Yolov4-Tiny EdgeTPU Missing Leaky Relu about pinto_model_zoo

[PINTO0309]

@benlemond
Try it if you have the time.

• EdgeTPU, YOLOv4-tiny, INT8
  https://drive.google.com/file/d/1NsJlF-MVxBw7JCw60K1uFLTAlwdv-C0p/view?usp=sharing

[PINTO0309]

I have converted the model in the following repositories, so please refer to the following for post-processing.
https://github.com/bubbliiiing/yolov4-tiny-keras.git

[PINTO0309]

@benlemond
I tried to generate the EdgeTPU model by replacing LeakyReLU with PReLU. Try it once.
https://github.com/PINTO0309/PINTO_model_zoo/tree/master/046_yolov4-tiny/20_Replace_LeakyReLU_to_PReLU

[benlemond]

@PINTO0309
main.ERROR - Only float32 and uint8 are supported currently, got -1695564548.Node number 5 (PRELU) failed to invoke.

[PINTO0309]

@benlemond

I found the source of the problem.
google-coral/edgetpu#187

It seems to be a bug when the TFLite converter converts ResizeNearestNeighbor.
half_pixel_centers should not be True, but will be forced to be converted to True.

I think it's a matter of the converter, so it's out of my hands at this point.

[gouliang1997]

I gave it a try and succeeded in getting these five outputs. But I don't know what to do with the output later.
I think I should get two outputs, one for detected_boxes and one for scores.

[gouliang1997]

I processed the output according to the process in the repo, but the x y w h value of the output of this model is between 40 and 60.(output_details[3][..., :4] and output_details[4][..., :4])
If I put these values into the sigmoid function, I will get the output all turn out to be 1 because these values are obviously too large.
I think I need to make some adjustments to the output of the model but I don't know how