Skip to content

Latest commit

 

History

History
187 lines (145 loc) · 8.61 KB

README.md

File metadata and controls

187 lines (145 loc) · 8.61 KB
Raw

Macaque logo

Macaque

Easy Kotlin/Java SDK for everyone who wants to start with Machine Learning on mobile devices.

Table of Contents

General Info

The main motivation behind the Macaque is to facilitate embedding ML models into mobile applications. We hope that with our SDK programmers with no AI experience will be able to achieve that goal in no time. Our solution comprises of macaque library and two sample projects.

There exists also an iOS implementation of Macaque which can be found here.

Tools and Requirements

  • Android Studio 3.x
  • Phone or Emulator with Android 5.0 (API level 21) - tested on Android 9.0

Setup

From JCenter

  1. You need to have jcenter() entry in repositories block of your build.gradle file - either in project or in module (currently default for all new Android projects).
  2. Add implementation 'com.enigmapattern.macaque:macaque:0.5.0' to dependencies block of your module build.gradle file.

Building from source

  1. Clone the repository
  2. Open it as a project in Android Studio
  3. Build 'macaque' module (Gradle -> macaque -> Tasks -> build -> build)
  4. Copy macaque-release.aar library file to 'libs' folder of your application's project
  5. Set up your build.gradle file to load .aar libraries (i.e. in dependencies block there is entry like implementation fileTree(dir: 'libs', include: ['*.aar'])
  6. After syncing gradle library is ready to use and can be found in com.enigmapattern.macaque package

Features

  • Support of TensorFlow Lite

  • Support for TensorFlow

  • Support for PyTorch 1.x

  • Effortless embedding of selected detection and classification models into a mobile application

    Models tested and used in example apps:

    • SSD object detector LINK
    • Mobilenet classifier LINK

  • Embedding of custom models with small boilerplate code

  • Support for quantized and unquantized models

  • Support for detection models returning locations relative to anchor boxes

  • Support for PoseNet models

  • Support for models not based on images.

Design

Concept of Submodules

Each instance of Macaque contains three submodules, each dedicated for other type of actions. There are following categories of submodules:

  • Input Data Providers - their role is taking and storing data from the user and providing it to the Macaque in format expected by the used Predictor.
  • Predictors - they take data from Input Data Provider, send it to library handling the model for inference then takes and returns raw results to the Output Data Provider
  • Output Data Providers - they take results from the model, parse it and return in form expected by the user.

Submodules Algorithm

Submodules implementations provided with library

Input Data Providers

  • BitmapToUINT8DataProvider - Expects user to provide Bitmap object to currentImage property, and converts it to form accepted by quantized models operating on images.
  • BitmapToFloatDataProvider - Expects user to provide Bitmap object to currentImage property, and converts it to form accepted by non-quantized models operating on images.

Output Data Providers

  • Detector4outputsBoundingBoxesDataProvider - Operates on ByteBuffers, takes results from detection models with 4 following output tensors: Locations, Classes, Scores, Number of detections (please note that tensors have to be in that order but do not have to have the same names). Returns list of DetectionResult objects.
  • ClassifierDataProvider - Operates pm ByteBuffers, takes results from classification models with single output tensor where to each index, representing class, there is assigned confidence score. Returns list with single ClassificationResult object.

Predictors

  • TensorFlowLitePredictor - default predictor for mobile usage. It utilizes TensorFlow Lite Interpreter. Requires use of properly converted models. Currently not all TensorFlow models can be converted to TensorFlow Lite. Initialization of this predictor is straightforward - only path to the model needs to be provided. Optionally flags to use GPU or NNApi (Android Neural Networks Api) can be set. There are optional constructors if further customization is required - look at the documentation in the code for more information.
  • TensorFlowPredictor - It utilizes TensorFlow Inference Interface. It can run all TensorFlow models. Initializing it requires from user providing model path as well as names and shapes of input and output tensors. There is optional constructor if further customization is required - look at the documentation in the code for more information.

Typical usage

  1. Select and instantiate proper input and output data providers eg.:
val inputDataProvider = BitmapToUINT8DataProvider()
val outputDataProvider = Detector4outputsBoundingBoxesDataProvider()
  1. Instantiate predictor object (provide model path, references to input and output data providers and, optionally, one of flags useGPU or useNNApi):
val modelPath = File(modelsDir, "coco_ssd_mobilenet_v1_1.0_quant_2018_06_29.tflite")
val predictor = Predictor(modelPath.absolutePath, inputDataProvider, outputDataProvider, useGPU = true)
  1. Provide data to input data provider:
inputDataProvider.currentImage = someBitmap
  1. Run prediction:
predictor.predict()
  1. Get properly parsed results from results provider:
val resultsList = outputDataProvider.getResults()

Example applications

  • MacaquePreview - uses simple classifier (eg. this) and classifies photos either taken by device camera application or loaded from memory.

  • MacaqueLivePreview - uses detection model (eg. this) and detects various objects on live preview.

Please note that in both applications model is not provided with source code. It has to be downloaded separately (e.g. from links above) and side-loaded to the files/models folder within app directory on external memory (e.g. /mnt/sdcard/Android/data/com.enigmapattern.macaquelivepreview/files/models).

Adapting data providers for custom models

Customizing input data providers

Custom input data providers have to implement InputDataProvider interface. They have to implement way of storing data provided by the user - in our implementations it is public var currentImage: Bitmap field. Moreover they have to override two methods of implemented interface:

  • data(forShapes: IntArray): Array<out Any> - converts data provided by user into the Array of objects, usually ByteBuffers
  • close() - cleans all resources utilized by the input data provider

Customizing output data providers

User's output data providers have to extend OutputDataProvider class and override the following method.

  • getResults(): List<Any> - converts model's output into format expected by the application. Note that if a single result is returned, it has to be in the form of a single-element list.

By default OutputDataProvider stores results in HashMap<Int, ByteBuffer>. It can be customized to utilize arrays (e.g. HashMap<Int, FloatArray>). To do that you have to override three methods:

  • initializeOutputHashMap(outputTensorsTypeSizes: Array<Int>, outputTensorsShapes: Array<IntArray>) - for approprite initialization of the array
  • resetOutputsHashMap() - originally it rewinds buffer's "head" when it is to be read again, for custom implementation with arrays instead of buffers it should be left empty.
  • close() - cleans the array

Call for contribution

Our Macaque is new on the market. There is still a lot to be done before it is mature. But the base is there and it is also for you to take your part in its development! We encourage everyone to tell their thoughts, request changes or features (e.g. via GitHub issues system) and of course by improving our codebase with pull requests. Let’s make Macaque better together!