def test_case1(self):
# refer to https://github.com/baidu-research/warp-ctc/blob/master/torch_binding/TUTORIAL.md
# this is the simplest case
# we have one sequence with probability: [0.2, 0.2, 0.2, 0.2, 0.2]
label_lengths_tensor = torch.tensor([1], dtype=torch.int32)
input_lengths_tensor = torch.tensor([1], dtype=torch.int32)
alphabet_size = 5
minibatch = 1
info = ctc.CtcOptions()
info.loc = ctc.CtcComputeLocation.CTC_CPU
info.num_threads = 0 # use default number of threads
info.blank_label = 0
label_lengths = kaldi.IntSubVectorFromDLPack(
to_dlpack(label_lengths_tensor))
input_lengths = kaldi.IntSubVectorFromDLPack(
to_dlpack(input_lengths_tensor))
status, size_in_bytes = ctc.GetWorkspaceSize(
label_lengths=label_lengths,
input_lengths=input_lengths,
alphabet_size=alphabet_size,
minibatch=minibatch,
info=info)
self.assertEqual(status, ctc.CtcStatus.CTC_STATUS_SUCCESS)
num_floats = size_in_bytes // 4 + 1
workspace_tensor = torch.empty(num_floats,
dtype=torch.float32).contiguous()
activations_tensor = torch.tensor([0.2, 0.2, 0.2, 0.2, 0.2],
dtype=torch.float32).contiguous()
gradients_tensor = torch.empty_like(activations_tensor)
flat_labels_tensor = torch.tensor([1], dtype=torch.int32)
costs_tensor = torch.empty(minibatch, dtype=torch.float32)
activations = kaldi.FloatSubVectorFromDLPack(
to_dlpack(activations_tensor))
gradients = kaldi.FloatSubVectorFromDLPack(to_dlpack(gradients_tensor))
flat_labels = kaldi.IntSubVectorFromDLPack(
to_dlpack(flat_labels_tensor))
costs = kaldi.FloatSubVectorFromDLPack(to_dlpack(costs_tensor))
workspace = kaldi.FloatSubVectorFromDLPack(to_dlpack(workspace_tensor))
status = ctc.ComputeCtcLossCpu(activations=activations,
gradients=gradients,
flat_labels=flat_labels,
label_lengths=label_lengths,
input_lengths=input_lengths,
alphabet_size=alphabet_size,
minibatch=minibatch,
costs=costs,
workspace=workspace,
options=info)
# 1.6094379425049 is copied from
# https://github.com/baidu-research/warp-ctc/blob/master/torch_binding/TUTORIAL.md
self.assertAlmostEqual(costs[0], 1.6094379425049)
def test_dlpack_subvector(self):
'''
kaldi.Hello accepts two parameters:
Param 1: reference to VectorBase<float>
Param 2: pointer to VectorBase<float>
This test shows that we can pass a DLPackSubVector
to `Hello`.
'''
tensor1 = torch.tensor([1, 2]).float()
v1 = kaldi.FloatSubVectorFromDLPack(to_dlpack(tensor1))
tensor2 = torch.tensor([10, 20, 30]).float()
v2 = kaldi.FloatSubVectorFromDLPack(to_dlpack(tensor2))
kaldi.Hello(v1, v2)
def test_pytorch_cpu_tensor_to_subvector(self):
'''
Note that we have two ways to convert a
PyTorch CPU tensor to kaldi's FloatSubVector:
Method 1:
v = kaldi.FloatSubVectorFromDLPack(to_dlpack(tensor))
Method 2:
v = kaldi.DLPackFloatSubVector.from_dlpack(to_dlpack(tensor))
'''
tensor = torch.arange(3).float()
# v and tensor share the same memory
# no data is copied
v = kaldi.FloatSubVectorFromDLPack(to_dlpack(tensor))
# since DLPackFloatSubVector is a subclass of FloatSubVector
# the following assertion is True
self.assertIsInstance(v, kaldi.FloatSubVector)
v[0] = 100
v[1] = 200
v[2] = 300
self.assertEqual(tensor[0], 100)
self.assertEqual(tensor[1], 200)
self.assertEqual(tensor[2], 300)
del v
# the destructor of DLPackFloatSubVector is invoked
# by the above `del v` statement, you should see the log message
# here if you have put it in the destructor.
# memory is shared between `v` and `tensor`
v = kaldi.DLPackFloatSubVector.from_dlpack(to_dlpack(tensor))
self.assertEqual(v[0], 100)
v[0] = 1 # also changes tensor
self.assertEqual(tensor[0], 1)
def test_case4(self):
# combine case1 to case3 to a minibatch
# the first example (a): input_length: 1, label_length: 1
# the second example (c, c): input_length: 3, label_length: 2
# the third example (b, c): input_length: 3, label_length: 2
label_lengths_tensor = torch.tensor([1, 2, 2], dtype=torch.int32)
input_lengths_tensor = torch.tensor([1, 3, 3], dtype=torch.int32)
alphabet_size = 5
minibatch = 3
info = ctc.CtcOptions()
info.loc = ctc.CtcComputeLocation.CTC_CPU
info.num_threads = 0 # use default number of threads
info.blank_label = 0
label_lengths = kaldi.IntSubVectorFromDLPack(
to_dlpack(label_lengths_tensor))
input_lengths = kaldi.IntSubVectorFromDLPack(
to_dlpack(input_lengths_tensor))
status, size_in_bytes = ctc.GetWorkspaceSize(
label_lengths=label_lengths,
input_lengths=input_lengths,
alphabet_size=alphabet_size,
minibatch=minibatch,
info=info)
self.assertEqual(status, ctc.CtcStatus.CTC_STATUS_SUCCESS)
num_floats = size_in_bytes // 4 + 1
workspace_tensor = torch.empty(num_floats,
dtype=torch.float32).contiguous()
ex1 = torch.tensor([[0.2, 0.2, 0.2, 0.2, 0.2]], dtype=torch.float32)
ex2 = torch.tensor(
[[1, 2, 3, 4, 5], [6, 7, 8, 9, 10], [11, 12, 13, 14, 15]],
dtype=torch.float32)
ex3 = torch.tensor([[-5, -4, -3, -2, -1], [-10, -9, -8, -7, -6],
[-15, -14, -13, -12, -11]],
dtype=torch.float32)
activations_tensor = pad_sequence([ex1, ex2, ex3], batch_first=False)
activations_tensor = activations_tensor.contiguous().view(-1)
gradients_tensor = torch.empty_like(activations_tensor)
# labels are: (a), (c, c) (b, c)
# which are: (1), (3, 3), (2, 3)
flat_labels_tensor = torch.tensor([1, 3, 3, 2, 3], dtype=torch.int32)
costs_tensor = torch.empty(minibatch, dtype=torch.float32)
activations = kaldi.FloatSubVectorFromDLPack(
to_dlpack(activations_tensor))
gradients = kaldi.FloatSubVectorFromDLPack(to_dlpack(gradients_tensor))
flat_labels = kaldi.IntSubVectorFromDLPack(
to_dlpack(flat_labels_tensor))
costs = kaldi.FloatSubVectorFromDLPack(to_dlpack(costs_tensor))
workspace = kaldi.FloatSubVectorFromDLPack(to_dlpack(workspace_tensor))
status = ctc.ComputeCtcLossCpu(activations=activations,
gradients=gradients,
flat_labels=flat_labels,
label_lengths=label_lengths,
input_lengths=input_lengths,
alphabet_size=alphabet_size,
minibatch=minibatch,
costs=costs,
workspace=workspace,
options=info)
self.assertAlmostEqual(costs[0], 1.6094379425049)
self.assertAlmostEqual(costs[1], 7.355742931366)
self.assertAlmostEqual(costs[2], 4.938850402832, places=6)
def test_case3(self):
# this is the third case
# we have 3 sequences with probability:
# [-5, -4, -3, -2, -1]
# [-10, -9, -8, -7, -6]
# [-15, -14, -13, -12, -11]
label_lengths_tensor = torch.tensor([2], dtype=torch.int32)
input_lengths_tensor = torch.tensor([3], dtype=torch.int32)
alphabet_size = 5
minibatch = 1
info = ctc.CtcOptions()
info.loc = ctc.CtcComputeLocation.CTC_CPU
info.num_threads = 0 # use default number of threads
info.blank_label = 0
label_lengths = kaldi.IntSubVectorFromDLPack(
to_dlpack(label_lengths_tensor))
input_lengths = kaldi.IntSubVectorFromDLPack(
to_dlpack(input_lengths_tensor))
status, size_in_bytes = ctc.GetWorkspaceSize(
label_lengths=label_lengths,
input_lengths=input_lengths,
alphabet_size=alphabet_size,
minibatch=minibatch,
info=info)
self.assertEqual(status, ctc.CtcStatus.CTC_STATUS_SUCCESS)
num_floats = size_in_bytes // 4 + 1
workspace_tensor = torch.empty(num_floats,
dtype=torch.float32).contiguous()
activations_tensor = torch.tensor(
[[-5, -4, -3, -2, -1], [-10, -9, -8, -7, -6],
[-15, -14, -13, -12, -11]],
dtype=torch.float32).contiguous().view(-1)
gradients_tensor = torch.empty_like(activations_tensor)
# the target sequence is b c, whichis 2 3
flat_labels_tensor = torch.tensor([2, 3], dtype=torch.int32)
costs_tensor = torch.empty(minibatch, dtype=torch.float32)
activations = kaldi.FloatSubVectorFromDLPack(
to_dlpack(activations_tensor))
gradients = kaldi.FloatSubVectorFromDLPack(to_dlpack(gradients_tensor))
flat_labels = kaldi.IntSubVectorFromDLPack(
to_dlpack(flat_labels_tensor))
costs = kaldi.FloatSubVectorFromDLPack(to_dlpack(costs_tensor))
workspace = kaldi.FloatSubVectorFromDLPack(to_dlpack(workspace_tensor))
status = ctc.ComputeCtcLossCpu(activations=activations,
gradients=gradients,
flat_labels=flat_labels,
label_lengths=label_lengths,
input_lengths=input_lengths,
alphabet_size=alphabet_size,
minibatch=minibatch,
costs=costs,
workspace=workspace,
options=info)
# 4.938850402832 is copied from
# https://github.com/baidu-research/warp-ctc/blob/master/torch_binding/TUTORIAL.md
self.assertAlmostEqual(costs[0], 4.938850402832, places=6)
def test_case2(self):
device = torch.device('cuda', device_id)
# this is the second case
# we have 3 sequences with probability:
# [1, 2, 3, 4, 5]
# [6, 7, 8, 9, 10]
# [11, 12, 13, 14, 15]
label_lengths_tensor = torch.tensor([2], dtype=torch.int32)
input_lengths_tensor = torch.tensor([3], dtype=torch.int32)
alphabet_size = 5
minibatch = 1
info = ctc.CtcOptions()
info.loc = ctc.CtcComputeLocation.CTC_GPU
info.blank_label = 0
label_lengths = kaldi.IntSubVectorFromDLPack(
to_dlpack(label_lengths_tensor))
input_lengths = kaldi.IntSubVectorFromDLPack(
to_dlpack(input_lengths_tensor))
status, size_in_bytes = ctc.GetWorkspaceSize(
label_lengths=label_lengths,
input_lengths=input_lengths,
alphabet_size=alphabet_size,
minibatch=minibatch,
info=info)
self.assertEqual(status, ctc.CtcStatus.CTC_STATUS_SUCCESS)
num_floats = size_in_bytes // 4 + 1
workspace_tensor = torch.empty(
num_floats, dtype=torch.float32).contiguous().to(device)
activations_tensor = torch.tensor(
[[1, 2, 3, 4, 5], [6, 7, 8, 9, 10], [11, 12, 13, 14, 15]],
dtype=torch.float32).contiguous().view(-1).to(device)
gradients_tensor = torch.empty_like(activations_tensor)
# the target sequence is cc, which is 3 3
flat_labels_tensor = torch.tensor([3, 3], dtype=torch.int32)
costs_tensor = torch.empty(minibatch, dtype=torch.float32)
activations = kaldi.CuSubVectorFromDLPack(to_dlpack(activations_tensor))
gradients = kaldi.CuSubVectorFromDLPack(to_dlpack(gradients_tensor))
flat_labels = kaldi.IntSubVectorFromDLPack(
to_dlpack(flat_labels_tensor))
costs = kaldi.FloatSubVectorFromDLPack(to_dlpack(costs_tensor))
workspace = kaldi.CuSubVectorFromDLPack(to_dlpack(workspace_tensor))
status = ctc.ComputeCtcLossGpu(activations=activations,
gradients=gradients,
flat_labels=flat_labels,
label_lengths=label_lengths,
input_lengths=input_lengths,
alphabet_size=alphabet_size,
minibatch=minibatch,
costs=costs,
workspace=workspace,
options=info)
# 7.355742931366 is copied from
# https://github.com/baidu-research/warp-ctc/blob/master/torch_binding/TUTORIAL.md
self.assertAlmostEqual(costs[0], 7.355742931366)
