Development Tools

-
In this blogpost I will cover the different tools I have used during the development process of this project.

General

The tool below is used in all implementations.

Github

This free hosting website was used for version control. It uses the git open-source VCS (Version Controll System) enabling a combination of both local and centralized version control. This allows for easy and efficient management of progress in development.

AIDE Implementation

AIDE or the AI Diagnosis Engine includes the pre-trained AI model. This model allows the diagnosis engine from the previous section to calculate the disease prediction rankings. It also handles the image pre-processing of incoming requests. The following components are involved in this process.

Anaconda Distribution

This tool is a python package manager and provides the ability to easily manage, test and deploy development environments.

Pycharm

This IDE made by the software company JetBrains, is used for Python development. It integrates nicely with Git version control and provides robust code completion and assistance.

Pykka

To implement this model in Python, a library called Pykka is used. It was inspired by the Scala framework “Akka”, which is a powerful implementation of the actor pattern. Pykka however, does not support as many features as Akka. It is not meant to be a python port of this framework but is capable enough for this implementation.



The actor model has a few basic rules in Pykka:
  •          An actor is an execution unit that executes concurrently with other actors.
  •         It does not share state with anybody else, but it can have its own state.
  •          It can only communicate with other actors by sending and receiving messages, if it has their address.
  •          When it receives a message, it may take actions like:

o   Altering its own state, e.g. so that it can react differently to a future message
o   Sending messages to other actors
o   Starting new actors
  •        It only processes one message at a time. Thus, a single actor does not give you any concurrency, and it does not need to use locks internally to protect its own state.

My implementation using this framework resulted in two important classes, which I will describe below.

v    Threading Actor

This implementation by the Pykka API is a class that provides the basic actor capabilities. This actor is not the same as the actor described in section 7.2. It refers to a class which, when initialized, is executed by its own thread i.e. a Python threading.Thread. Moreover, when it is asked for a response on a message as part of its implementation, it provides a “Future” object which is a thin wrapper around a Python queue.Queue. This object represents a value that will be available in the future. The value of this “Future” is calculated as soon as its “get” method is called. My two classes expanded on this base actor class to include the custom functionality necessary for this project.

I        Aid Actor

The first custom actor class represents the worker. Its goal is to process analysis requests parallel to other workers as they execute on their own threads. To do this, it includes the AI implementation explained in section 8.4. When it successfully identifies up to three diseases that match the same plant type as the request indicates. The diseases that are also part of the top 3 results in its analysis are returned encapsulated in a Future.

II       Actor Manager

The second custom actor focuses on creating and managing these aid actors. It distributes incoming requests for analysis among its initialized aid actors. When one of its aid actors returns a request result, it shapes these results into up to three classifications and places these on the database. The log related to that request is then updated by this actor, with references to up to three classifications. The user can compare these classifications on the PlantCare app with their own observation and pick the one that matches best as “chosen classification” for that log. However, if the aid actor returns no values at all, then the log is updated to trigger the retry system. The user can then choose to take a new picture and retry the analysis process.

Firebase Admin / Pyrebase

This is the Firebase library for server-side python applications. Just like the regular Firebase SDK described in 8.2.4, this library provides abstraction in python to easily implement communication with Firebase services. In this implementation, Firebase is used to send results back to the database and check for new analysis requests.
Pyrebase was first used to implement this functionality. However, during development this library was deprecated. It had a flaw that would cause this implementation to stop listening for new analysis request provided by the end-user after running for 30 minutes. This was caused by the lack of a re-authentication system as the services required the user, e.g. diagnosis engine, to send a refresh token after the 30-minute mark. Hence, the switch was made to the new official Firebase Admin SDK provided by Google. The new framework did not require many adjustments to be made and solved the problem.

Pytorch

One of those packages is a popular open-source deep learning platform for Python. one of the most popular among AI researchers/developers. It allows for the development, training and testing of AI models using dedicated graphical processing units (GPU) or central processing units (CPU). This framework was used to train a pre-existing model well known for image classification called AlexNet. This pre-trained model is then implemented into a class that is used in the diagnosis engine implementation.

Android Implementation

In the following paragraphs the tools used to develop the mobile app that is part of this solution, will be described in further detail.

Android Studio

This is the official IDE for Google’s Android operating system. It is built on JetBrains’ IntelliJ IDEA and was designed specifically for Android development. Just like PyCharm it integrates nicely with Git and provides robust code completion and assistance. On top of that, it enables android App simulation for testing.

Firebase SDK

Firebase is the chosen database, storage, authentication and remote configuration tool for this project. In android development, this is made possible using the Firebase Software development kit (SDK) for android. This kit provides abstraction to easily implement communication with firebase services.

v    FirebaseUI

FirebaseUI is an open source framework that provides another layer of abstraction above standard firebase sdk to make certain functionalities even easier to implement. For instance, an Android RecyclerView with special FirebaseRecyclerAdapter that makes loading firebase data into a RecyclerView much easier to implement.

v    Authentication

On top of fast and reliable database services, firebase offers authentication services. The authentication service can be used to securely log-in and register users as soon as they want to use the app. There are a few different forms of authentication, including email & password -, federated identity provider -, phone number -, anonymous and custom authentication. For this project, email and password authentication was chosen. This also stores the personal information of the user in a secure way and satisfies the security requirement.

v    Storage

Even more useful for my implementation is the “storage” service, firebase offers. This allows me to easily upload and download files that are stored in the Google cloud storage system. The syntax for this is almost identical to firebase database. In this project I use this service to pass images between architectural components and store them.

 Glide

This popular opensource android framework is a media management and image loading utility that handles media decoding, memory and disk caching, and resource pooling. It makes it easy to download and load images into Android ImageViews and is used in this implementation for that purpose.

 Image Picker

This library makes it easy to select images from the gallery and camera. It is used to pick images from local device storage or take new images using the device camera, that need to be uploaded and stored using firebase storage.

Comments

Post a Comment

Popular posts from this blog

AI Implementation

-
Image
To start with the AI implementation of this project, I researched some similar exisiting projects until I found an open-source project on GitHub  from Marko Arsenovic, that was comparing different results achieved by AI models using a plant disease dataset. This was an ideal starting point for my own project. A chart was also provided to clarify the results achieved by his project. I noticied that AlexNet achieved good results while not taking a lot of time to train as you can see in the chart below provided by Marko Arsenovic. This project used an open-source dataset provided by the R&D unit of Penn State University called PlantVillage, to train the neural networks in plant disease identification. This dataset includes 38 classifications each represented by more than 1000 pictures and spread over 14 plant types. In order to verify that our AI model is performing correctly it is important to keep some data for validation. Thus, 80% of plant disease images was used fo...
Read more

Future...

-
Recent developments in computer vision and artificial intelligence have made it possible to achieve classifications of complex patterns in images. These developments have made the successful implementation of this solution architecture achievable. My implementation does not represent a production-ready product as it is a PoC. However, it does prove that a product resulting from this solution is achievable. With this implementation it is safe to say that all technical requirements have been satisfied to an acceptable degree for this thesis. Future To achieve a production ready product a few improvements must be implemented.   Legal Compliance in general and specifically with regards to privacy/integrity (GDPR) and       Data Storage laws must be researched and implemented. 2.    A feature must be implemented in the AI implementation that allows for live training to support continuous improvement of the AI. This live training will lead to e...
Read more