chainer-fast-neuralstyle
chainer-fast-neuralstyle copied to clipboard
yusuketomoto
How could I get rid of the noise in the generated image?
No matter what style image I used, There are always some noise spots in the generated image. Anyone have the same issue????
@bucktoothsir WoW! pretty good result. the antifact may be removed by longer epochs
@bucktoothsir would u like share your training params? I really like this style!
@austingg sure. Firstly I used https://www.google.com/search?q=picabia+udnie&biw=960&bih=497&espv=2&tbm=isch&imgil=t4OUXK-A7MJhnM%253A%253BUiglBbl7HofR0M%253Bhttps%25253A%25252F%25252Fen.wikipedia.org%25252Fwiki%25252FFile%25253AFrancis_Picabia%25252C_1913%25252C_Udnie_(Young_American_Girl%25252C_The_Dance)%25252C_oil_on_canvas%25252C_290_x_300_cm%25252C_Mus%25252525C3%25252525A9e_National_d%25252525E2%2525252580%2525252599Art_Moderne%25252C_Centre_Georges_Pompidou%25252C_Paris..jpg&source=iu&pf=m&fir=t4OUXK-A7MJhnM%253A%252CUiglBbl7HofR0M%252C_&usg=__sJBDPFLB4UtviRImms__gsnP234%3D&dpr=1.5&ved=0ahUKEwiP9aK03JrOAhVG6GMKHX5uA7gQyjcINQ&ei=3V6cV8-SJMbQjwP-3I3ACw#imgrc=t4OUXK-A7MJhnM%3A as the style. Then I set the image size as 512.
@bucktoothsir would you please consider sharing your model? maybe here or on https://github.com/gafr/chainer-fast-neuralstyle-models
@bucktoothsir train it a bit further. Not necessarily a full additional epoch, but some 20-40k images might help. Make checkpoints every few thousand to see the progress. I noticed these artifacts come and go along the training.
@6o6o Is that means the training processing is not stable? maybe need some gradient clip
@benslaney sure. I would like to share it when I get a good model but not a semi-finished one.
@bucktoothsir could you fix the problem with the error? what params did you use with train.py? I have the same problem :(
@bucktoothsir hey :) try to reduce lambda_tv, like put it to 1e-5. Also did you preserve all other parameters as defaults ?
