AdaDeep:A DNN hyperparameter optimization framework

Adadeep integrates the adaptive model compression problem with a DNN hyperparameter optimization framework, using reinforcement learning to achieve adaptive lightweight model architecture search.

Model Architecture Diagram: your_image

Description: To address these challenges, we propose AdaDeep, an automated DNN compression framework that explores the trade-off between performance and resource constraints at a holistic system level. AdaDeep selects the most suitable combination of compression techniques and corresponding hyperparameters for a given DNN, layer-by-layer. Additionally, AdaDeep uncovers novel combinations of compression techniques. Evaluations show that AdaDeep achieves significant improvements in latency, energy efficiency, and storage blueuction in DNNs, while incurring negligible accuracy loss.

Source Code