Implemented the state-of-the-art style transfer Delete Retrieve Generate (https://arxiv.org/abs/1804.06437) paper from scratch.

Implemented all 4 models in the paper: DeleteOnly, DeleteAndRetrieve, TemplateBased, RetrieveOnly. Goal of this is to build everything from scratch (including e.g. ngram generation, beam search algorithms, etc...) with minimal help from the internet (& put #comments for where I did use the internet).

Learned (intuitively understood from first principles, not just memorize & paste some code):

• efficient ngrams generation
• TF-IDF (in this case TF-IDF word overlap distance function)
• style embeddings
• concatening latent space with embeddings
• Beam search (this took SOO LONG since i had to implement from scratch, was def worth the struggle, feel like my brain physically grew after)
• BLEU scoring