Open
Description
Both first order derivative of tfq layers and higher order gradient of non tfq layers work with GradientTape as follows:
x = tf.Variable(initial_value=0.2)
with tf.GradientTape() as t:
with tf.GradientTape() as t2:
y = x**3
g = t2.gradient(y, x)
print(g)
g2 = t.gradient(g, x)
print(g2)
##########
a = sy.Symbol("a")
c = cirq.Circuit()
c.append(cirq.rx(a)(cirq.GridQubit(0,0)))
model = tfq.layers.PQC(c, operators=[cirq.Z(cirq.GridQubit(0,0))])
with tf.GradientTape() as t:
o = model(tfq.convert_to_tensor([cirq.Circuit()]))[0,0]
g = t.gradient(o, model.variables)
print(g)However the following doesn't work, with tensorflow error as
InvalidArgumentError: Operation 'cond' has no attr named '_XlaCompile'. ValueError: Insufficient elements in branch_graphs[0].outputs. Expected: 6 Actual: 5
a = sy.Symbol("a")
c = cirq.Circuit()
c.append(cirq.rx(a)(cirq.GridQubit(0,0)))
model = tfq.layers.PQC(c, operators=[cirq.Z(cirq.GridQubit(0,0))])
with tf.GradientTape() as t:
with tf.GradientTape() as t2:
o = model(tfq.convert_to_tensor([cirq.Circuit()]))[0,0]
g = t2.gradient(o, model.variables)
print(g)
g2 = t.gradient(g, model.variables)
print(g2)Is my code wrong or tfq layers have special issues in terms of higher order gradients due to the way AD is implemented in these layers?