Unsupported gradient configurations¶
Device jacobian¶
In order to use diff_method="device" for a QNode, the device passed into
the constructor of a QNode must have its "provides_jacobian" capability set to True
and must contain a method jacobian(circuits, **kwargs) that returns the gradients for
each quantum circuit. This is not implemented for the default.qubit device because
the device doesn’t provide such a jacobian method (instead allows backpropagation to work).
See the custom plugins page for more detail.
An exception is raised if this configuration is used:
def print_grad():
dev = qml.device('default.qubit', wires=1, shots=None)
@qml.qnode(dev, diff_method='device')
def circuit(x):
qml.RX(x[0], wires=0)
return qml.expval(qml.Z(wires=0))
x = np.array([0.1], requires_grad=True)
print(qml.grad(circuit)(x))
>>> print_grad()
Traceback (most recent call last):
...
File "C:\pennylane\pennylane\qnode.py", line 448, in _validate_device_method
raise qml.QuantumFunctionError(
pennylane.QuantumFunctionError: The default.qubit device does not provide a native method for computing the jacobian.
Backpropagation¶
The backpropagation algorithm is analytic by nature, and hence passing shots=None
is the only supported configuration when diff_method="backprop" is used. Though
it is possible to always use the analytic gradient even when shots>0 (as is the case
with adjoint differentiation, see next section), in the current state of the code this would
break other things.
Currently an exception is raised if this invalid configuration is used:
def print_grad():
dev = qml.device('default.qubit', wires=1, shots=100)
@qml.qnode(dev, diff_method='backprop')
def circuit(x):
qml.RX(x[0], wires=0)
return qml.expval(qml.Z(wires=0))
x = np.array([0.1], requires_grad=True)
print(qml.grad(circuit)(x))
>>> print_grad()
Traceback (most recent call last):
...
File "C:\pennylane\pennylane\qnode.py", line 375, in _validate_backprop_method
raise qml.QuantumFunctionError("Backpropagation is only supported when shots=None.")
pennylane.QuantumFunctionError: Backpropagation is only supported when shots=None.
Changing to shots=None allows computing the analytic gradient:
def print_grad():
dev = qml.device('default.qubit', wires=1, shots=None)
@qml.qnode(dev, diff_method='backprop')
def circuit(x):
qml.RX(x[0], wires=0)
return qml.expval(qml.Z(wires=0))
x = np.array([0.1], requires_grad=True)
print(qml.grad(circuit)(x))
>>> print_grad()
[-0.09983342]
Adjoint differentiation¶
PennyLane implements the adjoint differentiation method from 2009.02823, which only discusses the gradient of expectation values of observables.
In particular, the following code works as expected:
def print_grad():
dev = qml.device('default.qubit', wires=1, shots=None)
@qml.qnode(dev, diff_method='adjoint')
def circuit(x):
qml.RX(x[0], wires=0)
return qml.expval(qml.Z(wires=0))
x = np.array([0.1], requires_grad=True)
print(qml.grad(circuit)(x))
>>> print_grad()
[-0.09983342]
default.qubit can differentiate any other measurement process as long as it
is in the Z measurement basis. In this case, we recommend using the device-provided vjp
(device_vjp=True) for improved performance scaling. This algorithm works
best when the final cost function only has a scalar value.
lightning.qubit only supports expectation values.
@qml.qnode(qml.device('default.qubit'), diff_method="adjoint", device_vjp=True)
def circuit(x):
qml.IsingXX(x, wires=(0,1))
return qml.probs(wires=(0,1))
def cost(x):
probs = circuit(x)
target = np.array([0, 0, 0, 1])
return qml.math.norm(probs-target)
>>> qml.grad(cost)(qml.numpy.array(0.1))
-0.07059288589999416
Furthermore, the adjoint differentiation algorithm is analytic by nature. If the an execution
has shots>0, an error is raised:
def print_grad_ok():
dev = qml.device('default.qubit', wires=1, shots=100)
@qml.qnode(dev, diff_method='adjoint')
def circuit(x):
qml.RX(x[0], wires=0)
return qml.expval(qml.Z(wires=0))
x = np.array([0.1], requires_grad=True)
print(qml.grad(circuit)(x))
>>> print_grad_ok()
DeviceError: Finite shots are not supported with adjoint + default.qubit
State gradients¶
In general, the state of a quantum circuit will be complex-valued, so differentiating the state directly is not possible without the use of complex analysis. Though complex gradients can be implemented for most “simple” functions, this is not supported in Autograd but is done in the other three interfaces.
Instead, in Autograd, real scalar-valued post-processing should be performed on the output state to allow the auto-differentiation frameworks to backpropagate through them. For example, the following code uses a scalar cost function dependent on the output state:
def state_scalar_grad():
dev = qml.device('default.qubit', wires=1, shots=None)
@qml.qnode(dev, diff_method='backprop')
def circuit(x):
qml.RX(x[0], wires=0)
return qml.state()
def cost_fn(x):
out = circuit(x)
return np.abs(out[0])
x = np.array([0.1], requires_grad=True)
print(qml.grad(cost_fn)(x))
>>> state_scalar_grad()
[-0.02498958]
However, changing from differentiating the scalar cost to differentiating the state directly will fail with an error:
def state_vector_grad():
dev = qml.device('default.qubit', wires=1, shots=None)
@qml.qnode(dev, diff_method='backprop')
def circuit(x):
qml.RX(x[0], wires=0)
return qml.state()
x = np.array([0.1], requires_grad=True)
print(qml.jacobian(circuit)(x))
>>> state_vector_grad()
Traceback (most recent call last):
...
File "C:\Python38\lib\site-packages\numpy\core\fromnumeric.py", line 57, in _wrapfunc
return bound(*args, **kwds)
ValueError: cannot reshape array of size 4 into shape (2,1)
Using a different interface that supports complex differentiation will fix this error:
def state_vector_grad_jax():
dev = qml.device('default.qubit', wires=1, shots=None)
@qml.qnode(dev, interface='jax', diff_method='backprop')
def circuit(x):
qml.RX(x[0], wires=0)
return qml.state()
x = jnp.array([0.1], dtype=np.complex64)
print(jax.jacrev(circuit, holomorphic=True)(x))
def state_vector_grad_tf():
dev = qml.device('default.qubit', wires=1, shots=None)
@qml.qnode(dev, interface='tf', diff_method='backprop')
def circuit(x):
qml.RX(x[0], wires=0)
return qml.state()
x = tf.Variable([0.1], trainable=True, dtype=np.complex64)
with tf.GradientTape() as tape:
out = circuit(x)
print(tape.jacobian(out, [x]))
def state_vector_grad_torch():
dev = qml.device('default.qubit', wires=1, shots=None)
@qml.qnode(dev, interface='torch', diff_method='backprop')
def circuit(x):
qml.RX(x[0], wires=0)
return qml.state()
x = torch.tensor([0.1], requires_grad=True, dtype=torch.complex64)
print(torch.autograd.functional.jacobian(circuit, (x,)))
>>> state_vector_grad_jax()
[[-0.02498958+0.j ]
[ 0. -0.49937513j]]
>>> state_vector_grad_tf()
[<tf.Tensor: shape=(2, 1), dtype=complex64, numpy=
array([[-0.02498958+0.j ],
[-0. +0.49937513j]], dtype=complex64)>]
>>> state_vector_grad_torch()
(tensor([[-0.0250+0.0000j],
[ 0.0000+0.4994j]]),)
Sample gradients¶
In PennyLane, samples are drawn from the eigenvalues of an observable, or from the computational basis states if no observable is provided. This process is not differentiable in general, so no gradient flow backwards through the sampling is allowed.
Currently, attempting to compute the gradient in this scenario will not raise an error, but the results will be incorrect:
def sample_backward():
dev = qml.device('default.qubit', wires=1, shots=20)
@qml.qnode(dev)
def circuit(x):
qml.RX(x[0], wires=0)
return qml.sample(wires=0)
x = np.array([np.pi / 2])
print(qml.jacobian(circuit)(x))
>>> sample_backward()
[[0.5]
[0.5]
[0.5]
[0.5]
[0.5]
[0.5]
[0.5]
[0.5]
[0.5]
[0.5]
[0.5]
[0.5]
[0.5]
[0.5]
[0.5]
[0.5]
[0.5]
[0.5]
[0.5]
[0.5]]
The forward pass is supported and will work as expected:
def sample_forward():
dev = qml.device('default.qubit', wires=1, shots=20)
@qml.qnode(dev)
def circuit(x):
qml.RX(x[0], wires=0)
return qml.sample(wires=0)
x = np.array([np.pi / 2])
print(circuit(x))
>>> sample_forward()
[0 1 0 0 0 1 1 0 0 1 1 1 0 0 0 1 1 0 0 0]