4172 shaares
Very nice progress on this type of architecture. It's definitely needed in part because it lowers the inference cost quite a lot. It's also nice to see it released with under the Apache 2 license and the training set be documented.
