parse02.py 文件源码

def train(args):
  trace('loading corpus ...')
  with open(args.source) as fp:
    trees = [make_tree(l) for l in fp]

  trace('extracting leaf nodes ...')
  word_lists = [extract_words(t) for t in trees]

  trace('extracting gold operations ...')
  op_lists = [make_operations(t) for t in trees]

  trace('making vocabulary ...')
  word_vocab = Vocabulary.new(word_lists, args.vocab)
  phrase_set = set()
  semi_set = set()
  for tree in trees:
    phrase_set |= set(extract_phrase_labels(tree))
    semi_set |= set(extract_semi_labels(tree))
  phrase_vocab = Vocabulary.new([list(phrase_set)], len(phrase_set), add_special_tokens=False)
  semi_vocab = Vocabulary.new([list(semi_set)], len(semi_set), add_special_tokens=False)

  trace('converting data ...')
  word_lists = [convert_word_list(x, word_vocab) for x in word_lists]
  op_lists = [convert_op_list(x, phrase_vocab, semi_vocab) for x in op_lists]

  trace('start training ...')
  parser = Parser(
      args.vocab, args.embed, args.queue, args.stack,
      len(phrase_set), len(semi_set),
  )
  if USE_GPU:
    parser.to_gpu()
  opt = optimizers.AdaGrad(lr = 0.005)
  opt.setup(parser)
  opt.add_hook(optimizer.GradientClipping(5))

  for epoch in range(args.epoch):
    n = 0

    for samples in batch(zip(word_lists, op_lists), args.minibatch):
      parser.zerograds()
      loss = my_zeros((), np.float32)

      for word_list, op_list in zip(*samples):
        trace('epoch %3d, sample %6d:' % (epoch + 1, n + 1))
        loss += parser.forward(word_list, op_list, 0)
        n += 1

      loss.backward()
      opt.update()

    trace('saving model ...')
    prefix = args.model + '.%03.d' % (epoch + 1)
    word_vocab.save(prefix + '.words')
    phrase_vocab.save(prefix + '.phrases')
    semi_vocab.save(prefix + '.semiterminals')
    parser.save_spec(prefix + '.spec')
    serializers.save_hdf5(prefix + '.weights', parser)

  trace('finished.')