def test_fsa_io(self):
s = '''
0 1 10 0.1
0 2 20 0.2
1 3 -1 0.3
2 3 -1 0.4
3
'''
fsa = k2.Fsa.from_str(s)
tensor = k2.to_tensor(fsa)
assert tensor.ndim == 2
assert tensor.dtype == torch.int32
del fsa # tensor is still accessible
fsa = k2.Fsa(tensor)
del tensor
assert torch.allclose(
fsa.scores, torch.tensor([0.1, 0.2, 0.3, 0.4],
dtype=torch.float32))
assert torch.allclose(
fsa.arcs.values()[:, :-1], # skip the last field `scores`
torch.tensor([[0, 1, 10], [0, 2, 20], [1, 3, -1], [2, 3, -1]],
dtype=torch.int32))
# now test vector of FSAs
ragged_arc = _k2.fsa_to_fsa_vec(fsa.arcs)
del fsa
fsa_vec = k2.Fsa(ragged_arc)
del ragged_arc
assert fsa_vec.shape == (1, None, None)
assert torch.allclose(
fsa_vec.scores,
torch.tensor([0.1, 0.2, 0.3, 0.4], dtype=torch.float32))
assert torch.allclose(
fsa_vec.arcs.values()[:, :-1], # skip the last field `scores`
torch.tensor([[0, 1, 10], [0, 2, 20], [1, 3, -1], [2, 3, -1]],
dtype=torch.int32))
tensor = k2.to_tensor(fsa_vec)
assert tensor.ndim == 1
assert tensor.dtype == torch.int32
del fsa_vec
fsa_vec = k2.Fsa(tensor)
assert torch.allclose(
fsa_vec.scores,
torch.tensor([0.1, 0.2, 0.3, 0.4], dtype=torch.float32))
assert torch.allclose(
fsa_vec.arcs.values()[:, :-1], # skip the last field `scores`
torch.tensor([[0, 1, 10], [0, 2, 20], [1, 3, -1], [2, 3, -1]],
dtype=torch.int32))
def test_symbol_table_and_dot(self):
isym_str = '''
<eps> 0
a 1
b 2
c 3
'''
osym_str = '''
<eps> 0
x 1
y 2
z 3
'''
isym = k2.SymbolTable.from_str(isym_str)
osym = k2.SymbolTable.from_str(osym_str)
rules = '''
0 1 1 1 0.5
0 1 2 2 1.5
1 2 3 3 2.5
2 3 -1 0 3.5
3
'''
fsa = k2.Fsa(_remove_leading_spaces(rules))
fsa.set_isymbol(isym)
fsa.set_osymbol(osym)
dot = fsa.to_dot()
dot.render('/tmp/fsa', format='pdf')
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(
fsa.to_str())
arcs = fsa.arcs
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.weights,
torch.tensor([-1.2, -2.2, -3.2, -4.2, -5.2, -6.2, -7.2, -8.2],
dtype=torch.float32))
fsa = fsa.to('cuda')
arcs[0][0] += 10
assert arcs[0][0] == 10, 'arcs should still be accessible'
arcs = fsa.arcs
assert arcs.dtype == torch.int32
assert arcs.device.type == 'cuda'
assert arcs.device.index == 0
assert arcs.shape == (8, 3), 'there should be 8 arcs'
assert torch.allclose(
arcs[1],
torch.tensor([0, 2, 10], dtype=torch.int32, device=arcs.device))
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.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 test_transducer_from_str(self):
s = '''
0 1 2 22 -1.2
0 2 10 100 -2.2
1 3 3 33 -3.2
1 6 -1 16 -4.2
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(_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, 33, 16, 26, 22, 36, 50], dtype=torch.int32))
expected_str = '''
0 1 2 22 -1.2
0 2 10 100 -2.2
1 3 3 33 -3.2
1 6 -1 16 -4.2
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(
fsa.to_str())
expected_str = '''
0 1 2 22 1.2
0 2 10 100 2.2
1 3 3 33 3.2
1 6 -1 16 4.2
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(
fsa.to_str(negate_scores=True))
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_transducer_from_tensor(self):
device_id = 0
device = torch.device('cuda', device_id)
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 == 'cuda'
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(
fsa.to('cpu').to_str()) # String IO supported on CPU only
def test_bad_case3(self):
# `arc_indexes` and `arcs` in this state are not consistent
arc_indexes = torch.IntTensor([0, 2, 2, 2])
arcs = torch.IntTensor([[0, 1, 0], [0, 2, 1], [1, 2, 0]])
fsa = k2.Fsa(arc_indexes, arcs)
self.assertFalse(k2.is_valid(fsa))
def test_acceptor_from_str(self):
s = '''
0 1 2 -1.2
0 2 10 -2.2
1 3 3 -3.2
1 6 -1 -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(_remove_leading_spaces(s))
expected_str = '''
0 1 2 -1.2
0 2 10 -2.2
1 3 3 -3.2
1 6 -1 -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(
fsa.to_str())
expected_str = '''
0 1 2 1.2
0 2 10 2.2
1 3 3 3.2
1 6 -1 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(
fsa.to_str(negate_scores=True))
arcs = fsa.arcs
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.weights,
torch.tensor([-1.2, -2.2, -3.2, -4.2, -5.2, -6.2, -7.2, -8.2],
dtype=torch.float32))
fsa = fsa.to('cuda')
arcs[0][0] += 10
assert arcs[0][0] == 10, 'arcs should still be accessible'
arcs = fsa.arcs
assert arcs.dtype == torch.int32
assert arcs.device.type == 'cuda'
assert arcs.device.index == 0
assert arcs.shape == (8, 3), 'there should be 8 arcs'
assert torch.allclose(
arcs[1],
torch.tensor([0, 2, 10], dtype=torch.int32, device=arcs.device))
