深度学习3 使用keras进行迁移学习（重新修订）

万里鹏程转瞬至

已于 2022-03-20 21:24:01 修改

阅读量2.1k

点赞数 4

分类专栏： tf2与keras深度学习

于 2021-01-06 21:06:16 首次发布

本文链接：https://blog.csdn.net/a486259/article/details/112297297

版权

深度学习

tf2与keras深度学习专栏收录该内容

18 篇文章 6 订阅

订阅专栏

修改说明：为了能使迁移学习所训练出的模型能适应后续的知识蒸馏，因此对迁移的代码做了细节上的修订。后续红色标明字样，皆为修改后的内容

采用模型分离的方式构造迁移学习，可以降低反向传播时卷积运算的工作量。

1、卷积基的提取

首先选定迁移的目标并提取其卷积基可按照不同的需求提取迁移特征。keras除此外还有 DenseNet121、MobileNet、Xception等网络可以用于迁移

注意：1、要按照个人需求和电脑配置调整 shape，batch_size, 和models数组内的模型名称

2、要从tensorflow.kreas中引入application库，这样子才能支持更多的模型库（如EfficientNet系列）

3、新增os环境设置，避免价值tensorflow时输出太多不必要的信息

import os
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3'  # or any {'0', '1', '2'}
from keras.preprocessing.image import ImageDataGenerator
from keras.utils import np_utils
#from tensorflow.python.keras.utils import np_utils  #上面语句报错可以改成这个
from tensorflow.keras import applications,Model
import numpy as np
import os,gc
batch_size=32#ResNet50,VGG16,InceptionV3 EfficientNetB7 NASNetLarge
shape=(226, 226)  
# DenseNet169 DenseNet201  #DenseNet121 ResNet152V2 ,"InceptionV3","MobileNetV3Small","ResNet50V2","VGG19"
models=["EfficientNetB0"]#
for modelname in models:
    mstr="applications.%s(include_top=False,input_shape=(114,114,3), weights='imagenet')"%modelname
    base = eval(mstr)
    print(modelname)
    print(modelname)
    model= Model(input=base.input, output=base.layers[-1].output)

2、数据集的划分

数据集的划分可以查看python工具方法 4 依据随机种子将图片文件划分为训练集、测试集、验证集_a486259的博客-CSDN博客，代码复制后即可运行

文件夹：flower 通过该代码可得到：flower-训练,flower-验证,flower-测试三个文件夹

3、特征的提取与保存

所提取的特征保存在features文件夹中，命名方式为：
modelname_data_训练、modelname_data_验证、modelname_data_测试

modelname_labels_训练、modelname_labels_验证、modelname_labels_测试

#代码复制过来不想调整缩进，所以使用了if True来控制
if True:
    datagen = ImageDataGenerator()#(rescale=1.0 / 255)
    root_path='./features/'
    if not os.path.exists(root_path):
        os.makedirs(root_path)
    dirList=['训练','验证','测试']
    data_path='flower//'
    for path in dirList:
        generator = datagen.flow_from_directory(data_path+'-'+path,
                                                target_size=shape,
                                                batch_size=batch_size,
                                                class_mode='categorical',
                                                shuffle=False)
        print(generator.class_indices)#输出数据的labels
        labels= np_utils.to_categorical(generator.classes)
        np.save(open(root_path+'label.npy', 'wb'), generator.class_indices)
        features= model.predict_generator(generator, len(generator.classes)/generator.batch_size)
        print('result.shape: ',features.shape)
        #import sys;sys.exit()
        print(len(generator.classes))
        print(path,'集-卷积池化数据固化成功！')
        np.save(open('%s%s_data_%s.npy'%(root_path,modelname,path), 'wb'), features)
        np.save(open('%s%s_labels_%s.npy'%(root_path,modelname,path), 'wb'), labels)
        np.save(open('%s%s_fileName_%s.npy'%(root_path,modelname,path), 'wb'), generator.filenames)
        gc.collect()

4、读取特征

import numpy as np
root_path='./features/VGG16_'
train_data = np.load(root_path+'data_训练.npy')
train_data=train_data.reshape(train_data.shape[0],-1)
train_label = np.load(root_path+'labels_训练.npy')
shuffle=True
if shuffle:
    import random
    randnum = random.randint(0,100)
    random.seed(randnum)
    random.shuffle(train_data)
    random.seed(randnum)
    random.shuffle(train_label)

validation_data = np.load(root_path+'data_验证.npy')
validation_data=validation_data.reshape(validation_data.shape[0],-1)
validation_label =np.load(root_path+'labels_验证.npy')

test_data = np.load(root_path+'data_测试.npy')
test_data=test_data.reshape(test_data.shape[0],-1)
test_label =np.load(root_path+'labels_测试.npy')

#数据的扁平化
inputShape=np.array(train_data.shape[1:]).prod()
train_data=train_data.reshape(-1,inputShape)
test_data=test_data.reshape(-1,inputShape)
validation_data=validation_data.reshape(-1,inputShape)

5、重新训练分类器

在这里修改了两行代码，移除了最后一个Dense中的激活函数，并为每一个dense设置了name，通过这种操作，可以已另一种方式实现模型拼接。需要进行更高级的知识蒸馏（实现kl loss，softmax升温）的朋友，请不要使用后续的模型拼接代码。

output = Dense(class_types,name='dense_out')(x)
output = Activation(tf.nn.softmax,name='Activation_out')(output)

import tensorflow as tf
from keras.models import Model
from keras.layers import Dense,Input,Dropout,Activation

inputs = Input(shape=(inputShape,))
x = Dense(100, activation='relu',name='dense1')(inputs)
x = Dense(50, activation='relu',name='dense2')(x)
x = Dropout(0.2)(x)
output = Dense(class_types,name='dense_out')(x)
output = Activation(tf.nn.softmax,name='Activation_out')(output)
model = Model(inputs=inputs, outputs=output)
model.compile(optimizer='adam',loss='categorical_crossentropy',metrics=['acc'])

from keras.callbacks import History,ModelCheckpoint,EarlyStopping
#最佳模型保存
ModelCheckpoint=ModelCheckpoint('best_model.h5', monitor='acc', verbose=0, save_best_only=True, save_weights_only=False, mode='auto', period=1)
#提前终止
EarlyStopping=EarlyStopping(monitor='val_loss', patience=2, verbose=2, mode='auto')
#历史数据记录
loss_history = History()
#模型训练
model.fit(train_data,train_label,epochs=100,batch_size=32,
          callbacks=[ModelCheckpoint,EarlyStopping,loss_history]
          validation_data=(validation_data, validation_label),
          verbose=2)
#模型评估
loss,acc=model.evaluate(test_data, test_label )
print('loss:',loss,'acc:',acc)
#历史数据绘图
import pandas as pd
d=pd.DataFrame(data=loss_history.history)
d.to_csv('history.csv')

6、全部代码

01.feature_save.py 复制代码保存为01.feature_save.py，然后执行，完成迁移特征的提取与保存

import os
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3'  # or any {'0', '1', '2'}
from tensorflow.keras.preprocessing.image import ImageDataGenerator
from tensorflow.python.keras.utils.np_utils import to_categorical
from tensorflow.keras import applications,Model
import numpy as np
import os,gc
batch_size=32#ResNet50,VGG16,InceptionV3 EfficientNetB7 NASNetLarge
shape=(56, 56)
# DenseNet169 DenseNet201  #DenseNet121 ResNet152V2 ,"InceptionV3","MobileNetV3Small","ResNet50V2","VGG19"
models=["EfficientNetB0"]#
for modelname in models:
    mstr="applications.%s(include_top=False,input_shape=(114,114,3), weights='imagenet')"%modelname
    base = eval(mstr)
    print(modelname)
    print(modelname)
    model= Model(inputs=base.input, outputs=base.layers[-1].output)
    datagen = ImageDataGenerator()#(rescale=1.0 / 255)
    root_path='./features/'
    if not os.path.exists(root_path):
        os.makedirs(root_path)
    dirList=['训练','验证','测试']
    data_path = r"./flower"
    for path in dirList:
        generator = datagen.flow_from_directory(data_path+'-'+path,
                                                target_size=shape,
                                                batch_size=batch_size,
                                                class_mode='categorical',
                                                shuffle=False)
        print(generator.class_indices)#输出数据的labels
        labels= to_categorical(generator.classes)
        np.save(open(root_path+'label.npy', 'wb'), generator.class_indices)
        features= model.predict_generator(generator, len(generator.classes)/generator.batch_size)
        print('result.shape: ',features.shape)
        #import sys;sys.exit()
        print(len(generator.classes))
        print(path,'集-卷积池化数据固化成功！')
        np.save(open('%s%s_data_%s.npy'%(root_path,modelname,path), 'wb'), features)
        np.save(open('%s%s_labels_%s.npy'%(root_path,modelname,path), 'wb'), labels)
        np.save(open('%s%s_fileName_%s.npy'%(root_path,modelname,path), 'wb'), generator.filenames)
        gc.collect()

02.new_cls.py 复制代码保存为02.new_cls.py，然后执行，完成分类器训练

#------------------------------特征数据读取---------------------------------------
import numpy as np
class_types=2
root_path='./features/VGG16_'
train_data = np.load(root_path+'data_训练.npy')
train_label = np.load(root_path+'labels_训练.npy')
shuffle=True
if shuffle:
    import random
    randnum = random.randint(0,100)
    random.seed(randnum)
    random.shuffle(train_data)
    random.seed(randnum)
    random.shuffle(train_label)
 
validation_data = np.load(root_path+'data_验证.npy')
validation_label =np.load(root_path+'labels_验证.npy')
 
test_data = np.load(root_path+'data_测试.npy')
test_label =np.load(root_path+'labels_测试.npy')
#-----------------------------------重新训练全连接分类器--------------------------
 
from keras.models import Model
from keras.layers import Dense,Input,Dropout,GlobalAveragePooling2D
 
inputs = Input(shape=test_data[0].shape)
x = GlobalAveragePooling2D()(inputs)#GlobalAveragePooling2D 为了减少对数据的形变操作，和降低模型参数量
x = Dense(100, activation='relu')(x)
x = Dense(50, activation='relu')(x)
x = Dropout(0.2)(x)
output = Dense(class_types, activation='softmax')(x)
model = Model(inputs=inputs, outputs=output)
model.compile(optimizer='adam',loss='categorical_crossentropy',metrics=['acc'])
 
from keras.callbacks import History,ModelCheckpoint,EarlyStopping
#最佳模型保存
ModelCheckpoint=ModelCheckpoint('gap_best_model.h5', monitor='acc', verbose=0, save_best_only=True, save_weights_only=False, mode='auto', period=1)
#提前终止
EarlyStopping=EarlyStopping(monitor='val_loss', patience=15, verbose=2, mode='auto')
#历史数据记录
loss_history = History()
#模型训练
model.fit(train_data,train_label,epochs=100,batch_size=32,
          callbacks=[ModelCheckpoint,EarlyStopping,loss_history],
          validation_data=(validation_data, validation_label),
          verbose=2)
#模型评估
loss,acc=model.evaluate(test_data, test_label)
print('test loss:',loss,'acc:',acc)
loss,acc=model.evaluate(train_data, train_label)
print('train loss:',loss,'acc:',acc)
loss,acc=model.evaluate(validation_data, validation_label)
print('valid loss:',loss,'acc:',acc)
#历史数据绘图
import pandas as pd
d=pd.DataFrame(data=loss_history.history)
d.to_csv(root_path+'history.csv')