DeepStyleX

DeepStyleX is an image transformation family of models that applies a style to the image. This model is inspired from these two papers:

• Perceptual Losses for Real-Time Style Transfer and Super-Resolution
• Instance Normalization: The Missing Ingredient for Fast Stylization

Tip

This project contains a Web UI to make it easy to use. For more information, reffer to this section.

Important

There is a suit of pretrained models, reffer to this section.

Getting started

python3 -m venv .venv
source .venv/bin/activate

To install the dependencies:

pip install -r requirements.txt

If it doesn't work, try

pip install torch pillow torchvision numpy tqdm pillow opencv-python

# For web-app
pip install fastapi python-multipart uvicorn

Example use case

Train

python3 scripts/model-cli.py train --dataset data/scenes/Coco/ --epochs 4 --style-path data/style/leaf-style.jpg --resume-path data/saves/model1.pth --output-path data/saves/model1.pth --checkpoint-dir data/check/ --checkpoint-interval 200

Test

python3 scripts/model-cli.py eval --model-path data/saves/model1.pth --input-image data/Central-Park.jpg --output-image output/central-leaf.jpg 

Web app

Launch the server

uvicorn scripts.api:app --reload --port 8080

Note: Your models should be in data/saves/ in order to be used in the web app

Download pretrained models

To download pretrained models, just run this python script:

python3 scripts/download-models.py