A Case for Generalizable DNN Cost Models for Mobile Devices

Abstract

Accurate workload characterization of Deep Neural Networks (DNNs) is challenged by both network and hardware diversity. Networks are being designed with newer motifs such as depthwise separable convolutions, bottleneck layers, etc., which have widely varying performance characteristics. Further, the adoption of Neural Architecture Search (NAS) is creating a Cambrian explosion of networks, greatly expanding the space of networks that must be modeled. On the hardware front, myriad accelerators are being built for DNNs, while compiler improvements are enabling more efficient execution of DNNs on a wide range of CPUs and GPUs. Clearly, characterizing each DNN on each hardware system is infeasible. We thus need cost models to estimate performance that generalize across both devices and networks. In this work, we address this challenge by building a cost model of DNNs on mobile devices. The modeling and evaluation are based on latency measurements of 118 networks on 105 mobile System-on-Chips (SoCs). As a key contribution, we propose that a hardware platform can be represented by its measured latencies on a judiciously chosen, small set of networks, which we call the signature set. We also design a machine learning model that takes as inputs (i) the target hardware representation (measured latencies of the signature set on the hardware) and (ii) a representation of the structure of the DNN to be evaluated, and predicts the latency of the DNN on the target hardware. We propose and evaluate different algorithms to select the signature set. Our results show that by carefully choosing the signature set, the network representation, and the machine learning algorithm, we can train accurate cost models that generalize well. We demonstrate the value of such a cost model in a collaborative workload characterization setup, wherein every mobile device contributes a small set of latency measurements to a centralized repository. With even a small number of measurements per new device, we show that the proposed cost model matches the accuracy of device-specific models trained on an order-of-magnitude larger number of measurements. The entire codebase is released at https://github.com/iitm-sysdl/Generalizable-DNN-cost-models.