def test_cuda_float_array1_to_tensor(self):
device_id = 0
_arr = _k2.get_cuda_float_array1(device_id)
arr = k2.Array(_arr)
tensor = arr.tensor()
assert tensor.ndim == 1
assert tensor.dtype == torch.float
assert tensor.device.type == 'cuda'
assert tensor.device.index == device_id
assert tensor[0] == _arr.get(0)
# now we change the tensor, `_arr` should also be changed
# since they share the underlying memory
tensor[0] += 100
assert tensor[0] == _arr.get(0)
val = tensor[0]
del _arr, arr
tensor[0] += 1
val += 1
assert tensor[0] == val, 'tensor should still be accessible'
del tensor
def test_cuda_int_array1_from_tensor(self):
device_id = 0
device = torch.device('cuda', device_id)
gt_tensor = torch.tensor([1, 2, 3], dtype=torch.int32).to(device)
array = k2.Array(gt_tensor)
actual_tensor = array.tensor()
assert actual_tensor.dtype == gt_tensor.dtype
assert actual_tensor.device == gt_tensor.device
assert actual_tensor.ndim == gt_tensor.ndim
assert torch.allclose(gt_tensor, actual_tensor)
gt_tensor += 100
assert torch.allclose(gt_tensor, actual_tensor), \
'actual_tensor should share the same memory with gt_tensor'
val = gt_tensor[0]
del gt_tensor, array
actual_tensor[0] += 1
val += 1
assert val == actual_tensor[0], \
'actual_tensor[0] should still be accessible'
del actual_tensor
def test_cpu_int_array2_to_tensor(self):
_arr = _k2.get_cpu_int_array2()
arr = k2.Array(_arr)
tensor = arr.tensor()
del _arr, arr
assert torch.allclose(
tensor, torch.tensor([[1, 2, 3], [4, 5, 6]], dtype=torch.int32))
def test_cpu_float_array2_from_tensor(self):
gt_tensor = torch.tensor([[1, 2, 3], [4, 5, 6]], dtype=torch.float32)
arr = k2.Array(gt_tensor)
actual_tensor = arr.tensor()
actual_tensor[0, 0] = 1000
assert torch.allclose(gt_tensor, actual_tensor)
del gt_tensor, arr
assert actual_tensor[0, 0] == 1000
def test_cuda_float_array2_to_tensor(self):
_arr = _k2.get_cuda_float_array2()
arr = k2.Array(_arr)
tensor = arr.tensor()
assert torch.allclose(
tensor,
torch.tensor([[1, 2, 3], [4, 5, 6]],
dtype=torch.float,
device='cuda'))
tensor[0, 0] = 100
assert _arr.get(0).get(0) == 100
def test_cuda_int32_array2_to_tensor(self):
gt_tensor = torch.tensor([[1, 2, 3], [4, 5, 6]],
dtype=torch.int32,
device='cuda')
arr = k2.Array(gt_tensor)
actual_tensor = arr.tensor()
assert torch.allclose(gt_tensor, actual_tensor)
gt_tensor[0, 0] = 100
del gt_tensor
assert actual_tensor[0, 0] == 100
def test_cuda_arc_array1_to_tensor(self):
device_id = 0
_arc_array = _k2.get_cuda_arc_array1(device_id)
arc_array = k2.Array(_arc_array)
tensor = arc_array.tensor()
assert tensor.ndim == 2
assert tensor.shape == (2, 4)
assert tensor.device.type == 'cuda'
assert tensor.device.index == device_id
assert tensor[0][0] == 1
assert tensor[0][3] == _k2._float_as_int(_arc_array.get(0).score)
tensor[0][0] = 10 # also change _arc_array
assert _arc_array.get(0).src_state == 10
del arc_array, _arc_array
tensor[0][0] += 10
assert tensor[0][0] == 20, 'tensor should still be accessible'
def test_cpu_int_array1_to_tensor(self):
_arr = _k2.get_cpu_int_array1()
arr = k2.Array(_arr)
tensor = arr.tensor()
assert tensor.ndim == 1
assert tensor.dtype == torch.int32
assert tensor.device.type == 'cpu'
assert tensor[0] == _arr.get(0)
# now we change the tensor, `_arr` should also be changed
# since they share the underlying memory
tensor[0] += 100
assert tensor[0] == _arr.get(0)
val = tensor[0]
del _arr, arr
assert tensor[0] == val, 'tensor should still be accessible'
del tensor
def test_cpu_arc_array1_from_tensor(self):
gt_tensor = torch.tensor(
[[1, 2, 3, _k2._float_as_int(1.5)],
[10, 20, 30, _k2._float_as_int(-2.5)]],
dtype=torch.int32).contiguous()
arc_array = k2.Array(gt_tensor, True)
arc0 = arc_array.data.get(0)
assert arc0.src_state == 1
assert arc0.dest_state == 2
assert arc0.symbol == 3
assert arc0.score == 1.5
actual_tensor = arc_array.tensor()
assert torch.allclose(gt_tensor, actual_tensor)
gt_tensor[0] += 10 # also change actual_tensor
assert torch.allclose(gt_tensor, actual_tensor)
assert arc_array.data.get(0).src_state == 11
del gt_tensor, actual_tensor
# arc_array is still accessible
assert arc_array.data.get(0).src_state == 11
