SGD Bug! about deepspeech.torch HOT 6 CLOSED

I think its probably a better idea as to find out why the gradOutput is 0, are you using a different dataset etc? I know when using a real system you probably don't want to add the max pooling layer (I do for the sake of reducing memory overhead).

By adding the max pooling layer we get very close to have less time steps available than in the labels which will cause all sorts of issues!

from deepspeech.torch.

Yes， i using a different dataset，This code speed up 10 times than i used before.I will do more test on my dataset.

from deepspeech.torch.

The code speeds up using a different dataset? May I ask which dataset you are using?

from deepspeech.torch.

Oh i means ,the code speeds up than eesen which i used before.The data set i use is a personal data set for Mandarin speech recognize.

from deepspeech.torch.

Ahh I see. Well removing the max pooling layer should hopefully prevent the gradOutputs from going to 0, in the DeepSpeech network (as I've read in the paper) it doesn't have max poolings. Should hopefully fix the grad outputs from being 0s!

from deepspeech.torch.

Hi @TangXing , some questions when I am using this code to train a Mandarin model, can i have your wechat and ask for some advice? thanks!

from deepspeech.torch.