This is a re-implementation of the paper "Universal style transfervia feature transforms" by Li et al.. The implementation follows most but not all details from the paper, including output activation, etc...
VGG pretrained weight file. Put this in the root folder.
Type help for more details. Basically there are two main functions as follows.
Download the MS COCO 2017 train and valid sets here. Next download the Wikiart train/test dataset.
To train a network using the default settings, use
python train.py path-to-MS-COCO-train path-to-Wikiart-train path-to-MS-COCO-val path-to-Wikiart-test
First, download the above VGG weight file. To test whether for a single image or folder of images, simply use
python test.py path-to-input path-to-style path_to_weight_files
path_to_weight_files
must contain 5 weight files with signature decoder-X-final.npz
where X= [1, 4, 7, 12, 17]
.
A model trained by this implementation can be downloaded here.
"Universal style transfervia feature transforms" by Li et al..
The pretrained VGG is taken from this Tensorflow implementation.