缘由可能以下:python
- 学习率可能太大
- batch size过小
- 样本分布不均匀
- 缺乏加入正则化
- 数据规模较小
一种很重要的缘由是数据split的时候没有shufflemysql
import numpy as np
index = np.arange(data.shape[0])
np.random.seed(1024)
np.random.shuffle(index)
data=data[index]
labels=labels[index]
缘由:输入的训练数据没有归一化形成
解决方法:把输入数值经过下面的函数过滤一遍,进行归一化web
#数据归一化
def data_in_one(inputdata):
inputdata = (inputdata-inputdata.min())/(inputdata.max()-inputdata.min())
return inputdata
- train loss 不断降低,test loss不断降低,说明网络仍在学习;
- train loss 不断降低,test loss趋于不变,说明网络过拟合;
- train loss 趋于不变,test loss不断降低,说明数据集100%有问题;
- train loss 趋于不变,test loss趋于不变,说明学习遇到瓶颈,须要减少学习率或批量数目;
- train loss 不断上升,test loss不断上升,说明网络结构设计不当,训练超参数设置不当,数据集通过清洗等问题。
通常来讲,较高的acc对应的loss较低,但这不是绝对,毕竟他们是两个不一样的东西,因此在实际实现中,咱们能够对二者进行一个微调。sql
关于epoch设置问题,咱们能够设置回调函数,选择验证集最高的acc做为最优模型。数据库
关于BN和dropout,其实这两个是两个彻底不一样的东西,BN针对数据分布,dropout是从模型结构方面优化,因此他们两个能够一块儿使用,对于BN来讲其不但能够防止过拟合,还能够防止梯度消失等问题,而且能够加快模型的收敛速度,可是加了BN,模型训练每每会变得慢些。json
代码示意:网络
……
from keras.layers import Concatenate,Dropout
……
concatenate = Concatenate(axis=2)([blstm,embedding_layer])
concatenate=Dropout(rate=0.1)(concatenate)
下面这段代码是我对本身实验数据作的augmentation,能够给你们提供一个参考。首先,个人数据集如图所示:
app
个人数据库中的essays表中有7列,每一行为一个数据样本,其中第一列AUTHID为样本编号,TEXT为文本内容,后面为文本的标记。对于文本的augmentation,一个比较合理的扩增数据集的方法就是将每个文本的句子循环移位,这样能够最大限度地保证文本总体的稳定。下面的代码读取essays表格中的样本信息,对文本进行循环移位后存入到table_augment表中。dom
代码示意:svg
#!/usr/bin/python
# -*- coding:utf8 -*-
from sqlalchemy import create_engine # mysql orm interface,better than mysqldb
import pandas as pd
import spacy # a NLP model like NLTK,but more industrial.
import json
TO_SQL='table_augment'
READ_SQL_TABLE='essays'
def cut_sentences(df):
all_text_name = df["AUTHID"] # type pandas.Series:get all text name(match the "#AUTHID" in essays)
all_text = df["TEXT"] # type pandas.Series:get all text(match the "TEXT" in essays)
all_label_cEXT=df["cEXT"]
all_label_cNEU=df["cNEU"]
all_label_cAGR=df["cAGR"]
all_label_cCON=df["cCON"]
all_label_cOPN=df["cOPN"]
all_number = all_text_name.index[-1] # from 0 to len(all_text_name)-1
for i in xrange(0,all_number+1,1):
print("start to deal with text ", i ," ...")
text = all_text[i] # type str:one of text in all_text
text_name = all_text_name[i] # type str:one of text_name in all_text_name
nlp = spacy.load('en')
test_doc = nlp(text)#.decode('utf8'))
cut_sentence = []
for sent in test_doc.sents: # get each line in the text
cut_sentence.append(sent.text)
""" type sent is spacy.tokens.span.Span, not a string, so, we call the member function Span.text to get its unicode form """
line_number = len(cut_sentence)
for itertor in range(line_number):
if itertor !=0:
cut_sentence=cut_sentence[1:]+cut_sentence[:1]
cut_sentence_json = json.dumps(cut_sentence)
input_data_dic = {'text_name': str(itertor)+"_"+text_name,
'line_number':line_number,
'line_text': cut_sentence_json,
'cEXT': all_label_cEXT[i],
'cNEU': all_label_cNEU[i],
'cAGR': all_label_cAGR[i],
'cCON': all_label_cCON[i],
'cOPN': all_label_cOPN[i]
}
input_data = pd.DataFrame(input_data_dic,index=[i],columns=['text_name',
'line_number',
'line_text',
'cEXT',
'cNEU',
'cAGR',
'cCON',
'cOPN'])
input_data.to_sql(TO_SQL, engine, if_exists='append', index=False, chunksize=100)
""" DataFrame.index will be insert to table by default. We don't want it, so we set the index = False(True default) """
print("text ", i ," finished")
if __name__ =='__main__':
engine = create_engine('mysql+pymysql://root:root@localhost:3306/personality_1?charset=utf8', echo=True,convert_unicode=True)
df = pd.read_sql_table(READ_SQL_TABLE, engine,chunksize=5) # read essays
for df_iter in df:
cut_sentences(df_iter)
具体来说就是model.load人家训练好的weight.hdf5,而后在这个基础上继续训练。具体能够见以后的博文中的断点训练。
调小学习速率(Learning Rate)以前已经讲过不在赘述
适当增大batch_size。以前已经讲过不在赘述
试一试别的优化器(optimizer)以前已经讲过不在赘述
Keras的回调函数EarlyStopping() 以前已经讲过,再也不赘述
正则化方法是指在进行目标函数或代价函数优化时,在目标函数或代价函数后面加上一个正则项,通常有L1正则与L2正则等。
代码片断示意:
from keras import regularizers
……
out = TimeDistributed(Dense(hidden_dim_2,
activation="relu",
kernel_regularizer=regularizers.l1_l2(0.01,0.01),
activity_regularizer=regularizers.l1_l2(0.01,0.01)
)
)(concatenate)
……
dense=Dense(200,
activation="relu",
kernel_regularizer=regularizers.l1_l2(0.01,0.01),
activity_regularizer=regularizers.l1_l2(0.01,0.01)
)(dense)
更多参考信息:
https://blog.csdn.net/mrgiovanni/article/details/52167016