def test_acceptor_wo_arcs_from_openfst(self):
s1 = '''
'''
s2 = '''
0 Inf
1 0.1
'''
s3 = '''
0 Inf
1 0.1
2 0.2
'''
for device in self.devices:
fsa1 = k2.Fsa.from_openfst(s1)
print("fsa1 = ", k2.to_str(fsa1))
self.assertEqual('', k2.to_str(fsa1))
fsa2 = k2.Fsa.from_openfst(s2)
self.assertEqual(_remove_leading_spaces(k2.to_str(fsa2)),
"1 2 -1 -0.1\n2")
arcs2 = fsa2.arcs.values()[:, :-1]
assert torch.all(
torch.eq(arcs2, torch.tensor([[1, 2, -1]], dtype=torch.int32)))
fsa3 = k2.Fsa.from_openfst(s3)
self.assertEqual(fsa3.arcs.dim0(), 4)
self.assertEqual(_remove_leading_spaces(k2.to_str(fsa3)),
"1 3 -1 -0.1\n2 3 -1 -0.2\n3")
def test_transducer2_ragged2_from_str(self):
s = '''
0 1 2 22 101 [] [] -1.2
0 2 10 100 102 [] [] -2.2
1 6 -1 16 103 [20 30] [40] -4.2
1 3 3 33 104 [] [] -3.2
2 6 -1 26 105 [] [] -5.2
2 4 2 22 106 [] [] -6.2
3 6 -1 36 107 [] [] -7.2
5 0 1 50 108 [] [] -8.2
6
'''
fsa = k2.Fsa.from_str(s,
aux_label_names=['aux_labels', 'aux_labels2'],
ragged_label_names=['ragged1', 'ragged2'])
assert fsa.aux_labels.dtype == torch.int32
assert fsa.aux_labels.device.type == 'cpu'
assert isinstance(fsa.ragged1, k2.RaggedTensor)
assert isinstance(fsa.ragged2, k2.RaggedTensor)
assert torch.all(
torch.eq(
fsa.aux_labels,
torch.tensor([22, 100, 16, 33, 26, 22, 36, 50],
dtype=torch.int32)))
assert torch.all(
torch.eq(
fsa.aux_labels2,
torch.tensor([101, 102, 103, 104, 105, 106, 107, 108],
dtype=torch.int32)))
assert torch.allclose(
fsa.scores,
torch.tensor([-1.2, -2.2, -4.2, -3.2, -5.2, -6.2, -7.2, -8.2],
dtype=torch.float32))
print("fsa.ragged1 = ", fsa.ragged1)
print("fsa.ragged2 = ", fsa.ragged2)
assert fsa.ragged1 == k2.RaggedTensor(
'[ [] [] [20 30] [] [] [] [] [] ]')
assert fsa.ragged2 == k2.RaggedTensor('[ [] [] [40] [] [] [] [] [] ]')
# only aux_labels will be printed right now..
expected_str = '''
0 1 2 22 101 [ ] [ ] -1.2
0 2 10 100 102 [ ] [ ] -2.2
1 6 -1 16 103 [ 20 30 ] [ 40 ] -4.2
1 3 3 33 104 [ ] [ ] -3.2
2 6 -1 26 105 [ ] [ ] -5.2
2 4 2 22 106 [ ] [ ] -6.2
3 6 -1 36 107 [ ] [ ] -7.2
5 0 1 50 108 [ ] [ ] -8.2
6
'''
print("fsa = ", _remove_leading_spaces(k2.to_str(fsa)))
assert _remove_leading_spaces(expected_str) == \
_remove_leading_spaces(k2.to_str(fsa))
def test_acceptor_from_openfst(self):
s = '''
0 1 2 -1.2
0 2 10 -2.2
1 6 1 -3.2
1 3 3 -4.2
2 6 2 -5.2
2 4 2 -6.2
3 6 3 -7.2
5 0 1 -8.2
7
6 -9.2
'''
for i in range(4):
if i == 0:
fsa = k2.Fsa.from_openfst(s)
elif i == 1:
fsa = k2.Fsa.from_openfst(s, acceptor=True)
elif i == 2:
fsa = k2.Fsa.from_openfst(s, num_aux_labels=0)
else:
fsa = k2.Fsa.from_openfst(s, aux_label_names=[])
expected_str = '''
0 1 2 -1.2
0 2 10 -2.2
1 6 1 -3.2
1 3 3 -4.2
2 6 2 -5.2
2 4 2 -6.2
3 6 3 -7.2
5 0 1 -8.2
6 8 -1 -9.2
7 8 -1 0
8
'''
assert _remove_leading_spaces(expected_str) == \
_remove_leading_spaces(k2.to_str(fsa, openfst=True))
arcs = fsa.arcs.values()[:, :-1]
assert isinstance(arcs, torch.Tensor)
assert arcs.dtype == torch.int32
assert arcs.device.type == 'cpu'
assert arcs.shape == (10, 3), 'there should be 10 arcs'
assert torch.all(
torch.eq(arcs[0], torch.tensor([0, 1, 2], dtype=torch.int32)))
assert torch.allclose(
fsa.scores,
torch.tensor([1.2, 2.2, 3.2, 4.2, 5.2, 6.2, 7.2, 8.2, 9.2, 0],
dtype=torch.float32))
fsa.scores *= -1
assert torch.allclose(
fsa.scores,
torch.tensor(
[-1.2, -2.2, -3.2, -4.2, -5.2, -6.2, -7.2, -8.2, -9.2, 0],
dtype=torch.float32))
def test_acceptor_from_openfst_ragged1(self):
s = '''
0 1 2 [] -1.2
0 2 10 [10] -2.2
1 6 1 [] -3.2
1 3 3 [11 12] -4.2
2 6 2 [] -5.2
2 4 2 [] -6.2
3 6 3 [] -7.2
5 0 1 [13] -8.2
7
6 -9.2
'''
fsa = k2.Fsa.from_openfst(s, num_aux_labels=0,
ragged_label_names=['ragged'])
expected_str = '''
0 1 2 [ ] -1.2
0 2 10 [ 10 ] -2.2
1 6 1 [ ] -3.2
1 3 3 [ 11 12 ] -4.2
2 6 2 [ ] -5.2
2 4 2 [ ] -6.2
3 6 3 [ ] -7.2
5 0 1 [ 13 ] -8.2
6 8 -1 [ ] -9.2
7 8 -1 [ ] 0
8
'''
string = _remove_leading_spaces(k2.to_str(fsa, openfst=True))
print("fsa=", string)
assert _remove_leading_spaces(expected_str) == string
def transform(args):
if args.normalise:
normalise(args)
symbols = symboletable(args)
with open(os.path.join(args.data_directory, 'emission_symbols'), 'w') as f:
f.write(symbols)
lpzscp = os.path.join(args.data_directory, 'lpz.scp') if not args.normalise else os.path.join(
args.data_directory, 'lpz_norm.scp')
fstpath = os.path.join(args.data_directory, 'efst') if not args.normalise else os.path.join(
args.data_directory, 'efst_norm')
os.makedirs(fstpath, exist_ok=True)
uttid2datafile = dict()
with open(lpzscp) as f:
for line in f:
lc = line.strip().split()
uttid2datafile[lc[0]] = lc[1]
fst_scp = os.path.join(args.data_directory, 'efst.scp') if not args.normalise else os.path.join(
args.data_directory, 'efst_norm.fst')
with open(fst_scp, 'w') as f:
fc = ''
for uttid in uttid2datafile.keys():
datafile = uttid2datafile[uttid]
data = np.load(datafile)
efst = transform_utt(data)
efst_str = k2.to_str(efst)
efst_file = os.path.join(fstpath, uttid)
with open(efst_file, 'w') as g:
g.write(efst_str)
fc += '%s %s\n' % (uttid, efst_file)
f.write(fc)
def lexicon_fst(args):
'''
This programme create lexicon.fst.pdf and lexicon.fst.txt based on args.word_file
input:
args: name_space
return:
lexicon: k2.Fsa, lexicon fst
output:
lexicon.fst.txt and lexicon.fst.pdf in args.data_directory
By lexicon fst, we compress the repeated chars in emission fst.
'''
symbols_str = symboletable(args)
symbols_paris = symbols_str.split('\n')
num_noneps = len(symbols_paris) - 1
symbol2fst = [None] # <eps> has no fst
for i in range(1, num_noneps + 1):
s = '''
0 1 %d %d 0.0
1 1 %d 0 0.0
1 2 -1 -1 0.0
2
''' % (i, i, i)
g = k2.Fsa.from_str(s, acceptor=False)
symbol2fst.append(g)
fst_vec = k2.create_fsa_vec(symbol2fst[1:])
fst_union = k2.union(fst_vec)
lexicon = k2.closure(fst_union)
lexicon.draw(os.path.join(args.data_directory, 'lexicon.fst.pdf'), title='lexicon')
# lexicon.symbols = k2.SymbolTable.from_str(symbols_str)
# lexicon.aux_symbols = k2.SymbolTable.from_str(symbols_str)
with open(os.path.join(args.data_directory, 'lexicon.fst.txt'), 'w') as f:
f.write(k2.to_str(lexicon))
def test_transducer_from_str(self):
s = '''
0 1 2 22 -1.2
0 2 10 100 -2.2
1 6 -1 16 -4.2
1 3 3 33 -3.2
2 6 -1 26 -5.2
2 4 2 22 -6.2
3 6 -1 36 -7.2
5 0 1 50 -8.2
6
'''
fsa = k2.Fsa.from_str(_remove_leading_spaces(s))
assert fsa.aux_labels.dtype == torch.int32
assert fsa.aux_labels.device.type == 'cpu'
assert torch.allclose(
fsa.aux_labels,
torch.tensor([22, 100, 16, 33, 26, 22, 36, 50], dtype=torch.int32))
expected_str = '''
0 1 2 22 -1.2
0 2 10 100 -2.2
1 6 -1 16 -4.2
1 3 3 33 -3.2
2 6 -1 26 -5.2
2 4 2 22 -6.2
3 6 -1 36 -7.2
5 0 1 50 -8.2
6
'''
assert _remove_leading_spaces(expected_str) == _remove_leading_spaces(
k2.to_str(fsa))
def test(self):
# for the symbol table
# <eps> 0
# a 0
# b 1
# c 2
# an FSA that recognizes a+(b|c)
s = '''
0 1 1 0.1
1 1 1 0.2
1 2 2 0.3
1 3 3 0.4
2 4 -1 0.5
3 4 -1 0.6
5
'''
a_fsa = k2.Fsa.from_str(s)
a_fsa.requires_grad_(True)
# an FSA that recognizes ab
s = '''
0 1 1 10
1 2 2 20
2 3 -1 30
3
'''
b_fsa = k2.Fsa.from_str(s)
b_fsa.requires_grad_(True)
fsa = k2.intersect(a_fsa, b_fsa)
assert len(fsa.shape) == 2
actual_str = k2.to_str(fsa)
expected_str = '\n'.join(
['0 1 1 10.1', '1 2 2 20.3', '2 3 -1 30.5', '3'])
assert actual_str.strip() == expected_str
loss = fsa.scores.sum()
loss.backward()
# arc 0, 2, and 4 of a_fsa are kept in the final intersected FSA
assert torch.allclose(
a_fsa.scores.grad,
torch.tensor([1, 0, 1, 0, 1, 0], dtype=torch.float32))
assert torch.allclose(b_fsa.scores.grad,
torch.tensor([1, 1, 1], dtype=torch.float32))
# if any of the input FSA is an FsaVec,
# the outupt FSA is also an FsaVec.
a_fsa.scores.grad = None
b_fsa.scores.grad = None
a_fsa = k2.create_fsa_vec([a_fsa])
fsa = k2.intersect(a_fsa, b_fsa)
assert len(fsa.shape) == 3
def test_treat_epsilon_specially_true(self):
# this version works only on CPU and requires
# arc-sorted inputs
# a_fsa recognizes `(1|3)?2*`
s1 = '''
0 1 3 0.0
0 1 1 0.2
0 1 0 0.1
1 1 2 0.3
1 2 -1 0.4
2
'''
a_fsa = k2.Fsa.from_str(s1)
a_fsa.requires_grad_(True)
# b_fsa recognizes `1|2|5`
s2 = '''
0 1 5 0
0 1 1 1
0 1 2 2
1 2 -1 3
2
'''
b_fsa = k2.Fsa.from_str(s2)
b_fsa.requires_grad_(True)
# fsa recognizes 1|2
fsa = k2.intersect(k2.arc_sort(a_fsa), k2.arc_sort(b_fsa))
assert len(fsa.shape) == 2
actual_str = k2.to_str(fsa)
expected_str = '\n'.join(
['0 1 0 0.1', '0 2 1 1.2', '1 2 2 2.3', '2 3 -1 3.4', '3'])
assert actual_str.strip() == expected_str
loss = fsa.scores.sum()
(-loss).backward()
# arc 1, 2, 3, and 4 of a_fsa are kept in the final intersected FSA
assert torch.allclose(a_fsa.grad,
torch.tensor([0, -1, -1, -1, -1]).to(a_fsa.grad))
# arc 1, 2, and 3 of b_fsa are kept in the final intersected FSA
assert torch.allclose(b_fsa.grad,
torch.tensor([0, -1, -1, -1]).to(b_fsa.grad))
# if any of the input FSA is an FsaVec,
# the outupt FSA is also an FsaVec.
a_fsa.scores.grad = None
b_fsa.scores.grad = None
a_fsa = k2.create_fsa_vec([a_fsa])
fsa = k2.intersect(k2.arc_sort(a_fsa), k2.arc_sort(b_fsa))
assert len(fsa.shape) == 3
def test_treat_epsilon_specially_false(self):
devices = [torch.device('cpu')]
if torch.cuda.is_available():
devices.append(torch.device('cuda'))
for device in devices:
# a_fsa recognizes `(0|1)2*`
s1 = '''
0 1 0 0.1
0 1 1 0.2
1 1 2 0.3
1 2 -1 0.4
2
'''
a_fsa = k2.Fsa.from_str(s1).to(device)
a_fsa.requires_grad_(True)
# b_fsa recognizes `1|2`
s2 = '''
0 1 1 1
0 1 2 2
1 2 -1 3
2
'''
b_fsa = k2.Fsa.from_str(s2).to(device)
b_fsa.requires_grad_(True)
# fsa recognizes `1`
fsa = k2.intersect(a_fsa, b_fsa, treat_epsilons_specially=False)
assert len(fsa.shape) == 2
actual_str = k2.to_str(fsa)
expected_str = '\n'.join(['0 1 1 1.2', '1 2 -1 3.4', '2'])
assert actual_str.strip() == expected_str
loss = fsa.scores.sum()
(-loss).backward()
# arc 1 and 3 of a_fsa are kept in the final intersected FSA
assert torch.allclose(a_fsa.grad,
torch.tensor([0, -1, 0, -1]).to(a_fsa.grad))
# arc 0 and 2 of b_fsa are kept in the final intersected FSA
assert torch.allclose(b_fsa.grad,
torch.tensor([-1, 0, -1]).to(b_fsa.grad))
# if any of the input FSA is an FsaVec,
# the outupt FSA is also an FsaVec.
a_fsa.scores.grad = None
b_fsa.scores.grad = None
a_fsa = k2.create_fsa_vec([a_fsa])
fsa = k2.intersect(a_fsa, b_fsa, treat_epsilons_specially=False)
assert len(fsa.shape) == 3
def test_transducer2_from_str(self):
s = '''
0 1 2 22 101 -1.2
0 2 10 100 102 -2.2
1 6 -1 16 103 -4.2
1 3 3 33 104 -3.2
2 6 -1 26 105 -5.2
2 4 2 22 106 -6.2
3 6 -1 36 107 -7.2
5 0 1 50 108 -8.2
6
'''
for i in range(2):
if i == 0:
fsa = k2.Fsa.from_str(s, num_aux_labels=2)
else:
fsa = k2.Fsa.from_str(
s, aux_label_names=['aux_labels', 'aux_labels2'])
assert fsa.aux_labels.dtype == torch.int32
assert fsa.aux_labels.device.type == 'cpu'
assert torch.all(
torch.eq(
fsa.aux_labels,
torch.tensor([22, 100, 16, 33, 26, 22, 36, 50],
dtype=torch.int32)))
assert torch.all(
torch.eq(
fsa.aux_labels2,
torch.tensor([101, 102, 103, 104, 105, 106, 107, 108],
dtype=torch.int32)))
assert torch.allclose(
fsa.scores,
torch.tensor([-1.2, -2.2, -4.2, -3.2, -5.2, -6.2, -7.2, -8.2],
dtype=torch.float32))
# only aux_labels will be printed right now..
expected_str = '''
0 1 2 22 -1.2
0 2 10 100 -2.2
1 6 -1 16 -4.2
1 3 3 33 -3.2
2 6 -1 26 -5.2
2 4 2 22 -6.2
3 6 -1 36 -7.2
5 0 1 50 -8.2
6
'''
assert _remove_leading_spaces(expected_str) == \
_remove_leading_spaces(k2.to_str(fsa))
def test_transducer_from_openfst(self):
s = '''
0 1 2 22 -1.2
0 2 10 100 -2.2
1 6 1 16 -4.2
1 3 3 33 -3.2
2 6 2 26 -5.2
2 4 2 22 -6.2
3 6 3 36 -7.2
5 0 1 50 -8.2
7 -9.2
6
'''
for i in range(3):
if i == 0:
fsa = k2.Fsa.from_openfst(s, acceptor=False)
elif i == 1:
fsa = k2.Fsa.from_openfst(s, num_aux_labels=1)
else:
fsa = k2.Fsa.from_openfst(s, aux_label_names=['aux_labels'])
assert fsa.aux_labels.dtype == torch.int32
assert fsa.aux_labels.device.type == 'cpu'
assert torch.all(
torch.eq(
fsa.aux_labels,
torch.tensor([22, 100, 16, 33, 26, 22, 36, 50, -1, -1],
dtype=torch.int32)))
assert torch.allclose(
fsa.scores,
torch.tensor([1.2, 2.2, 4.2, 3.2, 5.2, 6.2, 7.2, 8.2, 0, 9.2],
dtype=torch.float32))
expected_str = '''
0 1 2 22 -1.2
0 2 10 100 -2.2
1 6 1 16 -4.2
1 3 3 33 -3.2
2 6 2 26 -5.2
2 4 2 22 -6.2
3 6 3 36 -7.2
5 0 1 50 -8.2
6 8 -1 -1 0
7 8 -1 -1 -9.2
8
'''
assert _remove_leading_spaces(expected_str) == \
_remove_leading_spaces(k2.to_str(fsa, openfst=True))
def test_acceptor_from_str(self):
s = '''
0 1 2 -1.2
0 2 10 -2.2
1 6 -1 -3.2
1 3 3 -4.2
2 6 -1 -5.2
2 4 2 -6.2
3 6 -1 -7.2
5 0 1 -8.2
6
'''
fsa = k2.Fsa.from_str(_remove_leading_spaces(s))
expected_str = '''
0 1 2 -1.2
0 2 10 -2.2
1 6 -1 -3.2
1 3 3 -4.2
2 6 -1 -5.2
2 4 2 -6.2
3 6 -1 -7.2
5 0 1 -8.2
6
'''
assert _remove_leading_spaces(expected_str) == _remove_leading_spaces(
k2.to_str(fsa))
arcs = fsa.arcs.values()[:, :-1]
assert isinstance(arcs, torch.Tensor)
assert arcs.dtype == torch.int32
assert arcs.device.type == 'cpu'
assert arcs.shape == (8, 3), 'there should be 8 arcs'
assert torch.allclose(arcs[0],
torch.tensor([0, 1, 2], dtype=torch.int32))
assert torch.allclose(
fsa.scores,
torch.tensor([-1.2, -2.2, -3.2, -4.2, -5.2, -6.2, -7.2, -8.2],
dtype=torch.float32))
fsa.scores *= -1
assert torch.allclose(
fsa.scores,
torch.tensor([1.2, 2.2, 3.2, 4.2, 5.2, 6.2, 7.2, 8.2],
dtype=torch.float32))
def gfst(args):
'''
This programme is for debugging only.
Usually, for different task, we need diffrent gfst.
Imagine that we only have
'
<eps> 0
<blank> 1
<unk> 2
n 3
y 4
<eos> 5
'
these 6 different symbols and try to generate yyn, ynn, and 3gram gfst.
'''
symbols = symboletable(args)
yyn = '''
0 1 0 0 0.0
0 2 0 0 0.0
0 3 0 0 0.0
0 4 0 0 0.0
0 5 0 0 0.0
0 6 0 0 0.0
1 2 4 4 0.0
1 6 1 0 0.0
1 7 5 5 0.0
2 3 1 0 0.0
2 7 5 5 0.0
3 4 4 4 0.0
3 7 5 5 0.0
4 5 1 0 0.0
4 1 3 3 0.0
4 7 5 5 0.0
5 1 3 3 0.0
5 7 5 5 0.0
6 2 4 4 0.0
6 7 5 5 0.0
7 8 -1 -1 0.0
8
'''
yyn_fst = k2.Fsa.from_str(yyn, acceptor=False)
# yyn_fst.symbols = k2.SymbolTable.from_str(symbols)
# yyn_fst.aux_symbols = k2.SymbolTable.from_str(symbols)
gfst_dir = os.path.join(args.data_directory, 'G')
os.makedirs(gfst_dir, exist_ok=True)
yyn_fst.draw(os.path.join(gfst_dir, 'yyn.pdf'), 'yyn')
with open(os.path.join(gfst_dir, 'yyn.fst.txt'), 'w') as f:
f.write(k2.to_str(yyn_fst))
def test_acceptor_from_tensor(self):
fsa_tensor = torch.tensor(
[[0, 1, 2, _k2._float_as_int(-1.2)],
[0, 2, 10, _k2._float_as_int(-2.2)],
[1, 6, -1, _k2._float_as_int(-3.2)],
[1, 3, 3, _k2._float_as_int(-4.2)],
[2, 6, -1, _k2._float_as_int(-5.2)],
[2, 4, 2, _k2._float_as_int(-6.2)],
[3, 6, -1, _k2._float_as_int(-7.2)],
[5, 0, 1, _k2._float_as_int(-8.2)]],
dtype=torch.int32)
fsa = k2.Fsa(fsa_tensor)
expected_str = '''
0 1 2 -1.2
0 2 10 -2.2
1 6 -1 -3.2
1 3 3 -4.2
2 6 -1 -5.2
2 4 2 -6.2
3 6 -1 -7.2
5 0 1 -8.2
6
'''
assert _remove_leading_spaces(expected_str) == _remove_leading_spaces(
k2.to_str(fsa))
arcs = fsa.arcs.values()[:, :-1]
assert isinstance(arcs, torch.Tensor)
assert arcs.dtype == torch.int32
assert arcs.device.type == 'cpu'
assert arcs.shape == (8, 3), 'there should be 8 arcs'
assert torch.allclose(arcs[0],
torch.tensor([0, 1, 2], dtype=torch.int32))
assert torch.allclose(
fsa.score,
torch.tensor([-1.2, -2.2, -3.2, -4.2, -5.2, -6.2, -7.2, -8.2],
dtype=torch.float32))
fsa.score *= -1
assert torch.allclose(
fsa.score,
torch.tensor([1.2, 2.2, 3.2, 4.2, 5.2, 6.2, 7.2, 8.2],
dtype=torch.float32))
def test_transducer_from_str(self):
s = '''
0 1 2 22 -1.2
0 2 10 100 -2.2
1 6 -1 16 -4.2
1 3 3 33 -3.2
2 6 -1 26 -5.2
2 4 2 22 -6.2
3 6 -1 36 -7.2
5 0 1 50 -8.2
6
'''
for i in range(3):
if i == 0:
fsa = k2.Fsa.from_str(s, num_aux_labels=1)
elif i == 1:
fsa = k2.Fsa.from_str(s, acceptor=False)
else:
fsa = k2.Fsa.from_str(s, aux_label_names=['aux_labels'])
assert fsa.aux_labels.dtype == torch.int32
assert fsa.aux_labels.device.type == 'cpu'
assert torch.all(
torch.eq(
fsa.aux_labels,
torch.tensor([22, 100, 16, 33, 26, 22, 36, 50],
dtype=torch.int32)))
assert torch.allclose(
fsa.scores,
torch.tensor([-1.2, -2.2, -4.2, -3.2, -5.2, -6.2, -7.2, -8.2],
dtype=torch.float32))
expected_str = '''
0 1 2 22 -1.2
0 2 10 100 -2.2
1 6 -1 16 -4.2
1 3 3 33 -3.2
2 6 -1 26 -5.2
2 4 2 22 -6.2
3 6 -1 36 -7.2
5 0 1 50 -8.2
6
'''
assert _remove_leading_spaces(expected_str) == \
_remove_leading_spaces(k2.to_str(fsa))
def test(self):
s = '''
0 1 2 0.1
0 1 1 0.2
1 2 -1 0.3
2
'''
fsa = k2.Fsa.from_str(s)
fsa.requires_grad_(True)
sorted_fsa = k2.arc_sort(fsa)
actual_str = k2.to_str(sorted_fsa)
expected_str = '\n'.join(['0 1 1 0.2', '0 1 2 0.1', '1 2 -1 0.3', '2'])
assert actual_str.strip() == expected_str
loss = (sorted_fsa.scores[1] + sorted_fsa.scores[2]) / 2
loss.backward()
assert torch.allclose(fsa.scores.grad,
torch.tensor([0.5, 0, 0.5], dtype=torch.float32))
class TestConnect(unittest.TestCase):
s = '''
0 1 1 0.1
0 2 2 0.2
1 4 -1 0.3
3 4 -1 0.4
4
'''
fsa = k2.Fsa.from_str(s)
fsa.requires_grad_(True)
expected_str = '\n'.join(['0 1 1 0.1', '1 2 -1 0.3', '2'])
connected_fsa = k2.connect(fsa)
actual_str = k2.to_str(connected_fsa)
assert actual_str.strip() == expected_str
loss = connected_fsa.scores.sum()
loss.backward()
assert torch.allclose(fsa.scores.grad,
torch.tensor([1, 0, 1, 0], dtype=torch.float32))
def test_transducer_from_tensor(self):
for device in self.devices:
fsa_tensor = torch.tensor(
[[0, 1, 2, _k2.float_as_int(-1.2)],
[0, 2, 10, _k2.float_as_int(-2.2)],
[1, 6, -1, _k2.float_as_int(-4.2)],
[1, 3, 3, _k2.float_as_int(-3.2)],
[2, 6, -1, _k2.float_as_int(-5.2)],
[2, 4, 2, _k2.float_as_int(-6.2)],
[3, 6, -1, _k2.float_as_int(-7.2)],
[5, 0, 1, _k2.float_as_int(-8.2)]],
dtype=torch.int32).to(device)
aux_labels_tensor = torch.tensor([22, 100, 16, 33, 26, 22, 36, 50],
dtype=torch.int32).to(device)
fsa = k2.Fsa(fsa_tensor, aux_labels_tensor)
assert fsa.aux_labels.dtype == torch.int32
assert fsa.aux_labels.device.type == device.type
assert torch.all(
torch.eq(
fsa.aux_labels,
torch.tensor([22, 100, 16, 33, 26, 22, 36, 50],
dtype=torch.int32).to(device)))
assert torch.allclose(
fsa.scores,
torch.tensor([-1.2, -2.2, -4.2, -3.2, -5.2, -6.2, -7.2, -8.2],
dtype=torch.float32,
device=device))
expected_str = '''
0 1 2 22 -1.2
0 2 10 100 -2.2
1 6 -1 16 -4.2
1 3 3 33 -3.2
2 6 -1 26 -5.2
2 4 2 22 -6.2
3 6 -1 36 -7.2
5 0 1 50 -8.2
6
'''
assert _remove_leading_spaces(expected_str) == \
_remove_leading_spaces(k2.to_str(fsa))
def test_transducer_from_tensor(self):
devices = [torch.device('cpu')]
if torch.cuda.is_available():
devices.append(torch.device('cuda', 0))
for device in devices:
fsa_tensor = torch.tensor(
[[0, 1, 2, _k2.float_as_int(-1.2)],
[0, 2, 10, _k2.float_as_int(-2.2)],
[1, 6, -1, _k2.float_as_int(-4.2)],
[1, 3, 3, _k2.float_as_int(-3.2)],
[2, 6, -1, _k2.float_as_int(-5.2)],
[2, 4, 2, _k2.float_as_int(-6.2)],
[3, 6, -1, _k2.float_as_int(-7.2)],
[5, 0, 1, _k2.float_as_int(-8.2)]],
dtype=torch.int32).to(device)
aux_labels_tensor = torch.tensor([22, 100, 16, 33, 26, 22, 36, 50],
dtype=torch.int32).to(device)
fsa = k2.Fsa(fsa_tensor, aux_labels_tensor)
assert fsa.aux_labels.dtype == torch.int32
assert fsa.aux_labels.device.type == device.type
assert torch.allclose(
fsa.aux_labels,
torch.tensor([22, 100, 16, 33, 26, 22, 36, 50],
dtype=torch.int32).to(device))
expected_str = '''
0 1 2 22 -1.2
0 2 10 100 -2.2
1 6 -1 16 -4.2
1 3 3 33 -3.2
2 6 -1 26 -5.2
2 4 2 22 -6.2
3 6 -1 36 -7.2
5 0 1 50 -8.2
6
'''
assert _remove_leading_spaces(
expected_str) == _remove_leading_spaces(k2.to_str(fsa))
def test_acceptor_wo_arcs_from_str(self):
s1 = '''
'''
s2 = '''
0
1
'''
s3 = '''
1
'''
for device in self.devices:
fsa1 = k2.Fsa.from_str(s1)
self.assertEqual(k2.to_str(fsa1), '')
with self.assertRaises(ValueError):
_ = k2.Fsa.from_str(s2)
fsa3 = k2.Fsa.from_str(s3)
self.assertEqual(fsa3.arcs.dim0(), 0)
def test_transducer_from_openfst(self):
s = '''
0 1 2 22 -1.2
0 2 10 100 -2.2
1 6 1 16 -4.2
1 3 3 33 -3.2
2 6 2 26 -5.2
2 4 2 22 -6.2
3 6 3 36 -7.2
5 0 1 50 -8.2
7 -9.2
6
'''
fsa = k2.Fsa.from_openfst(_remove_leading_spaces(s), acceptor=False)
assert fsa.aux_labels.dtype == torch.int32
assert fsa.aux_labels.device.type == 'cpu'
assert torch.allclose(
fsa.aux_labels,
torch.tensor([22, 100, 16, 33, 26, 22, 36, 50, 0, 0],
dtype=torch.int32))
expected_str = '''
0 1 2 22 -1.2
0 2 10 100 -2.2
1 6 1 16 -4.2
1 3 3 33 -3.2
2 6 2 26 -5.2
2 4 2 22 -6.2
3 6 3 36 -7.2
5 0 1 50 -8.2
6 8 -1 0 -0
7 8 -1 0 -9.2
8
'''
assert _remove_leading_spaces(expected_str) == _remove_leading_spaces(
k2.to_str(fsa, openfst=True))
def test(self):
devices = [torch.device('cpu')]
if torch.cuda.is_available():
devices.append(torch.device('cuda'))
for device in devices:
for use_identity_map, sorted_match_a in [(True, True),
(False, True),
(True, False),
(False, False)]:
# recognizes (0|1)(0|2)
s1 = '''
0 1 0 0.1
0 1 1 0.2
1 2 0 0.4
1 2 2 0.3
2 3 -1 0.5
3
'''
# recognizes 02*
s2 = '''
0 1 0 1
1 1 2 2
1 2 -1 3
2
'''
# recognizes 1*0
s3 = '''
0 0 1 10
0 1 0 20
1 2 -1 30
2
'''
a_fsa = k2.Fsa.from_str(s1).to(device)
b_fsa_1 = k2.Fsa.from_str(s2).to(device)
b_fsa_2 = k2.Fsa.from_str(s3).to(device)
a_fsa.requires_grad_(True)
b_fsa_1.requires_grad_(True)
b_fsa_2.requires_grad_(True)
b_fsas = k2.create_fsa_vec([b_fsa_1, b_fsa_2])
if use_identity_map:
a_fsas = k2.create_fsa_vec([a_fsa, a_fsa])
b_to_a_map = torch.tensor([0, 1],
dtype=torch.int32).to(device)
else:
a_fsas = k2.create_fsa_vec([a_fsa])
b_to_a_map = torch.tensor([0, 0],
dtype=torch.int32).to(device)
c_fsas = k2.intersect_device(a_fsas, b_fsas, b_to_a_map,
sorted_match_a)
assert c_fsas.shape == (2, None, None)
c_fsas = k2.connect(c_fsas.to('cpu'))
# c_fsas[0] recognizes: 02
# c_fsas[1] recognizes: 10
actual_str_0 = k2.to_str(c_fsas[0])
expected_str_0 = '\n'.join(
['0 1 0 1.1', '1 2 2 2.3', '2 3 -1 3.5', '3'])
assert actual_str_0.strip() == expected_str_0
actual_str_1 = k2.to_str(c_fsas[1])
expected_str_1 = '\n'.join(
['0 1 1 10.2', '1 2 0 20.4', '2 3 -1 30.5', '3'])
assert actual_str_1.strip() == expected_str_1
loss = c_fsas.scores.sum()
(-loss).backward()
assert torch.allclose(
a_fsa.grad,
torch.tensor([-1, -1, -1, -1, -2]).to(a_fsa.grad))
assert torch.allclose(
b_fsa_1.grad,
torch.tensor([-1, -1, -1]).to(b_fsa_1.grad))
assert torch.allclose(
b_fsa_2.grad,
torch.tensor([-1, -1, -1]).to(b_fsa_2.grad))
def lexicon_fst_whole(args):
'''
This programme create lexicon.fst.pdf and lexicon.fst.txt based on args.word_file
input:
args: name_space
return:
lexicon: k2.Fsa, lexicon fst
output:
lexicon.fst.txt and lexicon.fst.pdf in args.data_directory
By lexicon fst, we compress the repeated chars in emission fst.
'''
symbols_str = symboletable(args)
symbols_paris = symbols_str.split('\n')
num_noneps = len(symbols_paris) - 1
s = ''
count = 1
for i in range(1, num_noneps + 1):
s += '''
0 %d %d %d 0.0
%d %d %d 0 0.0
%d %d -1 -1 0.0
%d 0 0 0 0.0''' % (
i, i, i, i, i, i, i, num_noneps + 1, i)
slines = s.strip().split('\n')
def extract_first_index(line):
line_content = line.strip().split()
return int(line_content[0])
slines = sorted(slines, key=lambda l: extract_first_index(l))
s = '\n'.join(slines)
s += '\n%d\n' % (num_noneps + 1)
# s1 = '''
# 0 1 1 1 0.0
# 0 2 2 2 0.0
# 0 3 3 3 0.0
# 0 4 4 4 0.0
# 0 5 5 5 0.0
# 1 1 1 0 0.0
# 1 6 -1 -1 0.0
# 1 0 0 0 0.0
# 2 2 2 0 0.0
# 2 6 -1 -1 0.0
# 2 0 0 0 0.0
# 3 3 3 0 0.0
# 3 6 -1 -1 0.0
# 3 0 0 0 0.0
# 4 4 4 0 0.0
# 4 6 -1 -1 0.0
# 4 0 0 0 0.0
# 5 5 5 0 0.0
# 5 6 -1 -1 0.0
# 5 0 0 0 0.0
# 6
# '''
with open('lex.txt', 'w') as f:
f.write(s)
g = k2.Fsa.from_str(s, acceptor=False)
# g.symbols = k2.SymbolTable.from_str(symbols_str)
# g.aux_symbols = k2.SymbolTable.from_str(symbols_str)
g.draw(os.path.join(args.data_directory, 'lexicon.newfst.pdf'), title='lexicon')
with open(os.path.join(args.data_directory, 'lexicon.newfst.txt'), 'w') as f:
f.write(k2.to_str(g))
def test_transducer3_from_openfst(self):
s = '''
0 1 2 22 33 44 -1.2
0 2 10 100 101 102 -2.2
1 6 1 16 17 18 -4.2
1 3 3 33 34 35 -3.2
2 6 2 26 27 28 -5.2
2 4 2 22 23 24 -6.2
3 6 3 36 37 38 -7.2
5 0 1 50 51 52 -8.2
7 -9.2
6
'''
for i in range(2):
if i == 0:
fsa = k2.Fsa.from_openfst(s, num_aux_labels=3)
else:
fsa = k2.Fsa.from_openfst(s,
aux_label_names=[
'aux_labels', 'aux_labels2',
'aux_labels3'
])
assert fsa.aux_labels.dtype == torch.int32
assert fsa.aux_labels.device.type == 'cpu'
assert torch.all(
torch.eq(
fsa.aux_labels,
torch.tensor([22, 100, 16, 33, 26, 22, 36, 50, -1, -1],
dtype=torch.int32)))
assert fsa.aux_labels2.dtype == torch.int32
assert fsa.aux_labels2.device.type == 'cpu'
assert torch.all(
torch.eq(
fsa.aux_labels2,
torch.tensor([33, 101, 17, 34, 27, 23, 37, 51, -1, -1],
dtype=torch.int32)))
assert fsa.aux_labels3.dtype == torch.int32
assert fsa.aux_labels3.device.type == 'cpu'
assert torch.all(
torch.eq(
fsa.aux_labels3,
torch.tensor([44, 102, 18, 35, 28, 24, 38, 52, -1, -1],
dtype=torch.int32)))
assert torch.allclose(
fsa.scores,
torch.tensor([1.2, 2.2, 4.2, 3.2, 5.2, 6.2, 7.2, 8.2, 0, 9.2],
dtype=torch.float32))
expected_str = '''
0 1 2 22 -1.2
0 2 10 100 -2.2
1 6 1 16 -4.2
1 3 3 33 -3.2
2 6 2 26 -5.2
2 4 2 22 -6.2
3 6 3 36 -7.2
5 0 1 50 -8.2
6 8 -1 -1 0
7 8 -1 -1 -9.2
8
'''
assert _remove_leading_spaces(expected_str) == \
_remove_leading_spaces(k2.to_str(fsa, openfst=True))
