Skip to content

Latest commit

 

History

History
71 lines (66 loc) · 3.27 KB

README.md

File metadata and controls

71 lines (66 loc) · 3.27 KB
Raw

Fast Style Transfer in Tensorflow 2

An implementation of fast style transfer, using Tensorflow 2 and many of the toolings native to it and TensorFlow Add Ons.

Dataset

Content Images

The COCO 2014 dataset was used for content images, which can be found here.

Style

Several style images are included in this repository. Please note that some images are preprocessed/cropped from the original artwork to abstract certain details.

Results

Results after 2 epochs. All of these samples were trained with the default hyper-parameters as a base line and can be tuned accordingly. For instance, "The Scream" model could use some tuning or addition training time, as there are untrained spots.

Running the code

Please note, this is not intended to be run on a local machine. With the availability of cloud notebooks, development was on a Colab runtime, which can be viewed here. Training time for 2 epochs was about 4 hours on a Colab instance with a GPU.

References

These are previous implementations that in Lau and TensorFlow that were referenced in migrating to TF2. Many thanks to their work.

@inproceedings{Johnson2016Perceptual,
  title={Perceptual losses for real-time style transfer and super-resolution},
  author={Johnson, Justin and Alahi, Alexandre and Fei-Fei, Li},
  booktitle={European Conference on Computer Vision},
  year={2016}
}

  @misc{engstrom2016faststyletransfer,
    author = {Logan Engstrom},
    title = {Fast Style Transfer},
    year = {2016},
    howpublished = {\url{https://github.com/lengstrom/fast-style-transfer/}},
    note = {commit xxxxxxx}
  }