Derived from this project written in Theano.
At this stage this is a rough draft, tested only with a single type of "punctuation" (actually, sentence boundaries). However, it is easily adapted to any number of punctuation markers.
Also added a little script to infer and visualize the attention weights for any sentence fed to a trained model.
Hyperparameters and paths to data and checkpoints are written in a bash file to be sourced before running the script, e.g.
source env.sh
python 01_train.python
See example_data to see how the data should be formatted.
Preprocessing scripts are not provided in this version.