使用Keras中的数据增强功能降低VGG16的精度

Question

我对Keras中具有数据增强功能的特征提取有疑问。我正在建立狗的品种分类器。

通过特征提取，我指的是通过在顶部添加Dense层并在输入数据上端到端运行整个对象来扩展模型（conv_base，VGG16）。这将使我能够使用数据增强，因为每一次输入图像都会在模型每次看到时都经过卷积基数。

训练集：属于133个类别的6680张图像

验证集：属于133个类别的835张图像

测试集：836个图像，属于133个类别

我能够彼此独立地成功实现数据增强和特征提取，但是当我尝试结合使用2时，由于某种原因，我的准确性变得非常低。为什么是这样？我的方法是否存在严重错误？

from keras.applications import VGG16

from keras.preprocessing.image import ImageDataGenerator

conv_base = VGG16(weights='imagenet',
                  include_top=False,
                  input_shape=(224, 224, 3))

model = Sequential()

model.add(conv_base)

conv_base.trainable = False

model.add(GlobalAveragePooling2D())

model.add(Dense(133, activation='softmax'))

model.compile(optimizer='rmsprop', loss='categorical_crossentropy', metrics=['accuracy'])

train_datagen_aug = 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,)

test_datagen_aug = ImageDataGenerator(rescale=1./255)

train_generator_aug = train_datagen_aug.flow_from_directory(
    'myImages/train',
    target_size=(224, 224),
    batch_size=50,
    class_mode='categorical')

validation_generator_aug = test_datagen_aug.flow_from_directory(
        'myImages/valid',
        target_size=(224, 224),
        batch_size=32,
        class_mode='categorical')

checkpointer_aug = ModelCheckpoint(filepath='saved_models/dogs_transfer_aug_model.h5', 
                            save_best_only=True)

history = model.fit_generator(
      train_generator_aug,
      steps_per_epoch=130,
      epochs=20,
      validation_data=validation_generator_aug,
      verbose=1,
      callbacks=[checkpointer_aug],
      validation_steps=26)

输出看起来像这样：

Epoch 1/20
130/130 [==============================] - 293s - loss: 15.9044 - acc: 0.0083 - val_loss: 16.0019 - val_acc: 0.0072
Epoch 2/20
130/130 [==============================] - 281s - loss: 15.9972 - acc: 0.0075 - val_loss: 15.9977 - val_acc: 0.0075
Epoch 3/20
130/130 [==============================] - 280s - loss: 16.0220 - acc: 0.0060 - val_loss: 15.9977 - val_acc: 0.0075
Epoch 4/20
130/130 [==============================] - 280s - loss: 15.9941 - acc: 0.0077 - val_loss: 16.0019 - val_acc: 0.0072

Answer 1

我建议这是模型超过拟合问题，如模型的损失和准确性所示。我们可以尝试使用较小版本的VGG16（减少层数）

from keras.models import Sequential
from keras.layers.core import Flatten, Dense, Dropout
from keras.layers.convolutional import MaxPooling2D, ZeroPadding2D

NUMBER_OF_TRAINING_SAMPLES = 6668 
NUMBER_OF_VALIDATION_SAMPLES = 835 # let's say you have 400 dogs and 400 cats
batch_size = 32
out_classes = 133
input_shape=(224, 224, 3)

def buildSmallVGG(out_classes, input_shape):
    model = Sequential()
    model.add(ZeroPadding2D((1,1),input_shape=input_shape))
    model.add(Conv2D(16, 3, 3, activation='relu'))
    model.add(MaxPooling2D((2,2), strides=(2,2)))

    model.add(ZeroPadding2D((1,1)))
    model.add(Conv2D(32, 3, 3, activation='relu'))
    model.add(MaxPooling2D((2,2), strides=(2,2)))


    model.add(ZeroPadding2D((1,1)))
    model.add(Conv2D(64, 3, 3, activation='relu'))
    model.add(MaxPooling2D((2,2), strides=(2,2)))

    model.add(Flatten())
    model.add(Dense(256, activation='relu'))
    model.add(Dropout(0.5))

    model.add(Dense(out_classes, activation='softmax'))
    return model

model = buildSmallVGG(out_classes, input_shape)

history = model.fit_generator(
      train_generator_aug,
      steps_per_epoch=NUMBER_OF_TRAINING_SAMPLES // batch_size,
      epochs=20,
      validation_data=validation_generator_aug,
      callbacks=[checkpointer_aug],
      validation_steps=NUMBER_OF_VALIDATION_SAMPLES // batch_size)

以上未经测试。如果您可以就损失，准确性等方面分享结果，那就太好了。