AdaIN

This work is supported by Cloud TPUs from Google's TPU Research Cloud (TRC)

Getting started

• Download the coco2017 and wikiart datasets
• Generate the tfrecords for training and validation.

python3 -m adain.dataset_utils.create_tfrecords --image_paths_pattern coco/train2017/* --prefix coco-train  --output_dir tfrecords
python3 -m adain.dataset_utils.create_tfrecords --image_paths_pattern coco/val2017/* --prefix coco-val  --output_dir tfrecords
python3 -m adain.dataset_utils.create_tfrecords --image_paths_pattern wikiart/train/* --prefix wikiart-train  --output_dir tfrecords

• Start training with:
  • python3 -m adain.main --config_path configs/coco-wikiart.json
• To export saved_model, use
  • python3 -m adain.export --config_path configs/coco-wikiart.json

Inference

content_images = glob('assets/images/content/*')
style_images = glob('assets/images/style/*')

saved_model = tf.saved_model.load('export')
inference_fn = saved_model.signatures['serving_default']

content_image = read_image(content_images[3])
style_image = read_image(style_images[16])

alpha = tf.constant(1.0)
resize = tf.constant(True)

serving_input = {
    'style_images': style_image,
    'content_images': content_image,
    'alpha': alpha,
    'resize': resize
}

stylized_image = inference_fn(**serving_input)['stylized_images'][0]
result = prepare_visualization_image(
    content_image[0], 
    style_image[0], 
    stylized_image, figsize=(20, 5))

imshow(result, figsize=(20, 10))

Results

Controlling Content-Style tradeoff by varying alpha

Training Logs

Tensorboard.Dev

tensorboard.dev

@article{DBLP:journals/corr/HuangB17,
  author    = {Xun Huang and
               Serge J. Belongie},
  title     = {Arbitrary Style Transfer in Real-time with Adaptive Instance Normalization},
  journal   = {CoRR},
  volume    = {abs/1703.06868},
  year      = {2017},
  url       = {http://arxiv.org/abs/1703.06868},
  archivePrefix = {arXiv},
  eprint    = {1703.06868},
  timestamp = {Mon, 13 Aug 2018 16:46:12 +0200},
  biburl    = {https://dblp.org/rec/journals/corr/HuangB17.bib},
  bibsource = {dblp computer science bibliography, https://dblp.org}
}