验证准确性没有增加训练 ResNet50

Question

我正在使用 ResNet50 模型进行微调以使用数据 agumentation 进行人脸识别，但观察到模型准确度在提高，但从一开始的验证准确度并没有提高，我不知道哪里出了问题，拜托查看我的代码。

我已尝试操作已添加的顶层，但没有帮助。

import os
os.environ['KERAS_BACKEND'] = 'tensorflow'
from keras.applications import ResNet50
from keras.models import Sequential
from keras.layers import Dense, Flatten, GlobalAveragePooling2D,Input,Dropout

num_classes = 13

base = ResNet50(include_top=False, weights='resnet50_weights_tf_dim_ordering_tf_kernels_notop.h5',input_tensor=Input(shape=(100,100,3)))
from keras.models import Model

x = base.output

#x = GlobalAveragePooling2D()(x)

x = Flatten()(x)

x = Dense(1024, activation = 'relu')(x)

x = Dropout(0.5)(x)

predictions = Dense(13, activation='softmax')(x)

model = Model(inputs=base.input, outputs=predictions)

for layers in base.layers:
    layers.trainable= False

model.compile(optimizer='sgd', loss='categorical_crossentropy', metrics=['accuracy'])
from keras.preprocessing.image import ImageDataGenerator

train_generator = ImageDataGenerator(featurewise_center=True,
                                rotation_range=20,
                                rescale=1./255,
                                shear_range=0.2,
                                zoom_range=0.2,
                                width_shift_range=0.2,
                                height_shift_range=0.2,
                                horizontal_flip=True)

test_generator = ImageDataGenerator(rescale=1./255)

from sklearn.model_selection import train_test_split

x_train,x_test,y_train,y_test = train_test_split(image,label,test_size=0.2,shuffle=True,random_state=0)

train_generator.fit(x_train)
test_generator.fit(x_test)

model.fit_generator(train_generator.flow(x_train,y_train,batch_size=32),
                       steps_per_epoch =10,epochs=50, 
                    validation_data=test_generator.flow(x_test,y_test))

输出：

Epoch 19/50
10/10 [==============================] - 105s 10s/step - loss: 1.9387 - acc: 0.3803 - val_loss: 2.6820 - val_acc: 0.0709
Epoch 20/50
10/10 [==============================] - 107s 11s/step - loss: 2.0725 - acc: 0.3230 - val_loss: 2.6689 - val_acc: 0.0709
Epoch 21/50
10/10 [==============================] - 103s 10s/step - loss: 1.8884 - acc: 0.3375 - val_loss: 2.6677 - val_acc: 0.0709
Epoch 22/50
10/10 [==============================] - 95s 10s/step - loss: 1.8265 - acc: 0.4051 - val_loss: 2.6799 - val_acc: 0.0709
Epoch 23/50
10/10 [==============================] - 100s 10s/step - loss: 1.8346 - acc: 0.3812 - val_loss: 2.6929 - val_acc: 0.0709
Epoch 24/50
10/10 [==============================] - 102s 10s/step - loss: 1.9547 - acc: 0.3352 - val_loss: 2.6952 - val_acc: 0.0709
Epoch 25/50
10/10 [==============================] - 104s 10s/step - loss: 1.9472 - acc: 0.3281 - val_loss: 2.7168 - val_acc: 0.0709
Epoch 26/50
10/10 [==============================] - 103s 10s/step - loss: 1.8818 - acc: 0.4063 - val_loss: 2.7071 - val_acc: 0.0709
Epoch 27/50
10/10 [==============================] - 106s 11s/step - loss: 1.8053 - acc: 0.4000 - val_loss: 2.7059 - val_acc: 0.0709
Epoch 28/50
10/10 [==============================] - 104s 10s/step - loss: 1.9601 - acc: 0.3493 - val_loss: 2.7104 - val_acc: 0.0709

Answer 1

发生这种情况是因为我只是直接添加了全连接层而没有先训练它，正如 keras 博客中提到的那样， https://blog.keras.io/building-powerful-image-classification-models-using-very-little-data.html

为了进行微调，所有层都应该从经过适当训练的权重开始：例如，您不应该在预训练的卷积基础之上添加一个随机初始化的全连接网络。这是因为由随机初始化的权重触发的大梯度更新会破坏卷积基中的学习权重。在我们的例子中，这就是为什么我们首先训练顶级分类器，然后才开始微调卷积权重。

所以答案是首先单独训练顶级模型，然后创建一个具有 ResNet50 模型及其权重的新模型，顶级模型及其权重位于 resnet 模型（基础模型）之上，然后首先通过冻结对其进行训练基础模型（ResNet50）和基础模型的最后一层。