiDFFeatures	Class contains tfidf features for text data.
TextBertFeatures	Features pipeline for BERT.

lightautoml.pipelines.ml

Pipelines that merge together single model training steps.

Base Classes

MLPipeline

Single ML pipeline.

Pipeline for Nested Cross-Validation

NestedTabularMLAlgo	Wrapper for MLAlgo to make it trainable over nested folds.
NestedTabularMLPipeline	Wrapper for MLPipeline to make it trainable over nested folds.

Pipeline for WhiteBox

WBPipeline

Special pipeline to handle WhiteBox model.

lightautoml.reader

Utils for reading, training and analysing data.

Readers

Reader	Abstract class for analyzing input data and creating inner LAMLDataset from raw data.
PandasToPandasReader	Reader to convert DataFrame to AutoML's PandasDataset.

Tabular Batch Generators

Batch Handler Classes

Data Read Functions

lightautoml.report

Report generators and templates.

lightautoml.tasks

Task Class

Task	Specify task (binary classification, multiclass classification, regression), metrics, losses.

Common Metrics

Classes

F1Factory	Wrapper for f1_score function.
BestClassBinaryWrapper	Metric wrapper to get best class prediction instead of probs.
BestClassMulticlassWrapper	Metric wrapper to get best class prediction instead of probs for multiclass.

Functions

mean_quantile_error	Computes Mean Quantile Error.
mean_huber_error	Computes Mean Huber Error.
mean_fair_error	Computes Mean Fair Error.
mean_absolute_percentage_error	Computes Mean Absolute Percentage error.
roc_auc_ovr	ROC-AUC One-Versus-Rest.
rmsle	Root mean squared log error.
auc_mu	Compute multi-class metric AUC-Mu.

lightautoml.tasks.losses

Wrappers of loss and metric functions for different machine learning algorithms.

Base Classes

MetricFunc	Wrapper for metric.
Loss	Loss function with target transformation.

Wrappers for LightGBM

Classes

LGBFunc	Wrapper of metric function for LightGBM.
LGBLoss	Loss used for LightGBM.

Functions

softmax_ax1	Softmax columnwise.
lgb_f1_loss_multiclass	Custom loss for optimizing f1.

Wrappers for CatBoost

Classes

CBLoss	Loss used for CatBoost.
CBCustomMetric	Metric wrapper class for CatBoost.
CBRegressionMetric	Regression metric wrapper for CatBoost.
CBClassificationMetric	Classification metric wrapper for CatBoost.
CBMulticlassMetric	Multiclassification metric wrapper for CatBoost.

Functions

cb_str_loss_wrapper

CatBoost loss name wrapper, if it has keyword args.

Wrappers for Sklearn

Classes

SKLoss

Loss used for scikit-learn.

Wrappers for Torch

Classes

TorchLossWrapper	Customize PyTorch-based loss.
TORCHLoss	Loss used for PyTorch.

Functions

torch_rmsle	Computes Root Mean Squared Logarithmic Error.
torch_quantile	Computes Mean Quantile Error.
torch_fair	Computes Mean Fair Error.
torch_huber	Computes Mean Huber Error.
torch_f1	Computes F1 macro.
torch_mape	Computes Mean Absolute Percentage Error.

lightautoml.text

Provides an internal interface for working with text features.

Sentence Embedders

DLTransformer	Deep Learning based sentence embeddings.
BOREP	Class to compute Bag of Random Embedding Projections sentence embeddings from words embeddings.
RandomLSTM	Class to compute Random LSTM sentence embeddings from words embeddings.
BertEmbedder	Class to compute HuggingFace transformers words or sentence embeddings.
WeightedAverageTransformer	Weighted average of word embeddings.

Torch Datasets for Text

BertDataset	Dataset class with transformers tokenization.
EmbedDataset	Dataset class for extracting word embeddings.

Tokenizers

BaseTokenizer	Base class for tokenizer method.
SimpleRuTokenizer	Russian tokenizer.
SimpleEnTokenizer	English tokenizer.

Pooling Strategies

SequenceAbstractPooler	Abstract pooling class.
SequenceClsPooler	CLS token pooling.
SequenceMaxPooler	Max value pooling.
SequenceSumPooler	Sum value pooling.
SequenceAvgPooler	Mean value pooling.
SequenceIndentityPooler	Identity pooling.

Utils

seed_everything	Set random seed and cudnn params.
parse_devices	Parse devices and convert first to the torch device.
custom_collate	Puts each data field into a tensor with outer dimension batch size.
single_text_hash	Get text hash.
get_textarr_hash	Get hash of array with texts.

lightautoml.transformers

Basic feature generation steps and helper utils.

Base Classes

LAMLTransformer	Base class for transformer method (like sklearn, but works with datasets).
SequentialTransformer	Transformer that contains the list of transformers and apply one by one sequentially.
UnionTransformer	Transformer that apply the sequence on transformers in parallel on dataset and concatenate the result.
ColumnsSelector	Select columns to pass to another transformers (or feature selection).
ColumnwiseUnion	Apply 1 columns transformer to all columns.
BestOfTransformers	Apply multiple transformers and select best.
ConvertDataset	Convert dataset to given type.
ChangeRoles	Change data roles (include dtypes etc).

Numeric

NaNFlags	Create NaN flags.
FillnaMedian	Fillna with median.
FillInf	Fill inf with nan to handle as nan value.
LogOdds	Convert probs to logodds.
StandardScaler	Classic StandardScaler.
QuantileBinning	Discretization of numeric features by quantiles.

Categorical

LabelEncoder	Simple LabelEncoder in order of frequency.
OHEEncoder	Simple OneHotEncoder over label encoded categories.
FreqEncoder	Labels are encoded with frequency in train data.
OrdinalEncoder	Encoding ordinal categories into numbers.
TargetEncoder	Out-of-fold target encoding.
MultiClassTargetEncoder	Out-of-fold target encoding for multiclass task.
CatIntersectstions	Build label encoded intertsections of categorical variables.

Datetime

TimeToNum	Basic conversion strategy, used in selection one-to-one transformers.
BaseDiff	Basic conversion strategy, used in selection one-to-one transformers.
DateSeasons	Basic conversion strategy, used in selection one-to-one transformers.

Decompositions

PCATransformer	PCA.
SVDTransformer	TruncatedSVD.

Text

TunableTransformer	Base class for ML transformers.
TfidfTextTransformer	Simple Tfidf vectorizer.
TokenizerTransformer	Simple tokenizer transformer.
OneToOneTransformer	Out-of-fold sgd model prediction to reduce dimension of encoded text data.
ConcatTextTransformer	Concat text features transformer.
AutoNLPWrap	Calculate text embeddings.

Image

ImageFeaturesTransformer	Simple image histogram.
AutoCVWrap	Calculate image embeddings.

lightautoml.utils

Common util tools.

Timer

Timer	Timer to limit the duration tasks.
PipelineTimer	Timer is used to control time over full automl run.
TaskTimer	Timer is used to control time over single ML task run.

lightautoml.validation

The module provide classes and functions for model validation.

Iterators

TrainValidIterator	Abstract class to train/validation iteration.
DummyIterator	Simple Iterator which use train data as validation.
HoldoutIterator	Iterator for classic holdout - just predefined train and valid samples.
CustomIterator	Iterator that uses function to create folds indexes.
FoldsIterator	Classic cv iterator.
TimeSeriesIterator	Time Series Iterator.

Iterators Getters and Utils

create_validation_iterator	Creates train-validation iterator.
get_numpy_iterator	Get iterator for np/sparse dataset.

Indices and Tables

• Index