This document is the API specification for the Deep Java Library (DJL).

The key design goal for the DJL is to simplify the use of deep learning for Java developers. Developers don't need to be machine learning/deep learning experts to get started. They can start with their existing Java expertise as a foundation to learn and use ML/DL. Developers can use their favorite IDE to build, train, and deploy their models. They can also easily integrate these models with their Java applications. They don't have to compromise on model performance, scale, or choice of GPU vs. CPUs.

DJL is deep learning engine agnostic. Developers don't have to make a choice between engines when they start their project. They can switch to a different engine at any time.

DJL provides a native Java development experience and functions just like any other regular Java library would. DJL's ergonomic API interface is designed to guide developers with best practices to accomplish deep learning tasks.

The following is an example of how to write inference code:

    // Assume user has a pre-trained already, they just need load it
    Model model = Model.load(modelDir, modelName);

    // User can implement Translator interface, read Translator for detail.
    Translator translator = new MyTranslator();

    // Next user need create a Predictor, and use Predictor.predict()
    // to get prediction.
    try (Predictor<Image, DetectedObjects> predictor = model.newPredictor(translator)) {
        DetectedObjects result = predictor.predict(img);
    }

More tutorials, documents, and examples can be on our GitHub repository.

Packages

Package	Description
ai.djl	Contains top level, engine-agnostic classes for both inference and training.
ai.djl.engine	Contains classes responsible for loading a deep learning engine.
ai.djl.inference	Contains classes to implement inference tasks.
ai.djl.metric	Contains classes to collect metrics information.
ai.djl.modality	Contains utility classes for each of the predefined modalities.
ai.djl.modality.cv	Contains utility classes for computer vision tasks and image processing.
ai.djl.modality.cv.output	Contains output types used in various computer vision applications.
ai.djl.modality.cv.transform	Contains Transforms for working with Images.
ai.djl.modality.cv.translator	Contains translators used for computer vision applications.
ai.djl.modality.cv.translator.wrapper	Contains wrappers to for multiple input formats to a BaseImageTranslator.
ai.djl.modality.cv.util	Contains utility classes for image manipulation.
ai.djl.modality.nlp	Contains utility classes for natural language processing tasks.
ai.djl.modality.nlp.bert	Contains classes that deal with BERT for natural language pre-processing tasks.
ai.djl.modality.nlp.embedding	Contains classes that deal with word embeddings for natural language pre-processing tasks.
ai.djl.modality.nlp.preprocess	Contains utility classes for natural language pre-processing tasks.
ai.djl.modality.nlp.qa	Contains utility classes for question and answer processing.
ai.djl.modality.nlp.translator	Contains utility classes for each of the predefined translator.
ai.djl.modality.rl	Contains utility classes for reinforcement learning.
ai.djl.modality.rl.agent	Contains agents to learn using reinforcement learning.
ai.djl.modality.rl.env	Contains environments to train reinforcement learning in.
ai.djl.ndarray	Contains classes and interfaces that define an n-dimensional array.
ai.djl.ndarray.index	Contains classes that help access NDArray's indices.
ai.djl.ndarray.index.dim	Contains classes that represent an index element in a NDArray's indices.
ai.djl.ndarray.index.full	Contains classes that represent simplified representations of an NDArray's indices.
ai.djl.ndarray.types	Contains classes that define n-dimensional array data types.
ai.djl.nn	Contains classes to construct neural networks.
ai.djl.nn.convolutional	Contains classes that define convolutional operations extending Convolution and Deconvolution.
ai.djl.nn.core	Contains classes that define simple neural network operations.
ai.djl.nn.norm	Contains classes that define normalizing neural network operations.
ai.djl.nn.pooling	Contains pooling neural network operations in Pool and helpers for it.
ai.djl.nn.recurrent	Contains classes for recurrent neural network operations.
ai.djl.nn.transformer	Contains blocks for transformer models.
ai.djl.repository	Contains a Maven-based Repository format for creating repositories of artifacts such as datasets and model zoos.
ai.djl.repository.zoo	Contains classes for working with model zoo repositories.
ai.djl.training	Contains classes and implementations to train a neural network.
ai.djl.training.dataset	Contains classes to download and prepare training and testing data.
ai.djl.training.evaluator	Contains classes for evaluating the effectiveness of models.
ai.djl.training.hyperparameter	Contains utilities to train, describe, and manipulate Hyperparameters.
ai.djl.training.hyperparameter.optimizer	Contains classes to optimize Hyperparameters.
ai.djl.training.hyperparameter.param	Contains different types of Hyperparameters.
ai.djl.training.initializer	Contains classes for initializing a Block's Parameters.
ai.djl.training.listener	Contains classes providing functionality during training through TrainingListener.
ai.djl.training.loss	Contains classes for measuring the Loss of a model.
ai.djl.training.optimizer	Contains classes for optimizing a neural network Block.
ai.djl.training.tracker	Contains classes for having a gradually changing hyper-parameter.
ai.djl.training.util	Contains utilities to use during training.
ai.djl.translate	Contains classes and interfaces that translate between java objects and NDArrays.