Skip to content

Latest commit

 

History

History
68 lines (50 loc) · 3.98 KB

README.md

File metadata and controls

68 lines (50 loc) · 3.98 KB

fast-neural-style-tensorflow

A tensorflow implementation for Perceptual Losses for Real-Time Style Transfer and Super-Resolution.

This code is based on Tensorflow-Slim and OlavHN/fast-neural-style.

Samples:

configuration style sample
wave.yml
cubist.yml
denoised_starry.yml
mosaic.yml
scream.yml
feathers.yml
udnie.yml

Requirements and Prerequisites:

  • Python 2.7.x
  • Now support Tensorflow >= 1.0

Attention: This code also supports Tensorflow == 0.11. If it is your version, use the commit 5309a2a (git reset --hard 5309a2a).

And make sure you installed pyyaml:

pip install pyyaml

Use Trained Models:

You can download all the 7 trained models from Baidu Drive.

To generate a sample from the model "wave.ckpt-done", run:

python eval.py --model_file <your path to wave.ckpt-done> --image_file img/test.jpg

Then check out generated/res.jpg.

Train a Model:

To train a model from scratch, you should first download VGG16 model from Tensorflow Slim. Extract the file vgg_16.ckpt. Then copy it to the folder pretrained/ :

cd <this repo>
mkdir pretrained
cp <your path to vgg_16.ckpt>  pretrained/

Then download the COCO dataset. Please unzip it, and you will have a folder named "train2014" with many raw images in it. Then create a symbol link to it:

cd <this repo>
ln -s <your path to the folder "train2014"> train2014

Train the model of "wave":

python train.py -c conf/wave.yml

(Optional) Use tensorboard:

tensorboard --logdir models/wave/

Checkpoints will be written to "models/wave/".

View the configuration file for details.