How to generate text with OpenAI's GPT-2, React and Flask

 - 3 min read


In this tutorial, we’ll build a Flask & React app with GPT-2 capabilities. We’ll go step by step, by tweaking the generator’s “interface”, then we’ll build the Flask server and finally the React frontend.

By the end of this tutorial, here’s what our app should look like:

GPT-2 Generator with React & Flask
GPT-2 Generator with React & Flask

Generating text with GPT-2

Thanks to pytorch-transformers, it’s actually really easy to play with state of the art NLP models. We’ll use a recipe found in pytorch-transformers/examples/ as a template for our app.

First let’s install our dependencies

# Download the transformers package
pip3 install pytorch-transformers

# Get only the text generation file from the repository

Now that we have our generation script, we need to change it a bit so it plays nice with our Flask app. The script itself normally as a CLI tool with arguments, like this.

python3 ./ \
    --model_type=gpt2 \
    --length=20 \
    --model_name_or_path=gpt2 \
    --promt="Hello world"

But since we want to call the text generation function from our Flask app, some changes are needed. First let’s rename main() to generate_text() and give it some arguments. Those arguments are exactly the some as the ones we normally give if we run it directly in a shell. Here’s what it looks like

def generate_text(
    # Set the seed manually
    if n_gpu > 0:

    # The rest of the old main() code
    # We just need to replace args.* with
    # the corresponding function's arguments

That’s it! Now we’re ready to expose our feature through a REST API with Flask.

Building the Flask app

Our server will be pretty minimalistic, with only one endpoint that handle a POST request. In the body, we will provide the text which will serve as a “prompt” for GPT-2 to generate stuff. We’ll also give a model which can be one the 3 GPT-2 models, namely the small (117M), medium (345M) and large (774M).

from flask import Flask, abort, jsonify, request
from flask_cors import CORS, cross_origin

from .run_generation import generate_text

app = Flask(__name__)
cors = CORS(app)
app.config['CORS_HEADERS'] = 'Content-Type'

@app.route("/generate", methods=['POST'])
def get_gen():
    data = request.get_json()

    if 'text' not in data or len(data['text']) == 0 or 'model' not in data:
        text = data['text']
        model = data['model']

        result = generate_text(

        return jsonify({'result': result})

Front-end work

Now that we set up our Flask server, it’s time to build to front-end. We’ll have a simple form that takes a model and a textarea that inputs the text (prompt). I also used material-ui to have fancy form controls. Alright, let’s set up our React app

# Create the app
create-react-app gpt2-frontend
cd gpt2-frontend

# Add some dependencies
yarn add @material-ui/core node-sass axios

We’ll also use React Hooks to handle the state. I shamelessly copy-pasted API related boilerplate from this very thourough article. Now here’s what our App.js looks like

function App() {
  const [text, setText] = useState("");
  const [model, setModel] = useState('gpt2');
  const [generatedText, postGenerateText] = postGenerateTextEndpoint();

  const generateText = () => {
    postGenerateText({ text, model, userId: 1 });

  return (
    <div className='app-container'>
      <form noValidate autoComplete='off'>
        <h1>React GPT-2</h1>
        <SelectBox model={model} setModel={setModel} />
        <TextBox text={text} setText={setText} />
        <Button onClick={generateText} />

      {generatedText.pending &&
        <div className='result pending'>Please wait</div>}

      {generatedText.complete &&
        (generatedText.error ?
          <div className='result error'>Bad Request</div> :
          <div className='result valid'>

Grab the code from Github