1. 原始：[005/005] Train Acc: 0.460762 Loss: 1.876458 | Val Acc: 0.457851 loss: 1.889790
2. Concatenate the neighboring phonemes for training.
   concat_nframes = 11
   hidden_layers = 1
   idden_dim = 256
   [005/005] Train Acc: 0.636576 Loss: 1.185562 | Val Acc: 0.629358 loss: 1.211024
3. narrower and deeper.
   concat_nframes = 11
   hidden_layers = 6
   hidden_dim = 1024
   [005/005] Train Acc: 0.719124 Loss: 0.879209 | Val Acc: 0.677173 loss: 1.045351
4. wider and shallower.
   concat_nframes = 11
   hidden_layers = 2
   hidden_dim = 1700
   [005/005] Train Acc: 0.736103 Loss: 0.816696 | Val Acc: 0.684846 loss: 1.009970
5. same as before.
   dropout rates equal to 0.25
   [005/005] Train Acc: 0.695657 Loss: 0.952147 | Val Acc: 0.691747 loss: 0.975430
6. same as before.
   dropout rates equal to 0.5
   [005/005] Train Acc: 0.656107 Loss: 1.091103 | Val Acc: 0.678647 loss: 1.020832
7. same as before.
   dropout rates equal to 0.75
   [005/005] Train Acc: 0.592512 Loss: 1.319299 | Val Acc: 0.641634 loss: 1.157008

观察出一些现象：

1. 从 2 可以看出，添加周围的数据是有用的。
2. 从 3 4 可以看出，矮胖版本的更好。
3. 从 4 5 6 7 可以看出，dropout 是有利的，但是不能太大。

关于最优的结果，也就是序号 5，再试一次。以下是结果。

[005/005] Train Acc: 0.695657 Loss: 0.952147 | Val Acc: 0.691747 loss: 0.975430

OK！真的是随机扔的么？

后面讲到了我应该会回来改 待填坑