Generalized Latency Performance Estimation for Once-For-All Neural Architecture Search
This work provides a more efficient and generalizable method for estimating neural network latency, reducing the significant manual effort and time required for building hardware-specific lookup tables for researchers and developers working on multi-platform deep learning deployment.
The paper addresses the lack of generalizability in Once-For-All (OFA) neural architecture search's latency estimation across diverse hardware platforms. It proposes a framework for building latency predictors using fine-tuning and GPU-generalization strategies, achieving over 50% lower RMSE loss compared to ProxylessNAS.
Neural Architecture Search (NAS) has enabled the possibility of automated machine learning by streamlining the manual development of deep neural network architectures defining a search space, search strategy, and performance estimation strategy. To solve the need for multi-platform deployment of Convolutional Neural Network (CNN) models, Once-For-All (OFA) proposed to decouple Training and Search to deliver a one-shot model of sub-networks that are constrained to various accuracy-latency tradeoffs. We find that the performance estimation strategy for OFA's search severely lacks generalizability of different hardware deployment platforms due to single hardware latency lookup tables that require significant amount of time and manual effort to build beforehand. In this work, we demonstrate the framework for building latency predictors for neural network architectures to address the need for heterogeneous hardware support and reduce the overhead of lookup tables altogether. We introduce two generalizability strategies which include fine-tuning using a base model trained on a specific hardware and NAS search space, and GPU-generalization which trains a model on GPU hardware parameters such as Number of Cores, RAM Size, and Memory Bandwidth. With this, we provide a family of latency prediction models that achieve over 50% lower RMSE loss as compared to with ProxylessNAS. We also show that the use of these latency predictors match the NAS performance of the lookup table baseline approach if not exceeding it in certain cases.