dnn深度神经网络，mlp多层感知器的实现

问题描述：

dnn深度神经网络，或者说mlp多层感知器的实现，通过加载excel的数据，即划分好的训练集和测试集（如何加载和划分，在我另一篇博客里），训练dnn，实现回归预测并将结果保存至指定excel文件中。使用时将修改下文件路径即可使用。
程序是在jupyter notebook下完成的，如想在pycharm下运行，需做少量修改。

#!/usr/bin/env python
# coding: utf-8

# In[1]:


import tensorflow  as tf
import xlrd
import os
from openpyxl import load_workbook
from xlutils.copy import copy


# In[2]:


import  numpy as np
import pandas as pd
import matplotlib.pyplot as plt
get_ipython().run_line_magic('matplotlib', 'inline')


# In[106]:


#dtrain = pd.read_csv('data/train_csv.csv',header=0)
#dtest  = pd.read_csv('data/test_csv.csv',header=0)
dtrain = pd.read_excel('data_4000/train_split.xls',header=0)
dtest = pd.read_excel('data_4000/test_split.xls',header=0)
x_train = dtrain.iloc[:,0:6]
y_train = dtrain.iloc[:,8:10]

x_test = dtest.iloc[:,0:6]
y_test = dtest.iloc[:,8:10]


# In[107]:


x_train_index = x_train.dtypes[x_train.dtypes != 'object'].index
x_train[x_train_index] = x_train[x_train_index].apply((lambda x: (x - x.mean()) / (x.std())))


# In[108]:


x_test_index = x_test.dtypes[x_test.dtypes != 'object'].index
x_test[x_test_index] = x_test[x_test_index].apply((lambda x: (x - x.mean()) / (x.std())))


# In[109]:


model = tf.keras.Sequential([tf.keras.layers.Dense(100,input_shape=(6,),activation='relu'),
                             tf.keras.layers.Dense(100,activation='relu'),
                             #tf.keras.layers.Dropout(0.2),
                             tf.keras.layers.Dense(500,activation='relu'),
                             #tf.keras.layers.Dropout(0.2),
                             tf.keras.layers.Dense(500,activation='relu'),
                             tf.keras.layers.Dropout(0.2),
                             tf.keras.layers.Dense(500,activation='relu'),
                             tf.keras.layers.Dropout(0.2),
                             tf.keras.layers.Dense(500,activation='relu'),
                             tf.keras.layers.Dropout(0.2),
                             tf.keras.layers.Dense(500,activation='relu'),
                             #tf.keras.layers.Dropout(0.1),
                             tf.keras.layers.Dense(2)])


# In[110]:


model.summary()


# In[ ]:


model.compile(optimizer='adam',
              loss='mse')
def scheduler(epoch):
    # 前50个epoch学习率保持不变，5个epoch后学习率按比例衰减
    if epoch < 10:
        return 0.001
    elif epoch < 50:
        return 0.0001
    else:
        lr = 0.0001 * tf.math.exp(0.01 * (50 - epoch))
        return lr.numpy()

callbacks = [
    tf.keras.callbacks.ModelCheckpoint('./models/weights.hdf5',
                                        monitor='loss',
                                        save_best_only=True),
    #tf.keras.callbacks.EarlyStopping(patience=10, min_delta=1e-4),
    tf.keras.callbacks.LearningRateScheduler(scheduler)]

history = model.fit(x_train,y_train,epochs=150, verbose=1, callbacks=callbacks)


# In[105]:


result = model.evaluate(x_test, y_test)
model.predict(x_test)


# In[65]:


if result <7.7:
    model.save('/models/')


# In[67]:


predict = model.predict(x_test)

print(predict.shape)
def write_excel_xls(path, value):
    index = len(value)  # 获取需要写入数据的行数
    workbook = xlrd.open_workbook(path)  # 打开工作簿
    sheets = workbook.sheet_names()  # 获取工作簿中的所有表格
    worksheet = workbook.sheet_by_name(sheets[0])  # 获取工作簿中所有表格中的的第一个表格
    rows_old = worksheet.nrows  # 获取表格中已存在的数据的行数
    new_workbook = copy(workbook)  # 将xlrd对象拷贝转化为xlwt对象
    new_worksheet = new_workbook.get_sheet(0)  # 获取转化后工作簿中的第一个表格
    
    for i in range(0,worksheet.nrows):
        for j in range(0,worksheet.ncols):
            new_worksheet.write(i, j, 0)
            
    for i in range(0, index):
        for j in range(0, len(value[i])):
            a= value[i][j].astype(np.float64)
            new_worksheet.write(i, j, a) # 追加写入数据，注意是从i+rows_old行开始写入
    new_workbook.save(path)  # 保存工作簿
aim_pred_file =r'C:\Users\gj7520\Desktop\pythob_files\dnn2\data_4000\pre_data_77.xls'
write_excel_xls(aim_pred_file,predict)


# In[68]:


def wirte_gt_test(aim_path, gt_test_path):
    gt_test = pd.read_excel(gt_test_path,header= 0)
    value = np.array(gt_test.iloc[:,8:10])
    print(value.shape)
    index = len(value)  # 获取需要写入数据的行数
    workbook = xlrd.open_workbook(aim_path)  # 打开工作簿
    sheets = workbook.sheet_names()  # 获取工作簿中的所有表格
    worksheet = workbook.sheet_by_name(sheets[0])  # 获取工作簿中所有表格中的的第一个表格
    rows_old = worksheet.nrows  # 获取表格中已存在的数据的行数
    new_workbook = copy(workbook)  # 将xlrd对象拷贝转化为xlwt对象
    new_worksheet = new_workbook.get_sheet(0)  # 获取转化后工作簿中的第一个表格
    
    for i in range(0,worksheet.nrows):
        for j in range(2,worksheet.ncols):
            new_worksheet.write(i, j, 0)
            
    for i in range(0, index):
        for j in range(0, len(value[i])):
            a = value[i][j].astype(np.float64)
            new_worksheet.write(i , j + 2, a)  # 追加写入数据，注意是从i+rows_old行开始写入
    new_workbook.save(aim_path)  # 保存工作簿

gt_test_file = 'data_4000/test_split.xls'
wirte_gt_test(aim_pred_file,gt_test_file)


# In[60]:


plt.plot(history.history['loss'])


# In[ ]:





# In[ ]: