The project heavily utilises GPU parallelism, so it is assumed that you have PyTorch with CUDA functionality installed and working.
Download and place the data in the .data folder. Download and place
pre-trained vector embeddings in the .vector_cache folder. Create
.serialization data folder where training data, together with the best/latest
trained models will be placed.
Currently, there is no GUI or command line interface. If you want to change the architecture/word embeddings, you have to modify the code. Thankfully, I have left comments pointing where to edit and examples how to edit.
Note that, the class implementing Rocktaschel et al's attention networks
RocktaschelEtAlAttention has word_by_word boolean flag in the constructor,
with which you can control what attention to use. Similarly,
RocktaschelEtAlConditionalEncoding has use_fastgrnn flag to switch between
gating mechanism.
For performing the stress tests use the instructions from
here. You can use
tensorboard to visualise results from .serialization_data folder.