Jesse Fredrickson 11/3/2020
This repo contains the code to operate simple a discord bot which, when prompted with a "claim waifu" message in chat, will dynamically generate and post a unique anime girlfriend. Who wouldn't want that
It is assumed the user already has python3. A basic understanding of python is useful but not necessary.
Python Environment
- Clone this repository
git clone https://github.com/jfreds91/waifu_bot - Navigate to cloned repo and create a new virtual environment
python3 -m venv env - Activate environment
source env/bin/activate - Update environment from requirements doc
pip install -r requirements.txt
System Installs
- At this point you already have the onnx runtime package(s) installed, but you may have to install additional dependencies, see below. On MacOS, I had to install OpenMP
brew install libomp - Follow the install instructions at https://github.com/Microsoft/onnxruntime
GAN Model download
- Download the TWDNE3.onnx model from https://hivemind-repo.s3-us-west-2.amazonaws.com/twdne3/twdne3.onnx and move it to your project repo
- If that's no longer active, look for an update on https://www.gwern.net/Faces#
Runtime
- You will need to create your own discord bot token. Instructions
- You will need to create your own credentials.cfg file containing the token to be read in by configparser
- You will need to edit the .onnx model filepaths. Currently hardcoded in discord_client.py
- You may adjust the "artistic liberty" of the generator by changing the TRUNCATION scalar from 0 (no liberty, converged to single average waifu) to any float. Gwern defaults to 0.7 for his work; anything over 1.2 or so results in Lovecraftian horror
The application is fun, but I wrote this project to explore a few interesting technical areas.
- ONNX. ONNX is a common model interoperability platform introduced by Facebook and Microsoft in 2017, and allows for users to run framework-agnostic models by converting to its own common format. This repo uses an onnx model and onnx runtime to perform inference and generate images.
- Discord API. I have a small but active discord which has helped me get through quarantine (COVID19), and improving it (for a loose definition of improve) is intriguing
- GANs. Though I did not train this model myself, learning how to (ab)use the generator model was a learning experience, and Gwern's blog was a fascinating read.
Thanks for reading, enjoy!
- Pull Image:
docker pull jfreds/waifu_docker_test:latest - Build Image from Dockerfile:
docker build -t jfreds/waifu_docker_test:v0 - Open shell in container:
docker run --entrypoint "/bin/bash" -it jfreds/waifu_docker_test:v0 - Remove all stopped containers:
docker rm $(docker ps -a -q)
- Run waifu generation:
docker run -v <local_dir>:<mounted_dir> jfreds/waifu_docker_test:v0 python3.7 main.py -dir <mounted_dir> -t <float> -s <string> - ex:
docker run -v "$(pwd)":/mnt/here jfreds/waifu_docker_test:v0 python3.7 main.py -dir /mnt/here -t 0.85 -s Elon Musk- saves: Elon Musk.jpeg
- where -v is local volume mount (allows container to save output to host storage)
- where -dir is save location (should be mounted location)
- where -t is truncation float; default is 0.7 (from git repo)
- where -s is seed string for waifu. Allows spaces, quotes not needed.