def compute_loss(y, t):
# softmax cross entropy
score = chx.log_softmax(y, axis=1)
mask = (t[:, chx.newaxis] == chx.arange(
10, dtype=t.dtype)).astype(score.dtype)
# TODO(beam2d): implement mean
return -(score * mask).sum() * (1 / y.shape[0])
def test_chainerx_to_cupy_noncontiguous():
dtype = 'float32'
a_chx = chainerx.arange(12, dtype=dtype, device='cuda:0').reshape(
(2, 6))[::-1, ::2]
offset = a_chx.offset
# test preconditions
assert offset > 0
assert not a_chx.is_contiguous
a_cupy = cupy.ndarray(
a_chx.shape,
cupy.dtype(a_chx.dtype.name),
cupy.cuda.MemoryPointer(
cupy.cuda.UnownedMemory(a_chx.data_ptr, a_chx.data_size, a_chx, 0),
offset),
strides=a_chx.strides,
)
assert a_chx.strides == a_cupy.strides
chainerx.testing.assert_array_equal_ex(a_chx,
a_cupy.get(),
strides_check=False)
a_cupy[1, 1] = 53
assert a_chx.strides == a_cupy.strides
chainerx.testing.assert_array_equal_ex(a_chx,
a_cupy.get(),
strides_check=False)
def test_chainerx_to_cupy_contiguous():
dtype = 'float32'
a_chx = chainerx.arange(6, dtype=dtype, device='cuda:0').reshape((2, 3))
a_cupy = cupy.ndarray(
a_chx.shape,
cupy.dtype(a_chx.dtype.name),
cupy.cuda.MemoryPointer(cupy.cuda.UnownedMemory(
a_chx.data_ptr + a_chx.offset,
a_chx.data_size,
a_chx,
0), 0),
strides=a_chx.strides,
)
assert a_cupy.device.id == 0
chainerx.testing.assert_array_equal_ex(a_chx, a_cupy.get())
# Write to a_cupy
a_cupy[0, 1] = 8
chainerx.testing.assert_array_equal_ex(
a_chx, numpy.array([[0, 8, 2], [3, 4, 5]], dtype))
# Write to a_chx
a_chx += 1
chainerx.testing.assert_array_equal_ex(
a_cupy.get(), numpy.array([[1, 9, 3], [4, 5, 6]], dtype))
def test_chainerx_to_cupy_noncontiguous():
dtype = 'float32'
a_chx = chainerx.arange(
12, dtype=dtype, device='cuda:0').reshape((2, 6))[::-1, ::2]
offset = a_chx.offset
# test preconditions
assert offset > 0
assert not a_chx.is_contiguous
a_cupy = cupy.ndarray(
a_chx.shape,
cupy.dtype(a_chx.dtype.name),
cupy.cuda.MemoryPointer(cupy.cuda.UnownedMemory(
a_chx.data_ptr,
a_chx.data_size,
a_chx,
0), offset),
strides=a_chx.strides,
)
assert a_chx.strides == a_cupy.strides
chainerx.testing.assert_array_equal_ex(
a_chx, a_cupy.get(), strides_check=False)
a_cupy[1, 1] = 53
assert a_chx.strides == a_cupy.strides
chainerx.testing.assert_array_equal_ex(
a_chx, a_cupy.get(), strides_check=False)
def compute_loss(y, t):
# softmax cross entropy
score = chx.log_softmax(y, axis=1)
mask = (t[:, chx.newaxis] == chx.arange(1000,
dtype=t.dtype)).astype(score.dtype)
# TODO(beam2d): implement mean
return -(score * mask).sum() * (1 / y.shape[0])
def test_ascontiguousarray_from_chainerx_array_device():
with chainerx.using_device(chainerx.get_device('native:0')):
dev = chainerx.get_device('native:1') # Non default one
assert chainerx.get_default_device() is not dev
a = chainerx.arange(10, device=dev)
b = chainerx.ascontiguousarray(a)
assert b.is_contiguous is True
assert b.device is dev
def forward_chainerx(self, inputs):
# TODO(niboshi): Current implementation is only intended to support
# MNIST example.
x, t = inputs
num_classes = x.shape[1]
score = chainerx.log_softmax(x, axis=1)
mask = (t[:, chainerx.newaxis] == chainerx.arange(
num_classes, dtype=t.dtype, device=x.device)).astype(score.dtype)
# TODO(beam2d): implement mean
y = -(score * mask).sum() * (1 / x.shape[0])
return y,
def forward_chainerx(self, inputs):
# TODO(niboshi): Current implementation is only intended to support
# MNIST example.
x, t = inputs
num_classes = x.shape[1]
score = chainerx.log_softmax(x, axis=1)
mask = (t[:, chainerx.newaxis] == chainerx.arange(
num_classes, dtype=t.dtype, device=x.device)).astype(score.dtype)
# TODO(beam2d): implement mean
y = -(score * mask).sum() * (1 / x.shape[0])
return y,
def test_scatter_dataset(self):
n = self.communicator.size
for shuffle in [True, False]:
for root in range(self.communicator.size):
self.check_scatter_dataset([], shuffle, root)
self.check_scatter_dataset([0], shuffle, root)
self.check_scatter_dataset(list(range(n)), shuffle, root)
self.check_scatter_dataset(list(range(n * 5 - 1)), shuffle,
root)
self.check_scatter_dataset(np.array([]), shuffle, root)
self.check_scatter_dataset(np.array([0]), shuffle, root)
self.check_scatter_dataset(np.arange(n), shuffle, root)
self.check_scatter_dataset(np.arange(n * 5 - 1), shuffle, root)
self.check_scatter_dataset(chx.array([]), shuffle, root)
self.check_scatter_dataset(chx.array([0]), shuffle, root)
self.check_scatter_dataset(chx.arange(n), shuffle, root)
self.check_scatter_dataset(chx.arange(n * 5 - 1), shuffle,
root)
def test_getitem_zero_sized_offsets(device):
a = chainerx.arange(6)
b = a[3:3]
# Test pre-conditions.
assert b.size == 0
assert b.offset == 12
# The offset of `c` should be the same as `b` since `b` is empty.
c = b[2:]
assert c.size == 0
assert c.offset == b.offset
def test_getitem_zero_sized_offsets(device):
a = chainerx.arange(6)
b = a[3:3]
# Test pre-conditions.
assert b.size == 0
assert b.offset == 12
# The offset of `c` should be the same as `b` since `b` is empty.
c = b[2:]
assert c.size == 0
assert c.offset == b.offset
def test_chainerx_to_cupy_nondefault_device():
dtype = 'float32'
a_chx = chainerx.arange(6, dtype=dtype, device='cuda:1').reshape((2, 3))
a_cupy = cupy.ndarray(
a_chx.shape,
cupy.dtype(a_chx.dtype.name),
cupy.cuda.MemoryPointer(
cupy.cuda.UnownedMemory(a_chx.data_ptr + a_chx.offset,
a_chx.data_size, a_chx, -1), 0),
strides=a_chx.strides,
)
assert a_cupy.device.id == 1
chainerx.testing.assert_array_equal_ex(a_chx, a_cupy.get())
def forward_chainerx(self, inputs):
if self.reduce == 'mean' and self.normalize:
x, t = inputs
n_classes = x.shape[1]
score = chainerx.log_softmax(x, axis=1)
mask = (t[:, chainerx.newaxis] == chainerx.arange(
n_classes, dtype=t.dtype, device=x.device)).astype(score.dtype)
y = (score * mask).sum() * (-1 / mask.sum())
return y,
x, t = inputs
y = chainerx.softmax_cross_entropy(x, t)
if self.reduce == 'mean':
return y.mean(),
return y,
def test_chainerx_to_cupy_nondefault_device():
dtype = 'float32'
a_chx = chainerx.arange(6, dtype=dtype, device='cuda:1').reshape((2, 3))
a_cupy = cupy.ndarray(
a_chx.shape,
cupy.dtype(a_chx.dtype.name),
cupy.cuda.MemoryPointer(cupy.cuda.UnownedMemory(
a_chx.data_ptr + a_chx.offset,
a_chx.data_size,
a_chx,
-1), 0),
strides=a_chx.strides,
)
assert a_cupy.device.id == 1
chainerx.testing.assert_array_equal_ex(a_chx, a_cupy.get())
def test_chainerx_to_cupy_delete_chainerx_first():
dtype = 'float32'
a_chx = chainerx.arange(6, dtype=dtype, device='cuda:0').reshape((2, 3))
a_cupy = cupy.ndarray(
a_chx.shape,
cupy.dtype(a_chx.dtype.name),
cupy.cuda.MemoryPointer(
cupy.cuda.UnownedMemory(a_chx.data_ptr + a_chx.offset,
a_chx.data_size, a_chx, 0), 0),
strides=a_chx.strides,
)
del a_chx
a_cupy += 1
chainerx.testing.assert_array_equal_ex(
a_cupy.get(), numpy.array([[1, 2, 3], [4, 5, 6]], dtype))
def test_chainerx_to_cupy_delete_chainerx_first():
dtype = 'float32'
a_chx = chainerx.arange(6, dtype=dtype, device='cuda:0').reshape((2, 3))
a_cupy = cupy.ndarray(
a_chx.shape,
cupy.dtype(a_chx.dtype.name),
cupy.cuda.MemoryPointer(cupy.cuda.UnownedMemory(
a_chx.data_ptr + a_chx.offset,
a_chx.data_size,
a_chx,
0), 0),
strides=a_chx.strides,
)
del a_chx
a_cupy += 1
chainerx.testing.assert_array_equal_ex(
a_cupy.get(), numpy.array([[1, 2, 3], [4, 5, 6]], dtype))
def test_chainerx_to_cupy_contiguous():
dtype = 'float32'
a_chx = chainerx.arange(6, dtype=dtype, device='cuda:0').reshape((2, 3))
a_cupy = cupy.ndarray(
a_chx.shape,
cupy.dtype(a_chx.dtype.name),
cupy.cuda.MemoryPointer(
cupy.cuda.UnownedMemory(a_chx.data_ptr + a_chx.offset,
a_chx.data_size, a_chx, 0), 0),
strides=a_chx.strides,
)
assert a_cupy.device.id == 0
chainerx.testing.assert_array_equal_ex(a_chx, a_cupy.get())
# Write to a_cupy
a_cupy[0, 1] = 8
chainerx.testing.assert_array_equal_ex(
a_chx, numpy.array([[0, 8, 2], [3, 4, 5]], dtype))
# Write to a_chx
a_chx += 1
chainerx.testing.assert_array_equal_ex(
a_cupy.get(), numpy.array([[1, 9, 3], [4, 5, 6]], dtype))
def check(*args, **kwargs):
a = chainerx.arange(*args, device=device, **kwargs)
b = chainerx.arange(*args, **kwargs)
array_utils.check_device(a, device)
chainerx.testing.assert_array_equal_ex(a, b)
def compute_loss(y, t): score = chx.log_softmax(y, axis=1) mask = (t[:, chx.newaxis]==chx.arange(10, dtype=t.dtype)).astype(score.dtype) return -(score * mask).sum() * (1 / y.shape[0])
def check(*args, **kwargs):
a = chainerx.arange(*args, device=device, **kwargs)
b = chainerx.arange(*args, **kwargs)
array_utils.check_device(a, device)
chainerx.testing.assert_array_equal_ex(a, b)
def test_to_numpy_positive_offset(device, copy):
a_chx = chainerx.arange(6).reshape(2, 3)[:, 1:]
a_np = chainerx.to_numpy(a_chx, copy)
_check_to_numpy(a_np, a_chx, device, copy)
def test_to_numpy_positive_offset(device, copy):
a_chx = chainerx.arange(6).reshape(2, 3)[:, 1:]
a_np = chainerx.to_numpy(a_chx, copy)
_check_to_numpy(a_np, a_chx, device, copy)
