在使用迁移学习训练用于对象检测的序列模型期间，训练和验证准确度保持不变

Question

我尝试训练一个二元分类模型，但经过 3 个 epoch 后，验证准确率仍然保持在 0.5000。

数据集包含两个类别的 1512 张图像，因此总共有 3024 张图像。我使用 keras 使用 VGG16 模型进行迁移学习。

from keras import models
from keras import layers
from keras import optimizers
from keras.preprocessing.image import ImageDataGenerator
from keras.callbacks import TensorBoard, EarlyStopping, ModelCheckpoint

# Stop when we stop learning.
early_stopper = EarlyStopping(patience=10)
# tensorboard
tensorboard = TensorBoard(log_dir='./logs')
model_check_point = ModelCheckpoint('vgg16.h5', save_best_only=True)
train_dir = 'dataset\\training_set'
validation_dir = 'dataset\\validation_set'
image_size_x = 360
image_size_y = 180
#Load the VGG model
vgg_conv = VGG16(weights='imagenet', include_top=False, input_shape=(image_size_x, image_size_y, 3))
# Freeze the layers except the last 4 layers
for layer in vgg_conv.layers[:-4]:
    layer.trainable = False

# Create the model
model = models.Sequential()

# Add the vgg convolutional base model
model.add(vgg_conv)

# Add new layers
model.add(layers.Flatten())
model.add(layers.Dense(1024, activation='relu'))
model.add(layers.Dropout(0.5))
model.add(layers.Dense(1, activation='sigmoid'))

train_datagen = ImageDataGenerator(
      rescale=1./255,
      rotation_range=20,
      width_shift_range=0.2,
      height_shift_range=0.2,
      horizontal_flip=True,
      fill_mode='nearest')

validation_datagen = ImageDataGenerator(rescale=1./255)

# Change the batchsize according to your system RAM
train_batchsize = 4
val_batchsize = 4

train_generator = train_datagen.flow_from_directory(
        train_dir,
        target_size=(image_size_x, image_size_y),
        batch_size=train_batchsize,
        class_mode='binary')

validation_generator = validation_datagen.flow_from_directory(
        validation_dir,
        target_size=(image_size_x, image_size_y),
        batch_size=val_batchsize,
        class_mode='binary')
# Compile the model
model.compile(loss='binary_crossentropy',
              optimizer=optimizers.Adam(lr=0.1),
              metrics=['acc'])
# Train the model
history = model.fit_generator(
      train_generator,
      steps_per_epoch=train_generator.samples/train_generator.batch_size ,
      epochs=7,
      validation_data=validation_generator,
      validation_steps=validation_generator.samples/validation_generator.batch_size,
      verbose=1,
      callbacks=[tensorboard, early_stopper, model_check_point])

结果是

Using TensorFlow backend.
2019-02-03 14:46:06.520723: I tensorflow/core/platform/cpu_feature_guard.cc:141] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2
Found 2416 images belonging to 2 classes.
Found 608 images belonging to 2 classes.
Epoch 1/7
604/604 [==============================] - 2046s 3s/step - loss: 8.0208 - acc: 0.5008 - val_loss: 8.0590 - val_acc: 0.5000
Epoch 2/7
604/604 [==============================] - 1798s 3s/step - loss: 8.0055 - acc: 0.5033 - val_loss: 8.0590 - val_acc: 0.5000
Epoch 3/7
604/604 [==============================] - 2500s 4s/step - loss: 8.0054 - acc: 0.5033 - val_loss: 8.0590 - val_acc: 0.5000

我尝试将学习率从 0.0001 提高到 0.01，尝试使用另一个优化器 (RMSprop)，但验证准确度仍然保持在 0.5000。 这是模型摘要

_________________________________________________________________
Layer (type)                 Output Shape              Param #
=================================================================
inception_v3 (Model)         (None, 9, 4, 2048)        21802784
=================================================================
Total params: 21,802,784
Trainable params: 0
Non-trainable params: 21,802,784
_________________________________________________________________
None

Answer 1

VGG 需要与rescale = 1/255. 不同的预处理。您可以从同一个模块中导入正确的预处理输入函数并将其提供给ImageDataGenerator：

from keras.applications.vgg16 import preprocess_input
train_datagen = ImageDataGenerator(preprocessing_function=preprocess_input)

请注意，您应该省略 rescale 参数，因为这是由预处理函数处理的。另见this answer。

如果这没有帮助，我建议您尝试冻结整个卷积基并只训练顶级分类器。卷积基包含预训练的权重，而顶部分类器中的权重是随机初始化的。这可能会导致对预训练权重的大梯度更新，从而使它们变得无用。此外，您没有那么多数据，因此冻结卷积堆栈可能是一个好主意，无论它是否会在此处引起您的问题。请参阅this answer 以了解有人遇到 VGG 问题的示例。

除此之外，您的模型定义看起来是正确的。