XiaoMi/mace
Mace
Mace es un framework de inferencia de deep learning móvil y motor de aceleración de hardware. Funciona como un runtime para ejecutar modelos de redes neuronales en dispositivos móviles, distribuyendo cálculos a través de CPUs, GPUs y NPUs.
El proyecto incluye un convertidor de modelos multiplataforma para transformar redes neuronales pre-entrenadas de varios formatos de la industria en representaciones optimizadas para móviles. También proporciona un ofuscador de redes neuronales que convierte los pesos del modelo en código fuente para proteger la propiedad intelectual contra la ingeniería inversa.
El framework gestiona los recursos del dispositivo optimizando la asignación de memoria y ajustando la configuración de energía del chip. Además, aborda el rendimiento de ejecución mediante la optimización de algoritmos matemáticos y la división de cálculos para mantener la capacidad de respuesta de la interfaz.
Features
- On-Device Inference - Implements local execution of pre-trained deep learning models directly on mobile client hardware.
- Deep Learning Inference Engines - Provides a high-performance system for executing predictions of deep learning models on mobile hardware.
- Inference Acceleration Engines - Distributes deep learning computations across mobile CPUs, GPUs, and NPUs for maximum performance.
- On-Device Model Runtimes - Provides an execution environment to run language and deep learning models entirely on local mobile hardware.
- Inference Speed Optimizers - Increases operation speed by applying hardware acceleration and optimized mathematical algorithms to complex calculations.
- Deep Learning Acceleration - Provides hardware acceleration specifically for deep learning tensor operations and matrix multiplications on mobile chips.
- Hardware Acceleration - Offloads intensive deep learning computations to specialized mobile hardware like GPUs and NPUs.
- Cross-Format Model Importers - Provides tools to import pre-trained models from various industry-standard formats into a mobile-optimized runtime.
- Memory Footprint Optimizers - Minimizes memory consumption by optimizing data buffer allocation and reuse throughout the processing graph.
- Mobile Model Format Converters - Transforms neural network checkpoints from various industry formats into optimized representations for mobile interpreters.
- Resource Optimization - Manages chip power settings and memory allocation to balance energy consumption and performance.
- On-Device Resource Optimizers - Manages chip power settings and memory allocation to balance energy consumption and performance.
- Hardware Power Controllers - Adjusts chip power settings and scheduling hints to balance energy consumption with processing speed.
- Model Obfuscation Tools - Secures neural network models by converting weights into obfuscated source code to prevent reverse engineering.
