ZFTurbo/Keras-inference-time-optimizer

non-standard layer

yyjabidintg opened this issue · 5 comments

yyjabidintg commented

what to do if my code has non-standard layer.
Any suggestion on how to modify the code?
Thanks a lot.

ZFTurbo commented

You can add it in function "get_copy_of_layer"
https://github.com/ZFTurbo/Keras-inference-time-optimizer/blob/master/kito.py#L72

There are 2 examples already added for "relu6" and "BilinearUpsampling" layers.

yyjabidintg commented

Thanks.
I find that you have ever used keras_retinanet.
But when I am trying to combine the Kito code into Keras_retinanet.
The layer regression_submodel output errors:
x = new_layer(prev_layer)
ValueError: Invalid input_shape argument [(None, None, None, 256), (None, None, None, 256), (None, None, None, 256), (None, None, None, 256), (None, None, None, 256)]: model has 1 tensor inputs.
But I having
Can you give any suggestions on which part of the code to moidfy?
Thank in advance.

ZFTurbo commented

Yes I used RetinaNet but wasn't able to create KITO code that supports it, because it has too complex structure. I still have plans to do it. When it'll be ready I'll post note in this thread.

ZFTurbo commented

I was able to optimze RetinaNet, but for some reason it works absolutely at the same speed as unoptimized version.

bfialkoff commented

Can you share how you were able to do it, I'm getting the same error mentioned above?