部分内容来自 博主史丹利复合田的Keras 入门课6 – 使用Inception V3模型进行迁移学习
地址：https://blog.csdn.net/tsyccnh/article/details/78889838
迁移学习主要分为两种
- 第一种即所谓的transfer learning，迁移训练时，移掉最顶层，比如ImageNet训练任务的顶层就是一个1000输出的全连接层，换上新的顶层，比如输出为10的全连接层，然后训练的时候，只训练最后两层，即原网络的倒数第二层和新换的全连接输出层。可以说transfer learning将底层的网络当做了一个特征提取器来使用。
- 第二种叫做fine tune，和transfer learning一样，换一个新的顶层，但是这一次在训练的过程中，所有的（或大部分）其它层都会经过训练。也就是底层的权重也会随着训练进行调整。
下载Inception V3相关数据
import os
from tensorflow.keras import layers
from tensorflow.keras import Model
!wget --no-check-certificate \
https://storage.googleapis.com/mledu-datasets/inception_v3_weights_tf_dim_ordering_tf_kernels_notop.h5 \
-O /tmp/inception_v3_weights_tf_dim_ordering_tf_kernels_notop.h5
from tensorflow.keras.applications.inception_v3 import InceptionV3
local_weights_file = '/tmp/inception_v3_weights_tf_dim_ordering_tf_kernels_notop.h5'
设置pre_model架构(也叫base_model)
InceptionV3模型，其两个参数比较重要，一个是weights，如果是’imagenet’,Keras就会自动下载已经在ImageNet上训练好的参数，如果是None，系统会通过随机的方式初始化参数，目前该参数只有这两个选择。另一个参数是include_top，如果是True会保留全连接层。如果是False，会去掉顶层的全连接网络
这里的input_shape = (150, 150, 3)是我们输入网络的猫狗图片结构
pre_trained_model = InceptionV3(input_shape = (150, 150, 3),
include_top = False,
weights = None)
pre_trained_model.load_weights(local_weights_file)
冻结pre_trained_model所有层，让骨架模型不再被训练
for layer in pre_trained_model.layers:
layer.trainable = False
由于网络结构太长这里只展示部分
pre_trained_model.summary()
这里我们不采用InceptionV3的全部层次，让网络取到mixed7层作为输出连接到我们新添加的网络
last_layer = pre_trained_model.get_layer('mixed7')
print('last layer output shape: ', last_layer.output_shape)
last_output = last_layer.output
last layer output shape: (None, 7, 7, 768)
给网络添加我们自己的几个层次，采用dropout减少过拟合
from tensorflow.keras.optimizers import RMSprop
# Flatten the output layer to 1 dimension
x = layers.Flatten()(last_output)
# Add a fully connected layer with 1,024 hidden units and ReLU activation
x = layers.Dense(1024, activation='relu')(x)
# Add a dropout rate of 0.2
x = layers.Dropout(0.2)(x)
# Add a final sigmoid layer for classification
x = layers.Dense (1, activation='sigmoid')(x)
model = Model( inputs=pre_trained_model.input, outputs=x)
model.compile(optimizer = RMSprop(lr=0.0001),
loss = 'binary_crossentropy',
metrics = ['acc'])
定义目录，采用数据增强
base_dir = '/tmp/cats_and_dogs_filtered'
train_dir = os.path.join( base_dir, 'train')
validation_dir = os.path.join( base_dir, 'validation')
train_cats_dir = os.path.join(train_dir, 'cats') # Directory with our training cat pictures
train_dogs_dir = os.path.join(train_dir, 'dogs') # Directory with our training dog pictures
validation_cats_dir = os.path.join(validation_dir, 'cats') # Directory with our validation cat pictures
validation_dogs_dir = os.path.join(validation_dir, 'dogs')# Directory with our validation dog pictures
train_cat_fnames = os.listdir(train_cats_dir)
train_dog_fnames = os.listdir(train_dogs_dir)
# Add our data-augmentation parameters to ImageDataGenerator
train_datagen = ImageDataGenerator(rescale = 1./255.,
rotation_range = 40,
width_shift_range = 0.2,
height_shift_range = 0.2,
shear_range = 0.2,
zoom_range = 0.2,
horizontal_flip = True)
# Note that the validation data should not be augmented!
test_datagen = ImageDataGenerator( rescale = 1.0/255. )
# Flow training images in batches of 20 using train_datagen generator
train_generator = train_datagen.flow_from_directory(train_dir,
batch_size = 20,
class_mode = 'binary',
target_size = (150, 150))
# Flow validation images in batches of 20 using test_datagen generator
validation_generator = test_datagen.flow_from_directory( validation_dir,
batch_size = 20,
class_mode = 'binary',
target_size = (150, 150))
Found 2000 images belonging to 2 classes.
Found 1000 images belonging to 2 classes.
训练网络
history = model.fit_generator(
train_generator,
validation_data = validation_generator,
steps_per_epoch = 100,
epochs = 20,
validation_steps = 50,
verbose = 2)
Epoch 1/20
100/100 - 29s - loss: 0.3360 - acc: 0.8655 - val_loss: 0.1211 - val_acc: 0.9470
Epoch 2/20
100/100 - 23s - loss: 0.2193 - acc: 0.9145 - val_loss: 0.1096 - val_acc: 0.9640
Epoch 3/20
100/100 - 23s - loss: 0.2038 - acc: 0.9290 - val_loss: 0.0888 - val_acc: 0.9660
Epoch 4/20
100/100 - 22s - loss: 0.1879 - acc: 0.9315 - val_loss: 0.1198 - val_acc: 0.9590
Epoch 5/20
100/100 - 23s - loss: 0.1760 - acc: 0.9415 - val_loss: 0.1155 - val_acc: 0.9660
Epoch 6/20
100/100 - 22s - loss: 0.1771 - acc: 0.9375 - val_loss: 0.1540 - val_acc: 0.9450
Epoch 7/20
100/100 - 23s - loss: 0.1916 - acc: 0.9370 - val_loss: 0.1616 - val_acc: 0.9550
Epoch 8/20
100/100 - 22s - loss: 0.1594 - acc: 0.9440 - val_loss: 0.1422 - val_acc: 0.9630
Epoch 9/20
100/100 - 23s - loss: 0.1669 - acc: 0.9465 - val_loss: 0.1099 - val_acc: 0.9650
Epoch 10/20
100/100 - 23s - loss: 0.1677 - acc: 0.9445 - val_loss: 0.1245 - val_acc: 0.9600
Epoch 11/20
100/100 - 22s - loss: 0.1653 - acc: 0.9470 - val_loss: 0.0918 - val_acc: 0.9730
Epoch 12/20
100/100 - 22s - loss: 0.1542 - acc: 0.9455 - val_loss: 0.1623 - val_acc: 0.9570
Epoch 13/20
100/100 - 22s - loss: 0.1525 - acc: 0.9520 - val_loss: 0.1087 - val_acc: 0.9670
Epoch 14/20
100/100 - 23s - loss: 0.1454 - acc: 0.9565 - val_loss: 0.1314 - val_acc: 0.9640
Epoch 15/20
100/100 - 22s - loss: 0.1279 - acc: 0.9525 - val_loss: 0.1515 - val_acc: 0.9630
Epoch 16/20
100/100 - 23s - loss: 0.1255 - acc: 0.9530 - val_loss: 0.1306 - val_acc: 0.9650
Epoch 17/20
100/100 - 22s - loss: 0.1430 - acc: 0.9575 - val_loss: 0.1226 - val_acc: 0.9660
Epoch 18/20
100/100 - 23s - loss: 0.1350 - acc: 0.9510 - val_loss: 0.1583 - val_acc: 0.9520
Epoch 19/20
100/100 - 22s - loss: 0.1288 - acc: 0.9580 - val_loss: 0.1170 - val_acc: 0.9710
Epoch 20/20
100/100 - 22s - loss: 0.1363 - acc: 0.9550 - val_loss: 0.1260 - val_acc: 0.9660
绘制损失和准确率图
import matplotlib.pyplot as plt
acc = history.history['acc']
val_acc = history.history['val_acc']
loss = history.history['loss']
val_loss = history.history['val_loss']
epochs = range(len(acc))
plt.plot(epochs, acc, 'r', label='Training accuracy')
plt.plot(epochs, val_acc, 'b', label='Validation accuracy')
plt.title('Training and validation accuracy')
plt.legend(loc=0)
plt.figure()
plt.show()
