Question

我已经从VGG16加载了权重，并将其添加到我的顺序模型中。我想通过冻结顶层（精细调整）来训练更轻量的VGG16。

一切都很好：我能够建立模型并预测新图像。但是现在我想加载模型，但是我无法做到。

这是我尝试显示的代码，如下所示：

model1 = applications.VGG16(weights='imagenet', 
include_top=False,input_shape=(img_width,img_height,3))

train_datagen = 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,
                                   fill_mode='nearest')

test_datagen = ImageDataGenerator(rescale=1./255)


 train_generator = train_datagen.flow_from_directory(train_data_dir,
                                        target_size=(img_width, img_height),
                                        batch_size=size_batch, 
                                        class_mode='binary', 
                                        shuffle=False) 


# repeat with the validation data

 test_generator = test_datagen.flow_from_directory(validation_data_dir,
                                       target_size=(img_width, img_height),
                                       batch_size=size_batch,
                                       class_mode='binary',
                                       shuffle=False)


 model = Sequential()
 model.add(Flatten(input_shape=model1.output_shape[1:]))
 model.add(Dense(256, activation='relu'))
 model.add(Dropout(0.2))
 model.add(Dense(1, activation='sigmoid'))

 new_model=Sequential()

 for l in model1.layers:
      new_model.add(l)

  new_model.add(model)

 for layer in new_model.layers[:25]:
       layer.trainable=False

 new_model.compile(optimizer=optimizers.SGD(lr=1e-3, 
 momentum=0.9),loss='binary_crossentropy',
              metrics=['accuracy'])
 checkpoint = ModelCheckpoint(fine_tuned_model_path, monitor='val_acc', 
                             verbose=1, save_best_only=True, 
                             save_weights_only=False, mode='auto')
# fine-tune the model
fit=new_model.fit_generator(train_generator,
                          steps_per_epoch=33,
                          nb_epoch=1,

                          validation_data=test_generator,

                          verbose=1,callbacks=[checkpoint])

然后我试图加载模型：

load_model("C:/Users/hi/POC/Fine_Tune/model.h5")

这是我收到的错误：

ValueError：您正试图加载包含14层的重量文件进入具有1层的模型。

Answer 1

我不明白为什么您必须定义一个新模型并将以前的VGG16层加载到新模型中。

我建议的最佳工作是冻结VGG16的层就像您在上一个for循环中所做的那样，将所需的架构作为可训练的层

最终，这将导致您删除两个for循环嵌入其中。

# the way you loaded your images and did not include_top layer 
model1 = applications.VGG16(weights='imagenet', include_top=False, input_shape = (img_width, img_height, 3))

#to see the structure of your architecture
model.summary() 

#freezing the layers you do not want for training in your architecture
for layer in model1.layers[:25]:
    layer.trainable = False

#the rest is the same from here on forth with the exclusion of the two for loops
#which you need to remove as they are no longer required.
train_datagen = 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,
                               fill_mode='nearest')

#etc...

Answer 2

根据Keras问题8898，可以通过编辑Keras代码keras/applications/vgg16.py

避免此错误

，以便用于读取的行 model.load_weights(weights_path)现在读为model.load_weights(weights_path, by_name=True)

我发现这也适用于其他应用程序模型的imagenet权重。 nasnet。

使用VGG16对模型进行微调后无法加载权重

2 个答案: