Pre-trained transformer models are awesome... unless latency is in any way important. I worked for an NLP consultancy and that probably applies to like half of the use cases I worked on - anything that has to be close to real time, anything that has to process a lot of documents, and doesn't have access to 8 GPUs. At that point this guide becomes pretty obsolete.
6 comments
[ 2.0 ms ] story [ 24.1 ms ] thread