anishathalye / neural-style

Neural style in TensorFlow! 🎨
https://anishathalye.com/an-ai-that-can-mimic-any-artist/
GNU General Public License v3.0
5.54k stars 1.52k forks source link

The effect of different layers of CNN #110

Closed limitforever closed 6 years ago

limitforever commented 6 years ago

Hello,anishathalye! In your code,I found that you defined the content network structure as CONTENT_LAYERS = ('relu4_2', 'relu5_2'),while the original paper defined it as 'con4_2',and there are also other different desighed such as 'relu4_2'. So my question is : what are the differences among them? Is it better for CONTENT_LAYERS = ('relu4_2', 'relu5_2')? I'll be appreciated if you could reply.

anishathalye commented 6 years ago

Uh I did some not-super-scientific experimentation a while ago and I think the thing I chose produced visually nice results. Feel free to experiment with different choices there. If you do some principled experiments and find that a particular setting is consistently better, feel free to submit a pull request :)