Hi, when I implement a quantum classifier using pennylane, it reported error: `ValueError: Cannot differentiate with respect to parameter`

. This is the example code:

```
dev = qml.device('qiskit.aer', wires=6, shots=10)
def cost(param, feat, label, model):
loss = []
for m in range(len(feat)):
predict = model(param, data=feat[m])
loss.append((label[m] - predict)**2)
return np.mean(np.array(loss))
# @qml.qnode(dev, interface='torch')
def circuit3(weights, data=None):
qml.templates.AmplitudeEmbedding(features=data, wires=range(6), normalize=True)
n_layers, n_qubits = weights.shape[0], weights.shape[1]
for i in range(n_layers):
for j in range(n_qubits):
qml.Rot(weights[i, j, 0], weights[i, j, 1], weights[i, j, 2], wires=j)
CNOT_layer(n_qubits)
return qml.expval(qml.PauliZ(0))
weights = np.random.random([4, 6, 3])
data = np.random.random([2, 64], requires_grad=False)
label = np.array([1, -1], requires_grad=False)
qnode = qml.QNode(circuit3, dev)
optimizer = GradientDescentOptimizer(1e-2)
weights = optimizer.step(lambda v: cost(v, data, label, qnode), weights)
```

I notice that function `AmplitudeEmbedding`

doesn’t support gradient calculation, does it cause the error? But I don’t need to calculate the gradient with respect to input feature. I am quite confused.

Any help and suggestions are welcome.

Thank you in advance!