SEEKING FEEDBACK: How should we structure TFLite samples

[Khanh LeViet]

Hi everyone,

tl;dr Please help us improve TFLite samples by answering this survey on how the apps should be structured.

We from the TFLite team would like to hear the community's opinion on how we should structure our sample apps to make it easier for TFLite users to learn how to use TFLite on different ML use cases.

There are currently several ways to run inference using a TFLite model:

• TFLite Task Library: easy to use and less code.
• TFLite Interpreter + TFLite Support Library: more complex but more flexible to customize.

First of all, we'd like to hear your opinion on whether you want to see both approaches in our samples, or just seeing the "easy" TFLite Task Library is enough.
If you prefer seeing both options, please let us know how the app should be structured to make it easier for TFLite users to understand and reuse components in the app.

We have prepared multiple implementation options that demonstrate both "Task Library" and "Interpreter + Support Library" inference path, and we'd like to hear your opinion on which style you prefer:

• Option 1: Task Library as the default option. 
• The app is optimized to use Task Library.
• Code: https://github.com/khanhlvg/TFLiteDemo/tree/master/bert_qa
  
• Task Library is used directly in the activity class to demonstrate the easiest way to integrate a TFLite model into the app.
• Interpreter + Support Lib is implemented as a separated Android module and can be enabled by changing a flag in the app code.
• Option 2: Task Library and "Interpreter + Support Lib" are each implemented in its own module. 
• The app is optimized to share UI code between the two modules without having a preference on any of the inference paths.
• Code: https://github.com/tensorflow/examples/tree/master/lite/examples/image_classification/android
• There's a shared interface for the Task Library path and Support Lib + Interpreter path. 
• You can switch between the two by changing the Android's product flavor.

Please let us know your opinion through this survey, and help us make the TFLite samples more useful to TFLite users.

Thank you very much in advance!

Khanh

--

    

Le Viet Gia Khanh (カン)

TensorFlow Developer Advocate

mail: khan...@google.com

tel: +81 3-6384-6867 

Shibuya Stream 

3-21-3 Shibuya, Shibuya-ku, Tokyo

150-0002, Japan