SpecAugment: max_num_masks is at least min_num_masks #37
Conversation
JackTemaki
requested review from
albertz,
curufinwe,
christophmluscher,
mmz33,
Marvin84,
Judyxujj and
kuacakuaca
| @@ -1,3 +1,4 @@ | |||
| import numpy as np | |||
There was a problem hiding this comment.
We mostly use import numpy in other code (consistent to import torch), here also in the repo (although not too much yet), but also in RETURNN.
Although this variant here is more the official recommended way.
So maybe for i6_models, we can independently decide now what variant to pick and then always keep it consistent.
There was a problem hiding this comment.
You merged without resolving this here, or commenting here. So what do you mean?
There was a problem hiding this comment.
Ah now I see #38. Shouldn't we have waited until this was decided?
There was a problem hiding this comment.
As this is an independent non-critical issue, I would not delay this PR by that. If we decide different to this PR we can just change it back.
albertz
changed the title
I triggered one "last batch" in Tedlium that had only one sequence that was apparently so short that max_masks got smaller than min_masks for the time.