Reproduce “AMC: AutoML for Model Compression and Acceleration on Mobile Devices” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yihui He, Ji Lin, Zhijian Liu, Hanrui Wang, Li-Jia Li, Song Han
AMC automates model compression using a reinforcement learning framework, outperforming hand-crafted methods by achieving higher compression ratios and faster inference on mobile devices with better accuracy preservation. The framework directly optimizes for hardware-aware metrics like latency.
Key insights
- Framing model compression as a sequential decision-making process with a continuous action space allows for discovering highly optimized, layer-wise compression policies that surpass human-designed heuristics.
- The strong correlation between validation accuracy before and after fine-tuning provides an efficient and effective reward signal, enabling rapid exploration of the vast compression policy space within practical timeframes.
- Directly optimizing for actual hardware metrics (like latency) during the RL search, rather than just theoretical FLOPs, ensures the learned policies yield tangible acceleration benefits on target mobile devices.
- The learned compression policies can discover non-intuitive sparsity patterns (e.g., 'sawtooth' patterns or differential pruning of 3x3 vs. 1x1 convolutions) that lead to superior accuracy-resource trade-offs.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.