The purpose of the project is to implement this paper and use the network to generate text.
Setup is initially based on this PyTorch language model example.
See more info about the course here.
The project is developed using Python 3.6 and PyTorch 0.4.
Training a language model is done in LanguageModel.ipynb and text generation can be done in either GenerateText.ipynb (for multinomial sampling) or GenerateTextBeamSearch.ipynb (for beam search).
Choosing the dataset is done in the notebooks. Here you can also choose which configuration to use. This points to an entry in config.yml that can easily be edited (or a new configuration can be added).
Models are saved using the name specified in config.yml every time a new best validation perplexity is reached. This allows easy loading for text generation, for example.