def forward_xp(self, inputs, xp):
a, = inputs
# This cast was introduced in order to avoid decreasing precision.
# ex.) numpy.sqrt(x) becomes a float16 array where x is an int8 array.
a = dtype_utils.cast_if_numpy_array(xp, a, self.out_dtype)
with IgnoreNumpyFloatingPointError():
y = self.func(xp, a)
y = dtype_utils.cast_if_numpy_array(xp, y, self.out_dtype)
return y,
def forward_xp(self, inputs, xp):
a, = inputs
# This cast was introduced in order to avoid decreasing precision.
# ex.) numpy.sqrt(x) becomes a float16 array where x is an int8 array.
a = dtype_utils.cast_if_numpy_array(xp, a, self.out_dtype)
with IgnoreNumpyFloatingPointError():
y = self.func(xp, a)
y = dtype_utils.cast_if_numpy_array(xp, y, self.out_dtype)
return y,
def test_where_scalar_scalar(xp, cond_shape, cond_dtype, in_types, out_dtype):
cond = xp.array(_random_condition(cond_shape, cond_dtype))
x_type, y_type = in_types
x = x_type(0)
y = y_type(2)
out = xp.where(cond, x, y)
return dtype_utils.cast_if_numpy_array(xp, out, out_dtype)
def test_eye(xp, N, M, k, dtype_spec, device):
if xp is numpy and isinstance(dtype_spec, chainerx.dtype):
dtype_spec = dtype_spec.name
out = xp.eye(N, M, k, dtype_spec)
if dtype_spec in (None, Unspecified):
out = dtype_utils.cast_if_numpy_array(xp, out, 'float32')
return out
def test_eye(xp, N, M, k, dtype_spec, device):
if xp is numpy and isinstance(dtype_spec, chainerx.dtype):
dtype_spec = dtype_spec.name
out = xp.eye(N, M, k, dtype_spec)
if dtype_spec in (None, Unspecified):
out = dtype_utils.cast_if_numpy_array(xp, out, 'float32')
return out
def forward_xp(self, inputs, xp):
a, b = inputs
if self.is_module:
y = xp.dot(a, b)
else:
y = a.dot(b)
y = dtype_utils.cast_if_numpy_array(xp, y, self.chx_expected_dtype)
return y,
def test_sqrt(xp, device, input, in_dtype, out_dtype):
if (input.size > 0 and not numpy.isfinite(input).all() and
numpy.dtype(in_dtype).kind != 'f'):
return chainerx.testing.ignore()
a = xp.array(input.astype(in_dtype))
a = dtype_utils.cast_if_numpy_array(xp, a, out_dtype)
return xp.sqrt(a)
def test_sqrt(xp, device, input, in_dtype, out_dtype):
if (input.size > 0 and not numpy.isfinite(input).all()
and numpy.dtype(in_dtype).kind != 'f'):
return chainerx.testing.ignore()
a = xp.array(input.astype(in_dtype))
a = dtype_utils.cast_if_numpy_array(xp, a, out_dtype)
return xp.sqrt(a)
def forward_xp(self, inputs, xp):
a, b = inputs
if self.is_module:
y = xp.dot(a, b)
else:
y = a.dot(b)
y = dtype_utils.cast_if_numpy_array(xp, y, self.chx_expected_dtype)
return y,
def forward_xp(self, inputs, xp):
if self.axis is _unspecified:
b = xp.concatenate(inputs)
else:
b = xp.concatenate(inputs, self.axis)
if self.chx_expected_dtype is not None:
b = dtype_utils.cast_if_numpy_array(xp, b, self.chx_expected_dtype)
return b,
def forward_xp(self, inputs, xp):
if self.axis is _unspecified:
b = xp.concatenate(inputs)
else:
b = xp.concatenate(inputs, self.axis)
if self.chx_expected_dtype is not None:
b = dtype_utils.cast_if_numpy_array(xp, b, self.chx_expected_dtype)
return b,
def test_where_scalar_scalar(xp, cond_shape, cond_dtype, in_types, out_dtype):
cond = _random_condition(cond_shape,
cond_dtype,
random_state=numpy.random.RandomState(seed=0))
cond = xp.array(cond)
x_type, y_type = in_types
x = x_type(0)
y = y_type(2)
out = xp.where(cond, x, y)
return dtype_utils.cast_if_numpy_array(xp, out, out_dtype)
def test_ones(xp, shape_as_tuple_or_int, dtype_spec, device):
if xp is numpy and isinstance(dtype_spec, chainerx.dtype):
dtype_spec = dtype_spec.name
if dtype_spec is Unspecified:
out = xp.ones(shape_as_tuple_or_int)
else:
out = xp.ones(shape_as_tuple_or_int, dtype_spec)
if dtype_spec in (None, Unspecified):
out = dtype_utils.cast_if_numpy_array(xp, out, 'float32')
return out
def test_ones(xp, shape_as_tuple_or_int, dtype_spec, device):
if xp is numpy and isinstance(dtype_spec, chainerx.dtype):
dtype_spec = dtype_spec.name
if dtype_spec is Unspecified:
out = xp.ones(shape_as_tuple_or_int)
else:
out = xp.ones(shape_as_tuple_or_int, dtype_spec)
if dtype_spec in (None, Unspecified):
out = dtype_utils.cast_if_numpy_array(xp, out, 'float32')
return out
def test_empty(xp, shape_as_tuple_or_int, dtype_spec, device):
if xp is numpy and isinstance(dtype_spec, chainerx.dtype):
dtype_spec = dtype_spec.name
if dtype_spec is Unspecified:
a = xp.empty(shape_as_tuple_or_int)
else:
a = xp.empty(shape_as_tuple_or_int, dtype_spec)
a.fill(0)
if dtype_spec in (None, Unspecified):
a = dtype_utils.cast_if_numpy_array(xp, a, 'float32')
return a
def test_empty(xp, shape_as_tuple_or_int, dtype_spec, device):
if xp is numpy and isinstance(dtype_spec, chainerx.dtype):
dtype_spec = dtype_spec.name
if dtype_spec is Unspecified:
a = xp.empty(shape_as_tuple_or_int)
else:
a = xp.empty(shape_as_tuple_or_int, dtype_spec)
a.fill(0)
if dtype_spec in (None, Unspecified):
a = dtype_utils.cast_if_numpy_array(xp, a, 'float32')
return a
def apply_func(is_module, func, xp, device, input, axis, dtypes):
(in_dtype, ), out_dtype = dtypes
try:
a_np = input.astype(in_dtype)
except (ValueError, OverflowError):
return xp.zeros(()) # invalid combination of data and dtype
a = xp.array(a_np)
a = func(is_module, xp, a, axis)
if xp is numpy:
a = dtype_utils.cast_if_numpy_array(xp, a, out_dtype)
return a
def forward_xp(self, inputs, xp):
a, b = inputs
b = dtype_utils.cast_if_numpy_array(xp, b, a.dtype)
if xp is chainerx:
a_ = a.as_grad_stopped().copy()
b_ = b.as_grad_stopped()
else:
a_ = a.copy()
b_ = b
with IgnoreNumpyFloatingPointError():
ret = self.func(xp, a_, b_)
assert ret is None # func should not return anything
return a_,
def test_arange_stop(xp, stop, dtype_spec, device):
# TODO(hvy): xp.arange(True) should return an ndarray of type int64
if xp is numpy and isinstance(dtype_spec, chainerx.dtype):
dtype_spec = dtype_spec.name
# Checked in test_invalid_arange_too_long_bool
if _is_bool_spec(dtype_spec) and stop > 2:
return chainerx.testing.ignore()
if isinstance(stop, bool) and dtype_spec is None:
# TODO(niboshi): This pattern needs dtype promotion.
return chainerx.testing.ignore()
out = xp.arange(stop, dtype=dtype_spec)
if dtype_spec in (None, Unspecified):
expected_dtype = _get_default_dtype(stop)
out = dtype_utils.cast_if_numpy_array(xp, out, expected_dtype)
return out
def test_arange_stop(xp, stop, dtype_spec, device):
# TODO(hvy): xp.arange(True) should return an ndarray of type int64
if xp is numpy and isinstance(dtype_spec, chainerx.dtype):
dtype_spec = dtype_spec.name
# Checked in test_invalid_arange_too_long_bool
if _is_bool_spec(dtype_spec) and stop > 2:
return chainerx.testing.ignore()
if isinstance(stop, bool) and dtype_spec is None:
# TODO(niboshi): This pattern needs dtype promotion.
return chainerx.testing.ignore()
out = xp.arange(stop, dtype=dtype_spec)
if dtype_spec in (None, Unspecified):
expected_dtype = _get_default_dtype(stop)
out = dtype_utils.cast_if_numpy_array(xp, out, expected_dtype)
return out
def test_arange_start_stop_step(xp, start, stop, step, dtype_spec, device):
if xp is numpy and isinstance(dtype_spec, chainerx.dtype):
dtype_spec = dtype_spec.name
# Checked in test_invalid_arange_too_long_bool
if _is_bool_spec(dtype_spec) and abs((stop - start) / step) > 2:
return chainerx.testing.ignore()
if ((isinstance(start, bool) or isinstance(stop, bool)
or isinstance(step, bool)) and dtype_spec is None):
# TODO(niboshi): This pattern needs dtype promotion.
return chainerx.testing.ignore()
out = xp.arange(start, stop, step, dtype=dtype_spec)
if dtype_spec in (None, Unspecified):
expected_dtype = _get_default_dtype(step)
out = dtype_utils.cast_if_numpy_array(xp, out, expected_dtype)
return out
def test_arange_start_stop_step(xp, start, stop, step, dtype_spec, device):
if xp is numpy and isinstance(dtype_spec, chainerx.dtype):
dtype_spec = dtype_spec.name
# Checked in test_invalid_arange_too_long_bool
if _is_bool_spec(dtype_spec) and abs((stop - start) / step) > 2:
return chainerx.testing.ignore()
if ((isinstance(start, bool)
or isinstance(stop, bool)
or isinstance(step, bool))
and dtype_spec is None):
# TODO(niboshi): This pattern needs dtype promotion.
return chainerx.testing.ignore()
out = xp.arange(start, stop, step, dtype=dtype_spec)
if dtype_spec in (None, Unspecified):
expected_dtype = _get_default_dtype(step)
out = dtype_utils.cast_if_numpy_array(xp, out, expected_dtype)
return out
def test_dot(is_module, xp, device, a_shape, b_shape, dtypes,
chx_expected_dtype):
# TODO(beam2d): Remove the skip after supporting non-float dot on CUDA
if (device.name == 'cuda:0'
and any(numpy.dtype(dtype).kind != 'f' for dtype in dtypes)):
pytest.skip('CUDA dot only supports floating kind dtypes.')
a_dtype, b_dtype = dtypes
a = array_utils.create_dummy_ndarray(xp, a_shape, a_dtype)
b = array_utils.create_dummy_ndarray(xp, b_shape, b_dtype)
if is_module:
y = xp.dot(a, b)
else:
y = a.dot(b)
return dtype_utils.cast_if_numpy_array(xp, y, chx_expected_dtype)
def test_rounding_routines(func, xp, device, input, dtypes):
(in_dtype, ), out_dtype = dtypes
a = xp.array(input.astype(in_dtype))
a = func(xp, a)
a = dtype_utils.cast_if_numpy_array(xp, a, out_dtype)
return a
def test_log(xp, device, input, in_dtype, out_dtype):
a = xp.array(input.astype(in_dtype))
a = dtype_utils.cast_if_numpy_array(xp, a, out_dtype)
return xp.log(a)
def test_concatenate_three_arrays_mixed_dtypes(
xp, shapes, axis, dtypes, chx_expected_dtype):
assert len(shapes) == 3
y = _concatenate(xp, shapes, dtypes, axis)
return dtype_utils.cast_if_numpy_array(xp, y, chx_expected_dtype)
def test_full(xp, shape_as_tuple_or_int, value, device):
out = xp.full(shape_as_tuple_or_int, value)
return dtype_utils.cast_if_numpy_array(xp, out, _get_default_dtype(value))
def forward_xp(self, inputs, xp):
x, = inputs
x = dtype_utils.cast_if_numpy_array(xp, x, self.chx_dtype)
return xp.tanh(x),
def test_full(xp, shape_as_tuple_or_int, value, device):
out = xp.full(shape_as_tuple_or_int, value)
return dtype_utils.cast_if_numpy_array(xp, out, _get_default_dtype(value))
def forward_xp(self, inputs, xp):
x, y = inputs
condition = xp.array(self.condition)
o = xp.where(condition, x, y)
o = dtype_utils.cast_if_numpy_array(xp, o, self.out_dtype)
return o,
def forward_xp(self, inputs, xp):
b = self.join(inputs, xp)
if self.chx_expected_dtype is not None:
b = dtype_utils.cast_if_numpy_array(xp, b, self.chx_expected_dtype)
return b,
def forward_xp(self, inputs, xp):
x, = inputs
x = dtype_utils.cast_if_numpy_array(xp, x, self.chx_dtype)
return xp.tanh(x),
def func(self, xp, a):
if xp is numpy:
out_dtype = dict(_expected_dtypes_math_functions).get(self.dtype)
a = dtype_utils.cast_if_numpy_array(xp, a, out_dtype)
return xp.cos(a)
def test_log(xp, device, input, in_dtype, out_dtype):
a = xp.array(input.astype(in_dtype))
a = dtype_utils.cast_if_numpy_array(xp, a, out_dtype)
return xp.log(a)
def forward_xp(self, inputs, xp):
a, = inputs
if self.out_dtype is not None:
a = dtype_utils.cast_if_numpy_array(xp, a, self.out_dtype)
y = self.func(xp, a)
return y,
