我在Keras中使用张量流作为后端安装LSTM网络时出现问题：
def build_recurrent(input_dim, output_dim):
model = Sequential()
model.add(LSTM(200,input_dim=input_dim,activation='tanh'))
#model.add(Dense)
model.add(Dropout(0.5))
model.add(Dense(output_dim,activation='softmax'))
return model
当我在输出图层上使用softmax（使用比例）时，我得到了这个：
Epoch 1/1500
0s - loss: nan - mean_squared_error: nan - mean_absolute_error: nan
Epoch 2/1500
0s - loss: nan - mean_squared_error: nan - mean_absolute_error: nan
Epoch 3/1500
0s - loss: nan - mean_squared_error: nan - mean_absolute_error: nan
Epoch 4/1500
0s - loss: nan - mean_squared_error: nan - mean_absolute_error: nan
Epoch 5/1500
0s - loss: nan - mean_squared_error: nan - mean_absolute_error: nan
while，当使用其他激活功能时，例如 tanh ：
Epoch 1/1500
0s - loss: 0.9173 - mean_squared_error: 0.9595 - mean_absolute_error: 0.9173
Epoch 2/1500
0s - loss: 1.0652 - mean_squared_error: 1.1457 - mean_absolute_error: 1.0652
Epoch 3/1500
0s - loss: 1.0652 - mean_squared_error: 1.1457 - mean_absolute_error: 1.0652
Epoch 4/1500
0s - loss: 1.0652 - mean_squared_error: 1.1457 - mean_absolute_error: 1.0652
Epoch 5/1500
0s - loss: 1.0652 - mean_squared_error: 1.1457 - mean_absolute_error: 1.0652
它会是什么样的问题？我应该更改隐藏层的激活功能吗？
另一个奇怪的事情是，这只发生在LSTM模型中，而当我使用简单的前馈网络时，我没有收到任何错误