quark0/darts
Darts
Darts es un framework y biblioteca de búsqueda de arquitectura diferenciable diseñado para automatizar el descubrimiento de estructuras óptimas de redes neuronales convolucionales y recurrentes. Sirve como una herramienta de investigación para encontrar topologías de celdas de alto rendimiento utilizando optimización basada en gradientes.
El framework emplea una super-red de celdas diferenciable y mecanismos de peso compartido para identificar la conectividad de red efectiva. Utiliza aproximación de segundo orden para estimar el rendimiento de candidatos arquitectónicos discretos y convierte los pesos continuos aprendidos en estructuras de grafos discretas mediante el mapeo de genotipo a modelo.
La biblioteca cubre la optimización de deep learning para capas convolucionales y recurrentes, incluyendo herramientas para entrenar y evaluar genotipos descubiertos frente a benchmarks estándar. También proporciona utilidades para cargar pesos preentrenados para inferencia y generar visualizaciones gráficas de la conectividad de celdas neuronales.
Features
- Neural Architecture Search - Provides a framework for automatically discovering optimal neural network architectures using gradient-based optimization.
- Neural Topology Optimizers - Finds high-performing cell topologies for convolutional and recurrent layers using second-order approximation.
- Differentiable Super-nets - Implements a differentiable cell super-net to identify effective network connectivity through parallel operation training.
- Second-Order Optimizers - Uses second-order approximation and Hessian-based information to estimate the performance of architectural candidates.
- Weight Sharing Strategies - Employs weight-sharing mechanisms where a single set of weights is used across all candidate operations to accelerate discovery.
- Architecture Search Frameworks - Provides a differentiable framework for automating the discovery and optimization of neural model architectures.
- Convolutional Neural Networks - Automates the discovery of cell structures specifically for convolutional neural networks.
- Neural Cell Primitives - Provides a set of basic convolutional and recurrent operations to serve as the search space for architecture discovery.
- Deep Learning Research - Serves as a research tool for training, evaluating, and visualizing discovered neural network genotypes.
- Genotype-to-Model Mappings - Converts learned continuous weights into discrete graph structures to instantiate full-sized neural networks.
- Deep Learning Optimization - Tunes the layers and connectivity of neural networks to optimize performance on specific datasets.
- Model Performance Benchmarking - Evaluates discovered neural genotypes against standard benchmarks to verify final accuracy and error rates.
- Recurrent Neural Networks - Searches for effective connectivity patterns in recurrent neural networks to optimize sequence processing.
- Image Classification Architectures - Differentiable architecture search framework.
- Model Optimization and Search - Differentiable architecture search for efficient neural network design.
- Neural Architecture Search - Differentiable architecture search for neural networks.
- Neural Network Architectures - Differentiable architecture search.
- Architecture Search - Listed in the “Architecture Search” section of the The Incredible Pytorch awesome list.
