AMC: AutoML for Model Compression and Acceleration on Mobile Devices

About

Model compression is a critical technique to efficiently deploy neural network models on mobile devices which have limited computation resources and tight power budgets. Conventional model compression techniques rely on hand-crafted heuristics and rule-based policies that require domain experts to explore the large design space trading off among model size, speed, and accuracy, which is usually sub-optimal and time-consuming. In this paper, we propose AutoML for Model Compression (AMC) which leverage reinforcement learning to provide the model compression policy. This learning-based compression policy outperforms conventional rule-based compression policy by having higher compression ratio, better preserving the accuracy and freeing human labor. Under 4x FLOPs reduction, we achieved 2.7% better accuracy than the handcrafted model compression policy for VGG-16 on ImageNet. We applied this automated, push-the-button compression pipeline to MobileNet and achieved 1.81x speedup of measured inference latency on an Android phone and 1.43x speedup on the Titan XP GPU, with only 0.1% loss of ImageNet Top-1 accuracy.

Yihui He, Ji Lin, Zhijian Liu, Hanrui Wang, Li-Jia Li, Song Han · 2018

Related benchmarks

Task	Dataset	Result
Image Classification	CIFAR-10 (test)	Accuracy—	3,381
Image Classification	CIFAR-10 (test)	Accuracy91.9	906
Image Classification	CIFAR-10	Accuracy—	564
Image Classification	ImageNet	Top-1 Accuracy70.8	431
Image Classification	ImageNet (val)	Top-1 Accuracy70.5	76
Image Classification	ImageNet (val)	Top-1 Accuracy70.5	68
Image Classification	ImageNet	Top-1 Accuracy71.8	60
Image Classification	ImageNet (val)	T170.8	45
Image Classification	CIFAR-10	Accuracy (Pruned)91.9	33
Image Classification	CIFAR-10	Top-1 Accuracy (Base)94.48	32