前面用一个简单的4层卷积网络，以猫狗共25000张图片作为训练数据，经过100 epochs的训练，最终得到的准确度为90%。

深度学习中有一种重要的学习方法是迁移学习，可以在现有训练好的模型基础上针对具体的问题进行学习训练，简化学习过程。

这里以imagenet的resnet50模型进行迁移学习训练猫狗分类模型。

importosfrom keras importlayers, optimizers, modelsfrom keras.applications.resnet50 importResNet50from keras.layers import *

from keras.models import Model

定义数据目录

src_path = r'D:\BaiduNetdiskDownload\train'dst_path= r'D:\BaiduNetdiskDownload\large'train_dir= os.path.join(dst_path, 'train')

validation_dir= os.path.join(dst_path, 'valid')

test_dir= os.path.join(dst_path, 'test')

class_name= ['cat', 'dog']

定义网络：

conv_base = ResNet50(weights='imagenet', include_top=False, input_shape=(150, 150, 3))

model=models.Sequential()

model.add(conv_base)

model.add(layers.Flatten())

model.add(layers.Dense(1, activation='sigmoid'))

conv_base.trainable=False

model.compile(loss='binary_crossentropy', optimizer=optimizers.RMSprop(lr=1e-4), metrics=['acc'])

model.summary()

_________________________________________________________________

Layer (type) Output Shape Param #

=================================================================

resnet50 (Model) (None, 5, 5, 2048) 23587712

_________________________________________________________________

flatten_1 (Flatten) (None, 51200) 0

_________________________________________________________________

dense_1 (Dense) (None, 1) 51201

=================================================================

Total params: 23,638,913

Trainable params: 51,201

Non-trainable params: 23,587,712

_________________________________________________________________

定义数据：

from keras.preprocessing.image importImageDataGenerator

batch_size= 64train_datagen=ImageDataGenerator(

rotation_range=40,

width_shift_range=0.2,

height_shift_range=0.2,

shear_range=0.2,

zoom_range=0.2,

horizontal_flip=True, )

test_datagen=ImageDataGenerator()

train_generator=train_datagen.flow_from_directory(#This is the target directory

train_dir,#All images will be resized to 150x150

target_size=(150, 150),

batch_size=batch_size,#Since we use binary_crossentropy loss, we need binary labels

class_mode='binary')

validation_generator=test_datagen.flow_from_directory(

validation_dir,

target_size=(150, 150),

batch_size=batch_size,

class_mode='binary')

训练：

history =model.fit_generator(

train_generator,

steps_per_epoch=train_generator.samples//batch_size,

epochs=20,

validation_data=validation_generator,

validation_steps=validation_generator.samples//batch_size)

训练过程：

Epoch 1/20

281/281 [==============================] - 155s 550ms/step - loss: 0.3354 - acc: 0.8644 - val_loss: 0.2028 - val_acc: 0.9433

Epoch 2/20

281/281 [==============================] - 79s 282ms/step - loss: 0.2502 - acc: 0.9008 - val_loss: 0.2067 - val_acc: 0.9432

Epoch 3/20

281/281 [==============================] - 79s 280ms/step - loss: 0.2318 - acc: 0.9125 - val_loss: 0.1934 - val_acc: 0.9484

Epoch 4/20

281/281 [==============================] - 79s 282ms/step - loss: 0.2179 - acc: 0.9147 - val_loss: 0.2026 - val_acc: 0.9459

......

281/281 [==============================] - 82s 292ms/step - loss: 0.1747 - acc: 0.9332 - val_loss: 0.2202 - val_acc: 0.9452

Epoch 16/20

281/281 [==============================] - 79s 283ms/step - loss: 0.1829 - acc: 0.9329 - val_loss: 0.2256 - val_acc: 0.9513

Epoch 17/20

281/281 [==============================] - 79s 280ms/step - loss: 0.1811 - acc: 0.9322 - val_loss: 0.2079 - val_acc: 0.9466

Epoch 18/20

281/281 [==============================] - 81s 288ms/step - loss: 0.1731 - acc: 0.9345 - val_loss: 0.2149 - val_acc: 0.9466

Epoch 19/20

281/281 [==============================] - 81s 289ms/step - loss: 0.1735 - acc: 0.9346 - val_loss: 0.2038 - val_acc: 0.9504

Epoch 20/20

281/281 [==============================] - 82s 291ms/step - loss: 0.1723 - acc: 0.9347 - val_loss: 0.2228 - val_acc: 0.9463

训练曲线：

     

可以看到在第3轮的时候，就得到最佳模型。

测试结果：

test_generator =test_datagen.flow_from_directory(

test_dir,

target_size=(150, 150),

batch_size=batch_size,

class_mode='binary')

test_loss, test_acc= model.evaluate_generator(test_generator, steps=test_generator.samples //batch_size)print('test acc:', test_acc)

Found 2500 images belonging to 2 classes.

test acc: 0.9536

可以看到迁移学习可以利用已有训练好的模型，进行特征提取，大大加快训练过程，提高模型精度。