Jitesh Jain^†, Divyansh Agarwal^†, Gagan Sharma^†, R Chinmay^†, Md Junaid Mahmood^†

[proposal] [mid-term report] [final report]

1. Overview
2. Setup Instructions
3. Experiments

This repo contains the code for our work Neural Style Transfer: A Technical Report done as a part of CSN-526: Machine Learning course project.

• Clone the repo:

  git clone https://github.com/praeclarumjj3/NST-Tech.git
  cd NST-Tech

• Create a conda environment:

  conda env create -f conda_env.yml
  conda activate nst

• Execute the following command to run style transfer:

  sh nst.sh

Note: There are arguments specified in the nst.sh script. Please modify them to run experiments under different settings.

• You may specify the content and style images to be used from the data/content and data/style folders respectively.

• We use the predictions from the AdaIn-Style method proposed in Arbitrary Style Transfer in Real-time with Adaptive Instance Normalization as ground truths while evaluating the performance of our method.

• Install image-similarity-measures:

  pip install image-similarity-measures[speedups]

• Execute the following command to calculate the PSNR, SSIM and RMSE scores:

  sh metrics.sh [path-to-gt] [path-to-our-prediction]

Note: The gts can be found in the data/gts/ directory. You may specify more metrics in the metrics.sh script.

This repo is a part of our course project for CSN-526: Machine Learning under Professor Pravendra Singh at CSE Department, IIT Roorkee. The code is open-sourced under the MIT License.

• Jitesh Jain: 19114039

• Divyansh Agarwal: 19115055

• Gagan Sharma: 19114032

• R Chinmay: 19114067

• Md Junaid Mahmood: 19116040