AdaIN

PyTorch implementation of this paper. Original implementation can be found at here.

Examples

These are generated images with this implementation.

Case 1

Content	Style

iter_10K	iter_20K	iter_30K	iter_40K

iter_50K	iter_60K	iter_70K	iter_80K

iter_90K	iter_100K	iter_110K	iter_120K

iter_130K	iter_140K	iter_150K	iter_160K

Case 2

Content	Style

iter_10K	iter_20K	iter_30K	iter_40K

iter_50K	iter_60K	iter_70K	iter_80K

iter_90K	iter_100K	iter_110K	iter_120K

iter_130K	iter_140K	iter_150K	iter_160K

Clone

via ssh

$ git clone [email protected]:kukosmos/AdaIN-pytorch-2019.git

via https

$ git clone https://github.com/kukosmos/AdaIN-pytorch-2019.git

Docker

We offer a docker image. If you prefer run this repository locally, please skip this section and follow the instructions from the following section. First, build the image.

$ docker build . --tag adain

Start a docker container with the created image.

$ docker run --rm -it -v $(pwd)/experiments:/workspace/experiments -v $(pwd)/data:/workspace/data -u $(id -u):$(id -g) adain

We provide two commands (adain-train, adain-test). Please see how to use them with following commands..

$ adain-train -h
$ adain-test -h

Requirments

Install following applications

• python3.6+
• unzip

Install following python libraries

• Pillow
• pytorch
• torchvision
• tensorboard
• tensorboardX
• tqdm

Train

To train a model yourself, do the following steps.

Prepare dataset

First, download the dataset with given script. For example, to download coco2017train dataset, type in the following command:

$ ./download.sh coco2017train

Detailed usage of download.sh script can be found as follows:

$ ./download.sh --help

Train model

Train the model with train.py. For example, to train model with coco2017train as content and wikiart as style, type in the following command:

$ python train.py \
>   --content-dir data/coco2017train \
>   --style-dir data/wikiart

To check the logs that saved in logs directory, type in the following command:

$ tensorboard --logdir logs

For more

Advanced options can be found with following command:

$ python train.py --help

Generate Results

To generate sytled images, use test.py. For example, to generate a styled image from trained model models/adain.pth with images/content.jpg as a content image and images/style.jpg as a style image, type in the following command:

$ python test.py \
>   --model models/adain.pth \
>   --content images/content.jpg \
>   --style images/style.jpg

Interpolation

To mix the two or more styles in one content image, specify the interpolation weights as follows:

$ python test.py \
>   --model models/adain.pth \
>   --content images/content.jpg \
>   --style images/style1.jpg images/style2.jpg images/style3.jpg \
>   --interpolation-weights 2 3 4

Note that all style images should have weights. In other words, the number of images should be equal to the number of weights.

For more

More options can be found with following command:

$ python test.py --help

References

• [1]: X. Huang and S. Belongie. "Arbitrary Style Transfer in Real-time with Adaptive Instance Normalization", in ICCV, 2017.
• [2]: PyTorch implementation of AdaIN
• [3]: COCO Dataset
• [4]: Wikiart dataset