search

tbornt / phoneme_ctc

Bidirectional dynamic RNN + CTC for phoneme recognition
44 stars 12 forks source link

Loss not decreasing #2

Open pranoot opened 5 years ago

pranoot commented 5 years ago

Hello,

Thank you sharing the codes. I am trying to train the model on TIMIT dataset from scratch but the loss is not decreasing. Can you please suggest something for the same?

Thanks.

Terrance82 commented 5 years ago

Hi pranoot, Could you share how do you train your model? Because I'm not so sure where or which is the TIMIT dataset.

Also is it possible that because both use the same dataset and hence the accuracy is not improving?

pranoot commented 5 years ago

Hello,

Thank you for the revert. This is link for the timit dataset. I used the same any hyper parameters and same architecture which you used in your codes. Can you tell me,

  1. Did you try with Adam optimizer as well or Momentum performed better.
  2. While training should I try reducing ctc_loss or should I try reducing the label error rate?

If my questions are not clear please let me know. Thanks and regards.