This is a runnning record of our issues making the android app for our AI assistant. This record begins Dec 6th 2016. At this stage we have a TF model that was installed in various android devices and google tablets. It works in Android 7 and above. The demo can be seen in this video

We are currently working with a development team to integrate the front end android interface with TF. Follow our progress.

Dec 6th 2017

Can't install the apk file in your phone? It may be because you're running a screen overlay app in your kernel that won't allow you to install the app. A screen overlay app adjusts the light on your screen. I like to use a screen overlay to reduce the blue light on my phone. If you want to install an apk file in your phone though, you need to turn off the screen overlay app before installing. Then it should install perfectly! You can turn your screen overlay back on after the installation and all apps should work normally.

-Amanda Ramcharan and Pete McCloskey

Dec 7th 2017

Exciting news! We now have the Tensorflow classification model integrated into the PV app! Here's a (rough) look at it operating. I put up the model parameters to see how it's working. In the bottom left corner, it says the inference time; here it says 677ms, that's less than one second to make a prediction, in the app, with no internet connection. WOW :o

- Amanda Ramcharan

Dec 8th 2017

After much debugging (Tensorflow needs input data to be in a particular format that was  tricky to build) we have the object detection model running! Here's an example of what an object detection model does from Pete McCLoskey's phone:

 

The model looks within an image and identifies different object and provides a measure of accuracy (0 being low and 1 being high) as to how confident it is about what it is "seeing".

These screens below make me happy. We're doing transfer learning again but this time to detect fall army worm damage on maize in pictures collected across the African continent. Looking forward to testing this model out in the Andriod app!

-Amanda Ramcharan

 

 

Dec 11th 2017

We're getting fancy with our cassava AI detection. We now have a meter to show how confident the model is in predicting a disease and we provide demo images for the user to confirm that the app made a correct diagnosis. We would like to use this last feature to teach users about disease phenotype.

 

 

-Amanda Ramcharan

Dec 12th 2017 

As we build this user interface a daily concern is who is the end user and what does that person need? We aim to have farmer trials in early 2018 in Tanzania and are lucky to have many experienced scientists from the International Institute of Tropical Agriculture (Tanzania) on our team. They have a lot of experience with farmers. But ultimately it comes down to what the farmers value and creating value for them 

-David Hughes  

Object Detection of Fall Army Worm is now working. We trained a model using transfer learning with a mere 12 images and here are some preliminary results from the TF Detect app trained to detect leaf damage on maize leaves. Exciting!

 

-Amanda Ramcharan and Pete McCloskey