src_ph = tf.placeholder('int32', [None, None])
prev_tokens = tf.placeholder_with_default([out_voc.bos_ix], [None])
h0 = model.encode(src_ph)
prev_state_ph = [tf.placeholder(var.dtype, var.shape) for var in h0]
new_state, new_logits = model.decode(prev_state_ph, prev_tokens)

def translate_line(src):
  _state = sess.run(h0, {src_ph: inp_voc.to_matrix([src])})
  tokens = [out_voc.bos_ix]
  output = []
  for i in range(100):
    output.append(out_voc.tokens[tokens[0]])
    feed = {prev_tokens: tokens}
    for var, val in zip(prev_state_ph, _state):
      feed[var] = val
    _state, _logits = sess.run([new_state, new_logits], feed)
    tokens = _logits.argmax(-1)
    if tokens[0] == model.out_voc.eos_ix:
      break
  return ' '.join(output)