Question about merge your normalization ops and postprocess ops

[sugartom]

Firstly, thanks a lot for your tutorial and code. It works perfectly!

And when reading your code, I found your comment in models/yolo.py:

We can merge normalization with other OPs, but we need to redefine input tensor for this.

So I would like to know whether you have any idea on how to do that?

I am now trying to export your version of YOLO from TF to TF-Serving, and if I chose self._raw_inp as input to export, then it will report Invalid argument: You must feed a value for placeholder tensor 'evaluation/input' with dtype float and shape [1,608,608,3] error. So instead, I exported self._eval_inp, and re-wrote the normalization part in numpy. It worked, but as you mentioned, it will be faster if we use tf's API to do the normalization. So I wonder whether you have any idea on how to merge normalization and postprocess in your _evaluate() in models/yolo.py.

Thanks! :-)