pjotrvr / neural-style-transform Goto Github PK

Creating fun images/videos by using Neural Style Transformation algorithms.

Some examples:

+ =

GIF of lion image being styled with cubist style:

This project is built upon PyTorch using Python 3.8+.

For GPU support install CUDA.

All modules used are in file requirements.txt.

Install: terminal -> pip3 install -r requirements.txt.

All parameters are setup in config.json file.

Parameters:

• image - content image - name of content image, eg. lion.jpg
  
• style - style image - name of style image, eg. cubist.jpg
  
• path_type - type of path - relative or absolute
• content_images_dir - path to content images folder, eg. data/content_images/
  
• style_images_dir - path to style images folder, eg. data/style_images/
  
• output_images_dir - path to output images folder, eg. data/output_images/
  
• freq - how frequently will image be saved while training (1 means after every step, 2 after every 2 steps, 0 means only final image will be saved)
  
• video - creating video from all training steps - True or False (only possible if frequency is >= 1)
  
• height - height of an image - int number, eg. 500
  
• image_format - in what format will image be saved - leave first value to be 4, second value can be .jpg, .png, etc.
  
• model - which model will be used for optimizing - vgg16 or vgg19
  
• optimizer - which optimizer will model use - currently only adam is supported
  
• init_method - initialization method for image - content orrandom
  
• iterations - number of iterations/epochs for training - int number, eg. 400