Simplified Web REST API of the StyleGAN using FastAPI

This is a web backend application to make the generator models trained by Precure StyleGAN ADA available on the web in a simplified way. StyleGAN Vue is a sample implementation of its frontend client and is deployed here connected to a demo backend with some models for immediate trial.

Sample Docker configurations for deploying StyleGAN models and the web-based user interface are also available.

Refer to the Swagger or ReDoc documentation on the sample backend or on your deployed instance at the paths /docs or /redoc.

To run this application, you need a Unix-like system with Python >= 3.12 installed.

Use the requirements.txt to install the minimal dependencies for serving and inferencing, including the dependencies of Precure StyleGAN ADA.

pip3 install -r requirements.txt

While Uvicorn is the recommended as ASGI middleware, other alternatives such as Hypercorn may also work. However, in a production environment, it is common to use process replication with a combination of Gunicorn and Uvicorn.

This repository contains the default server configuration using Uvicorn with some trained models. You can quickly try it out as follows:

git clone --recursive https://github.com/curegit/stylegan-fastapi.git
cd stylegan-fastapi
pip3 install -r requirements.txt
python3 -m uvicorn main:app

Please note that this setup is intended for trial or debugging purposes only and is not suitable for production use as it does not support process replication.

StyleGAN FastAPI can be customized by using a configuration file. StyleGAN FastAPI uses the TOML format for configuration, and the default server's TOML file is located in default/config.toml.

You can specify a different configuration TOML file to use by setting the STYLEGAN_TOML environment variable. STYLEGAN_TOML must be an absolute TOML file path or a TOML file path relative to the working directory. Note that the STYLEGAN_TOML environment variable must be set before importing the api/ Python package.

export STYLEGAN_TOML="./custom_config.toml"
uvicorn main:app

The configuration file follows the standard TOML format. A complete raw schema definition and its default values can be found in the api/conf.py Python file.

All properties are optional except the file field in ModelConfig to specify model files, and you must specify at least one model.

To merge the application log into the Gunicorn log, set logger = "gunicorn.error" in the StyleGAN FastAPI TOML configuration file.

You can customize how Gunicorn runs the application by overriding various options derived from sample Gunicorn settings.

from guni.conf import *

wsgi_app = "main:app"
workers = 3
raw_env = ["STYLEGAN_TOML=myconfig.toml"]

There are certain limitations when using CUDA/CuPy.

CUDA has a restriction that prevents processes from being forked after CUDA has been initialized. To address this, set preload = False in Gunicorn settings so that CUDA is initialized in each forked process after Gunicorn creates child processes. Note that this approach will consume more RAM/VRAM.

If RAM/VRAM usage is a concern, it may be advisable to use Uvicorn alone even in production environments.

[Unit]
Description=Gunicorn StyleGAN Daemon
Requires=stylegan.socket
After=network.target
After=syslog.target

[Service]
Type=notify
User=spam
Group=spam
WorkingDirectory=/home/spam/stylegan
ExecStart=/usr/local/bin/python3.12 -m gunicorn -c gunicorn.conf.py \
                                                --log-level info \
                                                --log-file /var/log/stylegan/system.log \
                                                --access-logfile /var/log/stylegan/access.log \
                                                --capture-output
RuntimeDirectory=stylegan
LogsDirectory=stylegan
PrivateTmp=true
ExecReload=/bin/kill -s HUP $MAINPID
KillMode=mixed
TimeoutStartSec=240
TimeoutStopSec=30
Restart=always

[Install]
WantedBy=multi-user.target

If the volatile /run location is available via RuntimeDirectory, it is recommended to set as tmp_dir = "/run/stylegan" in the StyleGAN FastAPI TOML for improved disk IO performance.

[Unit]
Description=Gunicorn StyleGAN Socket

[Socket]
ListenStream=/run/stylegan.sock
SocketUser=www-data
SocketMode=600

[Install]
WantedBy=sockets.target

/var/log/stylegan/*.log {
  monthly
  dateformat -%Y-%m-%d
  rotate 12
  compress
  ifempty
  missingok
  copytruncate
  maxsize 100M
}

• Ensure that proxy headers, if used, are set correctly (e.g., Forwarded, X-Forwarded-For) for production environments with a reverse proxy layer.
• It is also recommended to set header and body size limits via Gunicorn or a reverse proxy.

CC BY-NC 4.0