Refactored and Abstracted the GPT-2 model code while putting it beside a flask API and dockerized for easy deployment.

Code and samples from the paper "Language Models are Unsupervised Multitask Learners".

See more details in blog post.

Download the model data (needs gsutil):

• Download the model (Linux): sh download_model.sh 117M
• Download the model (Windows - Bash): sh download_model_windows.sh 117M

In case it fails to download the models/117M/checkpoint file, just create checkpoint file at models/117M/ and add these lines:

model_checkpoint_path: "model.ckpt"
all_model_checkpoint_paths: "model.ckpt"

Build the docker image:

docker-compose build

Start the services

docker-compose up

visit the external IP/localhost to see the live demo.