def test_mean_subtraction_forward_backward(seed, inshape, base_axis, ctx,
func_name):
from nbla_test_utils import function_tester
rng = np.random.RandomState(seed)
mean_shape = inshape[base_axis:] if base_axis >= 0 else inshape[base_axis +
len(inshape
):]
inputs = [
np.array(rng.randn(*inshape).astype(np.float32)),
np.zeros(mean_shape),
np.array([1000])
]
batch_stat = True
function_tester(rng,
F.mean_subtraction,
ref_mean_subtraction,
inputs,
func_args=[base_axis, batch_stat],
ctx=ctx,
func_name=func_name,
dstep=1e-2,
backward=[True, False, False],
atol_b=1e-2)
# Check if running mean works.
vinputs = []
for input in inputs:
vinputs.append(nn.Variable(input.shape, True))
vinputs[-1].d = input
for i in range(5):
inputs[0] = rng.randn(*inputs[0].shape)
vinputs[0].d[...] = inputs[0]
ref_y, rmean = ref_mean_subtraction(*(inputs +
[base_axis, batch_stat]))
with nn.auto_forward():
y = F.mean_subtraction(*(vinputs + [base_axis, batch_stat]))
# print('vinput[1].d', vinputs[1].d, vinputs[2].d)
# print('inputs[1]', inputs[1], inputs[2])
assert_allclose(vinputs[1].d, inputs[1])
# Check if global stat mode works
batch_stat = False
ref_y = ref_mean_subtraction(*(inputs + [base_axis, batch_stat]))
with nn.auto_forward():
y = F.mean_subtraction(*(vinputs + [base_axis, batch_stat]))
assert_allclose(ref_y, y.d)
def test_mean_subtraction_forward_backward(seed, inshape, base_axis, ctx, func_name):
from nbla_test_utils import function_tester
rng = np.random.RandomState(seed)
mean_shape = inshape[base_axis:]
inputs = [np.array(rng.randn(*inshape).astype(np.float32)),
np.zeros(mean_shape),
np.array([1000])]
batch_stat = True
function_tester(rng, F.mean_subtraction, ref_mean_subtraction,
inputs,
func_args=[base_axis, batch_stat],
ctx=ctx, func_name=func_name, dstep=1e-2,
backward=[True, False, False],
atol_b=1e-2)
# Check if running mean works.
vinputs = []
for input in inputs:
vinputs.append(nn.Variable(input.shape, True))
vinputs[-1].d = input
for i in range(5):
inputs[0] = rng.randn(*inputs[0].shape)
vinputs[0].d[...] = inputs[0]
ref_y, rmean = ref_mean_subtraction(
*(inputs + [base_axis, batch_stat]))
with nn.auto_forward():
y = F.mean_subtraction(*(vinputs + [base_axis, batch_stat]))
# print 'vinput[1].d', vinputs[1].d, vinputs[2].d
# print 'inputs[1]', inputs[1], inputs[2]
assert np.allclose(vinputs[1].d, inputs[1])
# Check if global stat mode works
batch_stat = False
ref_y = ref_mean_subtraction(*(inputs + [base_axis, batch_stat]))
with nn.auto_forward():
y = F.mean_subtraction(*(vinputs + [base_axis, batch_stat]))
assert np.allclose(ref_y, y.d)