Difference between using tensorflow-lite:0.0.0-nightly and tensorflow-lite in an Android App

Question

I am trying to debug an existing Android app that uses tensorflow-lite to detect objects. The app implements the tensorflow library like below :

implementation('org.tensorflow:tensorflow-lite:0.0.0-nightly') { changing = true }
implementation('org.tensorflow:tensorflow-lite-gpu:0.0.0-nightly') { changing = true }
implementation('org.tensorflow:tensorflow-lite-support:0.0.0-nightly') { changing = true }

But examples I have found online for object detection, have implemented tensorflow-lite in the following way :

implementation 'org.tensorflow:tensorflow-lite-task-vision:0.2.0'

My questions are:

• What is the difference between using the nightly snapshot and the "normal" library. From what I can gather online, the nightly build is an experimental branch and may contain bugs? I'm just confused about this because the existing app does not contain a reference to sonatype maven repository, which I understand was required to get the nightly builds to work in the app.

   allprojects { 
   mavenCentral  
    maven {       
      name 'ossrh-snapshot'
      url 'http://oss.sonatype.org/content/repositories/snapshots'
    }
  }
  

• My second question is what does this do line do : { changing = true } ?

PS: We are using our own custom trained model/tflite.

Answer 1

Changing, or snapshot versions are used when you need Gradle to get a new version of the dependency with the same name from time to time (once in 24 hours, unless specified explicitly otherwise).

I believe that whoever chose the nightly version of tensorflow, was wrong. As you say, this version may have bugs, and worse, these bugs will change overnight. Find some fixed version that you are comfortable with, study its changelog, and reset your implementation to refer to this version.