def test_qnode_gradient_fanout(self, qubit_device_1_wire, tol):
"Tests that the correct gradient is computed for qnodes which use the same parameter in multiple gates."
def expZ(state):
return np.abs(state[0])**2 - np.abs(state[1])**2
extra_param = 0.31
def circuit(reused_param, other_param):
qml.RX(extra_param, wires=[0])
qml.RY(reused_param, wires=[0])
qml.RZ(other_param, wires=[0])
qml.RX(reused_param, wires=[0])
return qml.expval(qml.PauliZ(0))
f = qml.QNode(circuit, qubit_device_1_wire)
zero_state = np.array([1., 0.])
for reused_p in thetas:
reused_p = reused_p**3 / 19
for other_p in thetas:
other_p = other_p**2 / 11
# autograd gradient
grad = autograd.grad(f)
grad_eval = grad(reused_p, other_p)
# manual gradient
grad_true0 = (expZ(Rx(reused_p) @ Rz(other_p) @ Ry(reused_p + np.pi / 2) @ Rx(extra_param) @ zero_state) \
-expZ(Rx(reused_p) @ Rz(other_p) @ Ry(reused_p - np.pi / 2) @ Rx(extra_param) @ zero_state)) / 2
grad_true1 = (expZ(Rx(reused_p + np.pi / 2) @ Rz(other_p) @ Ry(reused_p) @ Rx(extra_param) @ zero_state) \
-expZ(Rx(reused_p - np.pi / 2) @ Rz(other_p) @ Ry(reused_p) @ Rx(extra_param) @ zero_state)) / 2
grad_true = grad_true0 + grad_true1 # product rule
assert grad_eval == pytest.approx(grad_true, abs=tol)
def test_fanout_multiple_params(self, reused_p, other_p, tol):
"""Tests that the correct gradient is computed for qnodes which
use the same parameter in multiple gates."""
from gate_data import Rotx as Rx, Roty as Ry, Rotz as Rz
def expZ(state):
return np.abs(state[0]) ** 2 - np.abs(state[1]) ** 2
extra_param = 0.31
def circuit(reused_param, other_param):
qml.RX(extra_param, wires=[0])
qml.RY(reused_param, wires=[0])
qml.RZ(other_param, wires=[0])
qml.RX(reused_param, wires=[0])
return qml.expval(qml.PauliZ(0))
dev = qml.device("default.qubit", wires=1)
f = QubitQNode(circuit, dev)
zero_state = np.array([1., 0.])
# analytic gradient
grad_A = f.jacobian([reused_p, other_p])
# manual gradient
grad_true0 = (expZ(Rx(reused_p) @ Rz(other_p) @ Ry(reused_p + np.pi / 2) @ Rx(extra_param) @ zero_state) \
-expZ(Rx(reused_p) @ Rz(other_p) @ Ry(reused_p - np.pi / 2) @ Rx(extra_param) @ zero_state)) / 2
grad_true1 = (expZ(Rx(reused_p + np.pi / 2) @ Rz(other_p) @ Ry(reused_p) @ Rx(extra_param) @ zero_state) \
-expZ(Rx(reused_p - np.pi / 2) @ Rz(other_p) @ Ry(reused_p) @ Rx(extra_param) @ zero_state)) / 2
grad_true = grad_true0 + grad_true1 # product rule
assert grad_A[0, 0] == pytest.approx(grad_true, abs=tol)
def test_fanout_multiple_params(self, reused_p, other_p, tol, mocker, dev):
"""Tests that the correct gradient is computed for qnodes which
use the same parameter in multiple gates."""
from gate_data import Rotx as Rx, Roty as Ry, Rotz as Rz
def expZ(state):
return np.abs(state[0]) ** 2 - np.abs(state[1]) ** 2
extra_param = np.array(0.31, requires_grad=False)
@qnode(dev, diff_method="adjoint")
def cost(p1, p2):
qml.RX(extra_param, wires=[0])
qml.RY(p1, wires=[0])
qml.RZ(p2, wires=[0])
qml.RX(p1, wires=[0])
return qml.expval(qml.PauliZ(0))
zero_state = np.array([1.0, 0.0])
cost(reused_p, other_p)
spy = mocker.spy(dev, "adjoint_jacobian")
# analytic gradient
grad_fn = qml.grad(cost)
grad_D = grad_fn(reused_p, other_p)
spy.assert_called_once()
# manual gradient
grad_true0 = (
expZ(
Rx(reused_p) @ Rz(other_p) @ Ry(reused_p + np.pi / 2) @ Rx(extra_param) @ zero_state
)
- expZ(
Rx(reused_p) @ Rz(other_p) @ Ry(reused_p - np.pi / 2) @ Rx(extra_param) @ zero_state
)
) / 2
grad_true1 = (
expZ(
Rx(reused_p + np.pi / 2) @ Rz(other_p) @ Ry(reused_p) @ Rx(extra_param) @ zero_state
)
- expZ(
Rx(reused_p - np.pi / 2) @ Rz(other_p) @ Ry(reused_p) @ Rx(extra_param) @ zero_state
)
) / 2
expected = grad_true0 + grad_true1 # product rule
assert np.allclose(grad_D[0], expected, atol=tol, rtol=0)
def test_fanout_multiple_params(self, reused_p, other_p, tol):
"""Tests that the correct gradient is computed for qnodes which
use the same parameter in multiple gates."""
from gate_data import Rotx as Rx, Roty as Ry, Rotz as Rz
def expZ(state):
return np.abs(state[0])**2 - np.abs(state[1])**2
dev = qml.device("default.qubit", wires=1)
extra_param = np.array(0.31, requires_grad=False)
@qnode(dev)
def cost(p1, p2):
qml.RX(extra_param, wires=[0])
qml.RY(p1, wires=[0])
qml.RZ(p2, wires=[0])
qml.RX(p1, wires=[0])
return expval(qml.PauliZ(0))
zero_state = np.array([1.0, 0.0])
# analytic gradient
grad_fn = qml.grad(cost)
grad_A = grad_fn(reused_p, other_p)
# manual gradient
grad_true0 = (expZ(
Rx(reused_p) @ Rz(other_p) @ Ry(reused_p + np.pi / 2)
@ Rx(extra_param) @ zero_state) - expZ(
Rx(reused_p) @ Rz(other_p) @ Ry(reused_p - np.pi / 2)
@ Rx(extra_param) @ zero_state)) / 2
grad_true1 = (expZ(
Rx(reused_p + np.pi / 2) @ Rz(other_p) @ Ry(reused_p)
@ Rx(extra_param) @ zero_state) - expZ(
Rx(reused_p - np.pi / 2) @ Rz(other_p) @ Ry(reused_p)
@ Rx(extra_param) @ zero_state)) / 2
expected = grad_true0 + grad_true1 # product rule
assert np.allclose(grad_A[0], expected, atol=tol, rtol=0)
