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 keras.applications.resnet50 import ResNet50
from keras.layers import *    
from keras.models import Model

定义数据目录

src_path = r'D:BaiduNetdiskDownload	rain'
dst_path = r'D:BaiduNetdiskDownloadlarge'
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 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
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。