the predict text is longer than the labeled text

[SharonJin422]

Hello, I got a problem that the predict text is longer than the labeled text by non-sense words jointed the corrected words. What reasons may cause that?

I set the max seq_len to 128, I am not sure if it's that reason.

[MaybeShewill-CV]

@SharonJin422 Try to reduce seq_len

[SharonJin422]

Hello, I have taken contrast tests, with input size = [32, 512], and the max time step will be 128,but set different seq_len:
1)seq_len = 64
2) seq_len = 15

I find that the first test has more accurate predictions than the second, but i cannot explain what cause this, do you have any idea?

[MaybeShewill-CV]

@SharonJin422 You may look up something relative to ctc:)

[SharonJin422]

thanks, i finally figure out that it's the learning rate set too large.

[MaybeShewill-CV]

@SharonJin422 welcome:)