def test_example_training():
U = tq.gates.Rx('a', 0) + tq.gates.Rx('b', 1) + tq.gates.CNOT(
1, 3) + tq.gates.CNOT(0, 2) + tq.gates.CNOT(0, 1)
H1 = tq.paulis.Qm(1)
H2 = tq.paulis.Qm(2)
H3 = tq.paulis.Qm(3)
stackable = [
tq.ExpectationValue(U, H1),
tq.ExpectationValue(U, H2),
tq.ExpectationValue(U, H3)
]
stacked = tq.vectorize(stackable)
initial_values = {'a': 1.5, 'b': 2.}
cargs = {
'samples': None,
'backend': 'random',
'initial_values': initial_values
}
torched = tq.ml.to_platform(stacked, platform='torch', compile_args=cargs)
optimizer = optim.SGD(torched.parameters(), lr=.1, momentum=0.9)
for i in range(80):
optimizer.zero_grad()
out = torched()
loss = out.sum()
loss.backward()
optimizer.step()
called = torched().sum().detach().numpy()
assert np.isclose(called, 0.0, atol=1e-3)
def test_just_input_variables(initial_input_values):
U = tq.gates.Rx('a', 0) + tq.gates.Rx('b', 1) + tq.gates.CNOT(
1, 3) + tq.gates.CNOT(0, 2) + tq.gates.CNOT(0, 1)
H1 = tq.paulis.Qm(1)
H2 = tq.paulis.Qm(2)
H3 = tq.paulis.Qm(3)
stackable = [
tq.ExpectationValue(U, H1),
tq.ExpectationValue(U, H2),
tq.ExpectationValue(U, H3)
]
stacked = tq.vectorize(stackable)
cargs = {'samples': None, 'backend': None, 'initial_values': None}
input_vars = ['a', 'b']
tensorflowed = tq.ml.to_platform(stacked,
platform='tensorflow',
compile_args=cargs,
input_vars=input_vars)
tensorflowed.set_input_values(initial_input_values)
learning_rate = .1
momentum = 0.9
optimizer = optimizers.SGD(lr=learning_rate, momentum=momentum)
# We train all trainable variables (which are just input_variables)
var_list_fn = lambda: tensorflowed.trainable_variables
loss = lambda: tf.reduce_sum(tensorflowed())
for i in range(STEPS):
optimizer.minimize(loss, var_list_fn)
called = tf.math.reduce_sum(tensorflowed()).numpy().tolist()
assert np.isclose(called, 0.0, atol=1e-3)
def test_calls_correctly(angles):
U1 = tq.gates.Rx(angle='a', target=0)
H1 = tq.paulis.Y(0)
U2 = tq.gates.Ry(angle='b', target=0)
H2 = tq.paulis.X(0)
U3 = tq.gates.H(0) + tq.gates.Rz(angle='c', target=0) + tq.gates.H(0)
H3 = tq.paulis.Y(0)
evals = [
tq.ExpectationValue(U1, H1),
tq.ExpectationValue(U2, H2),
tq.ExpectationValue(U3, H3)
]
stacked = tq.vectorize(evals)
torched = tq.ml.to_platform(stacked,
platform='torch',
input_vars=['a', 'b', 'c'])
inputs = torch.from_numpy(angles)
output = torched(inputs)
summed = output.sum().detach().numpy()
analytic = -np.sin(angles[0]) + np.sin(angles[1]) - np.sin(angles[2])
assert np.isclose(summed, analytic, atol=1.e-3)
def test_calls_correctly(initial_input_values):
U1 = tq.gates.Rx(angle='a', target=0)
H1 = tq.paulis.Y(0)
U2 = tq.gates.Ry(angle='b', target=0)
H2 = tq.paulis.X(0)
U3 = tq.gates.H(0) + tq.gates.Rz(angle='c', target=0) + tq.gates.H(0)
H3 = tq.paulis.Y(0)
evals = [
tq.ExpectationValue(U1, H1),
tq.ExpectationValue(U2, H2),
tq.ExpectationValue(U3, H3)
]
stacked = tq.vectorize(evals)
tensorflowed = tq.ml.to_platform(stacked,
platform='tensorflow',
input_vars=['a', 'b', 'c'])
tensorflowed.set_input_values(initial_input_values)
output = tensorflowed()
summed = tf.math.reduce_sum(output)
detached = tf.stop_gradient(summed).numpy()
analytic = -np.sin(initial_input_values["a"]) + np.sin(
initial_input_values["b"]) - np.sin(initial_input_values["c"])
assert np.isclose(detached, analytic, atol=1.e-3)
def test_different_fixed_variables_cases(initial_params_values,
initial_input_values, fixed):
U = tq.gates.Rx('c', 0) + tq.gates.Rx('d', 1) + tq.gates.Rx('a', 0) + tq.gates.Rx('b', 1) + tq.gates.CNOT(1, 3) \
+ tq.gates.CNOT(0, 2) + tq.gates.CNOT(0, 1)
H1 = tq.paulis.Qm(1)
H2 = tq.paulis.Qm(2)
H3 = tq.paulis.Qm(3)
stackable = [
tq.ExpectationValue(U, H1),
tq.ExpectationValue(U, H2),
tq.ExpectationValue(U, H3)
]
stacked = tq.vectorize(stackable)
cargs = {
'samples': None,
'backend': None,
'initial_values': initial_params_values
}
input_vars = ['c', 'd']
tensorflowed = tq.ml.to_platform(stacked,
platform='tensorflow',
compile_args=cargs,
input_vars=input_vars)
tensorflowed.set_input_values(initial_input_values)
learning_rate = .1
momentum = 0.9
optimizer = optimizers.SGD(lr=learning_rate, momentum=momentum)
if fixed == "inputs":
var_list_fn = lambda: tensorflowed.get_params_variable(
) # Since we just want to train the parameters
elif fixed == "params":
var_list_fn = lambda: tensorflowed.get_inputs_variable(
) # Since we just want to train the inputs
else:
var_list_fn = lambda: tensorflowed.trainable_variables # We train all trainable variables
loss = lambda: tf.reduce_sum(tensorflowed())
for i in range(STEPS):
optimizer.minimize(loss, var_list_fn)
called = tf.math.reduce_sum(tensorflowed()).numpy().tolist()
assert np.isclose(called, 0.0, atol=1e-3)
# Check that what should have stayed fixed does stay fixed
# NOTE: by chance, these assertions may fail because they happen to not need to change. Just run again to make sure
if fixed == "inputs":
final_input_values = tensorflowed.get_input_values()
for input_var_name in input_vars:
assert np.isclose(initial_input_values[input_var_name],
final_input_values[input_var_name],
atol=1e-3)
elif fixed == "params":
final_params_values = tensorflowed.get_params_values()
for param_var_name in tensorflowed.weight_vars:
assert np.isclose(initial_params_values[param_var_name],
final_params_values[param_var_name],
atol=1e-3)