Hi all, In the demo scripts of HTS version 2.3.1, the variable "restoreCkpt" must be set considering size of training data, number of epochs, and mini-batch size for loading a trained neural network, otherwise a file read error may be occurred. In terms of avoiding the error, automatically loading a final checkpoint, namely a neural network whose parameters are updated at the end of training, seems preferable. I have made a patch file for automatically loading a final checkpoint. You can apply the attached patch file to the current demo scripts using the following commands: cd HTS-demo_CMU-ARCTIC-SLT; patch -p 1 -d . < HTS-demo_CMU-ARCTIC-SLT.patch If $restoreCkpt = 0, a final checkpoint is loaded, otherwise a user-specified checkpoint is used for parameter generation. Regards, Takenori -- Nagoya Institute of Technology Tokuda and Nankaku Laboratory Takenori Yoshimura takenori@xxxxxxxxxxxxxxx
Attachment:
HTS-demo_CMU-ARCTIC-SLT.patch
Description: Binary data