Toy implementation of the paper Sentence-State LSTM for Text Representation using PyTorch (only for classification) based on the tensorflow implemetation of the author
Right now there are two different models implemented. The S-LSTM model and a vanilla bidirectional RNN as a baseline. In addition, there is a simple attention module that can be put on top of either model.
At the moment only the ability of loading the Stanford sentiment treebank (all versions) via torchtext is available. So, in general adding support for any additional datasets should be straight forward enough with torchtext
One can use tensorboard-pytorch for plotting accuracies etc. More info can be found in this blogpost
the implementation seems kinda slow at the moment and because I don't have an access to a gpu right now I haven't benchmarked it there. Also, I have done almost zero hyperparameter tuning.