Do you want to generate a bunch of text with the enormous GPT-2 architecture? Then you got this powerful Python package aitextgen to do it for you at a very low cost of just $0.382/hr.

Yep, the aitextgen is a highly competent Python package that uses OpenAI's GPT-2 architecture, a massive transformer-based language model that could predict the next word in 40GB of Internet text.

GPT (GUID Partition Table), a Global Unique Identification Partition table used as a standard layout method to denote the physical computer storage device. GPT- 2 is a scaled-up form of GPT.

GPT-2 model with 1.5 billion parameters was typically trained on a massive dataset of 8 million web pages. GPT-2 is competent with a modest objective of predicting the next word, provided all of the previous words within some text.

Key Highlights

• aitextgen powerup the functionalities of Pytorch, Huggingface Transformers and Pytorch-lightening with precise optimisations of text generations comprising the best functions of textgenrnn and gpt-2-simple packages.

• It is capable of generating a ton of text much faster with utmost memory efficiency.

• It can fine-tune on a pretrained 124 million GPT-2 architecture from open AI.

• Anyone can create a very own GPT-2 model + tokeniser and could train it from scratch.

• The aitextgen is designed to preserve the compatibility with its base package of the transformers. That could facilitate the usage of custom models of various NLP tasks.

• The generate() function in this architecture provides massive control on the ton of text generated.

• aitextgen can train models on CPU, GPU, multiple GPUs and even on TPUs! Where it can add loggers optionally on the training progress bar.

• Here, input dataset is it's very own object, which helps in merging or cross-train multiple datasets to create a blended resulting dataset.

References:

• Website

• GitHub