MODI is a Cake Lab project to leverage dynamic inference location selection to meet SLAs and increase accuracy whenever possible. This benchmark application was developed to support the research presented in MODI: Mobile Deep Inference Made Efficient by Edge Computing.
Deep Learning Inference is increasing in popularity for mobile applications, allowing them to leverage new features like speech-to-text and image classification. In order to better explore this problem space this benchmark application was developed to test variations on local and remote inference.
Included are an Android application (in MODI_android) and a simple inference server (in Server).
MODI aims to improve the performance of deep learning inference for mobile devices. It accomplishes this by providing a number of models to dynamically select from at runtime, storing a wider selection of models at edge servers and managing a central repository of models that can be pushed to edge servers and then to mobile users as needed. In the MODI paper we introduce our vision for this system. This application was used for benchmarking on-device and cloud-based performance.
Mobile Network Latency and Variation Mobile Devices experience large variation between network types and even between different requests using the same network. This variation makes cloud-based inference potentially more unpredictable.
Communication from mobile devices to the cloud can be difficult due to the wide variety of networks that mobile devices use and the variation they experience. Therefore, in order to provide good user experience, it is important to make run-time decisions for inference location and the model used for inference.
On-device Inference Latency On-device inference latency depends heavily on the phone capabilities. Newer phones can perform better but many older devices experience large latency.
On-device inference is not a panacea solution as on-device inference can take unacceptably long times on mobile devices, as is shown above. Furthermore, on some devices on-device inference can have high variability. Much of this latency is due to model load time, so improving on-device management can improve overall performance greatly, but not entirely resolve the problems.
MODI System Design The proposed layout of the MODI system leverages three tiers of model locations and execution options to provide and pick the best model for a given inference.
The MODI system is designed to enable mobile devices to make the decision to run inference locally or remotely. It makes use of a tiered architecture to enable the most relevant models to be pushed to the edges and supplied to the mobile devices and models statistics to be filtered upwards to be centrally collated to provide the best model runtime decisions.
There are a number of test images included in MODI_android/app/src/main/assets/test_images.
These images are organized such that images are in a subfolder named after their ground-truth class.
Images are a subset of a larger set that consists of images from the TensorFlow For Poets project and other images in the same categories.
To use local inference models need to be added to the assets folder located at MODI_android/app/src/main/assets/models.
These models should be in TensorFlow's Frozen ProtoBuf file format, which are available on the TensorFlow website.
Additionally, the LocalClassifier.java file should be updated to reflect the models being used.
To run the remote inference run a server that presents a POST API that accepts images as inputs.
An example server has been included in Sample_Server/FlaskServer.py
This server runs using python3 and relevant requirements can be found in Sample_Server/requirements.txt
Local inference requires Android API 23 or above.
Other projects can be found at The Cake Lab's Project Site.
We ask that academic works which use this code reference the AdaptSize paper:
@inproceedings {216771,
author = {Samuel S. Ogden and Tian Guo},
title = {{MODI}: Mobile Deep Inference Made Efficient by Edge Computing},
booktitle = {{USENIX} Workshop on Hot Topics in Edge Computing (HotEdge 18)},
year = {2018},
address = {Boston, MA},
url = {https://www.usenix.org/conference/hotedge18/presentation/ogden},
publisher = {{USENIX} Association},
month = jul,
}
More publication details can be found on the conference page.
- Samuel S. Ogden ssogden@wpi.edu
This work is supported in part by NFS Grant CNS-1755659.