keras 的 example 文件 lstm_stateful.py 解析

该程序要通过一个LSTM来实现拟合窗口平均数的功能

先看输入输出数据，

print(x_train[:10])

[[[-0.08453234]]

 [[ 0.02169589]]

 [[ 0.07949955]]

 [[ 0.00898136]]

 [[ 0.0405444 ]]

 [[-0.0227726 ]]

 [[ 0.03033169]]

 [[ 0.03801032]]

 [[ 0.04372695]]

 [[ 0.03803725]]]

print(y_train[:10])

[[-0.03537864]
 [-0.03141822]
 [ 0.05059772]
 [ 0.04424045]
 [ 0.02476288]
 [ 0.0088859 ]
 [ 0.00377955]
 [ 0.03417101]
 [ 0.04086864]
 [ 0.0408821 ]]

 y_train就是 x_train 两两数的平均值，不过 x_train 的最初的第一个数舍去了，看起来 y_train 的第一个数没什么道理似的，这个不必关心

x_train.shape:  (800, 1, 1)
y_train.shape:  (800, 1)
x_test.shape:  (200, 1, 1)
y_test.shape:  (200, 1)

然后是神经网络结构，无论stateful是否为True，结构都是一样的：

_________________________________________________________________
Layer (type)                 Output Shape              Param #
=================================================================
lstm_1 (LSTM)                (1, 20)                   1760
_________________________________________________________________
dense_1 (Dense)              (1, 1)                    21
=================================================================
Total params: 1,781
Trainable params: 1,781
Non-trainable params: 0
_________________________________________________________________

注意：在stateful = True 时，我们要在fit中手动使得shuffle = False

预测结果：

 

从训练时的打印来看：

stateful=True 时，

Epoch 10 / 10
Train on 800 samples, validate on 200 samples
Epoch 1/1
800/800 [==============================] - 2s 2ms/step - loss: 4.9922e-06 - val_loss: 2.9957e-06

而 stateful=False 时，

Epoch 10/10
800/800 [==============================] - 2s 2ms/step - loss: 8.5024e-04 - val_loss: 9.6397e-04

原理部分介绍可以参考

https://blog.csdn.net/qq_27586341/article/details/88239404

——————————————————————

总目录

keras的example文件解析