search

6o6o / chainer-fast-neuralstyle

Chainer implementation of "Perceptual Losses for Real-Time Style Transfer and Super-Resolution".
28 stars 5 forks source link

Speed compared to fast-neuralstyle + instance normalization #3

Open crypdick opened 7 years ago

crypdick commented 7 years ago

Thanks for the repo and all your helpful comments upstream. I was wondering if you've done any speed comparison between your code and jcjohnson's implementation with instance normalization enabled?

6o6o commented 7 years ago

I never really got around to try out his repo. Should be a bit faster cause torch is more optimized in gpu handling. Some say that inst-norm is faster as well. I cannot confirm any of this as I'm currently on a different machine and do not have a chance to use it.

Divetoxx commented 7 years ago

I have tried, but only with CPU. And I can tell that jcjohnson's implementation with instance normalization is a lot slower than this code! For one picture 1350 x 1080 jcjohnson's implementation took 4 minutes on my CPU instead of 20 sec! 20 seconds is needed for this code. Maybe with GPU all look different, maybe...