Feed-forward networks
FNN, with configurable hidden layers and activations for classification and regression.
Java 20+ · CPU + optional Metal
Neural networks, built for Java.
Build, train, and run feed-forward, convolutional, recurrent, autoencoder, and Transformer networks on one shared tensor and autograd foundation.
- N-dimensional tensors
- Automatic differentiation
- MIT licensed
Supported networks
Choose a network. Keep the same foundation.
Compose layers, calculate gradients, and train different model types with DeepJ tensors, parameters, and optimizers.
Convolutional networks
Conv1D, Conv2D, max/average pooling, and global average pooling for channel-first inputs.
Recurrent networks
RNN, GRU, and LSTM, with explicit hidden state and per-sample sequence lengths.
Autoencoders
Compose an encoder and decoder, train a reconstruction objective, and access the latent representation separately.
Transformer encoders
Bidirectional attention, padding masks, pre-normalized blocks, and masked mean pooling.
Decoder Transformers
DeepJ Origin, Orbit, Prism, Nova, and Loop, with causal language-model training and text generation.
These model types use DeepJ layers, parameters, tensors, and autograd.
Core library
One foundation for every network.
Tensors, gradients, composition, training, and execution belong to the library as a whole.
N-dimensional tensors
One shape-aware tensor API for vectors, matrices, images, and sequences, including views and broadcasting.
Automatic differentiation
Record tensor operations and propagate first-order gradients through the same CPU/Metal backend.
Composable layers
Combine existing and new layers with Sequential, activation adapters, dropout, and BatchNorm.
Training utilities
Supervised matrix and ND batching, differentiable losses, AdamW, and explicit training/evaluation mode.
Saving and loading
Store network parameters and persistent buffers. Language-model files and resumable checkpoints have separate contracts.
CPU and optional Metal
Use Java CPU execution or select Metal on supported Apple Silicon. Network arithmetic shares these backend operations.
Training
Train locally or across machines.
Train DeepJ models on CPU or Apple Metal. To share training work across multiple Apple Silicon Macs or CPU machines, use the optional distributed module.
Training a model
Choose a training API for your model type, then configure minibatches, losses, optimizers, and evaluation mode. The guide also covers language models.
Read guideDistributed training
Train across multiple machines. Each worker trains a full model copy on different data, then synchronizes gradients before every update.
Read guideInstall
Bring DeepJ to your Java project.
JDK 20 or newer is required.
Add the Maven dependency to your project, or build DeepJ from source.
<dependency>
<groupId>io.github.kirstenali</groupId>
<artifactId>deepj</artifactId>
<version>0.13.0-alpha</version>
</dependency>Working with DeepJ
Build, train, and run.
Compose a network
Choose layers and activations, preserve the input shape contract, and share parameters when needed.
Read guideTrain with gradients
Compute a tensor loss, run backward, and update parameters. Select the appropriate batching and evaluation mode.
Read guideSave the state you need
Store parameters and persistent buffers, or use the language-model checkpoint APIs for their broader state.
Read guideDocumentation
Start with the library.
Find the shape, gradient, training, and persistence contracts before exploring specialist model tools.
Core guides
Tensors, autograd, model types, training, saving/loading, and backend execution.
Read guideAPI reference
Browse classes and methods for DeepJ networks, layers, and tensors.
Read guideVerification
Portable and full Metal/desktop builds, numerical checks, and coverage requirements.
Read guideLanguage-model tools
Text generation is part of DeepJ, too.
Decoder architectures, tokenizers, pretrained checkpoints, and chat interfaces build on the same neural-network library.
Decoder architectures
Origin, Orbit, Prism, Nova, and Loop support causal training and generation, with attention designs documented by their components.
Decoder referenceDesktop chat
The optional JavaFX app loads local language-model checkpoints, applies saved prompt formatting, and streams generated text.
Desktop guideCommand line
Streaming chat, generation, checkpoint inspection, themes, and backend diagnostics for supported language models.
CLI referencePublished checkpoints: Frisky 90M and DeepJ TinyStories. Their model cards describe training, loading requirements, and limitations.