pytorch-AdaIN

This is an unofficial pytorch implementation of Arbitrary Style Transfer in Real-time with Adaptive Instance Normalization.

Original torch implementation from the author can be found here.

Requirements

• Python 3.5+
• Pytorch 0.4+
• TorchVision
• tqdm

Anaconda environment recommended here!

(optional)

• GPU environment for training

Usage

Train

python train.py args

Possible ARGS are:

• -h, --help Show this help message and exit;
• --content_dir CONTENT_DIR Directory path to amount of content images;
• --style_dir STYLE_DIR Directory path to amount of style images;
• --save_models SAVE_MODELS Path to save the trained model (default=save_models);
• --batch_size BATCH_SIZE The size of batch to train (default=8);
• --alpha ALPHA a smooth transition between content-similarity and style-similarity can be observed by changing α from 0 to 1.0 (default=1.0);
• --lambda_weight LAMBDA_WEIGHT The degree of style transfer can be controlled during training by adjusting the style weight λ (default=10.0);
• --lr LR The learning rate of Adam (default=1e-4);
• --lr_decay LR_DECAY The decay rate of learning rate (default=5e-5);
• --max_epoch MAX_EPOCH Number of iterations (default=160000);
• --save_model_interval SAVE_MODEL_INVTERVAL The interval epoch to save trained model (default=10000)

---

Use --content_dir and --style_dir to provide the respective directory to the content and style images.

python train.py --content_dir input/content --style_dir input/style

Test

test.py

python test.py args

Possible ARGS are:

• -h, --help Show this help message and exit;
• --content CONTENT File path to content image;
• --style STYLE File path to style image;
• --content_dir CONTENT_DIR Directory path to a batch of content images;
• --style_dir STYLE_DIR Directory path to a batch of style images;
• --img_size IMG_SIZE Minimum size for images, keeping the original size if given 0 (default is 512);
• --output_dir OUTPUT_DIR Directory to save the stylized images (default is output);
• --style_interpolation_weights STYLE_INTERPOLATION_WEIGHTS The weight for blending the multiple style images;
• --decoder DECODER Path for the arguments of decoder (default is models/decoder.pth);
• --alpha ALPHA a smooth transition between content-similarity and style-similarity can be observed by changing α from 0 to 1.0 (default is 1.0);
• --perserve_color PERSERVE_COLOR If specified, preserve color of the content image (action=store_true);
• --crop CROP do center crop to create squared image (action=store_true);

---

Use --content and --style to provide the repective path to the content and style image.

python test.py --content input/content/brad_pitt.jpg --style input/style/sketch.png

You can also run the code on directories of content and style images using --content_dir and --style_dir. It will save every possible combination of content and styles to the output directory.

python test.py --content_dir input/content --style_dir input/style

This is an example of mixing four styles by specifying --style and --style_interpolation_weights option.

python test.py --content input/content/avril.jpg --style input/style/antimonocromatismo.jpg,input/style/asheville.jpg,input/style/sketch.png,input/style/impronte_d_artista.jpg --style_interpolation_weights 1,1,1,1 --crop

Results