本节基于Keras搭建全连接神经网络,实现对mnist手写数字的识别。通过设置不同的优化器、损失函数,观察模型训练效果。此外,为防止过拟合,在代码中向读者展示了正则化和dropout的使用方法。代码如下:
'''
通过全连接神经网络实现mnist手写数字分类
1 对比交叉熵损失与均方误差损失函数对训练结果的影响
2 对比不同优化算法对训练结果的影响
'''
from keras.models import Sequential
from keras.layers import Dense, Dropout
from keras.datasets import mnist
from keras.optimizers import SGD, Adam, Adagrad, RMSprop
from keras.regularizers import l2
from keras.utils import np_utils
import matplotlib.pyplot as plt
class Classification:
def __init__(self):
(self.x_train, self.y_train), (self.x_test, self.y_test) = mnist.load_data()
dim = self.x_train.shape[1] * self.x_train.shape[2]
self.x_train, self.x_test = self.x_train.reshape((self.x_train.shape[0], -1)) / 255, self.x_test.reshape((self.x_test.shape[0], -1)) / 255
self.y_train, self.y_test = np_utils.to_categorical(self.y_train, num_classes = 10), np_utils.to_categorical(self.y_test, num_classes = 10)
self.model = Sequential()
self.model.add(Dense(100, activation = 'relu', input_dim = dim, kernel_regularizer = l2(0.0001)))
self.model.add(Dropout(0.3))
self.model.add(Dense(80, activation = 'relu'))
self.model.add(Dense(30, activation = 'relu'))
self.model.add((Dense(20, activation = 'tanh')))
self.model.add(Dense(10, activation = 'softmax'))
# 评估和画图函数
def evaluate_and_plot(self, loss):
loss_1, acc_1 = self.model.evaluate(self.x_train, self.y_train)
loss_2, acc_2 = self.model.evaluate(self.x_test, self.y_test)
print('train loss:', loss_1, '\ntrain accuracy:', acc_1)
print('test loss:', loss_2, '\ntest accuracy:', acc_2)
plt.plot(loss, 'r')
plt.xlabel('episode')
plt.ylabel('loss')
plt.show()
# 优化器和学习率设置函数
def set_hyperparameter(self):
print('请设置优化器(1:SGD, 2:Adam, 3:Adagrad, 4:RMSprop)和学习率lr, 空格分隔:')
s = list(input().split(' '))
opt, learning_rate = int(s[0]), float(s[1])
if opt == 1:
optim = SGD(lr = learning_rate)
elif opt == 2:
optim = Adam(lr = learning_rate)
elif opt == 3:
optim = Adagrad(lr = learning_rate)
else:
optim = RMSprop(lr = learning_rate)
l = input('请设置损失函数, 交叉熵损失(C)和均方误差损失(M):')
loss_f = 'categorical_crossentropy' if l == 'C' else 'mse'
return optim, loss_f
def train(self, epoch, batchsize):
optim, loss_fun = self.set_hyperparameter()
self.model.compile(optimizer = optim, loss = loss_fun, metrics = ['accuracy'])
history = self.model.fit(self.x_train, self.y_train, batch_size = batchsize, epochs = epoch)
loss = history.history['loss']
return loss
if __name__ == '__main__':
Classify = Classification()
loss = Classify.train(10, 32)
Classify.evaluate_and_plot(loss)
使用SGD优化器,均方误差损失函数,学习率0.001。在训练过程中,loss变化如下。训练集识别精度0.21311,测试集识别精度0.21490。
使用SGD优化器,交叉熵损失函数,学习率0.001。在训练过程中,loss变化如下。训练集识别精度0.89151,测试集识别精度0.89780。
使用Adam优化器,均方误差损失函数,学习率0.001。训练过程中,loss变化如下。训练集识别精度0.96388,测试集识别精度0.96259。
使用Adam优化器,交叉熵损失函数,学习率为0.001。训练过程中,loss变化如下。训练集识别精度0.98274,测试集识别精度0.97229。
在损失函数初始值较大时,交叉熵损失函数可以获得更大的梯度,所以收敛速度要快于均方误差损失函数。此外,Adam优化器对梯度的一阶矩和二阶矩进行了综合考虑,训练效率也高SGD。在此时中,也发现有时训练集识别精度略高于测试集识别精度,即发生了轻微的过拟合。读者可自行尝试设置其他优化器,还可通过设置不同的batch_size、epoch等参数,对比训练结果。