神经网络用以变形文本矫正系列第八篇

0.前言

上一篇通过样本数555000进行了网络结构不同的测试，这一次通过样本数33000进行测试，查看结果。如果两者的结果一致，则说明样本数的影响对于网络结构的改变造不成影响。另一方面，也为了加快实验过程。

1.实验数据

样本数33000按顺序的实验结果：

23200/26400 [=========================>....] - ETA: 0s - loss: 4.5662e-04 - acc: 0.9544
24400/26400 [==========================>...] - ETA: 0s - loss: 4.5710e-04 - acc: 0.9543
25800/26400 [============================>.] - ETA: 0s - loss: 4.5744e-04 - acc: 0.9546
26400/26400 [==============================] - 1s 41us/step - loss: 4.5701e-04 - acc: 0.9543
Epoch 00180: early stopping

Testing ------------

 200/6600 [..............................] - ETA: 0s
2600/6600 [==========>...................] - ETA: 0s
4600/6600 [===================>..........] - ETA: 0s
6600/6600 [==============================] - 0s 25us/step
test cost: [0.0021578490043545821, 0.94696969335729431]

准确率为95.43%，测试的准确率为94.69%；

==============

开始训练。。。

1.1 网络结构更改

1）四层结构

25600/26400 [============================>.] - ETA: 0s - loss: 0.0012 - acc: 0.9624
26400/26400 [==============================] - 1s 43us/step - loss: 0.0012 - acc: 0.9622
Epoch 00036: early stopping

Testing ------------

 200/6600 [..............................] - ETA: 0s
2600/6600 [==========>...................] - ETA: 0s
4800/6600 [====================>.........] - ETA: 0s
6600/6600 [==============================] - 0s 26us/step
test cost: [0.00093834816696912503, 0.96606060410990857]

3层：准确率为95.43%，测试的准确率为94.69%；

四层：准确率为96.22%，测试的准确率为96.60%；

确实提高了！

2）五层结构

24600/26400 [==========================>...] - ETA: 0s - loss: 8.3841e-04 - acc: 0.9621
26000/26400 [============================>.] - ETA: 0s - loss: 8.2691e-04 - acc: 0.9619
26400/26400 [==============================] - 1s 45us/step - loss: 8.2514e-04 - acc: 0.9619
Epoch 00045: early stopping

Testing ------------

 200/6600 [..............................] - ETA: 0s
2200/6600 [=========>....................] - ETA: 0s
4400/6600 [===================>..........] - ETA: 0s
6600/6600 [==============================] - 0s 28us/step
test cost: [0.00072397346252039301, 0.96393939039923926]

四层：准确率为96.22%，测试的准确率为96.60%；

五层：准确率为96.19%，测试的准确率为96.39%

1.2 小结

可以证明，与样本数555000的结果一致.接下来就使用样本33000进行网络节点数的测试，分为四层网络和五层网络。

2.1四层网络节点数测试

firstLayerInputDim = 858
firstLayerOutput = 572

secondLayerOutput = 572

thirdLayerOutput = 225

lastLayerOutput = 5

更改为：

firstLayerInputDim = 858
firstLayerOutput = 642

secondLayerOutput = 321

thirdLayerOutput = 115


lastLayerOutput = 5

四层未改变节点数：准确率为96.22%，测试的准确率为96.60%；

开始测试：

25800/26400 [============================>.] - ETA: 0s - loss: 9.8590e-04 - acc: 0.9546
26400/26400 [==============================] - 1s 41us/step - loss: 9.8299e-04 - acc: 0.9545
Epoch 00045: early stopping

Testing ------------

 200/6600 [..............................] - ETA: 0s
2600/6600 [==========>...................] - ETA: 0s
5000/6600 [=====================>........] - ETA: 0s
6600/6600 [==============================] - 0s 25us/step
test cost: [0.00090097946072505283, 0.95681817423213611]

四层改变后节点数：准确率为95.45%，测试的准确率为95.68%；

不行，下降了！

更改为

firstLayerInputDim = 858
firstLayerOutput = 572

secondLayerOutput = 572

thirdLayerOutput = 335


lastLayerOutput = 5

训练过程缩略如下：

26000/26400 [============================>.] - ETA: 0s - loss: 0.0011 - acc: 0.9605
26400/26400 [==============================] - 1s 47us/step - loss: 0.0011 - acc: 0.9603
Epoch 00199: early stopping

Testing ------------

 200/6600 [..............................] - ETA: 0s
2400/6600 [=========>....................] - ETA: 0s
4400/6600 [===================>..........] - ETA: 0s
6600/6600 [==============================] - 0s 26us/step
test cost: [0.0010720568031749942, 0.96045454343159997]

四层未改变节点数：准确率为96.22%，测试的准确率为96.60%；

后：

准确率为96.03%，测试的准确率为96.04%；

下降了！

难道说，之前的四层网络及节点数是最好的选择吗？

==========

选择最大样本数进行四层网络训练

估计时间不会短

1067400/1069200 [============================>.] - ETA: 0s - loss: 2.7368e-05 - acc: 0.9915
1068600/1069200 [============================>.] - ETA: 0s - loss: 2.7454e-05 - acc: 0.9915
1069200/1069200 [==============================] - 50s 47us/step - loss: 2.7447e-05 - acc: 0.9915
Epoch 00045: early stopping

Testing ------------

   200/267300 [..............................] - ETA: 58s
266600/267300 [============================>.] - ETA: 0s
267300/267300 [==============================] - 9s 33us/step
test cost: [3.6917571197396347e-05, 0.99528245550391825]

三层：准确率达到了99.11%，测试的准确率是99.47%

四层：已经提高到了99.15%，测试的准确率是99.52%