Add support for unconstrained-egs for chain training + example recipes #2383
Conversation
| Cool. |
| Looks like the test code doesn't compile-- my branch has the same issue. Can you please address that? |
| I haven't followed closely the whole thing, but it seemed to me that the time constraint was found to be useful when LF-MMI was first made, why is it better without it now? |
hhadian changed the title 
| The time constraints are still used but indirectly, i.e. we can still change left/right tolerance but they are less likely to make a difference, I guess. The supervision FSTs are converted to unconstrained FSTs (i.e. e2e-style) only after time constraints are applied and after they are split up into chunks. |
| WriteBasicType(os, binary, label_dim); | ||
| KALDI_ASSERT(frames_per_sequence > 0 && label_dim > 0 && | ||
| num_sequences > 0); | ||
| WriteToken(os, binary, "<End2End>"); |
There was a problem hiding this comment.
(for context, Hossein said offline:)
The problem is in reading a non-binary supervision which starts like this: <Supervision> <Weight> 1 <NumSequences> 1 <FramesPerSeq> 16 <LabelDim> 981 0 1 498 498 0 2 497 497 0 3 496 496 I guess what happens is that after the label dim is read, we do if (PeekToken(is, binary) == 'E' ) to see if there is a <End2End> tag but this peeking removes the '\n' at the end of the first line which causes the next command which is ReadFstKaldi to fail (I guess because it expects a '\n' or whitespace but sees a '0'). I can fix ReadFstKaldi or PeekToken (e.g. to put back ws/newlines if no '<' was read). Not sure what is the best fix. What would you suggest? Likely the issue is that ungetc failed (which it sometimes can, but not always).
@hhadian, I think the easiest fix would be to revert this change here, to make it always write the <End2End> tag.
| I pushed a fix. I noticed there are no tests for the new features, i.e. e2e/unconstrained supervisions. Should I add tests for them? |
| If you can quickly add any tests, e.g. that the I/O works, then do it. Just don't make it a blocker. |
danpovey changed the title 
…example recipes (kaldi-asr#2383) This enables ignoring the supervision phone-to-frame alignment information inside chunks.
3 participants
This adds a few scripts (with results) which use unconstrained-egs (in PR #2341).