2019-06-22 14:28:21 +08:00
|
|
|
|
# -*- coding: UTF-8 -*-
|
|
|
|
|
# !/usr/bin/python
|
|
|
|
|
# @time :2019/6/8 14:37
|
|
|
|
|
# @author :Mo
|
|
|
|
|
# @function :train of VDCNN with baidu-qa-2019 in question title
|
2019-07-12 02:37:48 +08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
# 适配linux
|
2019-06-22 14:28:21 +08:00
|
|
|
|
import pathlib
|
|
|
|
|
import sys
|
|
|
|
|
import os
|
|
|
|
|
project_path = str(pathlib.Path(os.path.abspath(__file__)).parent.parent.parent)
|
|
|
|
|
sys.path.append(project_path)
|
2019-07-12 02:37:48 +08:00
|
|
|
|
# 地址
|
|
|
|
|
from keras_textclassification.conf.path_config import path_model, path_fineture, path_model_dir, path_hyper_parameters
|
|
|
|
|
# 训练验证数据地址
|
2019-06-22 14:28:21 +08:00
|
|
|
|
from keras_textclassification.conf.path_config import path_baidu_qa_2019_train, path_baidu_qa_2019_valid
|
2019-07-12 02:37:48 +08:00
|
|
|
|
# 数据预处理, 删除文件目录下文件
|
|
|
|
|
from keras_textclassification.data_preprocess.text_preprocess import PreprocessText, delete_file
|
|
|
|
|
# 模型图
|
2019-06-22 14:28:21 +08:00
|
|
|
|
from keras_textclassification.m08_TextVDCNN.graph import VDCNNGraph as Graph
|
2019-07-12 02:37:48 +08:00
|
|
|
|
# 计算时间
|
|
|
|
|
import time
|
2019-06-22 14:28:21 +08:00
|
|
|
|
|
|
|
|
|
|
2019-07-12 02:37:48 +08:00
|
|
|
|
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
|
2019-10-14 20:37:57 +08:00
|
|
|
|
'embedding_type': 'random', # 级别, 嵌入类型, 还可以填'xlnet'、'random'、 'bert'、 'albert' or 'word2vec"
|
2019-07-12 02:37:48 +08:00
|
|
|
|
'gpu_memory_fraction': 0.66, #gpu使用率
|
|
|
|
|
'model': {'label': 17, # 类别数
|
|
|
|
|
'batch_size': 32, # 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
|
|
|
|
|
# only VDCNN
|
|
|
|
|
'top_k': 2, # k-max pooling
|
|
|
|
|
'pool_type': 'max', # 池化选择, 可以选"max"、"avg"、"conv"
|
|
|
|
|
'shortcut': True, # resnet方案
|
|
|
|
|
# 论文参数, 长文本, long sentence, ,max_len=256 or 1024 ans so on
|
|
|
|
|
'filters': [[64, 1], [128, 1], [256, 1], [512, 1]], # 9 layer, 256 len max
|
|
|
|
|
# 'filters': [[64, 2], [128, 2], [256, 2], [512, 2]], # 17 layer
|
|
|
|
|
# 'filters': [[64, 5], [128, 5], [256, 2], [512, 2]], # 29 layer
|
|
|
|
|
# 'filters': [[64, 8], [128, 8], [256, 5], [512, 3]], # 49 layer, 1024 len max
|
2019-06-22 14:28:21 +08:00
|
|
|
|
|
2019-07-12 02:37:48 +08:00
|
|
|
|
# 自己设置的,效果不太佳, 短文本, short sentence, ,max_len=32 or 64 ans so on
|
|
|
|
|
# 'filters': [[3, 8], [6, 8], [12, 5], [24, 3]], # 49 layer
|
|
|
|
|
# 'filters': [[4, 8], [8, 8], [16, 5], [32, 3]], # 49 layer
|
|
|
|
|
# 'filters': [[3, 1], [6, 1], [12, 1], [24, 1]], # 9 layer
|
|
|
|
|
# 'filters': [[4, 1], [8, 1], [16, 1], [32, 1]], # 9 layer
|
|
|
|
|
'channel_size': 1, # CNN通道
|
|
|
|
|
'dropout': 0.32, # 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-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
|
|
|
|
|
'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
|
2019-06-22 14:28:21 +08:00
|
|
|
|
|
2019-07-12 02:37:48 +08:00
|
|
|
|
},
|
|
|
|
|
'embedding': {'layer_indexes': [12], # bert取的层数
|
2019-08-19 23:22:36 +08:00
|
|
|
|
# '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)
|
2019-07-12 02:37:48 +08:00
|
|
|
|
},
|
|
|
|
|
'data':{'train_data': path_baidu_qa_2019_train, # 训练数据
|
|
|
|
|
'val_data': path_baidu_qa_2019_valid # 验证数据
|
|
|
|
|
},
|
|
|
|
|
}
|
2019-06-22 14:28:21 +08:00
|
|
|
|
|
2019-07-12 02:37:48 +08:00
|
|
|
|
# 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
|
2019-06-22 14:28:21 +08:00
|
|
|
|
graph = Graph(hyper_parameters)
|
2019-07-12 02:37:48 +08:00
|
|
|
|
print("graph init ok!")
|
2019-06-22 14:28:21 +08:00
|
|
|
|
ra_ed = graph.word_embedding
|
2019-07-12 02:37:48 +08:00
|
|
|
|
# 数据预处理
|
2020-08-12 16:18:18 +08:00
|
|
|
|
pt = PreprocessText(path_model_dir)
|
2019-07-12 02:37:48 +08:00
|
|
|
|
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!")
|
2019-06-22 14:28:21 +08:00
|
|
|
|
print(len(y_train))
|
2019-07-12 02:37:48 +08:00
|
|
|
|
# 训练
|
2019-06-22 14:28:21 +08:00
|
|
|
|
graph.fit(x_train, y_train, x_val, y_val)
|
2019-07-12 02:37:48 +08:00
|
|
|
|
print("耗时:" + str(time.time()-time_start))
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
if __name__=="__main__":
|
2019-10-13 09:08:31 +08:00
|
|
|
|
train(rate=1)
|
2019-08-25 01:29:36 +08:00
|
|
|
|
# 注意: 4G的1050Ti的GPU、win10下batch_size=32,len_max=20, gpu<=0.87, 应该就可以bert-fineture了。
|
2019-07-12 02:37:48 +08:00
|
|
|
|
# 全量数据训练一轮(batch_size=32),就能达到80%准确率(验证集), 效果还是不错的
|
|
|
|
|
# win10下出现过错误,gpu、len_max、batch_size配小一点就好:ailed to allocate 3.56G (3822520832 bytes) from device: CUDA_ERROR_OUT_OF_MEMORY: out of memory
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
# 14251/14251 [==============================] - 47s 3ms/step - loss: 2.2805 - acc: 0.3443 - val_loss: 3.7807 - val_acc: 0.3808
|
|
|
|
|
# Epoch 00001: val_loss improved from inf to 3.78072, saving model to
|
|
|
|
|
# Epoch 2/20
|