train_multisplice_accel2参数优化

I trained chinese DNN model using steps/nnet2/train_multisplice_accel2.sh. Our training data is More than 4000 hours. Do we need to change The Default Parameters?If We want to do parameter optimization in DNN training, could you give us some advice? Which parameters can be adjusted? 我使用steps / nnet2 / train_multisplice_accel2.sh训练了中文DNN模型。我们的训练数据超过4000小时。我们是否需要更改默认参数？如果我们想在DNN训练中进行参数优化，您能给我们一些建议吗？哪些参数可以调整？

Fisher english recipe (1900 hours) could give you insight on the parameters. You will have to experiment anyway -- it seems the optimal
multisplice config depends significantly not only on the size of the corpus but on how noisy the data are. 费舍尔英语模型（1900小时）可以让您深入了解参数。无论如何，您都必须进行实验-最佳多拼接配置似乎不仅取决于语料库的大小，而且还取决于数据的噪音程度。

You could increase pnorm-input-dim/pnorm-output-dim values to 4000/400. The multi-splice config remains the same, unless the data is reverberant in which case the context can be increased with config ""layer0/-2:-1:0:1:2 layer1/-1:2 layer3/-3:3 layer4/-10:-7:2:5"". 您可以将pnorm-input-dim / pnorm-output-dim值增加到4000/400。除非数据是混响的，否则训练配置将保持不变，在这种情况下，可以通过配置“” layer0 / -2：-1：0：1：2 layer1 / -1：2 layer3 / -3来增加上下文： 3 layer4 / -10：-7：2：5“”。

In addition to these, increasing the num-hidden-layers or mix-up, Does this help? 除了这些，增加num-hidden-layers或混淆，这有帮助吗？

Yes, depending on you data and size of the corpus, increasing (or lowering) these might help.
是的，取决于您的数据和语料库的大小，增加（或降低）这些可能会有所帮助。