def test_var_ddof_backward(x, ddof):
y = Tensor(x)
def f(z):
return np.var(z, ddof=ddof)
o = var(y, ddof=ddof)
o.backward(2.)
g, = numerical_gradient_full(f, x, back_grad=np.asarray(2.))
assert_allclose(g, y.grad, rtol=1e-5, atol=1e-5)
def simple_batchnorm(x, gamma, beta, eps):
axes = [i for i in range(x.ndim)]
axes.pop(1) # every axis except 1
axes = tuple(axes)
keepdims_shape = tuple(1 if n != 1 else d for n, d in enumerate(x.shape))
mean = mg.mean(x, axis=axes, keepdims=True)
var = mg.var(x, axis=axes, keepdims=True)
norm = (x - mean) / mg.sqrt(var + eps)
if gamma is not None:
gamma = gamma.reshape(keepdims_shape)
norm *= gamma
if beta is not None:
beta = beta.reshape(keepdims_shape)
norm += beta
return norm
def test_var_no_axis_bkwrd(x):
import mygrad as mg
x = mg.Tensor(x, constant=False)
mg.var(x, axis=()).backward()
assert np.all(x.grad == np.zeros_like(x.data))
def test_var_no_axis_fwd(x):
import mygrad as mg
x = mg.Tensor(x, constant=False)
o = mg.var(x, axis=())
assert np.all(o.data == np.zeros_like(x.data))
def test_var_no_axis_bkwrd(x: mg.Tensor):
mg.var(x, axis=()).backward()
assert np.all(x.grad == mg.zeros_like(x))
def test_var_no_axis_fwd(x: mg.Tensor):
o = mg.var(x, axis=())
assert np.all(o == mg.zeros_like(x))
def test_var_ddof(x, ddof):
assert np.var(x, ddof=ddof) == var(x, ddof=ddof).data