Text Generator Powered by GPT-2-JA

Easy to use scripts to fine-tune GPT-2 Japanese model with your own text, to generate sentences, and to tweet them automatically.

Installation

1. Clone this repository

git clone https://github.com/discus0434/text-generation.git
cd text-generation

2. Install dependencies

pip

pip install -r requirements.txt

conda

conda create -n textgen python=3.9
echo "source activate textgen" > ~/.bashrc
conda run -n textgen pip install -r requirements.txt

...or build docker image

docker build text-generation/. -t textgen
docker run -it --runtime=nvidia -d --restart=always textgen:latest bash

3. Run setup.sh (not necessary for docker users)

Modify permission of shell scripts

chmod +x scripts/shell/setup.sh
chmod +x scripts/shell/finetune.sh
chmod +x scripts/shell/generate.sh

Run setup.sh to download a pretrained model

scripts/shell/setup.sh

4. Arrange Twitter API tokens (not necessary if you do not interact with Twitter)

Arrange .env file in the project directory. .env file should look like this:

API_KEY = XXX
API_SECRET_KEY = XXX
BEARER_TOKEN = XXX
ACCESS_TOKEN = XXX
ACCESS_TOKEN_SECRET = XXX

Usage

You can not only make a Twitter bot but just generate sentences.

Preparation

Arrange your text file you want to use in sample_texts/.

mv <your text file>.txt sample_texts/

...or download someone's tweets with Python script to use it

python ./scripts/python/download_tweets.py -u <Twitter user ID> -max <number of tweets you will get>

Usage 1: Just generate sentences

Fine-tuning

scripts/shell/finetune.sh

Generate 10 sentences with your fine-tuned model

scripts/shell/generate.sh

Usage 2: Manage a Twitter bot powered by fine-tuned GPT-2-JA

Fine-tuning and generate tweet

...Run all of cells in create_tweet.ipynb.

Fine-tuning and generate tweets per 30 min till interrupted

python ./scripts/python/create_tweets.py & disown

Acknowledgements

This code borrows from gpt2-japanese.