def pred_tet(path_hyper_parameter=path_hyper_parameters,
             path_test=None,
             rate=1.0):
    """
        测试集测试与模型评估
    :param hyper_parameters: json, 超参数
    :param path_test:str, path of test data, 测试集
    :param rate: 比率, 抽出rate比率语料取训练
    :return: None
    """
    hyper_parameters = load_json(path_hyper_parameter)
    if path_test:  # 从外部引入测试数据地址
        hyper_parameters['data']['val_data'] = path_test
    time_start = time.time()
    # graph初始化
    graph = Graph(hyper_parameters)
    print("graph init ok!")
    graph.load_model()
    print("graph load ok!")
    ra_ed = graph.word_embedding
    # 数据预处理
    pt = PreprocessText()
    y, x = read_and_process(hyper_parameters['data']['val_data'])
    # 取该数据集的百分之几的语料测试
    len_rate = int(len(y) * rate)
    x = x[1:len_rate]
    y = y[1:len_rate]
    y_pred = []
    count = 0
    for x_one in x:
        count += 1
        ques_embed = ra_ed.sentence2idx(x_one)
        if hyper_parameters['embedding_type'] == 'bert':  # bert数据处理, token
            x_val_1 = np.array([ques_embed[0]])
            x_val_2 = np.array([ques_embed[1]])
            x_val = [x_val_1, x_val_2]
        else:
            x_val = ques_embed
        # 预测
        pred = graph.predict(x_val)
        pre = pt.prereocess_idx(pred[0])
        label_pred = pre[0][0][0]
        if count % 1000 == 0:
            print(label_pred)
        y_pred.append(label_pred)

    print("data pred ok!")
    # 预测结果转为int类型
    index_y = [pt.l2i_i2l['l2i'][i] for i in y]
    index_pred = [pt.l2i_i2l['l2i'][i] for i in y_pred]
    target_names = [
        pt.l2i_i2l['i2l'][str(i)] for i in list(set((index_pred + index_y)))
    ]
    # 评估
    report_predict = classification_report(index_y,
                                           index_pred,
                                           target_names=target_names,
                                           digits=9)
    print(report_predict)
    print("耗时:" + str(time.time() - time_start))
示例#2
0
def train(hyper_parameters=None, rate=1.0):
    if not hyper_parameters:
        hyper_parameters = {
        'len_max': 2,  # 句子最大长度, 固定推荐20-50, bert越长会越慢, 占用空间也会变大, 小心OOM
        'embed_size': 768,  # 字/词向量维度, bert取768, word取300, char可以更小些
        'vocab_size': 20000,  # 这里随便填的,会根据代码里修改
        'trainable': False,  # embedding是静态的还是动态的, 即控制可不可以微调
        'level_type': 'char',  # 级别, 最小单元, 字/词, 填 'char' or 'word', 注意:word2vec模式下训练语料要首先切好
        'embedding_type': 'bert',  # 级别, 嵌入类型, 还可以填'xlnet'、'random'、 'bert'、 'albert' or 'word2vec"
        'gpu_memory_fraction': 0.86, #gpu使用率
        'model': {'label': 17,  # 类别数
                  'batch_size': 2,  # 批处理尺寸, 感觉原则上越大越好,尤其是样本不均衡的时候, batch_size设置影响比较大
                  'dropout': 0.5,  # 随机失活, 概率
                  'decay_step': 1000,  # 学习率衰减step, 每N个step衰减一次
                  'decay_rate': 0.9,  # 学习率衰减系数, 乘法
                  'epochs': 20,  # 训练最大轮次
                  'patience': 3, # 早停,2-3就好
                  'lr': 2e-5,  # 学习率,bert取5e-5,其他取1e-3, 对训练会有比较大的影响, 如果准确率一直上不去,可以考虑调这个参数
                  'l2': 1e-9,  # l2正则化
                  'activate_classify': 'softmax',  # 最后一个layer, 即分类激活函数
                  'loss': 'categorical_crossentropy',  # 损失函数
                  'metrics': 'accuracy',  # 保存更好模型的评价标准
                  'is_training': True,  # 训练后者是测试模型
                  'model_path': path_model,
                  # 模型地址, loss降低则保存的依据, save_best_only=True, save_weights_only=True
                  'path_hyper_parameters': path_hyper_parameters,  # 模型(包括embedding),超参数地址,
                  'path_fineture': path_fineture,  # 保存embedding trainable地址, 例如字向量、词向量、bert向量等
                  },
        'embedding': {'layer_indexes': [24], # bert取的层数
                      # 'corpus_path': '',     # embedding预训练数据地址,不配则会默认取conf里边默认的地址, keras-bert可以加载谷歌版bert,百度版ernie(需转换,https://github.com/ArthurRizar/tensorflow_ernie),哈工大版bert-wwm(tf框架,https://github.com/ymcui/Chinese-BERT-wwm)
                        },
        'data':{'train_data': path_baidu_qa_2019_train, # 训练数据
                'val_data': path_baidu_qa_2019_valid    # 验证数据
                },
    }

    # 删除先前存在的模型\embedding微调模型等
    delete_file(path_model_dir)
    time_start = time.time()
    # graph初始化
    graph = Graph(hyper_parameters)
    print("graph init ok!")
    ra_ed = graph.word_embedding
    # 数据预处理
    pt = PreprocessText()
    x_train, y_train = pt.preprocess_label_ques_to_idx(hyper_parameters['embedding_type'],
                                                       hyper_parameters['data']['train_data'],
                                                       ra_ed, rate=rate, shuffle=True)
    x_val, y_val = pt.preprocess_label_ques_to_idx(hyper_parameters['embedding_type'],
                                                   hyper_parameters['data']['val_data'],
                                                   ra_ed, rate=rate, shuffle=True)
    print("data propress ok!")
    print(len(y_train))
    # 训练
    graph.fit(x_train, y_train, x_val, y_val)
    print("耗时:" + str(time.time()-time_start))
示例#3
0
def pred_input(path_hyper_parameter=path_hyper_parameters):
    """
       输入预测
    :param path_hyper_parameter: str, 超参存放地址
    :return: None
    """
    # 加载超参数
    hyper_parameters = load_json(path_hyper_parameter)
    pt = PreprocessText(path_model_dir)
    # 模式初始化和加载
    graph = Graph(hyper_parameters)
    graph.load_model()
    ra_ed = graph.word_embedding
    ques = '我要打王者荣耀'
    # str to token
    ques_embed = ra_ed.sentence2idx(ques)
    if hyper_parameters['embedding_type'] in ['bert', 'albert']:
        x_val_1 = np.array([ques_embed[0]])
        x_val_2 = np.array([ques_embed[1]])
        x_val = [x_val_1, x_val_2]
    else:
        x_val = ques_embed
    # 预测
    pred = graph.predict(x_val)
    # 取id to label and pred
    pre = pt.prereocess_idx(pred[0])
    print(pre)
    while True:
        print("请输入: ")
        ques = input()
        ques_embed = ra_ed.sentence2idx(ques)
        print(ques_embed)
        if hyper_parameters['embedding_type'] in ['bert', 'albert']:
            x_val_1 = np.array([ques_embed[0]])
            x_val_2 = np.array([ques_embed[1]])
            x_val = [x_val_1, x_val_2]
        else:
            x_val = ques_embed
        pred = graph.predict(x_val)
        pre = pt.prereocess_idx(pred[0])
        print(pre)
            'lr': 1e-3,
            'l2': 1e-6,
            'activate_classify': 'softmax',
            'embedding_type': 'random',
            'is_training': True,
            'model_path': path_model_fast_text_baiduqa_2019,
        },
        'embedding': {
            'embedding_type': 'random',
            'corpus_path': path_embedding_random_char,
            'level_type': 'char',
            'embed_size': 300,
            'len_max': 50,
        },
    }
    graph = Graph(hyper_parameters)
    ra_ed = graph.word_embedding
    pt = PreprocessText()
    x_train, y_train = pt.preprocess_baidu_qa_2019_idx(
        path_baidu_qa_2019_train, ra_ed)
    x_val, y_val = pt.preprocess_baidu_qa_2019_idx(path_baidu_qa_2019_valid,
                                                   ra_ed)
    print(len(y_train))
    graph.fit(x_train, y_train, x_val, y_val)

# 1425170/1425170 [==============================] - 83s 58us/step - loss: 0.9383 - acc: 0.7106 - val_loss: 2.4205 - val_acc: 0.5029
# Epoch 00001: val_loss improved from inf to 2.42050, saving model to D:\workspace\pythonMyCode\django_project\ClassificationTextChinese/data/model/fast_text/model_fast_text.f5
# Epoch 2/20
# 验证集准确率50%左右
# time时间大约在2*4轮=8分钟左右
示例#5
0
def train(hyper_parameters=None, rate=1.0):
    if not hyper_parameters:
        hyper_parameters = {
        'len_max': 56,  # 句子最大长度, 固定推荐20-50, bert越长会越慢, 占用空间也会变大, 小心OOM
        'embed_size': 300,  # 字/词向量维度, bert取768, word取300, char可以更小些
        'vocab_size': 20000,  # 这里随便填的,会根据代码里修改
        'trainable': True,  # embedding是静态的还是动态的, 即控制可不可以微调
        'level_type': 'char',  # 级别, 最小单元, 字/词, 填 'char' or 'word', 注意:word2vec模式下训练语料要首先切好
        'embedding_type': 'random',  # 级别, 嵌入类型, 还可以填'xlnet'、'random'、 'bert'、 'albert' or 'word2vec"
        # 'gpu_memory_fraction': 0.76, #gpu使用率
        'model': {'label': 17,  # 类别数
                  'batch_size': 256,  # 批处理尺寸, 感觉原则上越大越好,尤其是样本不均衡的时候, batch_size设置影响比较大
                  'dropout': 0.5,  # 随机失活, 概率
                  'decay_step': 1000,  # 学习率衰减step, 每N个step衰减一次
                  'decay_rate': 0.999,  # 学习率衰减系数, 乘法
                  'filters': [3, 7, 7],
                  'filters_num': 300,  # 卷积个数 论文中 filters_num=150,300
                  'epochs': 20,  # 训练最大轮次
                  'patience': 3, # 早停,2-3就好
                  'lr': 1e-3,  # 学习率,bert取5e-5,其他取1e-3, 对训练会有比较大的影响, 如果准确率一直上不去,可以考虑调这个参数
                  'l2': 1e-9,  # l2正则化
                  'activate_classify': 'softmax',  # 最后一个layer, 即分类激活函数
                  'loss': 'categorical_crossentropy',  # 损失函数
                  'metrics': 'accuracy',  # 保存更好模型的评价标准
                  'optimizer_name': 'Adam', # 优化器, 可选['Adam', 'Radam', 'RAdam,Lookahead'], win10下必须使用GPU, 原因未知
                  'is_training': True,  # 训练后者是测试模型
                  'path_model_dir': path_model_dir, # 模型目录
                  'model_path': path_model,
                  # 模型地址, loss降低则保存的依据, save_best_only=True, save_weights_only=True
                  'path_hyper_parameters': path_hyper_parameters,  # 模型(包括embedding),超参数地址,
                  'path_fineture': path_fineture,  # 保存embedding trainable地址, 例如字向量、词向量、bert向量等
                  },
        'embedding': {'layer_indexes': [24], # bert取的层数
                      # 'ngram_ns': [3],
                      # 'corpus_path': path_baidu_qa_2019_train,
                        },
        'data':{'train_data': path_ccks_2020_el_cls_train, # 训练数据
                'val_data': path_ccks_2020_el_cls_dev    # 验证数据
                },
    }

    # 删除先前存在的模型\embedding微调模型等
    delete_file(path_model_dir)
    time_start = time.time()
    # graph初始化
    graph = Graph(hyper_parameters)
    print("graph init ok!")
    ra_ed = graph.word_embedding
    # 数据预处理
    pt = PreprocessSim(path_model_dir)
    x_train, y_train = pt.preprocess_label_ques_to_idx(hyper_parameters['embedding_type'],
                                                       hyper_parameters['data']['train_data'],
                                                       ra_ed, rate=rate, shuffle=True)
    x_val, y_val = pt.preprocess_label_ques_to_idx(hyper_parameters['embedding_type'],
                                                   hyper_parameters['data']['val_data'],
                                                   ra_ed, rate=rate, shuffle=True)
    print("data propress ok!")
    print(len(y_train))
    # 训练
    graph.fit(x_train, y_train, x_val, y_val)
    # 训练
    # graph.fit_generator(embed=ra_ed, rate=rate)
    print("耗时:" + str(time.time()-time_start))