Regarding the Composition in openfst-python, the Python Wrapper of OpenFST

March 14, 2024

This post is about the composition operate in openfst-python. It is demanded by the official documentation that before composition, the output labels of the first FST or the input labels of the second FST must be sorted. However, I find that, it seems not to be the case in the Python wrapper, openfst-python.

Here is a simple example to demonstrate it.

import openfst_python as fst

compiler = fst.Compiler()
print("0 1 1 8 0", file=compiler)
print("0 1 2 7 0", file=compiler)
print("0 1 3 6 0", file=compiler)
print("0 1 4 5 0", file=compiler)
print("0 1 5 9 0", file=compiler)
print("1 0", file=compiler)
f = compiler.compile()

f

compiler = fst.Compiler()
print("0 1 5 4 0", file=compiler)
print("0 1 6 3 0", file=compiler)
print("0 1 7 2 0", file=compiler)
print("0 1 8 1 0", file=compiler)
print("0 1 9 5 0", file=compiler)
print("1 0", file=compiler)
g = compiler.compile()

g

There is something worth noting in f and g. First, if the input labels f are sorted, then the output labels of f cannot be in sorted state. This is the same for g. Second, the composition result of f and g is simple as they are the inverse version of each other. 

fi = f.arcsort(sort_type="ilabel")
fo = f.arcsort(sort_type="olabel")
gi = g.arcsort(sort_type="ilabel")
go = g.arcsort(sort_type="olabel")

fi

fo

gi

go

h = fst.compose(f, g)
hio = fst.compose(fi, go)
hoo = fst.compose(fo, go)
hii = fst.compose(fi, gi)
hoi = fst.compose(fo, gi)

h

hio

hoi

hii

hoo

We can see that no matter how we do the arcsort for f and g, or even do not sort them at all, we can always do the composition without any error. This is different from what the official documentation described.

Here is the notebook file, if you’d like to play around.