Speed up inference

[copaah]

I'm back again. I have some pretty decent results on the Camvid dataset now thanks to your help. I have a question for you, that you might be able to anwer. I'm not able to reproduce the fast inference time. In the article they state that:

"For inference we merge batch normalization and dropout layers into the convolutional filters, to speed up all networks."

Do you know where I can find any related litterature on how to do so or perhabs you know how they do it?

[PavlosMelissinos]

I don't know, sorry. That sounds like a new layer that combines backend operations so you're more likely to find something like that as a tensorflow implementation, otherwise you could ask the original authors to post their predict network.

[copaah]

I should have checked the repo for the original implementation. Someone asked this exact question a couple of days ago :) Thanks for the reply nonetheless :)

[mrgloom]

Looks it was done like this: https://github.com/ZFTurbo/Keras-inference-time-optimizer