For bugs and other issues see the repository issues on GitHub.
-
Load training data from a file instead of storing it in memory to allow for quick resume of training without needing to regenerate the dataset from the source corpus.
-
Read training data from a file in chunks to allow for datasets beyond the size of available RAM.
-
Output attention alignments correctly in tensorboard summary logs.