MLPerf™ Mobile App

This project contains the MLPerf mobile app, an app-based implementation of MLPerf Inference tasks.

Table of Contents

• Overview
• Folder Structure
• Related Repositories
• Notes

Overview

The MLPerf app offers a simple mobile UI for executing MLPerf inference tasks and comparing results. The user can select a task, a supported reference model (float or quantized), and initiate both latency and accuracy validation for that task. As single-stream represents the most common inference execution on mobile devices, that is the default mode of inference measurement, with the results showing the 90%-ile latency and the task-specific accuracy metric result (e.g., top-1 accuracy for image classification).

Several important mobile-specific considerations are addressed in the app:

• Limited disk space - Certain datasets are quite large (multiple gigabytes), which makes an exhaustive evaluation difficult. By default, the app does not include the full dataset for validation. The client can optionally push part or all of the task validation datasets, depending on their use-case.
• Device variability - The number of CPU, GPU and DSP/NPU hardware permutations in the mobile ecosystem is quite large. To this end, the app affords the option to customize hardware execution, e.g., adjusting the number of threads for CPU inference, enabling GPU acceleration, or NN API acceleration (Android’s ML abstraction layer for accelerating inference).

The initial version of the app builds off of a lightweight, C++ task evaluation pipeline originally built for TensorFlow Lite. Most of the default MLPerf inference reference implementations are built in Python, which is generally incompatible with mobile deployment. This C++ evaluation pipeline has a minimal set of dependencies for pre-processing datasets and post-processing, is compatible with iOS and Android (as well as desktop platforms), and integrates with the standard MLPerf LoadGen library. While the initial version of the app uses TensorFlow Lite as the default inference engine, the plan is to support addition of alternative inference frameworks contributed by the broader MLPerf community.

Folder Structure

• datasets - Contains scripts to prepare test and calibration data used for accuracy evaluation and model quantization
• docs - contains documentation
• flutter - Contains the Flutter (iOS/Android/Windows) version of the app (for running the benchmarks on a certain device)
• react - Contains the React version of the app (for viewing the benchmark results on a website)
• mobile_back_apple - Apple (Core ML) backend for iOS
• mobile_back_pixel - Google Pixel backend for Android
• mobile_back_qti - QTI backend for Android
• mobile_back_samsung - Samsung backend for Android
• mobile_back_tflite - Combined TFLite / MediaTek backends for Android and TFLite backend for iOS
• tools - contains miscellaneous tools (e.g. formatting commands, code scanner)

Related Repositories

• inference: Reference implementations of MLPerf inference benchmarks.
• mobile_models: Storage for ML models and task definition file.
• mobile_open: MLPerf Mobile benchmarks.
• mobile_submission_3.1: MLPerf Mobile inference v3.1 results.
• mobile_datasets: Scripts to create performance-only test sets for MLPerf Mobile
• mobile_app_open/flutter/cpp/datasets: Instruction and scripts to create dataset for this MLPerf Mobile App.

Notes

The MLPerf™ Mobile App is now further developed as a Flutter app, which is maintained in the main branch master.

The branch android-v2 is used to maintain the legacy version of the MLPerf™ Mobile App, which is a native Android app.