Keras猫狗⼤战六：⽤resnet50预训练模型进⾏迁移学习，精度 提⾼到95.3%
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。
import os
from keras import layers, optimizers, models
from snet50 import ResNet50
from keras.layers import *
dels import Model
定义数据⽬录
src_path = r'D:\BaiduNetdiskDownload\train'
txs
dst_path = r'D:\BaiduNetdiskDownload\large'
李记泽train_dir = os.path.join(dst_path, 'train')
反分裂国家法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'))
ainable = False
modelpile(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
=================================================================
www.tangentTotal params: 23,638,913
华北水利水电学院学报Trainable params: 51,201
Non-trainable params: 23,587,712
_________________________________________________________________
定义数据：
from keras.preprocessing.image import ImageDataGenerator
batch_size = 64
train_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
2012诺贝尔生理学奖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。
豫公网安备41160202000603
站长QQ:729038198
关于我们
投诉建议