def train(hyper_parameters=None, rate=1.0):
if not hyper_parameters:
hyper_parameters = {
'len_max': 50, # 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, 本地win10-4G设为20就好, 过大小心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'random'、 'bert' or 'word2vec"
'gpu_memory_fraction': 0.66, #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': 100, # 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': 1e-2, # 学习率, bert取5e-5,其他取1e-3,如果acc较低或者一直不变,优先调这个, 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
'num_rnn_layers': 1, # rnn层数
'rnn_type': 'GRU', # rnn类型,可以填"LSTM","GRU","CuDNNLSTM","CuDNNGRU"
'rnn_units': 256, # rnn隐藏元
},
'embedding': {'layer_indexes': [12], # 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
},
'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和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))
def pred_tet(path_hyper_parameter=path_hyper_parameters,
path_test=None,
rate=1.0):
# 测试集的准确率
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(path_model_dir)
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'] in ['bert',
'albert']: # 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))
def pred_input(path_hyper_parameter=path_hyper_parameters):
# 输入预测
# 加载超参数
hyper_parameters = load_json(path_hyper_parameter)
pt = PreprocessText()
# 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'] == 'bert':
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'] == 'bert':
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)
def train(hyper_parameters=None, rate=1.0):
if not hyper_parameters:
hyper_parameters = {
'len_max':
160, # 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, 本地win10-4G设为20就好, 过大小心OOM
'embed_size': 200, # 字/词向量维度, bert取768, word取300, char可以更小些
'vocab_size': 21128, # 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':
'word2vec', # 级别, 嵌入类型, 还可以填'xlnet'、'random'、 'bert'、 'albert' or 'word2vec"
'gpu_memory_fraction': 0.66, # gpu使用率
'ifChangeOutput': True,
'model': {
'label': 19, # 类别数
'batch_size':
100, # 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': 100, # 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': 50, # 训练最大轮次
'patience': 3, # 早停,2-3就好
'lr':
4e-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':
'sigmoid', # 'sigmoid', # 最后一个layer, 即分类激活函数
'loss': 'binary_crossentropy', # 损失函数, 可能有问题, 可以自己定义
# 'metrics': 'top_k_categorical_accuracy', # 1070个类, 太多了先用topk, 这里数据k设置为最大:33
'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
'num_rnn_layers': 1, # rnn层数
'rnn_type':
'LSTM', # rnn类型,可以填"LSTM","GRU","CuDNNLSTM","CuDNNGRU"
'rnn_units': 256, # rnn隐藏元
},
'embedding': {
'layer_indexes': [12], # 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': "./data/train.csv", # 训练数据
'val_data': "./data/val.csv", # 验证数据
'test_data': "./data/test.csv"
},
}
# 删除先前存在的模型和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 = PreprocessTextMulti()
print(ra_ed, rate)
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)
print('train data progress ok!')
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 progress 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))
def evaluate(path_hyper_parameter=path_hyper_parameters, rate=1.0):
# 输入预测
# 加载超参数
hyper_parameters = load_json(path_hyper_parameter)
pt = PreprocessTextMulti()
# 模式初始化和加载
graph = Graph(hyper_parameters)
graph.load_model()
ra_ed = graph.word_embedding
# init confusion table
dict_all = initConfusion()
# get validation data
ques_list, val_list, que, val = pt.preprocess_label_ques_to_idx(
hyper_parameters['embedding_type'],
hyper_parameters['data']['test_data'],
ra_ed,
rate=rate,
shuffle=True)
print(len(ques_list))
print("que:", len(que))
# print(val)
# str to token
ques_embed_list = []
count = 0
acc_count = 0
not_none_count = 0
not_none_acc_count = 0
sum_iou = 0
sum_all_iou = 0
for index, que___ in enumerate(que):
# print("原句 ", index, que[index])
# print("真实分类 ", index, val[index])
# print("ques: ", ques)
ques_embed = ra_ed.sentence2idx(que[index])
if hyper_parameters['embedding_type'] == 'albert':
x_val_1 = np.array([ques_embed[0]])
x_val_2 = np.array([ques_embed[1]])
ques_embed = [x_val_1, x_val_2]
else:
x_val = ques_embed
# print("ques_embed: ", ques_embed)
if hyper_parameters['embedding_type'] == 'bert':
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
# print("x_val", x_val)
ques_embed_list.append(x_val)
# 预测
pred = graph.predict(x_val)
# print(pred)
# 取id to label and pred
pre = pt.prereocess_idx(pred[0])
# print("pre",pre)
ls_nulti = []
threshold = 0.65
has_scope = False
has_dense = False
for i, ls in enumerate(pre[0]):
if ls[0] in ['多发', '散发', '无']:
if not has_scope:
has_scope = True
ls_nulti.append(ls)
if ls[0] in val[index].split(","):
dict_all[ls[0]]['TP'] += 1
else:
dict_all[ls[0]]['FN'] += 1
else:
if ls[0] in val[index].split(","):
dict_all[ls[0]]['FP'] += 1
else:
dict_all[ls[0]]['TN'] += 1
if ls[0] not in ['多发', '散发', '无']:
if ls[1] > threshold or not has_dense:
ls_nulti.append(ls)
if ls[0] in val[index].split(","):
dict_all[ls[0]]['TP'] += 1
else:
dict_all[ls[0]]['FP'] += 1
has_dense = True
else:
if ls[0] in val[index].split(","):
dict_all[ls[0]]['FN'] += 1
else:
dict_all[ls[0]]['TN'] += 1
# print("预测结果", index, pre[0])
# print(ls_nulti)
res = cal_acc(ls_nulti, val[index].split(","))
res_iou = cal_iou(ls_nulti, val[index].split(","))
sum_iou += res_iou
# sum_all_iou+=res_all_iou
if res:
# if val[index] != "无":
# not_none_acc_count += 1
acc_count += 1
else:
print("原句 ", index, que[index])
print("真实分类 ", index, val[index])
print("pre ", pre)
print("iou ", res_iou)
print(ls_nulti)
count += 1
if val[index] != "无":
not_none_count += 1
print("acc: ", acc_count / count)
# print("not none acc: ", not_none_acc_count / not_none_count)
print("average iou: ", sum_iou / count)
import prettytable as pt
tb = pt.PrettyTable()
tb.field_names = [" ", "Recall", "Precision", "TP", "FP", "TN", "FN"]
for item in dict_all:
if dict_all[item]['TP'] + dict_all[item]['FN'] == 0:
recall = 1
else:
recall = dict_all[item]['TP'] / (dict_all[item]['TP'] +
dict_all[item]['FN'])
if dict_all[item]['TP'] + dict_all[item]['FP'] == 0:
precision = 1
else:
precision = dict_all[item]['TP'] / (dict_all[item]['TP'] +
dict_all[item]['FP'])
# print(item,recall,precision)
tb.add_row([
item, recall, precision, dict_all[item]['TP'],
dict_all[item]['FP'], dict_all[item]['TN'], dict_all[item]['FN']
])
print(tb)
# log
append_log(hyper_parameters, acc_count / count,
not_none_acc_count / not_none_count, threshold)
'is_training': True,
'model_path': path_model_fast_text_baiduqa_2019,
'num_rnn_layers': 1, # 论文是2，但训练实在是太慢了
'rnn_type': 'GRU', # type of rnn, select 'LSTM', 'GRU', 'CuDNNGRU', 'CuDNNLSTM', 'Bidirectional-LSTM', 'Bidirectional-GRU'
'rnn_units': 256, # large 650, small is 300
},
'embedding':{ 'embedding_type': 'random',
'corpus_path': path_embedding_random_char,
'level_type': 'char',
'embed_size': 30,
'len_max': 50,
},
}
import time
time_start = time.time()
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)
print("耗时:" + str(time.time()-time_start))
# indexs = [ids for ids in range(len(y_train))]
# random.shuffle(indexs)
# x_train, y_train = x_train[indexs], y_train[indexs]
# graph.fit(x_train[0:32000], y_train[0:32000], x_val[0:3200], y_val[0:3200])
graph.fit(x_train, y_train, x_val, y_val)