Abstract: Deep neural networks are widely used in various domains, but the prohibitive computational complexity prevents their deployment on mobile devices. Numerous model compression algorithms have been proposed, however, it is often difficult and time-consuming to choose proper hyper-parameters to obtain an efficient compressed model. In this paper, we propose an automated framework for model compression and acceleration, namely PocketFlow. This is an easy-to-use toolkit that integrates a series of model compression algorithms and embeds a hyper-parameter optimization module to automatically search for the optimal combination of hyper-parameters. Furthermore, the compressed model can be converted into the TensorFlow Lite format and easily deployed on mobile devices to speed-up the inference. PocketFlow is now open-source and publicly available at https://github.com/Tencent/PocketFlow.
TL;DR: We propose PocketFlow, an automated framework for model compression and acceleration, to facilitate deep learning models' deployment on mobile devices.
Keywords: model compression, hyper-parameter optimization, mobile devices