为什么深度学习中Keras的准确性总是0.00％且损失很大

Question

我已经在Keras中建立了一个代码来训练神经网络，以模仿我在MATLAB中开发的系统的行为。我将MATLAB的输出和输入数据导出到Keras。每次训练时，精度为0.00％，损耗始终为382.9722 ....

我尝试了所有操作（增加隐藏层，激活功能，批处理大小，纪元等等），似乎没有什么可以解决此问题。如果有人可以告诉我代码或数据有问题，我将不胜感激。

data = pd.read_csv('testkeras.txt')
print(data.head())

Y = data.output
X = data.drop('output', axis=1)

xtrain, xtest, ytrain, ytest = train_test_split(X,Y,test_size=0.5)


model = Sequential()
model.add(Dense(units = 64, input_dim = 6, init = 'uniform',             
activation='relu'))
model.add(Dense(units = 32, activation='relu'))
model.add(Dense(units = 16, activation='relu'))
model.add(Dense(1, activation='sigmoid')) #output layer


model.compile(optimizer = 'rmsprop', loss = 'mean_absolute_error', 
metrics=['acc'])


history = model.fit(xtrain, ytrain, batch_size = 2048, epochs = 20,      
validation_split= 0.2, verbose=1)


score = model.evaluate(xtest, ytest, batch_size=2048)
print(score)

来自matlab的示例原始数据为（输入前6列，输出最后一列）

2,2,2,2,2,2,2.5404e+05
2,2,2,2,2,2,2.5404e+05
2,2,1.9998,1.9998,1.9998,1.9998,2.5404e+05
2,2,1.9988,1.9988,1.9988,1.9988,2.5404e+05
2,2,1.9938,1.9938,1.9938,1.9938,2.5404e+05
2,2,1.9687,1.9687,1.9687,1.9687,2.5403e+05
2,2,1.8431,1.8431,1.8431,1.8431,2.5401e+05
2,2,1.2153,1.2153,1.2153,1.2153,2.5388e+05
2,2,-1.9186,-1.9186,-1.9186,-1.9186,2.5324e+05
2,2,-17.469,-17.469,-17.469,-17.469,2.5007e+05
2,1.9997,-92.331,-92.331,-92.331,-92.331,2.3481e+05
2,1.9936,-402.94,-402.94,-402.94,-402.94,1.7135e+05
2,1.9724,-723.02,-723.02,-723.02,-723.02,1.0558e+05
2,1.9373,-938.65,-938.65,-938.65,-938.65,60759
1.9999,1.8683,-1105.7,-1105.7,-1105.7,-1105.7,24988
1.9999,1.8212,-1152.8,-1152.8,-1152.8,-1152.8,14210
1.9997,1.7097,-1190.6,-1190.6,-1190.6,-1190.6,3712
1.9996,1.6936,-1192.1,-1192.1,-1192.1,-1192.1,3012.4
1.9994,1.6126,-1192.5,-1192.5,-1192.5,-1192.5,898.37
1.9992,1.5645,-1189.5,-1189.5,-1189.5,-1189.5,291.6
1.9987,1.4363,-1176.9,-1176.9,-1176.9,-1176.9,-362.02
1.9981,1.3097,-1161.9,-1161.9,-1161.9,-1161.9,-523.72
1.9974,1.1848,-1146.5,-1146.5,-1146.5,-1146.5,-564.79
1.9965,1.0615,-1131.1,-1131.1,-1131.1,-1131.1,-576.24
1.9955,0.93983,-1115.8,-1115.8,-1115.8,-1115.8,-580.39
1.9944,0.81985,-1100.6,-1100.6,-1100.6,-1100.6,-582.7
1.9931,0.70149,-1085.6,-1085.6,-1085.6,-1085.6,-584.53
1.9918,0.58475,-1070.7,-1070.7,-1070.7,-1070.7,-586.19
1.9903,0.46962,-1055.9,-1055.9,-1055.9,-1055.9,-587.78
1.9887,0.35607,-1041.3,-1041.3,-1041.3,-1041.3,-589.31
1.987,0.2441,-1026.8,-1026.8,-1026.8,-1026.8,-590.78
1.9852,0.13368,-1012.4,-1012.4,-1012.4,-1012.4,-592.21
1.9833,0.024813,-998.22,-998.22,-998.22,-998.22,-593.58
1.9813,-0.082527,-984.13,-984.13,-984.13,-984.13,-594.9
1.9791,-0.18835,-970.17,-970.17,-970.17,-970.17,-596.17
1.9769,-0.29267,-956.34,-956.34,-956.34,-956.34,-597.4
1.9745,-0.39551,-942.64,-942.64,-942.64,-942.64,-598.57
1.9721,-0.49687,-929.07,-929.07,-929.07,-929.07,-599.7
1.9695,-0.59677,-915.62,-915.62,-915.62,-915.62,-600.78

X火车数据是

3492 -0.49055   2.0     2.0     2.0     2.0     2.0
9730 -0.49055   2.0     2.0     2.0     2.0     2.0
3027 -0.49055   2.0     2.0     2.0     2.0     2.0
4307 -0.49055   2.0     2.0     2.0     2.0     2.0
3364 -0.49055   2.0     2.0     2.0     2.0     2.0
(5008, 6)

Y轴数据是

3492   -1.333700e-06
9730    5.215400e-08
3027    4.209600e-06
4307    5.215400e-08
3364    5.215400e-08
Name: output, dtype: float64
(5008,) 

Answer 1

如评论中所述，这是一个回归问题，因此准确性没有意义。

但是您的代码中还有另一个问题。您的最后一层激活功能为S型：

model.add(Dense(1, activation='sigmoid')) #output layer

Sigmoid Function的定义是介于0和1之间，这意味着网络的输出将永远不会小于0或大于1。 因此，您永远也不会接近负输出。我看到2个解决此问题的方法

• 在0到1之间缩放输出数据，例如与Sklearns MinMaxScaler
• 使用激活功能，它可以输出任何数字，例如linar，或者如果您确定数据介于-1和1之间（例如tanh）。

如果您缩放输入数据，它也可能会提高性能（减少错误，加快学习速度）。通常以均值为0且方差为1的方式进行缩放。这就是标准化。您可以这样做，例如与sklearns StandardScaler

您的训练数据似乎也有问题：

3492 -0.49055   2.0     2.0     2.0     2.0     2.0
9730 -0.49055   2.0     2.0     2.0     2.0     2.0
3027 -0.49055   2.0     2.0     2.0     2.0     2.0
4307 -0.49055   2.0     2.0     2.0     2.0     2.0
3364 -0.49055   2.0     2.0     2.0     2.0     2.0

此处的每一行都是相同的，而标签（y）有所不同。您无法建立将相同输入映射到不同输出的网络。