Package Structure Overview
- Package Description
- Util libraries
- Multi-class Classification
- Multi-label Classification
- Anomaly Detection
- Third party models
- Other modules
Tribuo has a modular structure to allow minimal dependencies for any specific deployment. We describe the overall package structure below.
The top level project has core modules which define the API, data interactions, a math library, and common modules shared across prediction types.
- Core - (artifactID:
tribuo-core, package root:
org.tribuo) Provides the main classes and interfaces:
Dataset- A list of
Examples plus associated feature information, such as the number of categories for categorical features, or the mean and variance in the case of real-valued features.
DataSource- A list of
Examples processed from some other format and accompanied by the provenance describing the source and processing of these
Example- An array or list of String and value pairs. The
Exampleis typed with a subclass of Output that represents the appropriate type of response.
Feature- An immutable tuple of String and value. The String is the feature name, which is used as the feature’s unique identifier.
FeatureMap- A map from String to
VariableInfoobjects. When immutable, it also contains feature id numbers, although these should be treated as an implementation detail and not relied upon.
Model- A class that can make predictions of a specific
Output- An interface denoting the type of output: regression, multi-label, multi-class, clustering, or anomaly detection.
OutputInfo- An interface representing information about the output.
Trainer- A class that generates
Models based on a
Datasetof a specific output type.
Prediction- A class that stores the output of a
Modelwhen presented with an
Examplefor labeling. It contains scores for each of the predicted labels. These scores may optionally be a probability distribution.
VariableInfo- A class representing information about the feature, e.g., the number of times it occurred in the dataset.
- Core contains several other packages.
Datasets which provide a view on another dataset, either subsampling it or excluding features below a threshold.
datasource- Implementations of
DataSourcewhich operate on in-memory or simple on-disk formats.
ensemble- Base classes and interfaces for ensemble models.
evaluation- Base evaluation classes for all output types. This package also includes evaluation-related classes for cross-validation and train-test splitting.
hash- An implementation of feature hashing which obfuscates any feature names that are stored in a
Model. Hashing prevents feature names from leaking out of the training data.
provenance- Provenance classes for Tribuo. Provenance tracks the location and transformations of datasets, the parameters of trainers, and other useful information.
sequence- A sequence prediction API.
transform- A feature transformation package that can be applied to a full dataset or to individual features matched via regexes. It also contains wrapping trainers (trainers that wrap another trainer to provide additional functionality) and wrapping models to ensure that the same transformations are always applied at prediction time.
util- Utilities for basic operations such as for working with arrays and random samples.
- Data - (artifactID:
tribuo-data, package root:
org.tribuo.data) provides classes which deal with sampled data, columnar data, csv files and text inputs. The user is encouraged to provide their own text processing infrastructure implementation, as the one here is fairly basic.
columnar- The columnar package provides many useful base classes for extracting features from columnar data.
csv- Builds on the columnar package and supplies infrastructure for working with CSV and other delimiter separated data.
sql- Builds on the columnar package and supplies infrastructure for working with JDBC sources.
text- Text processing infrastructure interfaces and an example implementation.
- Json - (artifactID:
tribuo-json, package root:
org.tribuo.json) provides functionality for loading from json data sources, and for stripping provenance out of a model.
- Math - (artifactID:
tribuo-math, package root:
org.tribuo.math) provides a linear algebra library for working with both sparse and dense vectors and matrices.
kernel- a set of kernel functions for use in the SGD package (and elsewhere).
la- a linear algebra library containing functions used in the SGD implementation. It is not a full BLAS.
optimisers- a set of stochastic gradient descent algorithms, including
RMSPropand several others.
AdaGradshould be considered the default algorithm since it works best across the widest range of linear SGD problems.
util- various util classes for working with arrays, vectors and matrices.
There are 3 utility libraries which are used by Tribuo but do not depend on other parts of it.
- InformationTheory - (artifactID:
tribuo-util-infotheory, package root:
org.tribuo.util.infotheory) provides discrete information theoretic functions suitable for computing clustering metrics, feature selection and structure learning.
- ONNXExport - (artifactID:
tribuo-util-onnx, package root:
org.tribuo.util.onnx) provides infrastructure for building ONNX graphs from Java. This package is suitable for use in other JVM libraries which want to write ONNX models, and provides additional type safety and usability over directly writing the protobufs.
- Tokenization - (artifactID:
tribuo-util-tokenization, package root:
org.tribuo.util.tokens) provides a tokenization API suitable for feature extraction or information retrieval, along with several tokenizer implementations, including a wordpiece implementation suitable for use with models like BERT.
Multi-class classification is the act of assigning a single label from a set of labels to a test example. The classification module has several submodules:
||Contains an Output subclass for use with multi-class classification tasks, evaluation code for checking model performance, and an implementation of Adaboost.SAMME. It also contains simple baseline classifiers.|
||An implementation of CART decision trees.|
||A set of main functions for training & testing models on any supported dataset. This submodule depends on all the classifiers and allows easy comparison between them. It should not be imported into other projects since it is intended purely for development and testing.|
||An implementation of LIME for classification tasks. If you use the columnar data loader, LIME can extract more information about the feature domain and provide better explanations.|
||An implementation of several information theoretic feature selection algorithms for classification problems.|
||A wrapper around the LibLinear-java library. This provides linear-SVMs and other l1 or l2 regularised linear classifiers.|
||A wrapper around the Java version of LibSVM. This provides linear & kernel SVMs with sigmoid, gaussian and polynomial kernels.|
|Multinomial Naive Bayes||
||An implementation of a multinomial naive bayes classifier. Since it aims to store a compact in-memory representation of the model, it only keeps track of weights for observed feature/class pairs.|
||An implementation of stochastic gradient descent based classifiers. It includes a linear package for logistic regression and linear-SVM (using log and hinge losses, respectively), a kernel package for training a kernel-SVM using the Pegasos algorithm, a crf package for training a linear-chain CRF, and a fm package for training pairwise factorization machines. These implementations depend upon the stochastic gradient optimisers in the main Math package. The linear, fm, and crf packages can use any of the provided gradient optimisers, which enforce various different kinds of regularisation or convergence metrics. This is the preferred package for linear classification and for sequence classification due to the speed and scalability of the SGD approach.|
||A wrapper around the XGBoost Java API. XGBoost requires a C library accessed via JNI. XGBoost is a scalable implementation of gradient boosted trees.|
Multi-label classification is the task of predicting a set of labels for a test example rather than just a single label.
The independent binary predictor breaks each multi-label prediction into n binary predictions, one for each possible label. To achieve this, the supplied trainer takes a classification trainer and uses it to build n models, one per label, which are then run in sequence on a test example to produce the final multi-label output. A similar approach is used in the classifier chains to convert a classification trainer into a multi-label trainer.
||Contains an Output subclass for multi-label prediction, evaluation code for checking the performance of a multi-label model, and a basic implementation of independent binary predictions. It also contains implementations of Classifier Chains and Classifier Chain Ensembles, which are more powerful ensemble techniques for multi-label prediction tasks.|
||An implementation of stochastic gradient descent based classifiers. It includes a linear package for independent logistic regression and linear-SVM (using log and hinge losses, respectively), along with factorization machines using either loss for each output label. These implementations depend upon the stochastic gradient optimisers in the main Math package. The linear and fm packages can use any of the provided gradient optimisers, which enforce various different kinds of regularisation or convergence metrics.|
Regression is the task of predicting real-valued outputs for a test example. This package provides several modules:
||contains an Output subclass for use with regression data, as well as evaluation code for checking model performance using standard regression metrics (R^2, explained variance, RMSE, and mean absolute error). The module also contains simple baseline regressions.|
||A wrapper around the LibLinear-java library. This provides linear-SVMs and other l1 or l2 regularised linear regressions.|
||A wrapper around the Java version of LibSVM. This provides linear & kernel SVRs with sigmoid, gaussian and polynomial kernels.|
||An implementation of two types of CART regression trees. The first type builds a separate tree per output dimension, while the second type builds a single tree for all outputs.|
||An implementation of stochastic gradient descent for linear regression and factorization machine regression. It uses the main Math package’s set of gradient optimisers, which allow for various regularisation and descent algorithms.|
||An implementation of sparse linear models. It includes a co-ordinate descent implementation of ElasticNet, a LARS implementation, a LASSO implementation using LARS, and a couple of sequential forward selection algorithms.|
||A wrapper around the XGBoost Java API. XGBoost requires a C library accessed via JNI.|
Clustering is the task of grouping input data. The clustering system implemented is single membership – each datapoint is assigned to one and only one cluster. This package provides two modules:
||Contains the Output subclass for use with clustering data, as well as the evaluation code for measuring clustering performance.|
||An implementation of HDBSCAN, a non-parametric density based clustering algorithm.|
||An implementation of K-Means using the Java 8 Stream API for parallelisation, along with the K-Means++ initialization algorithm.|
Anomaly detection is the task of finding outliers or anomalies at prediction time using a model trained on non-anomalous data. This package provides two modules:
||Contains the Output subclass for use with anomaly detection data.|
||A wrapper around the Java version of LibLinear, which provides a one-class SVM.|
||A wrapper around the Java version of LibSVM, which provides a one-class SVM.|
The common module shares code across multiple prediction types. It provides the base support for LibLinear, LibSVM, nearest neighbour, tree, and XGBoost models. The nearest neighbour submodule is standalone, however the rest of the submodules require the prediction specific implementation modules. The common tree package contains the implementations of Random Forests and Extremely Randomized Trees (ExtraTrees).
Third party models
Tribuo supports loading a number of third party models which were trained outside the system (even in other programming languages) and scoring them from Java using Tribuo’s infrastructure. Currently, we support loading ONNX, TensorFlow and XGBoost models. Additionally we support wrapping an OCI Data Science model deployment in a Tribuo model.
- OCI - Supports deploying Tribuo models to OCI Data Science, and wrapping OCI Data Science models in Tribuo external models to allow them to be served with other Tribuo models.
- ONNX - ONNX (Open Neural Network eXchange) format is used
by several deep learning systems as an export format, and there are
converters from systems like scikit-learn to the ONNX format. Tribuo provides
a wrapper around Microsoft’s ONNX Runtime that can
score ONNX models on both CPU and GPU platforms. ONNX support is found in the
tribuo-onnxartifact in the
org.tribuo.interop.onnxpackage which also provides a feature extractor that uses BERT embedding models. This package can load Tribuo-exported ONNX models and extract the stored Tribuo provenance objects from those models.
- TensorFlow - Tribuo supports loading TensorFlow’s frozen graphs and saved models and scoring them.
- XGBoost - Tribuo supports loading XGBoost classification and regression models.
Tribuo includes experimental support for TensorFlow-Java 0.4.0 (using
TensorFlow 2.7.0) in the
tribuo-tensorflow artifact in the
org.tribuo.interop.tensorflow package. Models can be defined using
TensorFlow-Java’s graph construction mechanisms, and Tribuo will manage the
gradient optimizer output function and loss function. It includes a Java
serialisation system so that all TensorFlow models can be serialised and
deserialised in the same way as other Tribuo models. TensorFlow models run by
default on GPU if one is available and the appropriate GPU jar is on the
This support remains experimental while the TF JVM SIG rewrites the TensorFlow Java API. We participate in the TensorFlow JVM SIG, and are working to improve TensorFlow not just for Tribuo but for the Java community as a whole.
Tribuo demonstrates the TensorFlow interop by including an example config file, several example model generation functions and protobuf for an MNIST model graph.
Tribuo has a number of other modules:
||Contains support for reading and writing Json formatted data, along with a program for inspecting and removing provenance information from models.|
||Contains support for reading and writing model cards in Json format, using the provenance information in Tribuo models to guide the card construction.|
||A utility for reproducing Tribuo models and datasets.|