Keras：在预训练的网络上fine-tune

Keras：自建数据集图像分类的模型训练、保存与恢复
Keras：使用预训练网络的bottleneck特征

准备

fine-tune的三个步骤:

• 搭建vgg-16并载入权重;
• 将之前定义的全连接网络加载到模型顶部,并载入权重;
• 冻结vgg16网络的一部分参数.

在之前的Keras：自建数据集图像分类的模型训练、保存与恢复里制作了实验用的数据集并初步进行了训练.然后在Keras：使用预训练网络的bottleneck特征中定义并训练了要使用全连接网络,并将网络权重保存到了bottleneck_fc_model.h5文件中.

fine-tune过程

根据keras中…/keras/applications/vgg16.py的VGG16模型形式,构造VGG16模型的卷积部分,并载入权重(vgg16_weights_tf_dim_ordering_tf_kernels_notop.h5).然后添加预训练好的模型.训练时冻结最后一个卷积块前的卷基层参数.

示例:

#!/usr/bin/python
# coding:utf8

from keras.models import Sequential
from keras import optimizers
from keras.preprocessing.image import ImageDataGenerator
from keras.layers import Flatten, Dense, Dropout, Conv2D, MaxPooling2D
from keras import backend as K
K.set_image_dim_ordering('th')


# 构造VGG16模型
model = Sequential()

# Block 1
model.add(Conv2D(64, (3, 3), activation='relu', padding='same', name='block1_conv1', input_shape=(3, 150, 150)))
model.add(Conv2D(64, (3, 3), activation='relu', padding='same', name='block1_conv2'))
model.add(MaxPooling2D((2, 2), strides=(2, 2), name='block1_pool'))

# Block 2
model.add(Conv2D(128, (3, 3), activation='relu', padding='same', name='block2_conv1'))
model.add(Conv2D(128, (3, 3), activation='relu', padding='same', name='block2_conv2'))
model.add(MaxPooling2D((2, 2), strides=(2, 2), name='block2_pool'))

# Block 3
model.add(Conv2D(256, (3, 3), activation='relu', padding='same', name='block3_conv1'))
model.add(Conv2D(256, (3, 3), activation='relu', padding='same', name='block3_conv2'))
model.add(Conv2D(256, (3, 3), activation='relu', padding='same', name='block3_conv3'))
model.add(MaxPooling2D((2, 2), strides=(2, 2), name='block3_pool'))

# Block 4
model.add(Conv2D(512, (3, 3), activation='relu', padding='same', name='block4_conv1'))
model.add(Conv2D(512, (3, 3), activation='relu', padding='same', name='block4_conv2'))
model.add(Conv2D(512, (3, 3), activation='relu', padding='same', name='block4_conv3'))
model.add(MaxPooling2D((2, 2), strides=(2, 2), name='block4_pool'))

# Block 5
model.add(Conv2D(512, (3, 3), activation='relu', padding='same', name='block5_conv1'))
model.add(Conv2D(512, (3, 3), activation='relu', padding='same', name='block5_conv2'))
model.add(MaxPooling2D((2, 2), strides=(2, 2), name='block5_pool'))

model.load_weights('vgg16_weights_tf_dim_ordering_tf_kernels_notop.h5',by_name=True)
model.summary()

# 在初始化好的VGG网络上添加预训练好的模型
top_model = Sequential()
top_model.add(Flatten(input_shape=model.output_shape[1:])) #  (4,4,512)
top_model.add(Dense(256, activation='relu'))
top_model.add(Dropout(0.5))
top_model.add(Dense(1, activation='sigmoid'))

top_model.load_weights('bottleneck_fc_model.h5',by_name=True)
model.add(top_model)

# 将最后一个卷积块前的卷基层参数冻结,把随后卷积块前的权重设置为不可训练（权重不会更新）
for layer in model.layers[:25]:
    layer.trainable = False

model.compile(loss='binary_crossentropy',
              optimizer=optimizers.SGD(lr=1e-4, momentum=0.9),
              metrics=['accuracy'])

# 以低学习率进行训练
train_datagen = ImageDataGenerator(rescale=1./255,
                                   shear_range=0.2,
                                   zoom_range=0.2,
                                   horizontal_flip=True)

test_datagen = ImageDataGenerator(rescale=1./255)
train_generator = train_datagen.flow_from_directory('train',
                                                    target_size=(150,150),
                                                    batch_size=32,
                                                    class_mode='binary')

validation_generator = test_datagen.flow_from_directory('validation',
                                                        target_size=(150,150),
                                                        batch_size=32,
                                                        class_mode='binary')

model.fit_generator(train_generator,
                    steps_per_epoch=10,
                    epochs=50,
                    validation_data=validation_generator,
                    validation_steps=10)

输出:

_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
block1_conv1 (Conv2D)        (None, 64, 150, 150)      1792      
_________________________________________________________________
block1_conv2 (Conv2D)        (None, 64, 150, 150)      36928     
_________________________________________________________________
block1_pool (MaxPooling2D)   (None, 64, 75, 75)        0         
_________________________________________________________________
block2_conv1 (Conv2D)        (None, 128, 75, 75)       73856     
_________________________________________________________________
block2_conv2 (Conv2D)        (None, 128, 75, 75)       147584    
_________________________________________________________________
block2_pool (MaxPooling2D)   (None, 128, 37, 37)       0         
_________________________________________________________________
block3_conv1 (Conv2D)        (None, 256, 37, 37)       295168    
_________________________________________________________________
block3_conv2 (Conv2D)        (None, 256, 37, 37)       590080    
_________________________________________________________________
block3_conv3 (Conv2D)        (None, 256, 37, 37)       590080    
_________________________________________________________________
block3_pool (MaxPooling2D)   (None, 256, 18, 18)       0         
_________________________________________________________________
block4_conv1 (Conv2D)        (None, 512, 18, 18)       1180160   
_________________________________________________________________
block4_conv2 (Conv2D)        (None, 512, 18, 18)       2359808   
_________________________________________________________________
block4_conv3 (Conv2D)        (None, 512, 18, 18)       2359808   
_________________________________________________________________
block4_pool (MaxPooling2D)   (None, 512, 9, 9)         0         
_________________________________________________________________
block5_conv1 (Conv2D)        (None, 512, 9, 9)         2359808   
_________________________________________________________________
block5_conv2 (Conv2D)        (None, 512, 9, 9)         2359808   
_________________________________________________________________
block5_pool (MaxPooling2D)   (None, 512, 4, 4)         0         
=================================================================
Total params: 12,354,880
Trainable params: 12,354,880
Non-trainable params: 0
_________________________________________________________________
Found 60 images belonging to 2 classes.
Found 60 images belonging to 2 classes.
Epoch 1/50

 1/10 [==>...........................] - ETA: 6:57 - loss: 0.7880 - acc: 0.3929
 2/10 [=====>........................] - ETA: 6:23 - loss: 0.7920 - acc: 0.4152
 3/10 [========>.....................] - ETA: 5:25 - loss: 0.8292 - acc: 0.3839
 4/10 [===========>..................] - ETA: 4:47 - loss: 0.8184 - acc: 0.3895
 5/10 [==============>...............] - ETA: 3:59 - loss: 0.8159 - acc: 0.3929
 6/10 [=================>............] - ETA: 3:08 - loss: 0.8001 - acc: 0.4048
 7/10 [====================>.........] - ETA: 2:18 - loss: 0.8094 - acc: 0.4184
 8/10 [=======================>......] - ETA: 1:32 - loss: 0.8031 - acc: 0.4247
 9/10 [==========================>...] - ETA: 46s - loss: 0.8041 - acc: 0.4296 
10/10 [==============================] - 899s 90s/step - loss: 0.8125 - acc: 0.4260 - val_loss: 0.8145 - val_acc: 0.4000
Epoch 2/50

 1/10 [==>...........................] - ETA: 6:55 - loss: 0.8487 - acc: 0.4062
 2/10 [=====>........................] - ETA: 5:50 - loss: 0.8443 - acc: 0.4353
 3/10 [========>.....................] - ETA: 5:08 - loss: 0.8430 - acc: 0.4256
 4/10 [===========>..................] - ETA: 4:18 - loss: 0.8258 - acc: 0.4263
 5/10 [==============>...............] - ETA: 3:32 - loss: 0.8310 - acc: 0.4339
 6/10 [=================>............] - ETA: 2:53 - loss: 0.8266 - acc: 0.4397
 7/10 [====================>.........] - ETA: 2:11 - loss: 0.8270 - acc: 0.4305
 8/10 [=======================>......] - ETA: 1:26 - loss: 0.8220 - acc: 0.4347
  9/10 [==========================>...] - ETA: 43s - loss: 0.8311 - acc: 0.4340 

 ......
 ......