To Do

baseline:

• attn decoder: cuda + minibatch capable
• cross-entropy loss
• initialize decoder with mean encoder hidden instead of last (? i vote try with and without)
• linear between embeds and hidden (personally I'd like to try this with and without to compare)
• maxibatches
• beam search (beam size=5)

not baseline:

• morph-tag data, bpe it