Skip to content

Latest commit

 

History

History
86 lines (66 loc) · 3.65 KB

README.md

File metadata and controls

86 lines (66 loc) · 3.65 KB
Raw

PyTorch Implementation of Neural Style Transfer

Implementation of the neural style transfer algorithm (Gatys, Ecker, Bethge) using PyTorch. See accompanying blog post.

Generally follows the PyTorch tutorial, with some modifications.

Style transfer can run either on a single content image, or a directory of content images.

Example

Style image + Content image -> Result

The style image used here is Nocturne in Black and Gold, the Falling Rocket by James Abbott McNeill Whistler.

Setup

Create and activate a new virtual environment:

$ conda create -n style_transfer python=3
$ conda activate style_transfer

Install dependencies:

$ pip install -r requirements.txt

Usage

Neural style transfer on a single image

$ python single_neural_style_transfer.py

Settings

In single_neural_style_transfer.py:

  • Setting the style image

    • On line 258: style_path = 'style_imgs/marbled_paint_1.jpeg'
    • Replace this path with the path to the image you want to use as style input.

  • Setting the content image

    • On line 259: content_path = 'content_imgs/clouds/clouds12.jpg'
    • Replace this path with the path to the image you want to use as content input.

  • Setting the output path

    • On line 275: torchvision.utils.save_image(output, 'output.png')
    • By default, the result will save to a file named output.png in the same directory.
    • If you want, replace this with the filepath where you would like the output image to go.

Neural style transfer on a directory of images

$ python batch_neural_style_transfer.py

Settings

In batch_neural_style_transfer.py:

  • Setting the style image

    • On line 249: style_path = 'style_imgs/turner_sunset.jpg'
    • Replace this path with the path to the image you want to use as style input.

  • Setting the content image directory

    • On line 250: content_paths = 'content_imgs/*/*'
    • Replace this path with the directory of images you'd like to use as content input.
    • content_images/*/* will grab nested directories in the content_images directory. content_images/* will grab all images inside the content_images directory (unnested).

  • Setting the output path

    • On line 251: output_folder = 'results/turner_sunset/'
    • Replace this with the directory you would like the output image results to go.

Other options

  • Image size: imsize = 256
    • By default, the resulting images will be 256 x 256 pixels.
  • Style and content weights, number of iterations
    • Can tweak style_weight, content_weight, and num_steps.
    • More steps -> more stylized.
def run_style_transfer(cnn, norm_mean, norm_std,
        content_img, style_img, input_img,
        num_steps=300, style_weight=1000000, content_weight=1):

File Overview

  • single_neural_style_transfer.py: Run neural style transfer on a single content image.
  • batch_neural_style_transfer.py: Run neural style transfer on a directory of content images.
  • requirements.txt: Contains dependencies for project to be installed.

Optional:

  • style_imgs/: Directory containing images I thought would be nice to use as style input.
  • content_images/: Directory containing images I thought would be nice to use as content input.
  • results/: Directory containing results from previous runs.