This is a Keras implementation of Fast-Neural-Style (link)

Results

You could also use it to train a network combining two different styles. Here are some results:

These images are generated by combining two styles above.

How to use it

For training the network:

python main.py -c ./configs/[config_name] -m train

For predicting:

python main.py -c ./configs/[config_name] -m predict -i [image_read_path] -o [image_save_path]

For viewing the baseline: (link)

python main.py -c ./configs/[config_name] -m temp_view -i [image_read_path] -o [image_save_path] --iters [ITER]

The implementation of viewing base line is a little different from the original paper and shares the same loss net with training process, so you could use this temp_view function to adjust the hyperparameters.

For training the network, you should download COCO dataset first and unzip it to the train_image_path, which is specified in the config files.

Here is the explanation for parameters in config files:

• "net_name": the name of your style transferring net. weights will be saved according to the name of the network;
• "learning_rate": learning rate of the Adam optimizer. It is set to 1e-3, as recommended in the original paper;
• "content_weight": the weight of content loss;
• "style_weight": the weight of style loss;
• "total_variation_weight": the weight of total variation loss;
• "train_image_height" / "train_image_width": All of images in the training set will be resized to pre-defined height and width. It's set to 256/256 as recommended in the original paper;
• "plot_model": If it's set to true, the program will use pyplot to plot the model and save the graph;
• "content_layer": the layer for computing content loss. Please do not change it;
• "style_layer": a list, specified the layer for computing style loss. Please do not change it;
• "style_image_path": the path to style image;
• "style_image_path_2": the path to another style image. If exists the program will get results by combining two different styles;
• "test_image_path": the path to test image. You could use this option to validate the training process;
• "test_res_save_path": All test results will be saved to this path.

Performance

It takes about 8 hours to train a network on a Nvidia K80 GPU. After training, predicting will be really fast and only use less than one second.