A simple Python package that wraps existing model fine-tuning and generation scripts for OpenAI's GPT-2 text generation model (specifically the "small" 124M and "medium" 355M hyperparameter versions). Additionally, this package allows easier generation of text, generating to a file for easy curation, allowing for prefixes to force the text to start with a given phrase. For finetuning, it is strongly recommended to use a GPU, although you can generate using a CPU (albeit much more slowly). If you are training in the cloud, using a Colaboratory notebook or a Google Compute Engine VM w/ the TensorFlow Deep Learning image is strongly recommended. (as the GPT-2 model is hosted on GCP) You can use gpt-2-simple to retrain a model using a GPU for free in this Colaboratory notebook, which also demos additional features of the package. Note: Development on gpt-2-simple has mostly been superceded by aitextgen, which has similar AI text generation capabilities with more efficient training time.
Features
- Model management from OpenAI's official GPT-2 repo (MIT License)
- Model finetuning from Neil Shepperd's fork of GPT-2 (MIT License)
- Text generation output management from textgenrnn (MIT License / also created by me)
- gpt-2-simple can be installed via PyPI
- The original GPT-2 model was trained on a very large variety of sources, allowing the model to incorporate idioms not seen in the input text
- GPT-2 can only generate a maximum of 1024 tokens per request (about 3-4 paragraphs of English text)
- A simple Python package that wraps existing model fine-tuning and generation scripts for OpenAI's GPT-2 text generation model