gafr / chainer-fast-neuralstyle-models Goto Github PK

How would you like for me to contribute? I did Starry Night for the new code version. It has some improvements.

Hi guys, first of all thanks for the awesome work!

I'm new to all of this and I'm not sure exactly which parameters would yield the best results for each style. Also went through the comments, based on what I understood, also from @6o6o comment on the starry night model there's a lot of trial and error involved.

So I have this idea (I'm relatively good at distributing work across clusters, spawning machines on demand etc); if we could pick X number popular styles (from Prisma or similar) and someone with more experience could indicate possible ranges for each training parameter (say for example we get 125 possible combinations for a single style) I could dynamically launch 125 * X instances and get them to do the training, report back to us (email, upload whatever) and shut them down when finished.

Yeah one time potentially high cost (I'll take it on me) but then we would've filled up this repo. Just a thought, what do you think?

Hi, there!
I'm a Tsinghua University Student learning Neural Networks this semester, and I'm focusing on Neural Style Transfer right now.
When I was studying on the results of this paper, I found something strange:
The result picture using the model of StarryNight refered by DylanAlloy has some discordant part on the top right of the highest tower. And when I generated by the given model on my computer, the discordant part becomes even more obvious.(as the picture showed below, in the red circles)
Also, I've tried Gatys's Image Style Transfer Using Convolutional Neural Networks and there was nothing wrong on that part.
I'm wondering what may cause the result? If anyone is still interested in the error part, could you please give me some fair explanations? Thank you!
p.s. Sorry for my poor English.

Thank you for your hard work on this project. But currently it is not usable due to the changes on the recent version. As you wrote "change on how the models are trained which makes them not suitable for the most recent version relased. Will update here soon." -- this issue is a request for an estimated time. Thanks again!

Hi there!
Could anyone please train a model using this style image and with --image_size 1024 ?

I think the resulting pictures could be beautiful. I tried to use this style with this online service: https://dreamscopeapp.com/ and got this result:

Sad that this service gives only 512x512 results :( I think 1024x1024 would be more beautiful...
Could anyone please help? I have no powerful video card :(

I've been trying to train a few models, but they all seem to come out pretty weird.

So, example:
http://imgur.com/a/bLaTj

based on 512x512 crop of illegal beauty

https://static12.insales.ru/images/products/1/7694/42319374/Beamused_.jpg

options

-d train2014 --image_size 512 --lambda_tv 10e-4 --lambda_feat 1e0 --lambda_style 1e1 --lr 1e-3

and that took over a day of course..

So, I'm wondering I guess how to smooth it out or any advice on better options?

Maybe I choose I difficult source, but art I've tried has also come out quite textured, with not as much style.

Why the images with the model has noise?
I used "python train.py -s style1.jpg -d train2014 -g 0 "

what is the correct params?

First off, you guys are champions. I can't wait to contribute.

Secondly did you notice that your model filesizes are way smaller? Why do you think that is?

Check this out, I wrote this software for easy styling

NeuralStyler- Turn Your Videos/Photos/GIF into Art
http://neuralstyler.com/

*Style videos,gif animation and photos
*No need to upload videos (Offline processing)
*Faster AI styling algorithm
*Extensible styling system(Plugin)

Simply copy the model to the styles directory
edtaonisl.model
edtaonisl-style.jpg

use source:
https://github.com/yusuketomoto/chainer-fast-neuralstyle

I get http://msvocds.blob.core.windows.net/coco2014/train2014.zip
unzip train2014.zip
then

suker@suker:~/chainer-fast-neuralstyle$ python train.py -s rio.jpg -d ./train2014 -g 0
/usr/local/lib/python2.7/dist-packages/pkg_resources/init.py:1298: UserWarning: /home/gtx1080/.python-eggs is writable by group/others and vulnerable to attack when used with get_resource_filename. Consider a more secure location (set with .set_extraction_path or the PYTHON_EGG_CACHE environment variable).
warnings.warn(msg, UserWarning)
num traning images: 82783
82783 iterations, 2 epochs
/usr/local/lib/python2.7/dist-packages/chainer-1.14.0-py2.7-linux-x86_64.egg/chainer/cuda.py:87: UserWarning: cuDNN is not enabled.
Please reinstall chainer after you install cudnn
(see https://github.com/pfnet/chainer#installation).
'cuDNN is not enabled.\n'
epoch 0
(epoch 0) batch 0/82783... training loss is...235437792.0
(epoch 0) batch 1/82783... training loss is...329452736.0
(epoch 0) batch 2/82783... training loss is...154218272.0
(epoch 0) batch 3/82783... training loss is...63537960.0
(epoch 0) batch 4/82783... training loss is...58619268.0
(epoch 0) batch 5/82783... training loss is...55183628.0
(epoch 0) batch 6/82783... training loss is...54082024.0

is operate OK?

after more then 5 hours, the train finish.

the output
rio_0.model
rio_0.state

rio_1.model
rio_1.state

rio.model
rio.state

I use rio.model, to gen picture, it seems not ok, is any step i'm wrong?

rio-train.zip

rio.jpg
rio.model
rio-dst.jpg

thanks

The animated gifs have been created with a simple "real-time" video capture wrapper written in python, will release the code shortly.

Would really like to see it to replicate the results!

I love the styles you have! Thank you so much!

I'm trying to make some models of my own, but having trouble with the exact training
options.
What do you mean by (crop) in your examples? Is that a flag? Are you cropping the original art?
Also, your training set, does "full coco" mean the full train2014 w/ 80k images? or a different set?

Can anyone train a model using this style image?

I really like how it looks on Prisma (the app), and I don't think I have a powerful enough computer to train this model.

Thanks!

Hi,
Just wondering what happened to some of the models, there used to be more.
Hoping we can get more models in here.