使用 ResNet50，验证准确性和损失没有改变

Question

我正尝试在 Python (keras) 中使用 ResNet50 进行图像识别。我尝试用 VGG16 完成同样的任务，我得到了一些像这样的结果（对我来说似乎没问题）： resultsVGG16。训练和验证 accuracy/loss 功能在每一步都变得更好，因此网络必须学习。

然而，ResNet50 训练函数下注更好，而验证函数没有改变：resultsResNet

我两次都使用相同的代码和数据，只是模型发生了变化。

那么ResNet50只在训练数据上学习的原因是什么？

我的 ResNet 模型如下所示：

'''python

model = Sequential()
base_model = VGG16(weights='imagenet', include_top=False,input_shape= 
(image_size,image_size,3))
for layer in base_model.layers[:-4]:
    layer.trainable=False

model.add(base_model)
model.add(Flatten())
model.add(Dense(256, activation='relu'))
model.add(Dropout(0.4))
model.add(Dense(NUM_CLASSES, activation='softmax'))

VGG 非常相似：

model = Sequential()
base_model = ResNet50(include_top=False, weights='imagenet', input_shape= 
(image_size,image_size,3))
for layer in base_model.layers[:-8]:
     layer.trainable=False

model.add(base_model)
model.add(Flatten())
model.add(Dense(256, activation='relu'))
model.add(Dropout(0.4))
model.add(Dense(NUM_CLASSES, activation='softmax'))

Answer 1

您的模型没有错误，但这可能是 ResNet 本身的问题，因为在 Github 和 Stack Overflow 中提出了很多问题，1,2,，已经考虑过这个预训练模型。

话虽如此，我找到了一个解决方法，它对我有用，希望对你也有用。

解决方法是替换数据扩充步骤，

Train_Datagen = ImageDataGenerator(rescale=1./255,  rotation_range=40, width_shift_range=0.2,
    height_shift_range=0.2, brightness_range=(0.2, 0.7), shear_range=45.0, zoom_range=60.0,
    horizontal_flip=True, vertical_flip=True)

Val_Datagen = ImageDataGenerator(rescale=1./255,  rotation_range=40, width_shift_range=0.2,
    height_shift_range=0.2, brightness_range=(0.2, 0.7), shear_range=45.0, zoom_range=60.0,
    horizontal_flip=True, vertical_flip=True)

和tf.keras.applications.resnet.preprocess_input，如下图：

Train_Datagen = ImageDataGenerator(dtype = 'float32', preprocessing_function=tf.keras.applications.resnet.preprocess_input)
Val_Datagen = ImageDataGenerator(dtype = 'float32', preprocessing_function=tf.keras.applications.resnet.preprocess_input)

通过如上所示修改Data Augmentation，我卡在50%的验证准确率逐渐提高到97%。这样做的原因可能是 ResNet 可能需要特定的预处理操作（不太确定）。

使用 ResNet50 产生超过 95% 的训练和验证准确度（对于猫和狗数据集）的完整工作代码如下所示：

import tensorflow as tf
from tensorflow.keras.applications import ResNet50
import os
import numpy as np
from keras.preprocessing.image import ImageDataGenerator
from tensorflow.keras.layers import Dense, Dropout, Flatten
from tensorflow.keras.optimizers import Adam
from tensorflow.keras.models import Sequential

# The Convolutional Base of the Pre-Trained Model will be added as a Layer in this Model
Conv_Base = ResNet50(include_top = False, weights = 'imagenet', input_shape = (150,150, 3))

for layer in Conv_Base.layers[:-8]:
    layer.trainable = False

model = Sequential()
model.add(Conv_Base)
model.add(Flatten())
model.add(Dense(units = 256, activation = 'relu'))
model.add(Dropout(0.5))
model.add(Dense(units = 1, activation = 'sigmoid'))

model.summary()

base_dir = 'Deep_Learning_With_Python_Book/Dogs_Vs_Cats_Small'
if os.path.exists(base_dir):    
    train_dir = os.path.join(base_dir, 'train')
    validation_dir = os.path.join(base_dir, 'validation')
    test_dir = os.path.join(base_dir, 'test')
else:
    print("The Folder, {}, doesn't exist'".format(base_dir))

batch_size = 20

Train_Datagen = ImageDataGenerator(dtype = 'float32', preprocessing_function=tf.keras.applications.resnet.preprocess_input)
Val_Datagen = ImageDataGenerator(dtype = 'float32', preprocessing_function=tf.keras.applications.resnet.preprocess_input)

train_gen = Train_Datagen.flow_from_directory(directory = train_dir, target_size = (150,150), 
                                       batch_size = batch_size, class_mode = 'binary')

val_gen = Val_Datagen.flow_from_directory(directory = validation_dir, target_size = (150,150), 
                                       batch_size = batch_size, class_mode = 'binary')

epochs = 15
Number_Of_Training_Images = train_gen.classes.shape[0]
steps_per_epoch = Number_Of_Training_Images/batch_size

model.compile(optimizer = 'Adam', loss = 'binary_crossentropy', metrics = ['accuracy'])

history = model.fit(train_gen, epochs = epochs, 
                    #batch_size = batch_size,
                    validation_data = val_gen, steps_per_epoch = steps_per_epoch)

import matplotlib.pyplot as plt

train_acc = history.history['accuracy']
val_acc = history.history['val_accuracy']
train_loss = history.history['loss']
val_loss = history.history['val_loss']
No_Of_Epochs = range(epochs)

plt.plot(No_Of_Epochs, train_acc, marker = 'o', color = 'blue', markersize = 12, 
                 linewidth = 2, label = 'Training Accuracy')
plt.plot(No_Of_Epochs, val_acc, marker = '.', color = 'red', markersize = 12, 
                 linewidth = 2, label = 'Validation Accuracy')

plt.title('Training Accuracy and Testing Accuracy w.r.t Number of Epochs')

plt.legend()

plt.figure()

plt.plot(No_Of_Epochs, train_loss, marker = 'o', color = 'blue', markersize = 12, 
                 linewidth = 2, label = 'Training Loss')
plt.plot(No_Of_Epochs, val_acc, marker = '.', color = 'red', markersize = 12, 
                 linewidth = 2, label = 'Validation Loss')

plt.title('Training Loss and Testing Loss w.r.t Number of Epochs')

plt.legend()

plt.show()

指标如下图所示，

使用 ResNet50，验证准确性和损失没有改变

With ResNet50 the validation accuracy and loss is not changing

python

tensorflow

keras

resnet

conv-neural-network