experiments to run

• baseline

• 2 layers for encoder

• run on bpe'd but not tokenized data

• linear layer between embeddings and hidden

• averaging hidden states for init to decoder

• beam search

• once beam search is working, run on morphgen-ed data