Faking images with stylegan2

Scripts for faking images with with stylegan2

Overview

These scripts are meant to automate the process of setting up an environment and making fake images from a corpus of input images on a machine with a discrete GPU. Stylegan2 is very user-friendly, but it still takes a bit of setup, which I would forget and have to re-learn, so I just automated them. Since tasks run for a long time, these scripts make it easy to start and then restart whenever the process stops for one reason or another, such as a power outage, system or program crash. They also make it easy to log in, start the scripts, and log out without killing the process (by using nohup).

Scripts

Except for config.sh, which is meant to be modified by the user, the scripts are in /scripts/ within this repo.

config.sh

Start here. You'll want to set some variables here that point to your files. However, not all of the scripts use these vars, and a nice improvement to this repo would be to replace hardcoded paths to pull these from the env vars set in this file.

get-max-dimensions.py

If you don't know how big the source images are, and you have a huge folder of them, this script goes through each image and keeps track of the largest dimension so you can choose an appropriate IMAGE_DIMENSION value.

setup.sh

Start here. This script sets up a Python virtual env and installs requirements.txt, and it clones a copy of StyleGAN 2. Currently we are using this modified fork that adds features to resume training. Hopefully this will be merged into the official branch.

Now you need to activate your virtual environment to use the rest of the scripts

. venv/bin/activate

You should see your prompt changed with (venv) added to it.

resize-squares.py

You need squares images that 2^n in either dimension, meaning n = 128, 256, 512, 1024x1024, for example. If your input images are not in this format, use resize-squares.py. This will take images from your RAW_IMAGE_DIR with the RAW_IMAGE_EXT file extension and create an IMAGE_DIR filled with NxN crops.

If you already have these in this format and don't need to square your images, just manually set the IMAGE_DIR and don't run resize-squares.py.

This script requires the env vars set by config.sh, so you can run:

bash config.sh ; python scripts/resize-squares.py

create-datasets.sh

Creates the data sets from the data directory.

bash scripts/create-datasets.sh

This uses nohup, so you can safely disconnect your terminal session.

train-on-data.sh

You'll want to delete the line:

    --resume-pkl="$LATEST_SNAPSHOT" \

when you first start. You'll want to add a line line this when you need to resume training, such as after a crash, but you'll need to point it to an appropriate file in your own results dir.

Usage:

bash scripts/train-on-data.sh

compile-fakes-video.py

Turn the results fake pngs into a video using openCV2.

bash scripts/compile-fakes-video.py

This assumes your results are in ./results, which they should be by default.