Conversion of TFLite model precision and original model files.

[Criminal-9527]

I have downloaded the hand_landmarks_detector.tflite model, but this is a TensorFlow float16 model. I hope to obtain the original float32 model, or directly get an int8 model is also possible. Is there a way to directly convert it, or can I get the original .pb file of this model? The official documentation of hand_landmark only has .pbtxt files, and they cannot be directly converted into .pb files with TensorFlow's API. Are there any MediaPipe-specific APIs that can convert it?

[ayushgdev]

Hello @Criminal-9527

Unfortunately, we do not have any float32 model at our disposal. Regarding the APIs for conversion of .pbtxt files, we are checking with the team for the possibility.

[Criminal-9527]

Are there any original model files available for use? We ultimately want to get an ONNX format model, but the tools I currently know for converting from TFLite format to ONNX format (such as tflite2onnx) do not support the conversion of float16. The warning is as follows: Data type float16 not supported/tested yet, the generated model may contain errors. Although an ONNX format model will eventually be converted, it is basically unrecognizable. @ayushgdev