Implementation of Arbitrary Style Transfer in Real-time with Adaptive Instance Normalization in Tensorflow
https://arxiv.org/abs/1703.06868
git clone https://github.com/ftokarev/tf-adain
cd style-transfer
virtualenv venv --python /usr/bin/python3
source venv/bin/activate
pip install -r requirements.txt
bash models/get_vgg.sh
Use --content and --style to provide the respective path to the content and style image, for example:
./test.py --content input/content/cornell.jpg --style input/style/woman_with_hat_matisse.jpg --content_size 0 --style_size 0
To run the code on directories of content and style images use --content_dir and --style_dir options. It will save every possible
combination of content and styles to the output directory.
./test.py --content_dir input/content --style_dir input/style
Other options:
--crop: Center crop both content and style images beforehand.--content_size: Scale the content image to this size, keeping proportions. Set to zero to keep the original size.--style_size: Scale the style image to this size, keeping proportions. Set to zero to keep the original size.
To see all available options, type:
./test.py -h
Use --alpha to adjust the degree of stylization. It should be a value between 0 and 1 (default).
Add --preserve_color to preserve the color of the content image. Example usage:
./test.py --content input/content/newyork.jpg --style input/style/brushstrokes.jpg --content_size 0 --style_size 0 --preserve_color
To interpolate between several style images, provide them as a comma-separated list. You can control the relative weight of each style by using the --style_interp_weights option. Example usage:
./test.py --content input/content/avril.jpg --style input/style/picasso_self_portrait.jpg,input/style/impronte_d_artista.jpg --style_interp_weights 1,1 --crop --style_size 400
Use --mask to provide the path to a binary foreground mask. You can transfer the foreground and background of the content image to different styles.
Note that you also to need to provide two style images separated be comma, in which the first one is applied to foreground and the second one is applied to background. Example usage:
./test.py --content input/content/blonde_girl.jpg --style input/style/woman_in_peasant_dress_cropped.jpg,input/style/mondrian_cropped.jpg --mask input/mask/mask.png --content_size 0 --style_size 0
- Download the wikiart dataset from Kaggle (https://www.kaggle.com/c/painter-by-numbers/download/train.zip), unzip images into
datasets/wikiart/train - Download the MS COCO dataset using the included downloader:
./datasets/get_coco_train.sh
- Prepare datasets for training
./prepare_dataset.py datasets/coco/train2014/ datasets/coco/
./prepare_dataset.py datasets/wikiart/train/ datasets/wikiart/
- Launch the training script
./train.py
To see all available options, type:
./train.py -h
- This implementation is based on the original Torch implementation (https://github.com/xunhuang1995/AdaIN-style)
[X] make sure the default works:
python train.py
python test.py --content_dir input/content --style_dir input/style --output_dir outputs_c --decoder_weights checkpoints/adain-2000
[X] try elu -- seems not good train with elu in both encoder and decoder
python train.py --encoder_activation elu --decoder_activation elu --checkpoint_dir checkpoints_elu
python test.py --content_dir input/content --style_dir input/style --encoder_activation elu --decoder_activation elu --decoder_weights checkpoints_elu/adain-2000 --output_dir outputs_elu
[ ] try to train on 512 patch size
python train.py --random_crop_size 512 --batch_size 4 --checkpoint_dir checkpoints_512
python test.py --content_dir input/content --style_dir input/style --output_dir outputs_c512 --decoder_weights checkpoints_512/adain-2000
Generate content-style outputs, using prebuild .h5 models
python test.py --content_dir input/content --style_dir input/style
Train:
python train.py
python test.py --content_dir input/content --style_dir input/style --output_dir outputs_c --decoder_weights checkpoints/adain-2000