Add --builtins_only Flag to CLI for generating TFLITE_BUILTINS-Only TensorFlow Lite Models

[JihwanEom]

Issue Type

Feature Request

OS

Linux

onnx2tf version number

1.26.0

onnx version number

1.16.1

onnxruntime version number

1.18.1

onnxsim (onnx_simplifier) version number

0.4.33

tensorflow version number

2.17.0

Download URL for ONNX

N/A

Parameter Replacement JSON

N/A

Description

Hello!

I have been using onnx2tf very effectively it's extremely useful.

After converting a model from ONNX to TensorFlow Lite, I use TensorFlow Lite via GMS(Google Mobile Services) to minimize the required additional size of the app. Specifically, I use:

• com.google.android.gms:play-services-tflite-java:16.1.0
• com.google.android.gms:play-services-tflite-support:16.1.0

However, GMS only supports models built with tf.lite.OpsSet.TFLITE_BUILTINS. To work around this, I manually comment out the parts that include tf.lite.OpsSet.SELECT_TF_OPS (including other code that selects TensorFlow ops) in the following lines: https://github.com/PINTO0309/onnx2tf/blob/main/onnx2tf/onnx2tf.py#L1425-L1428

It would be very useful if a flag like --builtins_only could be supported in the CLI to enable this functionality easily.

Have a great weekend!

[PINTO0309]

Any untrained model will do, just provide us with an ONNX file whose structure we can understand.

[JihwanEom]

Apologies for not providing a clear explanation earlier.
I would like to propose adding an option to onnx2tf that allows conversion to TensorFlow Lite using a --builtins-only flag.
This would enable users to convert models without selecting specific operations.
This is not a bug report, but rather a feature suggestion to offer more granular options for users.

In my case, to prevent an increase in app size, I intended to use the TensorFlow Lite built into Android (GMS). This requires converting the model using only built-in operations, which is why I am suggesting this feature.

Thank you for considering this proposal!

[PINTO0309]

Is it your intention that a model will be generated whether you specify tf.lite.OpsSet.SELECT_TF_OPS or not, but that it's okay if a conversion error occurs if you don't specify it? Or do you mean combining multiple conversion options is too tedious?

I'm not rejecting your suggestion, just that I don't really understand the benefit of adding this new conversion option. This might make sense if you want to abort on an error if your model contains non-built-in operations.

I don't know the specifications of TensorFlow Lite via GMS(Google Mobile Services), so please tell me in detail what kind of difference in behavior occurs between models generated with and without tf.lite.OpsSet.SELECT_TF_OPS. This will be useful information for other onnx2tf users.

I'm currently out and about, so I can't check what difference there is in the definition information of the .tflite file depending on whether or not the options are specified.