TODO

switch to tensorflow 1.0:              done
replicate the original results:        done
add prediction:                        done
add evaluation on test set by eval.pl: done
refactor code with more encapsulation: done
add checkpointing:                     done
add logging:                           done/simple enough for tensorboard graphs
implement model without bicond_tf10:   done
early_stopping:                        not done/not for course project
tfdbg:                                 not done/not stable
tf.train.Saver vs
  tf.summary.FileWriter:               not done/not for course project
F1 measure:                            not done
can we optimize wrt to F1?:            not done/not for course project
Rocktaschel et al. (2016):             done
matching LSTM:                         done
Xavier init:                           not done
add dropout:                           not done
implement model without utils:         not done
bidirectional + Bengio Attention:      not done
two stage classification:              not done
unit testing with numpy:               bad idea



Note: it is possible to compute lengths of sequences inside the graph using
tf.nn.embedding_lookup_sparse.