Hugging Face, a startup specializing in natural language processing, today released a landmark update to their popular Transformers library, offering unprecedented compatibility between two major deep learning frameworks, PyTorch and TensorFlow 2.0.
Transformers 2.0 embraces the “best of both worlds”, combining PyTorch’s ease of use with TensorFlow’s production-grade ecosystem. The new library makes it easier for scientists and practitioners to select different frameworks for the training, evaluation and production phases of developing the same language model.
“This is a lot deeper than what people usually think when they talk about compatibility,” said Thomas Wolf, who leads Hugging Face’s science team. “It’s not only about being able to use the library separately in PyTorch and TensorFlow. We’re talking about being able to seamlessly move from one framework to the other dynamically during the life of the model.”
“It’s the number one feature that companies asked for since the launch of the library last year” said Clement Delangue, CEO of Hugging Face.
Notable features of the new release:
- 8 architectures with over 30 pretrained models, some in more than 100 languages.
- Load a model and pre-process a dataset in less than 10 lines of code.
- Train a state-of-the-art language model in a single line with the tf.keras fit function.
- Share pretrained models, reducing compute costs and carbon footprint.
About Hugging Face Transformers
With half a million installs since January 2019, Transformers is the most popular open-source NLP library. More than 1,000 companies including Bing, Apple or Stichfix are using it in production for text classification, question-answering, intent detection, text generation or conversational. Hugging Face, the creators of Transformers, have raised US$5M so far from Betaworks; Ronny Conway; deep learning directors at Salesforce, Amazon, Apple; and Kevin Durant.
(Press release provided by Hugging Face)