This project contains a notebook which explores Johnson et. al's (2016) Perceptual Losses for Real-Time Style Transfer and Super Resolution paper. Since it promises similar results as in Gatys et. al's (2015) A Neural Algorithm of Artistic Style paper in less time, I want to try it out for myself.

A closer look into this project can be found in my blog post.

The Jupyter notebook called perceptual_losses.ipynb can be opened. Note that the following frameworks must exist in your PIP or Anaconda environment:

• PyTorch
• NumPy
• Python Imaging Library
• tqdm

In this project, I used COCO dataset's 2014 training images (83K/13GB). Download this dataset and update DATASET_PATH to the location of the dataset.

After downloading the dataset, run the notebook from top to bottom. Training the dataset after two epochs takes around four hours.

Find any input image and update TEST_IMAGE_PATH to the image's location. Note that there is already a pre-trained model. Running the following line of code:

load_model_checkpoint(image_transformation_network, optimizer, "style-transfer-alpha1e5-beta1e10-epoch-2")

will load the pre-trained model and invoke the subsequent lines to style the input image.