在keras多变量神经网络中获取nan训练成本

Question

我正在训练一个神经网络，让它有6个输入和2个输出。我正在使用带有Tensorflow后端的Keras。经过预处理后，下面是我的代码：

training_examples = features.head(2584)
training_targets = targets.head(2584)

validation_examples = features.tail(650)
validation_targets = targets.tail(650)

model = Sequential()
model.add(Dense(12, input_dim=6))
model.add(Dense(8))
model.add(Dense(8))
model.add(Dense(2))
model.compile(loss='mse', optimizer='sgd')
print("Training--------")
for step in range(500):
  cost = model.train_on_batch(training_examples, training_targets)
  if step % 100 == 0:
    print('train cost: ', cost)

每次我运行这个命令都会得到如下输出

Training--------
train cost:  6670.4097
train cost:  nan
train cost:  nan
train cost:  nan
train cost:  nan

第一次培训的成本通常在2,000- 14000之间。特征和目标的数值都小于100。我不确定为什么会发生这种事。

编辑:我添加了features.info()和targets.info()来检查空值，数据帧中没有空值。

<class 'pandas.core.frame.DataFrame'> Int64Index: 3231 entries, 0 to 3230 Data columns (total 6 columns): TBRG_Rain_infield 3231 non-null float64 numRange_infield 3231 non-null float64 Air_T_edge 3231 non-null float64 RH_edge 3231 non-null float64 TBRG_Rain_edge 3231 non-null float64 numRange_edge 3231 non-null float64 dtypes: float64(6) memory usage: 176.7 KB <class 'pandas.core.frame.DataFrame'> Int64Index: 3231 entries, 0 to 3230 Data columns (total 2 columns): Air_T 3231 non-null float64 RH 3231 non-null float64 dtypes: float64(2) memory usage: 75.7 KB

Answer 1

您的数据帧看起来是正确的，但是您可能应该将您的输入特征缩放到0到1之间，或者具有均值0和单位方差。我试着重现你的例子，一次有伸缩，一次没有缩放。

不需要扩展：

from keras.models import Sequential
from keras.layers import Dense
import pandas as pd
import numpy as np

features = pd.DataFrame(np.random.randint(0, 100, size=(1000, 6)).astype(float))
targets = pd.DataFrame(np.random.rand(1000, 2), dtype=np.float64)

training_examples = features.head(100)  
training_targets = targets.tail(100)

model = Sequential()
model.add(Dense(12, input_dim=6))
model.add(Dense(8))
model.add(Dense(8))
model.add(Dense(2))
model.compile(loss='mse', optimizer='sgd')
print("Training--------")
for step in range(500):
    cost = model.train_on_batch(training_examples, training_targets)
    if step % 100 == 0:
        print('train cost: ', cost)

以输出形式给出：

Training--------
train cost:  6834.277
train cost:  nan
train cost:  nan
train cost:  nan
train cost:  nan

但是，如果我将特征初始化为介于0和1之间：

features = pd.DataFrame(np.random.rand(1000, 6), dtype=np.float64)

这是输出：

Training--------
train cost:  1.1240386
train cost:  0.09793612
train cost:  0.08868038
train cost:  0.084703445
train cost:  0.0826226

您可以从scikit StandardScaler -learn to scale your data查看一下。