A question about sequence length

[ruiann]

Hello,

it looks like the sequence length in your code is fixed, but the length of CASIA handwriting data varies in a big range, I guess the handwriting data generated by your code may end up with many [0, 0, 0, 0, 1] for a short Chinese character like '一'. I try to use bucket to group them by sequence length, the classification for CASIA handwriting works well, but the generation doesn't. Can you give me some information about the training loss decay of your code?

Thanks a lot.