Style Transfer with Style Attentional networks

Image style transfer using a style attentional network.

Training

• install requirements with pip install -r requirements.txt
• get your kaggle API key from here
• start training with python3 train.py

Arguments

• content_dir directory to content images (default to data/content)
• style_dir directory to style images (default to data/style)
• vgg pretrained vgg file (default to pretrained/vgg.pth)
• lr leaning rate
• lr_decay leaning rate decay
• max_iter maximum number of iterations (default to 160,000)
• batch_size number of image pairs in a batch
• style_weight style images weight (default to 3.0)
• content_weight content images weight (default to 1.0)
• n_threads number of threads used for the data loader (default to 2)
• save_model_interval interval to save models (default to 1000)
• start iter number of iteration to start training from (in case training was stopped)

Testing

• run test.py with appropriate arguments

Arguments

• content content image
• style style image
• alpha style content trade-off
• vgg path to pretrained vgg model
• decoder path to pretrained decoder
• transformer path to pretrained transfomer model

More results

References

• D.Y. Park and K.H. Lee "Arbitrary Style Transfer with Style-Attentional Networks.", in CVPR, 2019.
• original paper implementation
• AdaIN implementation