[英]Keras Logistic Regression returns nan on first epoch
I have a really simple data set. 我有一个非常简单的数据集。 I cleaned the data (one hot encoding, normalizing the data and check for missing values or NaNs) and my learning rate is pretty small. 我清理了数据(一个热编码,规范化数据并检查缺失值或NaN),我的学习率非常小。 But when tried to run a simple logistic regression using Keras and Theano as backend 但是当试图使用Keras和Theano作为后端进行简单的逻辑回归时
model = Sequential()
model.add(Dense(input_dim=84, activation='softmax',
bias_initializer='normal', units=6))
rms = RMSprop(lr=0.001, rho=0.9, epsilon=1e-08, decay=0.0)
model.compile(optimizer=rms, loss='categorical_crossentropy')
batch_size = 100
nb_epoch = 1
n = X.shape[0] # number of training examples
history = model.fit(X, Y_oh, batch_size=batch_size, epochs=nb_epoch)
Error: 错误:
Epoch 1/1
5459/5459 [==============================] - 0s - loss: nan
I checked here and tried to downgrade Theano to the version mentioned but it still gives the same error 我在这里检查并试图将Theano降级到提到的版本,但它仍然给出了相同的错误
Here is what X looks like 这是X的样子
[[ 0.35755179 0.13747887 0.3 ..., 0. 0. 0. ]
[ 0.36401758 0.14963742 0.55 ..., 0. 0. 0. ]
[ 0.37889517 0.13775149 0.275 ..., 0. 0. 0. ]
...,
[ 0.34387947 0.18706723 0.05 ..., 0. 0. 0. ]
[ 0.35708726 0.12905512 0.75 ..., 0. 0. 0. ]
[ 0.37915882 0.08061174 0.05 ..., 0. 1. 0. ]]
and Y_oh (generated using the following code): 和Y_oh(使用以下代码生成):
Y_oh = np_utils.to_categorical(Y.T[0],6)
[[ 0. 0. 0. 0. 0. 1.]
[ 0. 0. 0. 0. 0. 1.]
[ 0. 0. 0. 0. 0. 1.]
...,
[ 1. 0. 0. 0. 0. 0.]
[ 1. 0. 0. 0. 0. 0.]
[ 1. 0. 0. 0. 0. 0.]]
The problem is your activation here 问题是你在这里激活
model.add(Dense(input_dim=84, activation='softmax',
bias_initializer='normal', units=6))
Swap 'softmax' for 'sigmoid' (or 'tanh') and you should be good. 交换'ssmoid'(或'tanh')的'softmax',你应该很好。
Softmax has the property that the sum of its outputs is 1 , so that the network's output has a probability interpretation. Softmax具有其输出之和为1的特性 ,因此网络的输出具有概率解释。 The problem is, since you have only 1 output, it would either will never train (since the output will always be the same) or get unreasonable gradients trying to do so. 问题是,由于你只有1个输出,它将永远不会训练(因为输出总是相同的)或者得到不合理的渐变试图这样做。
eg Fix 例如修复
model.add(Dense(input_dim=84, activation='sigmoid',
bias_initializer='normal', units=6))
Try to apply Log(0) and see what is the result. 尝试应用Log(0)并查看结果是什么。 it is undefined "NAN" I answered this question in details here NAN in neural network 它是未定义的“NAN”我在这里详细回答了这个问题NAN在神经网络中
声明:本站的技术帖子网页,遵循CC BY-SA 4.0协议,如果您需要转载,请注明本站网址或者原文地址。任何问题请咨询:yoyou2525@163.com.