Create separate internal reading functions for K2 FSAs and OpenFST FSAs about k2 HOT 12 CLOSED

from k2.

Can it be implemented in Python?

from k2.

from k2.

from k2.

from k2.

Can we provide some command line tools similar to lhotse that convert openfst format to k2 format offline so
that we can keep the c++ code simple?

from k2.

from k2.

@danpovey double checking, for WFST

0 1 1 1 0.1
0 2 2 2 0.2
1 2 3 3 0.3
2

We would still create a super final state, because there is no arc entering the original final state with label -1, right?

Regarding the start state, is it required by K2 that there will only be one start state with state 0? If yes we may have to add a super start state as well?

from k2.

OK, we can require that in the text form, the final-state must be listed. However I'm not sure what to do about the empty FSA, i.e. with no states. That is actually supported by the tools as-is, but perhaps we could settle on always having at least 2 states (and just supporting an empty FSA as one not having any arcs). (?). Thinking of the pros and cons of this.

There is only one start state, this is the same as in OpenFST.

from k2.

... one possibility is to say that for empty FSAs, we just have no arcs, and to make it possible to represent the empty FSA in a vector of FSAs (and write to disk) we actually store the FSA on disk as 2 tensors: one being the row_splits of FSAs -> arcs, equivalent to fsa_vec.RowSplits(2)[fsa_vec.RowSplits(1)], and the 2nd being the tensor of arcs. And the equivalent in text form (although I dont recommend using text form much) would be to represent the empty FSA as just a blank line, with regular FSAs terminated by a single blank line if we serialize multiple FSAs to a single text string. Again, I'm not sure how much we'll want to serialize vectors of FSAs to text strings, i.e. whether this is needed.

from k2.

from k2.

from k2.