Transformers: a pretrained model becomes an import
On 17 November 2018 Hugging Face released the library first named pytorch-pretrained-bert. It gave access to pretrained transformer models through one interface, two lines of code instead of an implementation of your own.
Why it matters
This is where a pretrained model stopped being a research artefact and became a project dependency. The atlas holds TensorFlow, PyTorch, Theano, Caffe and Keras, libraries in which models are written. This is the first one in which models are simply taken.
The 2019 paper states the intent: state-of-the-art Transformer architectures under a unified API with a curated collection of pretrained models, extensible for researchers, simple for practitioners, and fast and robust in industrial deployment. There are twenty-two authors, the first being Thomas Wolf. The date is easy to get wrong, and the error costs two years. The name transformers has existed on the Python package index since 17 August 2016, but that version 0.1 has nothing to do with this library: Hugging Face appears under that name only from version 2.0.0 of 26 September 2019. The real chain is pytorch-pretrained-bert from 17 November 2018, then pytorch-transformers from 5 July 2019, then transformers from 26 September 2019. The record sits on the first of those, because that is when the library became available. The repository was created on 29 October 2018, three weeks before the first release.