Style Transfer on Fashion Products

This is the implementation of the Neural Style algorithm by Gatys et al. and
DeepObjStyle (case of m=n) by Mastan and Raman for the Fashion Product Images Dataset.

Requirements

For downloading the libraries via conda, try:

conda install python=3.9
conda install pytorch=1.10 torchvision=0.11 cudatoolkit=11.3 -c pytorch
conda install pandas matplotlib
conda install scikit-learn opencv -c conda-forge 
conda install seaborn scikit-image -c anaconda

Running

First create a folder named results to store the outputs.
For running the optimization, please set method to 'neural_style' or 'deepobjstyle', choose input images, and try:

python main.py --method='deepobjstyle' --style_image_id=0 --content_image_id=0 

Hyperparameters

For setting the weights of the loss terms, please refer to the notation of the papers mentioned in the beginning and try the following for DeepObjStyle:

python main.py --method='deepobjstyle' --alpha11=1e4 --alpha12=5e6 -alpha13=1e-4 --alpha2=1e-2

And the following for Neural Style:

python main.py --method='neural_style'--alpha=1 --beta=5e6

Data and Segmentation Preprocessing

To randomly sample images from the Fashion Product Images Dataset, download the dataset from kaggle and add the style images to the folder 'images/raw'. For preprocessing the images, please try:

python main.py --method='deepobjstyle' --prepare_data --path_style_raw='images\raw' --path_content_raw='fashion-product-images-dataset/fashion-dataset/images'

For creating respective segmentation images, please try:

python main.py --method='deepobjstyle' --prepare_segmentation --path_segmentation='images/segmentation'

Acknowlegements

The code is based on NEURAL TRANSFER USING PYTORCH and, yagudin/PyTorch-ddp-photo-styletransfer.