def test_ungroup_irregular_length_tensors():
x = np.array([[0, 1], [2, 3], [4, 5], [6, 7], [8, 9]], dtype=fX)
lengths = np.array([2, 1, 0, 2])
ans = np.array([[[0, 1], [2, 3]], [[4, 5], [0, 0]], [[0, 0], [0, 0]],
[[6, 7], [8, 9]]])
res = ungroup_irregular_length_tensors(x, lengths).eval()
np.testing.assert_equal(ans, res)
def compute_output(self, network, in_vw, lengths_vw):
network.create_vw(
"default",
variable=il.ungroup_irregular_length_tensors(
in_vw.variable, lengths_vw.variable),
shape=(None, None) + in_vw.shape[1:],
tags={"output"},
)
def compute_output(self, network, in_vw, lengths_vw):
network.create_vw(
"default",
variable=il.ungroup_irregular_length_tensors(in_vw.variable,
lengths_vw.variable),
shape=(None, None) + in_vw.shape[1:],
tags={"output"},
)
def test_ungroup_irregular_length_tensors_grad():
x = np.array([[0, 1], [2, 3], [4, 5], [6, 7], [8, 9]], dtype=fX)
x = T.constant(x)
lengths = np.array([2, 1, 0, 2])
ans = np.array([[1, 1], [1, 1], [2, 2], [0, 0], [0, 0]], dtype=fX)
ungrouped = ungroup_irregular_length_tensors(x, lengths)
out = ungrouped[0].sum() + 2 * ungrouped[1].sum()
grad = T.grad(out, x).eval()
np.testing.assert_equal(ans, grad)
def test_ungroup_irregular_length_tensors():
x = np.array([[0, 1],
[2, 3],
[4, 5],
[6, 7],
[8, 9]],
dtype=fX)
lengths = np.array([2, 1, 0, 2])
ans = np.array([[[0, 1],
[2, 3]],
[[4, 5],
[0, 0]],
[[0, 0],
[0, 0]],
[[6, 7],
[8, 9]]])
res = ungroup_irregular_length_tensors(x, lengths).eval()
np.testing.assert_equal(ans, res)
def test_ungroup_irregular_length_tensors_grad():
x = np.array([[0, 1],
[2, 3],
[4, 5],
[6, 7],
[8, 9]],
dtype=fX)
x = T.constant(x)
lengths = np.array([2, 1, 0, 2])
ans = np.array([[1, 1],
[1, 1],
[2, 2],
[0, 0],
[0, 0]],
dtype=fX)
ungrouped = ungroup_irregular_length_tensors(x, lengths)
out = ungrouped[0].sum() + 2 * ungrouped[1].sum()
grad = T.grad(out, x).eval()
np.testing.assert_equal(ans, grad)
def test_ungroup_irregular_length_tensors_numeric_gradient():
lengths = np.array([2, 3, 4, 5, 7, 2], dtype=fX)
T.verify_grad(lambda x: ungroup_irregular_length_tensors(x, lengths),
[np.random.randn(23, 10).astype(fX)],
rng=np.random)
def test_ungroup_irregular_length_tensors_numeric_gradient():
lengths = np.array([2, 3, 4, 5, 7, 2], dtype=fX)
T.verify_grad(lambda x: ungroup_irregular_length_tensors(x, lengths),
[np.random.randn(23, 10).astype(fX)],
rng=np.random)
