Пример #1
0
def check_wrong(output_key, w2v_key='ntua_ek'):
    mode = TEST
    path = data_config.output_path(output_key, mode, LABEL_PREDICT)
    pred = load_label_list(path)

    path = data_config.path(mode, LABEL)
    gold = load_label_list(path)

    w2v_model_path = data_config.path(ALL, WORD2VEC, w2v_key)
    vocab_train_path = data_config.path(TRAIN, VOCAB, 'ek')

    # 加载字典集
    # 在模型中会采用所有模型中支持的词向量, 并为有足够出现次数的单词随机生成词向量
    vocab_meta_list = load_vocab_list(vocab_train_path)
    vocabs = [_meta['t'] for _meta in vocab_meta_list if _meta['tf'] >= 2]

    # 加载词向量与相关数据
    lookup_table, vocab_id_mapping, embedding_dim = load_lookup_table(
        w2v_model_path=w2v_model_path, vocabs=vocabs)

    tokens_0 = load_tokenized_list(data_config.path(mode, TURN, '0.ek'))
    tokens_1 = load_tokenized_list(data_config.path(mode, TURN, '1.ek'))
    tokens_2 = load_tokenized_list(data_config.path(mode, TURN, '2.ek'))
    tid_list_0 = tokenized_to_tid_list(tokens_0, vocab_id_mapping)
    tid_list_1 = tokenized_to_tid_list(tokens_1, vocab_id_mapping)
    tid_list_2 = tokenized_to_tid_list(
        load_tokenized_list(data_config.path(mode, TURN, '2.ek')),
        vocab_id_mapping)

    max_seq_len = 0
    for p, g, tid_0, tid_1, tid_2, tk_0, tk_1, tk_2 in zip(
            pred, gold, tid_list_0, tid_list_1, tid_list_2, tokens_0, tokens_1,
            tokens_2):
        if p != g and (len(tid_0) > 30 or len(tid_1) > 30 or len(tid_2) > 30):
            print('pred: {}, gold: {}'.format(p, g))
            print('turn0: {}'.format(' '.join(tk_0)))
            print('turn1: {}'.format(' '.join(tk_1)))
            print('turn2: {}'.format(' '.join(tk_2)))

        if p != g:
            max_seq_len = max(max_seq_len, len(tid_0), len(tid_1), len(tid_2))
    print(max_seq_len)
Пример #2
0
def train(dataset_key,
          text_version,
          label_version=None,
          config_path='config.yaml'):
    """
    python algo/main.py train semeval2018_task3 -l A -t ek
    python algo/main.py train semeval2018_task3 -l A -t ek -c config_ntua.yaml
    python algo/main.py train semeval2018_task3 -l A -t raw -c config_ntua_char.yaml

    python algo/main.py train semeval2019_task3_dev -t ek

    python algo/main.py train semeval2018_task1 -l love
    python algo/main.py train semeval2014_task9

    :param dataset_key: string
    :param text_version: string
    :param label_version: string
    :param config_path: string
    :return:
    """
    pos_label = None
    if dataset_key == 'semeval2018_task3' and label_version == 'A':
        pos_label = 1

    config_data = yaml.load(open(config_path))

    data_config = getattr(
        importlib.import_module('dataset.{}.config'.format(dataset_key)),
        'config')

    output_key = '{}_{}_{}'.format(config_data['module'].rsplit('.', 1)[1],
                                   text_version, int(time.time()))
    if label_version is not None:
        output_key = '{}_{}'.format(label_version, output_key)
    print('OUTPUT_KEY: {}'.format(output_key))

    # 准备输出路径的文件夹
    data_config.prepare_output_folder(output_key=output_key)
    data_config.prepare_model_folder(output_key=output_key)

    shutil.copy(config_path, data_config.output_path(output_key, ALL, CONFIG))

    # 根据配置加载模块
    module_relative_path = config_data['module']
    NNModel = getattr(importlib.import_module(module_relative_path), 'NNModel')
    NNConfig = getattr(importlib.import_module(module_relative_path),
                       'NNConfig')

    if config_data['analyzer'] == WORD:
        w2v_key = '{}_{}'.format(config_data['word']['w2v_version'],
                                 text_version)
        w2v_model_path = data_config.path(ALL, WORD2VEC, w2v_key)
        vocab_train_path = data_config.path(TRAIN, VOCAB, text_version)

        # 加载字典集
        # 在模型中会采用所有模型中支持的词向量, 并为有足够出现次数的单词随机生成词向量
        vocab_meta_list = load_vocab_list(vocab_train_path)
        vocab_meta_list += load_vocab_list(
            semeval2018_task3_date_config.path(TRAIN, VOCAB, text_version))
        vocabs = [
            _meta['t'] for _meta in vocab_meta_list
            if _meta['tf'] >= config_data[WORD]['min_tf']
        ]

        # 加载词向量与相关数据
        lookup_table, vocab_id_mapping, embedding_dim = load_lookup_table(
            w2v_model_path=w2v_model_path, vocabs=vocabs)
        json.dump(
            vocab_id_mapping,
            open(data_config.output_path(output_key, ALL, VOCAB_ID_MAPPING),
                 'w'))
        max_seq_len = MAX_WORD_SEQ_LEN
    elif config_data['analyzer'] == CHAR:
        texts = load_text_list(data_config.path(TRAIN, TEXT))
        char_set = set()
        for text in texts:
            char_set |= set(text)
        lookup_table, vocab_id_mapping, embedding_dim = build_random_lookup_table(
            vocabs=char_set, dim=config_data['char']['embedding_dim'])
        max_seq_len = MAX_CHAR_SEQ_LEN
    else:
        raise ValueError('invalid analyzer: {}'.format(
            config_data['analyzer']))

    # 加载训练数据
    datasets, output_dim = load_dataset(data_config=data_config,
                                        analyzer=config_data['analyzer'],
                                        vocab_id_mapping=vocab_id_mapping,
                                        seq_len=max_seq_len,
                                        with_label=True,
                                        label_version=label_version,
                                        text_version=text_version)

    # 加载配置
    nn_config = NNConfig(config_data)
    train_config = TrainConfig(config_data['train'])

    # 初始化数据集的检索
    index_iterators = {
        mode: IndexIterator(datasets[mode][LABEL_GOLD])
        for mode in [TRAIN, TEST]
    }
    # 按配置将训练数据切割成训练集和验证集
    index_iterators[TRAIN].split_train_valid(train_config.valid_rate)

    # 计算各个类的权重
    if train_config.use_class_weights:
        label_weight = {
            # 参考 sklearn 中 class_weight='balanced'的公式, 实验显示效果显着
            _label: float(index_iterators[TRAIN].n_sample()) /
            (index_iterators[TRAIN].dim * len(_index))
            for _label, _index in index_iterators[TRAIN].label_index.items()
        }
    else:
        label_weight = {
            _label: 1.
            for _label in range(index_iterators[TRAIN].dim)
        }

    # 基于加载的数据更新配置
    nn_config.set_embedding_dim(embedding_dim)
    nn_config.set_output_dim(output_dim)
    nn_config.set_seq_len(max_seq_len)
    # 搭建神经网络
    nn = NNModel(config=nn_config)
    nn.build_neural_network(lookup_table=lookup_table)

    batch_size = train_config.batch_size
    fetches = {
        mode: {_key: nn.var(_key)
               for _key in fetch_key[mode]}
        for mode in [TRAIN, TEST]
    }
    last_eval = {TRAIN: None, VALID: None, TEST: None}

    model_output_prefix = data_config.model_path(key=output_key) + '/model'

    best_res = {mode: None for mode in [TRAIN, VALID]}
    no_update_count = {mode: 0 for mode in [TRAIN, VALID]}
    max_no_update_count = 10

    with tf.Session() as sess:
        sess.run(tf.global_variables_initializer())
        saver = tf.train.Saver(tf.global_variables())

        dataset = datasets[TRAIN]
        index_iterator = index_iterators[TRAIN]

        # 训练开始 ##########################################################################
        for epoch in range(train_config.epoch):
            print('== epoch {} =='.format(epoch))

            # 利用训练集进行训练
            print('TRAIN')
            n_sample = index_iterator.n_sample(TRAIN)
            labels_predict = list()
            labels_gold = list()

            for batch_index in index_iterator.iterate(batch_size,
                                                      mode=TRAIN,
                                                      shuffle=True):
                feed_dict = {
                    nn.var(_key): dataset[_key][batch_index]
                    for _key in feed_key[TRAIN]
                }
                feed_dict[nn.var(SAMPLE_WEIGHTS)] = list(
                    map(label_weight.get, feed_dict[nn.var(LABEL_GOLD)]))
                feed_dict[nn.var(TEST_MODE)] = 0
                res = sess.run(fetches=fetches[TRAIN], feed_dict=feed_dict)

                labels_predict += res[LABEL_PREDICT].tolist()
                labels_gold += dataset[LABEL_GOLD][batch_index].tolist()

            labels_predict, labels_gold = labels_predict[:
                                                         n_sample], labels_gold[:
                                                                                n_sample]
            labels_predict, labels_gold = labels_predict[:
                                                         n_sample], labels_gold[:
                                                                                n_sample]
            res = basic_evaluate(gold=labels_gold,
                                 pred=labels_predict,
                                 pos_label=pos_label)
            last_eval[TRAIN] = res
            print_evaluation(res)

            global_step = tf.train.global_step(sess, nn.var(GLOBAL_STEP))

            if train_config.valid_rate == 0.:
                if best_res[TRAIN] is None or res[F1_SCORE] > best_res[TRAIN][
                        F1_SCORE]:
                    best_res[TRAIN] = res
                    no_update_count[TRAIN] = 0
                    saver.save(sess,
                               save_path=model_output_prefix,
                               global_step=global_step)
                else:
                    no_update_count[TRAIN] += 1
            else:
                if best_res[TRAIN] is None or res[F1_SCORE] > best_res[TRAIN][
                        F1_SCORE]:
                    best_res[TRAIN] = res
                    no_update_count[TRAIN] = 0
                else:
                    no_update_count[TRAIN] += 1

                # 计算在验证集上的表现, 不更新模型参数
                print('VALID')
                n_sample = index_iterator.n_sample(VALID)
                labels_predict = list()
                labels_gold = list()

                for batch_index in index_iterator.iterate(batch_size,
                                                          mode=VALID,
                                                          shuffle=False):
                    feed_dict = {
                        nn.var(_key): dataset[_key][batch_index]
                        for _key in feed_key[TEST]
                    }
                    feed_dict[nn.var(TEST_MODE)] = 1
                    res = sess.run(fetches=fetches[TEST], feed_dict=feed_dict)
                    labels_predict += res[LABEL_PREDICT].tolist()
                    labels_gold += dataset[LABEL_GOLD][batch_index].tolist()

                labels_predict, labels_gold = labels_predict[:
                                                             n_sample], labels_gold[:
                                                                                    n_sample]
                res = basic_evaluate(gold=labels_gold,
                                     pred=labels_predict,
                                     pos_label=pos_label)
                last_eval[VALID] = res
                print_evaluation(res)

                # Early Stop
                if best_res[VALID] is None or res[F1_SCORE] > best_res[VALID][
                        F1_SCORE]:
                    saver.save(sess,
                               save_path=model_output_prefix,
                               global_step=global_step)
                    best_res[VALID] = res
                    no_update_count[VALID] = 0
                else:
                    no_update_count[VALID] += 1

            if no_update_count[TRAIN] >= max_no_update_count:
                break

        # 训练结束 ##########################################################################
        # 确保输出文件夹存在

    print(
        '========================= BEST ROUND EVALUATION ========================='
    )

    with tf.Session() as sess:
        prefix_checkpoint = tf.train.latest_checkpoint(
            data_config.model_path(key=output_key))
        saver = tf.train.import_meta_graph('{}.meta'.format(prefix_checkpoint))
        saver.restore(sess, prefix_checkpoint)

        nn = BaseNNModel(config=None)
        nn.set_graph(tf.get_default_graph())

        for mode in [TRAIN, TEST]:
            dataset = datasets[mode]
            index_iterator = index_iterators[mode]
            n_sample = index_iterator.n_sample()

            prob_predict = list()
            labels_predict = list()
            labels_gold = list()
            hidden_feats = list()

            for batch_index in index_iterator.iterate(batch_size,
                                                      shuffle=False):
                feed_dict = {
                    nn.var(_key): dataset[_key][batch_index]
                    for _key in feed_key[TEST]
                }
                feed_dict[nn.var(TEST_MODE)] = 1
                res = sess.run(fetches=fetches[TEST], feed_dict=feed_dict)
                prob_predict += res[PROB_PREDICT].tolist()
                labels_predict += res[LABEL_PREDICT].tolist()
                hidden_feats += res[HIDDEN_FEAT].tolist()
                labels_gold += dataset[LABEL_GOLD][batch_index].tolist()

            prob_predict = prob_predict[:n_sample]
            labels_predict = labels_predict[:n_sample]
            labels_gold = labels_gold[:n_sample]
            hidden_feats = hidden_feats[:n_sample]

            if mode == TEST:
                res = basic_evaluate(gold=labels_gold,
                                     pred=labels_predict,
                                     pos_label=pos_label)
                best_res[TEST] = res

            # 导出隐藏层
            with open(data_config.output_path(output_key, mode, HIDDEN_FEAT),
                      'w') as file_obj:
                for _feat in hidden_feats:
                    file_obj.write('\t'.join(map(str, _feat)) + '\n')
            # 导出预测的label
            with open(data_config.output_path(output_key, mode, LABEL_PREDICT),
                      'w') as file_obj:
                for _label in labels_predict:
                    file_obj.write('{}\n'.format(_label))
            with open(data_config.output_path(output_key, mode, PROB_PREDICT),
                      'w') as file_obj:
                for _prob in prob_predict:
                    file_obj.write('\t'.join(map(str, _prob)) + '\n')

        for mode in [TRAIN, VALID, TEST]:
            if mode == VALID and train_config.valid_rate == 0.:
                continue
            res = best_res[mode]
            print(mode)
            print_evaluation(res)

            json.dump(
                res,
                open(data_config.output_path(output_key, mode, EVALUATION),
                     'w'))
            print()

    print('OUTPUT_KEY: {}'.format(output_key))