def test_backward_on_X(self):
downstream_gradient = np.ones_like(self.x)
self.batch_norm(self.x)
d_X = self.batch_norm.backward(downstream_gradient)
np.testing.assert_almost_equal(
d_X, numeric_gradient(self.batch_norm, self.x))
def test_backward_on_X(self):
self.linear(self.matrix)
d_X = self.linear.backward(self.downstream_gradient)
np.testing.assert_almost_equal(
numeric_gradient(self.linear, self.matrix),
d_X
)
def test_gradient(self):
y_hat = np.array([.8, .75, .25, .01, .99])
y = np.array([1, 0, 0, 1, 1])
bce = BinaryCrossEntropy()
bce(y_hat, y)
np.testing.assert_almost_equal(
bce.gradient(), numeric_gradient(partial(bce, y=y), y_hat))
def test_backward_on_beta(self):
downstream_gradient = np.ones_like(self.x)
self.batch_norm(self.x)
self.batch_norm.backward(downstream_gradient)
def _forward_wrt_beta(new_beta):
self.batch_norm.beta[:] = new_beta
return self.batch_norm(self.x)
np.testing.assert_almost_equal(
self.batch_norm.beta.grad,
numeric_gradient(_forward_wrt_beta, self.batch_norm.beta).ravel())
def test_backward_on_b(self):
self.linear(self.matrix)
self.linear.backward(self.downstream_gradient)
def forward_wrt_b(new_b):
self.linear.b[:] = new_b
return self.linear(self.matrix)
np.testing.assert_almost_equal(
numeric_gradient(forward_wrt_b, self.linear.b),
self.linear.b.grad
)
def test_backward_on_W(self):
self.linear(self.matrix)
self.linear.backward(self.downstream_gradient)
def forward_wrt_W(new_W):
self.linear.W[:] = new_W
return self.linear(self.matrix)
np.testing.assert_almost_equal(
numeric_gradient(forward_wrt_W, self.linear.W),
self.linear.W.grad
)
def test_gradient(self):
y_hat = np.array([
[.2, .2, .2, .2, .2],
[.1, .1, .1, .6, .1],
[.1, .1, .1, .1, .6],
],
dtype=np.float64)
y = np.array([0, 3, 4], dtype=np.float64)
cce = CategoricalCrossEntropy()
cce(y_hat, y)
np.testing.assert_almost_equal(
cce.gradient(), numeric_gradient(partial(cce, y=y), y_hat))