Motivation for tflite model maker

[Ben B]

Dear colleagues/friends

I was having a look at TFLite Model Maker and don't understand the motivation for it. It looks to be a subset of Tensorflow, and might not add any extra features (e.g. on device training)?

I guess its main reason to avoid dealing with a larger complex API like tensorflow?

https://www.tensorflow.org/lite/guide/model_maker

Thanks

[Margaret Maynard-Reid]

Hi Ben,

Model maker makes model training easy - train a model with just a few lines of code with input data from folders of data, TFDS or Kaggle. No need to worry about data preprocessing, how model training or transfer learning works or post processing etc.

Model maker also outputs tflite models with metadata which enables features such as ML Model Binding in Android Studio - inputs a tflite model easily into Android project, and auto-generated code for loading tflite model and run inference etc.

I wrote a two part tutorials earlier last year:

• Icon Classifier with Model Maker
• Implement the Classifier in an Android App

So the purpose of model maker as I see it, is to help app devs with very little ML / TensorFlow experience. 

Hope this helps.

Margaret

--
You received this message because you are subscribed to the Google Groups "TensorFlow Lite" group.
To unsubscribe from this group and stop receiving emails from it, send an email to tflite+un...@tensorflow.org.
To view this discussion on the web visit https://groups.google.com/a/tensorflow.org/d/msgid/tflite/c1e5b2a5-e582-4fa4-bacf-9760546b32c5n%40tensorflow.org.

[Ben B]

Thanks Margaret, that’s great, ill take a look

Just a thought about model binding: I used model binding a few months ago, and found that debugging was impossible with it, since you can’t read or put breakpoints on the codegen code (its hidden). Apart from that, it was quite ergonomic. Im not sure if that has been improved on. I prefer the support/task library and tflite code generator now, depending on the task.