Because the original tf.pow function cannot be quatilized. So I want to adjust the implementation method of the pow function so that tflite can support the pow function.
The original method of Pow function is
Po = Pi ^ gamma
changed to
ln (Po) = gamma * ln (Pi)
so,
Po = exp ^ ( gamma * ln (Pi) )
In this way, the power function can be replaced by exp function and ln function. Because Taylor expansion only needs multiplication and addition to calculate exp function and ln function, tflite provides multiplication and addition.
I just need the range 0~1. The range of Taylor expansion of ln function (at x=1) is as follows:
I need all ranges less than 0. So I tried the exponential function of taylor expenditure in the subsection, as follows:
I also tried linear approximation of exponential function.
I still can't speed up the model with both methods.
Can anyone help me?
quantized tflite support function list: https://github.com/tensorflow/tensorflow/tree/f9bdcd6d9c714fc5dde232ba3b8b0a5128a07516/tensorflow/lite/delegates/hexagon