pennylane（pennylane 搜索）

PennyLane - Quantum Machine Learning

PennyLane treats quantum computers like neural network layers. It allows for the calculation of gradients of quantum circuits (using the parameter-shift rule or backpropagation), enabling the optimization of hybrid classical-quantum models.

When to Use

• Developing and training Quantum Neural Networks (QNNs)
• Variational Quantum Algorithms (VQE, QAOA)
• Hybrid classical-quantum machine learning (e.g., Quantum CNNs)
• Quantum chemistry calculations in a differentiable framework
• Benchmarking quantum algorithms across different hardware (IBM, Rigetti, Xanadu, IonQ)
• Optimizing quantum control pulses
• Investigating Barren Plateaus and other QML-specific phenomena

Reference Documentation

Official docs: https://docs.pennylane.ai/ Demos/Tutorials: https://pennylane.ai/qml/demonstrations.html Search patterns: qml.qnode, qml.device, qml.expval, qml.grad, qml.templates

Core Principles

The QNode (Quantum Node)

A QNode is a quantum circuit bound to a device, which can be called like a standard Python function. It is the fundamental unit that PennyLane can differentiate.

Hardware Agnosticism

PennyLane provides a unified interface. The same code can run on a high-performance simulator (default.qubit), a GPU-accelerated backend (lightning.qubit), or real quantum hardware.

Automatic Differentiation

Quantum circuits in PennyLane are "aware" of their gradients. You can use standard optimizers (Adam, SGD) to tune rotation angles in the circuit.

Quick Reference

Installation

pip install pennylane
# For GPU support
pip install pennylane-lightning[gpu]

Standard Imports

import pennylane as qml
from pennylane import numpy as np # Use PennyLane's wrapped NumPy for gradients

Basic Pattern - Differentiable Circuit

import pennylane as qml
from pennylane import numpy as np

# 1. Define Device
dev = qml.device("default.qubit", wires=2)

# 2. Define QNode (Quantum Node)
@qml.qnode(dev)
def circuit(params):
    qml.RX(params[0], wires=0)
    qml.RY(params[1], wires=1)
    qml.CNOT(wires=[0, 1])
    return qml.expval(qml.PauliZ(1))

# 3. Calculate Gradient
params = np.array([0.1, 0.2], requires_grad=True)
grad_fn = qml.grad(circuit)
print(f"Gradient: {grad_fn(params)}")

Critical Rules

✅ DO

• Use qml.numpy - Always use the PennyLane-wrapped NumPy for parameters you want to differentiate
• Use Templates - Instead of building layers manually, use qml.templates (like StronglyEntanglingLayers) for efficient QML architectures
• Set requires_grad - Explicitly mark trainable parameters with requires_grad=True
• Prefer lightning.qubit - For simulations with >15 qubits, use the lightning device for significantly better performance
• Use Broadcasting - Many gates support broadcasting over input arrays, which is faster than loops
• Batch Circuits - Use qml.batch_input or qml.map for processing multiple inputs simultaneously

❌ DON'T

• Mix NumPy versions - Using standard import numpy as np for trainable parameters will break the autograd engine
• Overcomplicate Ansätze - Start with simple circuits to avoid Barren Plateaus (where gradients vanish)
• Use too many qubits on simulators - Memory scales as 2^N. 30 qubits require ~16GB of RAM for a single statevector
• Hardcode Wire Indices - Use variables for wires to make your code reusable and scalable

Anti-Patterns (NEVER)

import pennylane as qml
# ❌ BAD: Standard numpy for trainable parameters
import numpy as np

# ✅ GOOD: PennyLane's wrapped numpy
from pennylane import numpy as np

# ❌ BAD: Creating a device inside a loop
for i in range(100):
    dev = qml.device("default.qubit", wires=2) # Expensive initialization

# ✅ GOOD: Define device once
dev = qml.device("default.qubit", wires=2)

# ❌ BAD: Manual parameter shifting
# (shift = 0.5 * pi, calc f(x+s) - f(x-s)...)
# ✅ GOOD: Let PennyLane handle it automatically
grad = qml.grad(circuit)(params)

Circuit Templates (qml.templates)

Built-in QML Layers

import pennylane as qml

@qml.qnode(dev)
def qnn_layer(inputs, weights):
    # Encoding classical data into quantum state
    qml.AngleEmbedding(inputs, wires=range(n_qubits))

    # Trainable entangling layer
    qml.StronglyEntanglingLayers(weights, wires=range(n_qubits))

    return qml.expval(qml.PauliZ(0))

Hybrid Optimization (PyTorch/TF)

Integration with Classical Frameworks

import torch
import pennylane as qml

dev = qml.device("default.qubit", wires=2)

@qml.qnode(dev, interface="torch")
def quantum_layer(phi, theta):
    qml.RX(phi, wires=0)
    qml.RZ(theta, wires=1)
    return qml.expval(qml.PauliZ(0))

# Now this QNode can be used in a Torch Module
phi = torch.tensor(0.1, requires_grad=True)
theta = torch.tensor(0.2, requires_grad=True)
result = quantum_layer(phi, theta)
result.backward()
print(phi.grad)

Quantum Chemistry (qml.qchem)

VQE Workflow

from pennylane import qchem

# 1. Define molecule
symbols = ["H", "H"]
coordinates = np.array([0.0, 0.0, -0.6614, 0.0, 0.0, 0.6614])

# 2. Build Hamiltonian
H, qubits = qchem.molecular_hamiltonian(symbols, coordinates)

# 3. Define VQE circuit
dev = qml.device("default.qubit", wires=qubits)
@qml.qnode(dev)
def vqe_circuit(params):
    # Simple Hartree-Fock state + rotations
    qml.BasisState(np.array([1, 1, 0, 0]), wires=range(qubits))
    qml.DoubleExcitation(params[0], wires=[0, 1, 2, 3])
    return qml.expval(H)

Measurements and Observables

Beyond Expectation Values

@qml.qnode(dev)
def multi_measure(params):
    qml.RX(params[0], wires=0)
    return (
        qml.expval(qml.PauliZ(0)), # Expectation value
        qml.var(qml.PauliZ(0)),    # Variance
        qml.probs(wires=[0, 1]),   # Probabilities
        qml.sample(qml.PauliX(1))  # Raw samples (shots)
    )

Practical Workflows

1. Quantum Variational Classifier

def variational_classifier(weights, bias, x):
    """A simple QNN classifier."""
    return circuit(weights, x) + bias

def cost(weights, bias, X, Y):
    predictions = [variational_classifier(weights, bias, x) for x in X]
    return square_loss(Y, predictions)

# Optimizer
opt = qml.AdamOptimizer(stepsize=0.1)
# ... loop with opt.step(cost, ...)

2. Computing Gradients on Hardware (Parameter-Shift)

# When running on real hardware, PennyLane automatically uses
# the parameter-shift rule to calculate gradients via multiple executions.
dev_remote = qml.device("braket.aws.qubit", device_arn="...", wires=2)

@qml.qnode(dev_remote, diff_method="parameter-shift")
def hardware_circuit(params):
    qml.RY(params[0], wires=0)
    return qml.expval(qml.PauliZ(0))

3. Noise Simulation

dev_noisy = qml.device("default.mixed", wires=2)

@qml.qnode(dev_noisy)
def noisy_circuit(p):
    qml.Hadamard(wires=0)
    # Apply depolarizing noise to wire 0
    qml.DepolarizingChannel(p, wires=0)
    return qml.state()

Performance Optimization

JAX Integration for Speed

JAX is often the fastest interface for large-scale simulations due to its Just-In-Time (JIT) compilation.

import jax

@qml.qnode(dev, interface="jax")
def fast_circuit(x):
    qml.RX(x, wires=0)
    return qml.expval(qml.PauliZ(0))

jit_circuit = jax.jit(fast_circuit)

Adjoint Differentiation

For large circuits, diff_method="adjoint" is much more memory-efficient than backpropagation.

dev = qml.device("lightning.qubit", wires=20)
@qml.qnode(dev, diff_method="adjoint")
def large_circuit(params):
    ...

Common Pitfalls and Solutions

The "Vanishing Gradient" (Barren Plateaus)

As the number of qubits and layers increases, the gradient often becomes exponentially small.

# ✅ Solution:
# 1. Use better initialization for weights.
# 2. Use local observables (PauliZ(i)) instead of global ones.
# 3. Use identity-block initialization.

Non-Trainable Inputs

Sometimes you want to pass data (like images) that shouldn't be optimized.

# ✅ Solution: Use the 'argnum' in qml.grad or use non-array types
# Or explicitly:
params = np.array(0.1, requires_grad=True)
data = np.array(0.5, requires_grad=False)

Output Shape Mismatch

qml.probs returns an array of size 2^N.

# ❌ Problem: Using probs(wires=[0,1,2]) in a loss function expecting a scalar.
# ✅ Solution: Use expval() for a single scalar or handle the distribution.

PennyLane bridges the gap between quantum physics and artificial intelligence. By making quantum circuits differentiable, it transforms them into powerful, trainable tools for the next generation of scientific computing.