Release Notes

0.4.0

Released: 2019-07-07

• Added stylish train --layer-weights option to initialize weights for each layer from STYLE_LAYERS. The default value was initially hard-coded to 0.2, but has now be changed to 1.0 as it produces better results.

• Updated stylish, stylish.transform and stylish.vgg to uses name scopes as much as possible in order to improve the graph visibility within Tensorboard.

• Improved time display during training.

0.3.0

Released: 2019-07-05

• Added stylish train --limit option to set a maximum number of files to use from the training dataset folder.

• Record style loss, content loss, total variation loss and the sum of all losses to generate scalar curves within Tensorboard.

0.2.0

Released: 2019-05-27

• Added stylish download command line option to download elements necessary for the training (Vgg19 model and training data).

• Added stylish.logging to manage logger using sawmill for convenience.

• Removed stylish.train and moved logic within stylish to increase code readability.

• [command line] Updated stylish.command_line to use click instead of argparse to manage the command line interface for convenience.

• Fixed stylish.transform.instance_normalization() logic.

0.1.4

Released: 2018-05-19

• Always use GPU for the training when available.

0.1.3

Released: 2018-05-19

• Updated stylish.train module to prevent fixing the shape of the input placeholder.

0.1.2

Released: 2018-05-18

• Updated stylish.transform module to let the size of the images unknown when processing the checkpoint.

• Updated stylish.train.extract_model() to increase verbosity.

0.1.1

Released: 2018-05-09

• Fixed --content-target command line option as it should take a single value, not a list of values.

• Fixed stylish.train.extract_model() to pass the correct placeholder identifier to the session.

0.1.0

Released: 2018-05-08

• Initial release.