dromara/Omega-AI

Omega-AI：基于java打造的深度学习框架，帮助你快速搭建神经网络，实现模型推理与训练，引擎支持自动求导，多线程与GPU运算，GPU支持CUDA，CUDNN。

What it solves

Omega-AI is a deep learning framework built entirely in Java. It aims to make AI development more accessible to Java developers and provides a way to deeply understand the underlying principles of algorithms and models by implementing them from scratch.

How it works

The framework provides a comprehensive set of building blocks for neural networks, including various layer types (Fully Connected, Convolutional, RNN, LSTM, Transformer, and Yolo layers), activation functions (ReLU, Softmax, Sigmoid), and optimizers (Adam, SGD, RMSProp). It supports GPU acceleration via CUDA and cuDNN through the JCuda library, allowing for multi-GPU training. The engine is designed to be independent of third-party APIs for its core model and algorithm implementations.

Who it’s for

Java developers who want to build, train, and test deep learning models without relying on Python-based frameworks, as well as students or researchers interested in the internal mechanics of AI models.

Highlights

Broad Model Support: Supports a wide range of architectures including BP, CNN, RNN, VGG16, ResNet, YOLO, LSTM, Transformer, GPT, Llama, and Diffusion/Stable Diffusion.
GPU Acceleration: Integrated support for CUDA and cuDNN for high-performance training and inference.
Comprehensive Tooling: Includes built-in data loaders for .bin, .txt, and .idx3-ubyte formats, and various learning rate update strategies.
End-to-End Demos: Provides practical examples ranging from MNIST digit recognition and YOLO object detection to GPT-based chatbots and Stable Diffusion text-to-image generation.