https://github.com/YaleDHLab/stylegan2-helpers
Automates a bunch of steps to go from a folder of images to a trained network using stylegan2
https://github.com/YaleDHLab/stylegan2-helpers
gans helper-tool stylegan2
Last synced: 12 days ago
JSON representation
Automates a bunch of steps to go from a folder of images to a trained network using stylegan2
- Host: GitHub
- URL: https://github.com/YaleDHLab/stylegan2-helpers
- Owner: YaleDHLab
- License: gpl-3.0
- Archived: true
- Created: 2020-08-27T17:29:45.000Z (over 4 years ago)
- Default Branch: main
- Last Pushed: 2024-09-04T18:58:38.000Z (9 months ago)
- Last Synced: 2024-11-15T06:32:35.953Z (6 months ago)
- Topics: gans, helper-tool, stylegan2
- Language: Python
- Homepage:
- Size: 37.1 KB
- Stars: 3
- Watchers: 4
- Forks: 1
- Open Issues: 3
-
Metadata Files:
- Readme: README.md
- License: LICENSE
Awesome Lists containing this project
README
# Note: This repository has been archived
This project was developed under a previous phase of the Yale Digital Humanities Lab. Now a part of Yale Library’s Computational Methods and Data department, the Lab no longer includes this project in its scope of work. As such, it will receive no further updates.
# Faking images with stylegan2
> Scripts for faking images with with [stylegan2](https://github.com/NVlabs/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](https://github.com/ashirviskas/stylegan2.git) 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
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
bash scripts/train-on-data.sh
```
### compile-fakes-video.py
Turn the results fake pngs into a video using openCV2.
```bash
bash scripts/compile-fakes-video.py
```
This assumes your results are in ./results, which they should be by default.