Compute Aware Multi-Objective Neural Network Compression Overview:This project implements a multi-objective optimization framework for deep neural network compression using weight sharing and the NSGA-II evolutionary algorithm.
The goal is to simultaneously optimize:Classification Accuracy, Model Size (Parameters),Computational Cost (FLOPs)
The framework identifies Pareto-optimal solutions, enabling flexible trade-offs between performance and efficiency.
Key Features are Multi-objective optimization using NSGA-II, FLOPs-aware compression (beyond memory-only methods) Supports multiple architectures: ResNet-18, MobileNetV2, EfficientNet-B0 Weight-sharing-based compression,Pareto front visualization,Model export and evaluation pipeline Methodology: Train baseline models on CIFAR-10, Apply weight-sharing compression, Use NSGA-II to optimize: Accuracy Parameter count FLOPs Extract Pareto-optimal models Analyze trade-offs between efficiency and performance 📊 Results Summary 🔹 ResNet-18 Accuracy: 94.46% → 93.18% Params: -12.0% FLOPs: -14.5% Model Size: -93.9% 🔹 EfficientNet-B0 Accuracy: 89.46% → 89.07% Params: -1.2% FLOPs: -19.2% Model Size: -97.1% 🔹 MobileNetV2 Accuracy: 90.53% → 84.91% Params: +226.3% FLOPs: +1310.4% Model Size: -88.7%