Lyken17/pytorch-OpCounter
Pytorch OpCounter
pytorch-OpCounter es una utilidad de perfilado para redes neuronales de PyTorch diseñada para cuantificar la eficiencia del modelo calculando operaciones de punto flotante y conteos de multiplicación-acumulación. Funciona como un analizador de complejidad para medir el costo computacional y la carga de trabajo teórica de diferentes arquitecturas de modelos.
La herramienta permite la definición de reglas de conteo de operaciones personalizadas para admitir módulos de terceros no cubiertos por defecto. Utiliza hooks de avance (forward hooks) para interceptar llamadas de módulos y recorrido recursivo del árbol de módulos para agregar operaciones a través de sub-módulos hijos.
El proyecto proporciona capacidades para auditoría de rendimiento y análisis de complejidad, transformando conteos de operaciones sin procesar en formatos de métricas escalados legibles por humanos.
Features
- Model Complexity Analysis - Quantifies the theoretical workload of a model to estimate its efficiency and resource requirements before deployment.
- Model Complexity Calculators - Calculates multiply-accumulate and floating point operations to quantify neural network computational cost.
- Deep Learning Optimization - Evaluates the operational overhead of different model architectures to optimize for speed and hardware constraints.
- Computational Cost Profilers - Measures the computational cost of PyTorch neural networks by calculating floating point operations and multiply-accumulate counts.
- FLOPs Counters - Measures floating point operations to evaluate the efficiency and performance of PyTorch model architectures.
- MACs and FLOPs Analyzers - Calculates the number of MACs and FLOPs to quantify the computational cost of PyTorch neural networks.
- Recursive Module Operations - Walks through the hierarchical structure of a neural network to aggregate total operations across all child sub-modules.
- Model Execution Hooks - Intercepts PyTorch module calls using forward hooks to record and accumulate operation counts during a single pass.
- Operation Count Profilers - Provides a profiling tool for counting multiply-accumulate operations and defining custom counting rules for PyTorch modules.
- Custom Operation Counting Rules - Defining specific operation counting logic for specialized PyTorch modules not supported by default profiling tools.
- Operational State Accumulators - Implements a global state to sum multiply-accumulate and floating point operations across disparate layers of the model.
- Complexity Counting Rules - Provides a mechanism to specify how to count operations for third-party modules not supported by default.
- Computational Complexity Mappings - Maps specific module types or function names to predefined mathematical formulas for calculating computational complexity.
- Herramientas de desarrollo - FLOPs counting utility.
- More to explore - Tool for counting MACs and FLOPs.
- PyTorch Utilities - Listed in the “PyTorch Utilities” section of the The Incredible Pytorch awesome list.
