search

titu1994 / Neural-Style-Transfer

Keras Implementation of Neural Style Transfer from the paper "A Neural Algorithm of Artistic Style" (http://arxiv.org/abs/1508.06576) in Keras 2.0+
Apache License 2.0
2.28k stars 482 forks source link

Not able to reproduce the results #50

Closed tj198505 closed 6 years ago

tj198505 commented 6 years ago

Hi,

With all default settings in the jupyter notebook, the generated results does not look any similar to the results you posted. I did not use google.colab to import content and style images, instead, the image are directly set by CONTENT_IMAGE_FN = "../deepArt.ai/test/blue-moon-lake.jpg" STYLE_IMAGE_FN = "../deepArt.ai/test/starry_night.jpg"

gen_at_iteration_10

titu1994 commented 6 years ago

You need to run it for 100 epochs. It's quite slow so I set the default as 10 epochs.

tj198505 commented 6 years ago

Here is the result with 100 epochs gen_at_iteration_100

titu1994 commented 6 years ago

Hmm interesting. I can't recall the set of combinations that I ran to get the result, but it definitely wasn't the default settings I've posted here. I do remember epochs being 100, and changing a few other parameters.

The defaults are good baselines, but each image generally needs specific settings of parameters to get great results.

tj198505 commented 6 years ago

Which parameters you have change to achieve better results? What is your strategy to change them?

titu1994 commented 6 years ago

Well, I used Inetwork, and used more style weights. It's a trial and error process, and that's from nearly two years ago so I don't remember.

tj198505 commented 6 years ago

Should this issue be 'open' then? Someone else may have the same issue.

titu1994 commented 6 years ago

You could in essence try variations of the settings on Colab and see which setting gets you the result.

This issue is closed because it is very challenging to get the exact setting which gets the best output. Sadly, I have a folder of over 300+ images which were failed attempts of just 1 single image.