Artifact for MLPerf Mobile Inference Benchmark for MLSys 2022

Artifact for MLPerf Mobile Inference Benchmark for MLSys 2022

Please consult the full paper for the full author list.

The MLPerf Mobile App artifact is an Android app which provides latency and accuracy benchmarking capabilities for four mobile-targeted neural network models, evaluated on three computer vision tasks and one natural language processing task. The app can be run standalone to evaluate performance ("performance mode") or optionally on full validation datasets to evaluate both accuracy and performance ("accuracy mode"). Additionally, smartphone vendors can provide optimized back-end implementations to take advantage of architecture-specific features to increase inference speed. The Artifact Evaluation Appendix in the full paper provides instructions for reproducing performance mode evaluations without a custom back-end.