jonathanonglq/test

Neural Style Transfer

This is a simple Jupyter Notebook which allows users to recombine the content and style of two arbitrary images, and capture the transition in the form of a gif. The algorithm used is based on the original paper, A Neural Algorithm of Artistic Style, and references this TensorFlow tutorial.

Adjustable Parameters

The notebook allows users to customise certain parameters to generate the desired output. These parameters are as follows:

• style_path: path of the style image
• content_path : path of the content image
• feature_maps : Can be set to True/False. If True, the notebook will display and save sample feature maps of the convolutional layers used (refer to style_layers below for the specific layers used). Useful for investigating the features learned by the model when implementing the style transfer.
• create_gif : Can be set to True/False. If True, the notebook will display and save a gif of the style transfer process.
• gif_fps : Defines the frame per second (fps) of the output gif. Default set as 2.
• plot_loss : Can be set to True/False. If True, the notebook will plot the graph of total variational loss vs epoch/steps.
• style_layers and content_layers: The convolutional layers used to extract the style and content of the style and content images respectively. The default layers are based on the layers recommended in the original paper - 'block1_conv1','block2_conv1','block3_conv1','block4_conv1', and 'block5_conv1' for the style image, and 'block5_conv2' for the content image.
• style_weight and content_weight : Weights which can be adjusted to customise the "amount" of style/content to transfer to the final image.
• total_variation_weight : Can be adjusted to modify the rate of convergence of the total variation loss of the model.
• opt : The optimizer used to train the model. The paper recommends LBFGS, but the Adam optimizer set as the default here works just as well.

Sample Images (1/2)

Content Image

Style Image

Output Gif

Sample Images (2/2)

Content Image

Style Image

Output Gif