Apache MXNet
Apache MXNet is a retired deep learning framework (now in the Apache Attic) designed for both efficiency and flexibility. It provided a multi-language API for building and training deep neural networks with support for distributed training, the Gluon high-level API, and deployment on edge devices. MXNet supported Python, Scala, Java, C++, R, Julia, and Perl.
APIs
Apache MXNet
MXNet provides APIs in Python, Scala, Java, C++, R, Julia, and Perl for deep learning model development, with the Gluon high-level API for imperative model building, Symbol/NDAr...
Features
Seamlessly transitions between Gluon eager imperative mode and symbolic execution for research flexibility and production efficiency.
Supports Parameter Server and Horovod for scalable distributed training across multiple GPUs and nodes.
Native APIs in Python, Scala, Java, C++, R, Julia, Clojure, and Perl for broad developer accessibility.
Intuitive Gluon API for imperative model building with automatic differentiation and dynamic computation graphs.
NumPy-like array operations for GPU-accelerated numerical computing as the foundation of MXNet computations.
Symbolic computation graph API for efficient inference and production deployment.
Pre-trained models for computer vision, NLP, and other tasks accessible via the Gluon model zoo.
Lightweight deployment support for edge devices and mobile platforms via TVM and ONNX export.
Use Cases
Build and train image classification, object detection, and segmentation models using GluonCV toolkit.
Develop NLP models for text classification, sentiment analysis, and language modeling using GluonNLP.
Build time series forecasting models using the GluonTS toolkit for probabilistic forecasting.
Train large neural networks across multiple GPUs and nodes using Parameter Server or Horovod.
Rapid prototyping of novel deep learning architectures using the Gluon imperative API.
Integrations
Computer vision toolkit built on MXNet providing pre-trained models and training utilities for vision tasks.
NLP toolkit built on MXNet with pre-trained language models and text processing utilities.
Time series modeling toolkit built on MXNet for probabilistic forecasting.
ONNX model format support for importing and exporting models to/from other frameworks.
Apache TVM deep learning compiler for optimizing MXNet model deployment on diverse hardware targets.
Horovod distributed training framework integration for efficient multi-GPU and multi-node training.
Dive into Deep Learning interactive textbook using MXNet for teaching deep learning concepts.