1.Tesorflow訓練模型的數(shù)據(jù)加載

• ?將tensorflow的訓練數(shù)據(jù)數(shù)組（矩陣）保存為.npy的數(shù)據(jù)格式。為后續(xù)的模型訓練提供便捷的方法。例如如下：

import numpy as np
x=np.random.rand(100,7,9)#x是訓練數(shù)據(jù)，這有100條數(shù)據(jù)，每一條有7*9個特征
np.save(r"C:\結果\y_train_feature.npy",feature)#feature是訓練數(shù)據(jù)矩陣

• ?加載.npy訓練數(shù)據(jù)和測試數(shù)組（矩陣），加載后需要調整數(shù)據(jù)的形狀以滿足設計模型的輸入輸出需求，不然無法訓練模型。

import numpy as np

 '''加載訓練和測試數(shù)據(jù)'''
y_train_feature=np.load('C:\結果\y_train_feature.npy')
y_test_feature=np.load('C:\結果\y_test_feature.npy')

'''加載訓練和測試的標簽'''
y_train=np.load('C:\結果\y_train.npy')
y_test=np.load('C:\結果\y_test.npy')
    
'''調整數(shù)據(jù)為模型輸入輸出所需要的形狀'''
y_test_feature=y_test_feature.reshape(-1,7,129)
y_train_feature=y_train_feature.reshape(-1,7,129)
y_test=y_test.reshape(-1,129)
y_train=y_train.reshape(-1,129)
  

?2.Tensorflow模型搭建的訓練和保存

• 這里可以采用自定義層和tensorflow的API搭建網(wǎng)絡模型，以API為例子搭建多層LSTM網(wǎng)絡模型用于訓練。

import tensorflow as tf
from tensorflow import  keras
from keras.models import  Model
from keras.layers import Input, MaxPooling1D,Dense, Conv1D, Conv2D, Dropout, Flatten,\
 BatchNormalization, Reshape, Activation, concatenate,SeparableConv2D,LSTM,GRU,Conv2DTranspose,SimpleRNN
from keras.layers.pooling import MaxPooling2D


 '''模型輸入'''
input_feature = Input(shape=y_train_feature.shape[1:], name='input_feature')

  '''模型搭建LSTM'''
fc1=LSTM(512,return_sequences=True)(input_feature)
fc2=LSTM(512,return_sequences=True)(fc1)
fc3=LSTM(512,return_sequences=True)(fc2)
fc4=LSTM(512)(fc3)
'''全連接的輸出層'''
output_2 = Dense(129,activation='linear',kernel_initializer='random_uniform',\
                     bias_initializer='random_uniform',name='output1')(fc4)
   """以Model來組合整個網(wǎng)絡"""
model = Model(inputs= input_feature, outputs=output_2)
model.summary()#察看網(wǎng)絡架構

• ?tensorflow模型的訓練，優(yōu)化器的選擇，損失函數(shù)的設置，批量大小，學習率大小，訓練epoch次數(shù)，最優(yōu)模型的保存等操作。

adam=Adam(lr=0.0001,decay=1e-6)#優(yōu)化器的選擇
model.compile(optimizer=adam,
          loss='mae',#均方誤差
          metrics=['mae'])#平均絕對誤差
'''訓練網(wǎng)絡'''

earlyStopping = callbacks.EarlyStopping(monitor='val_mae',patience=5,mode='min')

#保存模型的路徑
filepath = "weights-improvement-{epoch:02d}-{val_mae:.2f}.h5"


#在每個訓練期（epoch）后保存模型
#period：檢查點之間的間隔（epoch數(shù)）
checkpoint = callbacks.ModelCheckpoint(filepath, monitor='val_mae', verbose=1, save_best_only=True,
                            mode='min',save_weights_only=False,period=1)

print('Training------')
   
model.fit(y_train_feature,y_train,epochs=800,batch_size=128,verbose=1,validation_data=(y_test_feature,y_test),shuffle=True,callbacks=[earlyStopping,checkpoint])

model.save(r'new_models_ripleinputV7.h5')

?3.Tensorflow模型的調用

• 訓練好并保存后的模型，調用時可以采用 load_model() 方法獲得保存后的模型。

from keras.models import load_model
import numpy as np

model = tf.keras.models.load_model(r"保存模型的路徑\LSTM.h5")
'''輸入數(shù)據(jù)獲得模型輸出結果'''
output=model.predict(input_feature)#input_feature形狀需要滿足模型輸入的形狀要求

'''將張量形式的輸出轉化為numpy數(shù)組的形式'''
output=np.array(output)

4.整體代碼實現(xiàn)

• tensorflow網(wǎng)絡模型的訓練完整實現(xiàn)（API網(wǎng)絡搭建）的形式。自定義的網(wǎng)絡模型可采用同樣的方法訓練模型。

# -*- coding: utf-8 -*-
"""
Created on Sat Sep 11 19:47:21 2021

@author: 茶墨先生
"""

'''導入需要的包'''
import tensorflow as tf
from tensorflow import  keras
from tensorflow.keras import  layers,Sequential,datasets
from keras import regularizers,callbacks 
tf.compat.v1.disable_eager_execution()
from keras import backend as K
from keras.datasets import mnist
from keras.optimizers import Adam
import numpy as np
from keras.models import  Model
from keras.layers import Input, MaxPooling1D,Dense, Conv1D, Conv2D, Dropout, Flatten,\
 BatchNormalization, Reshape, Activation, concatenate,SeparableConv2D,LSTM,GRU,Conv2DTranspose,SimpleRNN
from keras.layers.pooling import MaxPooling2D
  
#自定義模型
'''訓練'''
def Train():
    '''加載訓練數(shù)據(jù)'''
    y_train_feature=np.load('C:\結果\y_train_speech_feature.npy')
    y_test_feature=np.load('C:\結果\y_test_speech_feature.npy')
    y_train=np.load('C:\結果\y_train.npy')
    y_test=np.load('C:\結果\y_test.npy')
    
    y_test_feature=y_test_feature.reshape(-1,7,129)
    y_train_feature=y_train_feature.reshape(-1,7,129)
    y_test=y_test.reshape(-1,129)
    y_train=y_train.reshape(-1,129)
    print("y_train: ",y_train.shape)
    print("y_train_speech_feature: ", y_train_feature.shape)
    
    
    '''構建需要的網(wǎng)絡模型'''
    input_feature = Input(shape=y_train_feature.shape[1:], name='input_feature')
    fc1=LSTM(512,return_sequences=True)(input_feature)
    fc2=LSTM(512,return_sequences=True)(fc1)
    fc3=LSTM(512,return_sequences=True)(fc2)
    fc4=LSTM(512)(fc3)
    output_2 = Dense(129,activation='linear',kernel_initializer='random_uniform',\
                     bias_initializer='random_uniform',name='output1')(fc4)
    """以Model來組合整個網(wǎng)絡"""
    model = Model(inputs= input_feature, outputs=output_2)
    model.summary()
    
    adam=Adam(lr=0.0001,decay=1e-6)
    model.compile(optimizer=adam,
              loss='mae',#均方誤差
              metrics=['mae'])#平均絕對誤差
    '''訓練網(wǎng)絡'''
    
    earlyStopping = callbacks.EarlyStopping(monitor='val_mae',patience=5,mode='min')
    
    #保存模型的路徑
    filepath = "weights-improvement-{epoch:02d}-{val_mae:.2f}.h5"
    
    
    #在每個訓練期（epoch）后保存模型
    #period：檢查點之間的間隔（epoch數(shù)）
    checkpoint = callbacks.ModelCheckpoint(filepath, monitor='val_mae', verbose=1, save_best_only=True,
                                mode='min',save_weights_only=False,period=1)
    
    print('Training------')
   
    model.fit(y_train_feature,y_train,epochs=800,batch_size=128,verbose=1,validation_data=(y_test_feature,y_test),shuffle=True,callbacks=[earlyStopping,checkpoint])
    
    model.save(r'new_models_ripleinputV7.h5')

'''主函數(shù)調用'''
if __name__ =='__main__':
    
    '''訓練'''
    Train()


    

讀書，生活，旅行。感謝關注和支持！文章來源地址http://www.zghlxwxcb.cn/news/detail-742708.html

到了這里，關于Tensorflow實現(xiàn)訓練數(shù)據(jù)的加載—模型搭建訓練保存—模型調用和加載全流程的文章就介紹完了。如果您還想了解更多內容，請在右上角搜索TOY模板網(wǎng)以前的文章或繼續(xù)瀏覽下面的相關文章，希望大家以后多多支持TOY模板網(wǎng)！