huawei-noah/ghostnet
Ghostnet
GhostNet ofrece un conjunto de arquitecturas de modelos de IA eficientes y patrones de diseño de redes neuronales creados para reducir el consumo de cómputo y memoria. Funciona como un backbone de visión artificial y un vision transformer ligero, optimizando el equilibrio entre la precisión predictiva y la velocidad de inferencia.
El proyecto se centra en reducir el consumo de recursos para el despliegue en dispositivos móviles y hardware de borde (edge). Esto se logra mediante implementaciones de vision transformers ligeros y arquitecturas que minimizan el número total de parámetros.
El código base cubre una amplia gama de capacidades para la optimización de inferencia, incluyendo la reducción de costes computacionales y el uso de memoria. Implementa patrones de diseño estructurales como bloques convolucionales de profundidad separable (depthwise-separable), convoluciones de profundidad con cuello de botella lineal y bloques de transformer con pesos vinculados para reducir la latencia de inferencia.
Features
- Vision Backbones and Classification - Provides high-efficiency computer vision backbones for image recognition and feature extraction.
- Depthwise Separable Convolutions - Provides depthwise-separable convolutional blocks that split standard convolutions into spatial depthwise and pointwise channel mixing filters.
- Cheap-Operation-Based Convolutions - Implements convolutional layers that use linear transformations to generate redundant feature maps and reduce floating point operations.
- Inference Cost Optimizations - Employs pre-defined neural network architectures that minimize computational costs on resource-limited hardware.
- Linear Bottlenecks - Implements linear-bottleneck depthwise convolutions to generate more feature maps with fewer parameters.
- Edge AI Model Deployment - Provides lightweight neural network architectures optimized for efficient deployment on local hardware and edge devices.
- Resource-Constrained Optimizations - Structures neural network backbones to minimize memory access and compute cycles for low-power edge hardware deployment.
- Inference Latency Reduction - Provides optimized model structures that reduce memory usage and computation time to lower inference latency.
- Inference Speed Optimizers - Optimizes the execution speed and memory usage of neural network inference for faster live predictions.
- Vision Transformer Implementations - Implements Vision Transformer architectures optimized for reduced parameter counts and high inference efficiency.
- Efficient Neural Architectures - Offers a set of lightweight neural network backbones designed to reduce computational and memory overhead.
- Efficient Vision Transformers - Implements efficient vision transformers with optimized attention mechanisms to reduce resource requirements.
- Mobile AI Architectures - Develops AI backbones optimized for the strict energy and memory limits of mobile device hardware.
- MLP-Mixer Architectures - Utilizes MLP-Mixer architectures to process visual data using multi-layer perceptrons for high-performance feature extraction.
- Neural Network Design Frameworks - Provides a collection of optimized layers and structural design patterns to minimize parameter counts in neural networks.
- Weight Tying Strategies - Implements tied-weight transformer blocks that share parameters across layers to decrease total model size and memory footprint.
- Neural Network Architectures - Lightweight architecture using cheap operations for feature generation.
