Open zezic opened 8 years ago
Same question: I have prefect results on paid ostagram and really average results in custom render:
I got much better results with -init=image Instead of the default random setting On May 2, 2016 3:44 PM, "Andrew G. Kuznetsov" notifications@github.com wrote:
Same question: I have prefect results on paid ostagram and really average results in custom render:
[image: result] https://cloud.githubusercontent.com/assets/435258/14954632/a9fab14e-107c-11e6-924d-ec1d9c69e3b6.jpg
— You are receiving this because you commented. Reply to this email directly or view it on GitHub https://github.com/SergeyMorugin/ostagram/issues/2#issuecomment-216225597
@GabiGrin oh thank you! it's probably used in ostagram! I'm just get working my macbook gt650m with NIN-Imagenet model, and it's compatible to make 650px wide images with 384 gpu cores! but i still can't run cudd backend on mac(
@GabiGrin here are 5000 cycles 650px wide on NIN:
@cavinsmith Looks good 👯 I also couldn't get cuda running, seems like there's not enough gpu memory on my mac. However, with 650 cycles in CPU mode I get nice results in 25-30 minutes. Do you see any difference from the first 5/6 hundred cycles up to the 6k ones? I output a new file on each 50 cycles, and stop seeing major differences after 400-500 iterations.
@GabiGrin try NIN network - works on mac. I see difference maximum up to 2000 cycles
it looks like the init_params
are loaded from a secret config file. Shame.
@j2kun -init image
and all default works very nice. I also have good recipe for NIN network (works for up to 700px on 1gb ram on macbook)
I have found NIN to be awful with my tested settings, so yeah please let me know what settings you use for that. (Do you also have examples?)
@j2kun one of my NIN config is here in comments
Results are different with different styles, but I can show some interesting (my best results are NSFW, write me a letter to cav@cav.ru and I will send it to you):
1000 cycles
4600 cycles
1200 cycles
1900 cycles
Can you share neural_style configs used at ostagram? I can't get results with clean neural_style from repo similar to "wow" ostagram ouput. Are you using special model or just VGG_ILSVRC_19_layers?