deepjavalibrary/djl

Deep Java Library (DJL) – Quick Overview

What it is – DJL is a Java‑first, high‑level deep‑learning framework that sits on top of popular native engines (TensorFlow, PyTorch, MXNet, etc.). It lets Java developers load, train, and run inference with neural networks without learning a new language or switching away from their usual IDEs.

Key ideas

• Engine‑agnostic – you pick an engine at runtime; you can swap TensorFlow ↔ PyTorch ↔ MXNet without changing your code.
• Automatic CPU/GPU selection – DJL detects the hardware and chooses the best backend.
• Model zoo integration – pre‑trained models are discovered via a simple Criteria builder.
• Full training support – you can define a network with built‑in blocks, create a Trainer, and run classic training loops.

Typical usage

// Inference – pick a pre‑trained object‑detection model
Criteria<Image, Classifications> criteria = Criteria.builder()
    .optApplication(Application.CV.OBJECT_DETECTION)
    .setTypes(Image.class, Classifications.class)
    .optFilter("backbone", "resnet50")
    .build();
Image img = ImageFactory.getInstance().fromUrl("http://…");
try (ZooModel<Image, Classifications> model = criteria.loadModel();
     Predictor<Image, Classifications> pred = model.newPredictor()) {
    Classifications result = pred.predict(img);
    // …process result…
}

// Training – a tiny MLP on MNIST
Block block = new Mlp(28*28, 10, new int[]{128,64});
Model model = Model.newInstance("mlp");
model.setBlock(block);
Dataset train = new Mnist.Builder().setUsage(Usage.TRAIN).build();
Dataset test  = new Mnist.Builder().setUsage(Usage.TEST).build();
TrainingConfig cfg = setupTrainingConfig();
Trainer trainer = model.newTrainer(cfg);
trainer.initialize(new Shape(1, 28*28));
EasyTrain.fit(trainer, epochs, train, test);
model.save(modelDir, "mlp");

Who it’s for – Java developers who want to experiment with deep learning, embed inference in enterprise services, or build end‑to‑end ML pipelines without leaving the JVM ecosystem.

Where to learn more –

• Official docs: https://docs.djl.ai/
• DJL’s “Dive into Deep Learning” book (Java version)
• Javadoc API reference
• Community Slack, GitHub discussions, and Zhihu column for Chinese‑language updates.

License – Apache‑2.0, so you can use it in both open‑source and commercial projects.

---

All details are taken directly from the repository’s README.