deeplearning4j/deeplearning4j-examples

low FPS when running pretrained TinyYolo on Android

Open

#882 opened on Aug 27, 2019

View on GitHub
 (0 comments) (0 reactions) (0 assignees)Java (2,395 stars) (1,841 forks)batch import
help wantedquestion

Description

Issue Description

Hi, I am trying to run TinyYOLO object detection in real time on Android. For my inference code block, I am experiencing around 2800ms per inference, which is around 0.5FPS.

Few Methods tried:

  1. Increase android VM Heap size from 256MB to 512MB, memory during app running is floating around 300MB, performance not improved significantly.
  2. thought of using Threading for inference but the prediction will not be synced with the actual frame

Can you share some ideas on how to further improve the inference time?

The objdetection activity code: https://gist.github.com/yptheangel/f1e3c3dfd64c470d151890be00465c7a

Gradle dependecies: https://gist.github.com/yptheangel/6af7ed1febd21b5b4d2339eb8ce985b5

Version Information

Phone: Huawei P9+ OS: Android OS 7. CPU Specs : arm64 Octa-core (4x2.5 GHz Cortex-A72 & 4x1.8 GHz Cortex-A53) Model used : TinyYOLOv2 from DL4J Model zoo (59MB)

Contributor guide

Tech stack
javaandroid
Domain
mobilemachine learning
Issue type
research
DifficultyEstimated implementation difficulty for a new contributor, from 1 for very small changes to 5 for expert-level work.
4
Estimated timeA rough time range for an experienced contributor to investigate, implement, test, and prepare a pull request.
3-5 days
Activity statusHow available the issue appears right now: fresh, active, stale, blocked, or waiting on maintainer input.
stale
ClarityHow clearly the issue explains the expected change, acceptance criteria, and next step.
mostly clear
Prerequisites
Basic Android developmentFamiliarity with Deeplearning4j
Newbie friendlinessA 1-100 score estimating how approachable this issue is for first-time contributors.
20
Research direction
The user is experiencing low FPS (0.5 FPS) with TinyYOLO on Android using DL4J. Investigate Android specific optimizations such as using native inference (C++ backend), reducing model size, or leveraging GPU acceleration (OpenCL or RenderScript). Review the provided inference code (gist) for potential bottlenecks like repeated memory allocations or inefficient pre/post processing. Consider using Android Profiler to identify CPU/GPU usage. Also check if the model can be quantized or pruned for mobile deployment.