def test_fixed_batch_norm_invalid_dimensions(
device, x_shape, gamma_shape, beta_shape, mean_shape, var_shape, axis,
float_dtype):
x, gamma, beta, mean, var = _create_batch_norm_ndarray_args(
chainerx, device, x_shape, gamma_shape, beta_shape, mean_shape,
var_shape, float_dtype)
with pytest.raises(chainerx.DimensionError):
chainerx.fixed_batch_norm(
x, gamma, beta, mean=mean, var=var, eps=1e-2, axis=axis)
def test_fixed_batch_norm_invalid_dimensions(
device, x_shape, gamma_shape, beta_shape, mean_shape, var_shape, axis,
float_dtype):
x, gamma, beta, mean, var = _create_batch_norm_ndarray_args(
chainerx, device, x_shape, gamma_shape, beta_shape, mean_shape,
var_shape, float_dtype)
with pytest.raises(chainerx.DimensionError):
chainerx.fixed_batch_norm(
x, gamma, beta, mean=mean, var=var, eps=1e-2, axis=axis)
def test_fixed_batch_norm(device, x_shape, reduced_shape, eps, axis,
float_dtype):
def create_args(xp):
return _create_batch_norm_ndarray_args(xp, device, x_shape,
reduced_shape, reduced_shape,
reduced_shape, reduced_shape,
float_dtype)
x_chx, gamma_chx, beta_chx, mean_chx, var_chx = create_args(chainerx)
x_np, gamma_np, beta_np, mean_np, var_np = create_args(numpy)
optional_args = {}
if eps is not None:
optional_args['eps'] = eps
if axis is not None:
optional_args['axis'] = axis
y_chx = chainerx.fixed_batch_norm(x_chx,
gamma_chx,
beta_chx,
mean=mean_chx,
var=var_chx,
**optional_args)
y_np = chainer.functions.fixed_batch_normalization(x_np,
gamma_np,
beta_np,
mean=mean_np,
var=var_np,
**optional_args).data
chainerx.testing.assert_allclose_ex(y_chx, y_np, rtol=1e-6, atol=1e-5)
def test_fixed_batch_norm(
device, x_shape, reduced_shape, eps, axis, float_dtype):
def create_args(xp):
return _create_batch_norm_ndarray_args(
xp, device, x_shape, reduced_shape, reduced_shape, reduced_shape,
reduced_shape, float_dtype)
x_chx, gamma_chx, beta_chx, mean_chx, var_chx = create_args(chainerx)
x_np, gamma_np, beta_np, mean_np, var_np = create_args(numpy)
optional_args = {}
if eps is not None:
optional_args['eps'] = eps
if axis is not None:
optional_args['axis'] = axis
y_chx = chainerx.fixed_batch_norm(
x_chx, gamma_chx, beta_chx, mean=mean_chx, var=var_chx,
**optional_args)
y_np = chainer.functions.fixed_batch_normalization(
x_np, gamma_np, beta_np, mean=mean_np, var=var_np,
**optional_args).data
chainerx.testing.assert_allclose_ex(
y_chx, y_np, rtol=1e-6, atol=1e-5,
float16_rtol=1e-2, float16_atol=1e-2)
def forward_chainerx(self, inputs):
x, gamma, beta, mean, var = inputs
y = chainerx.fixed_batch_norm(x,
gamma,
beta,
mean=mean,
var=var,
**self.optional_args)
return y,
def forward_chainerx(self, inputs):
# TODO(niboshi): Support conditions implemented as fallback
# TODO(niboshi): chainerx.fixed_batch_norm does not support backward
if chainer.config.enable_backprop:
return chainer.Fallback
x, gamma, beta, mean, var = inputs
axis_chx = _chainerx_compute_axis(x.ndim, gamma.ndim, self.axis)
if not _chainerx_is_supported(x.device, axis_chx):
return chainer.Fallback
y = chainerx.fixed_batch_norm(
x, gamma, beta, mean, var, self.eps, axis_chx)
return y,
def forward_chainerx(self, inputs):
# TODO(niboshi): Support conditions implemented as fallback
# TODO(niboshi): chainerx.fixed_batch_norm does not support backward
if chainer.config.enable_backprop:
return chainer.Fallback
x, gamma, beta, mean, var = inputs
axis_chx = _chainerx_compute_axis(x.ndim, gamma.ndim, self.axis)
if not _chainerx_is_supported(x.device, axis_chx):
return chainer.Fallback
y = chainerx.fixed_batch_norm(
x, gamma, beta, mean, var, self.eps, axis_chx)
return y,
def forward_chainerx(self, inputs):
x, gamma, beta, mean, var = inputs
y = chainerx.fixed_batch_norm(
x, gamma, beta, mean=mean, var=var, **self.optional_args)
return y,
