# Question of parameter-shift method in broadcasting **URL:** \ **Category:** PennyLane Help\ **Created:** [July 28, 2024, 3:36am UTC](https://discuss.pennylane.ai/t/question-of-parameter-shift-method-in-broadcasting/4940 "2024-07-28T03:36:34Z")\ **Posts on this page:** 6\ **Page:** 1
**Author:** ![Noah\_niu](https://yyz2.discourse-cdn.com/flex012/user_avatar/discuss.pennylane.ai/noah_niu/32/2709_2.png) [@Noah\_niu](https://discuss.pennylane.ai/u/Noah_niu)\ **Post date:** [July 28, 2024, 3:36am UTC](https://discuss.pennylane.ai/t/question-of-parameter-shift-method-in-broadcasting/4940/1 "2024-07-28T03:36:34Z")
Hello! If applicable, put your complete code example down below. Make sure that your code: - is 100% self-contained — someone can copy-paste exactly what is here and run it to reproduce the behaviour you are observing - includes comments I want to run a hybrid quantum CNN model. However if I use the parameter-shift method @qml.qnode(dev, diff\_method=“parameter-shift”), an error returns, that is ------ ‘NotImplementedError: Computing the gradient of broadcasted tapes with respect to the broadcasted parameters using the parameter-shift rule gradient transform is currently not supported. See #4462 for details.’ Is there any method to fix the error? Thank you! ```auto # Put code here ``` # Define the quantum circuit using PennyLane n\_qubits = 5 dev = qml.device(“default.qubit”, wires=n\_qubits) @qml.qnode(dev, diff\_method=“parameter-shift”) def qnode(inputs, weights): qml.AngleEmbedding(inputs, wires=range(n\_qubits)) qml.BasicEntanglerLayers(weights, wires=range(n\_qubits)) return [qml.expval(qml.PauliZ(wires=i)) for i in range(n\_qubits)] # Define the QLayer n\_layers = 1 weight\_shapes = {“weights”: (n\_layers, n\_qubits)} # Define a simple CNN architecture class Net(nn.Module): def **init** (self): super(Net, self). **init** () # Convolutional layer 1 with 1 input channels (for greyscale images), 16 output channels, and 5x5 kernel self.conv1 = nn.Conv2d(1, 16, 5, stride=1, padding=2) # Batch normalization after convolutional layer 1 self.bn1 = nn.BatchNorm2d(16) # Max pooling layer with a 2x2 window self.pool = nn.MaxPool2d(2, 2) # Convolutional layer 2 with 16 input channels (from the previous layer), 32 output channels, and 5x5 kernel self.conv2 = nn.Conv2d(16, 32, 5, stride=1, padding=2) # Batch normalization after convolutional layer 2 self.bn2 = nn.BatchNorm2d(32) # Quantum layer self.qlayer1 = qml.qnn.TorchLayer(qnode, weight\_shapes) self.qlayer2 = qml.qnn.TorchLayer(qnode, weight\_shapes) self.qlayer3 = qml.qnn.TorchLayer(qnode, weight\_shapes) self.qlayer4 = qml.qnn.TorchLayer(qnode, weight\_shapes) # Fully connected layers self.fc1 = nn.Linear(32 \* 7 \* 7, 120) self.fc2 = nn.Linear(120, 20) self.fc3 = nn.Linear(20, 10) ``` def forward(self, x): # Propagate the input through the CNN layers x = self.pool(F.relu(self.bn1(self.conv1(x)))) x = self.pool(F.relu(self.bn2(self.conv2(x)))) # Flatten the output from the convolutional layers x = x.view(-1, 32 * 7 * 7) # Pass the output to the quantum layer x = F.relu(self.fc1(x)) x = F.relu(self.fc2(x)) x_1, x_2, x_3, x_4 = torch.split(x, 5, dim=1) x_1 = self.qlayer1(x_1) x_2 = self.qlayer2(x_2) x_3 = self.qlayer3(x_3) x_4 = self.qlayer4(x_4) x = torch.cat([x_1, x_2, x_3, x_4], axis=1) x = self.fc3(x) return x ``` device = torch.device(“cuda” if torch.cuda.is\_available() else “cpu”) model = Net().to(device) criterion = nn.CrossEntropyLoss() optimizer = torch.optim.Adam(model.parameters(), lr=0.01, weight\_decay=0.01) from torchvision.transforms import ToTensor from torch.utils.data import DataLoader, Subset num\_train\_images = 1000 num\_test\_images = 200 # Download and load the MNIST dataset train\_data = datasets.MNIST( root=‘data’, train=True, transform=ToTensor(), download=True, ) test\_data = datasets.MNIST( root=‘data’, train=False, transform=ToTensor() ) # Create subsets of the data train\_subset = Subset(train\_data, range(num\_train\_images)) test\_subset = Subset(test\_data, range(num\_test\_images)) # Create data loaders for the subsets train\_loader = DataLoader(train\_subset, batch\_size=32, shuffle=True) test\_loader = DataLoader(test\_subset, batch\_size=32, shuffle=False) # Example: Print the number of images in the subsets print(f’Number of training images: {len(train\_loader.dataset)}‘) print(f’Number of test images: {len(test\_loader.dataset)}’) dataset = train\_data # Initialize your CNN model cnn = Net() # Define loss function and optimizer criterion = nn.CrossEntropyLoss() # Cross-entropy loss for classification optimizer = torch.optim.SGD(cnn.parameters(), lr=0.001, momentum=0.9) # Stochastic Gradient Descent optimizer # Split your data into training and validation sets train\_size = int(0.8 \* len(dataset)) train\_set, val\_set = torch.utils.data.random\_split(dataset, [train\_size, len(dataset) - train\_size]) train\_loader = torch.utils.data.DataLoader(train\_set, batch\_size=4, shuffle=True) #val\_loader = torch.utils.data.DataLoader(val\_set, batch\_size=4, shuffle=False) # Training loop num\_epochs = 10 for epoch in range(num\_epochs): ct = datetime.datetime.now() print(f"{epoch=}, {ct}") running\_loss = 0.0 for i, data in enumerate(train\_loader, 0): inputs, labels = data optimizer.zero\_grad() # Zero the parameter gradients to avoid accumulation outputs = cnn(inputs) # Forward pass loss = criterion(outputs, labels) # Compute the loss loss.backward() # Backpropagation optimizer.step() # Update the model parameters print(‘Finished Training’) If you want help with diagnosing an error, please put the full error message below: ```auto # Put full error message here ``` And, finally, make sure to include the versions of your packages. Specifically, show us the output of `qml.about()`. Name: PennyLane Version: 0.37.0 Summary: PennyLane is a cross-platform Python library for quantum computing, quantum machine learning, and quantum chemistry. Train a quantum computer the same way as a neural network. Home-page: [GitHub - PennyLaneAI/pennylane: PennyLane is a cross-platform Python library for quantum computing, quantum machine learning, and quantum chemistry. Built by researchers, for research.](https://github.com/PennyLaneAI/pennylane) Author: Author-email: License: Apache License 2.0 Location: d:\anaconda3\lib\site-packages Requires: appdirs, autograd, autoray, cachetools, networkx, numpy, packaging, pennylane-lightning, requests, rustworkx, scipy, semantic-version, toml, typing-extensions Required-by: PennyLane\_Lightning Platform info: Windows-10-10.0.22631-SP0 Python version: 3.10.9 Numpy version: 1.26.4 Scipy version: 1.10.0 Installed devices: - default.clifford (PennyLane-0.37.0) - default.gaussian (PennyLane-0.37.0) - default.mixed (PennyLane-0.37.0) - default.qubit (PennyLane-0.37.0) - default.qubit.autograd (PennyLane-0.37.0) - default.qubit.jax (PennyLane-0.37.0) - default.qubit.legacy (PennyLane-0.37.0) - default.qubit.tf (PennyLane-0.37.0) - default.qubit.torch (PennyLane-0.37.0) - default.qutrit (PennyLane-0.37.0) - default.qutrit.mixed (PennyLane-0.37.0) - default.tensor (PennyLane-0.37.0) - null.qubit (PennyLane-0.37.0) - lightning.qubit (PennyLane\_Lightning-0.37.0) ---
**Author:** ![CatalinaAlbornoz](https://yyz2.discourse-cdn.com/flex012/user_avatar/discuss.pennylane.ai/catalinaalbornoz/32/1196_2.png) [@CatalinaAlbornoz](https://discuss.pennylane.ai/u/CatalinaAlbornoz)\ **Post date:** [July 30, 2024, 10:14pm UTC](https://discuss.pennylane.ai/t/question-of-parameter-shift-method-in-broadcasting/4940/2 "2024-07-30T22:14:01Z")
Hi @Noah_niu , welcome to the Forum! My guess is that your code is a modified version of one of our demos. Is this right? If so, which demo did you use as your base? On the other hand, it looks like `diff_method=“parameter-shift”` is not supported for your workflow. Are you able to remove this? For all devices if you don’t set any diff\_method PennyLane will choose the best one for that device. In the case of default.qubit it will try to choose ‘backprop’. My guess is that this will fix your issues. Let me know if this works for you! ---
**Author:** ![Noah\_niu](https://yyz2.discourse-cdn.com/flex012/user_avatar/discuss.pennylane.ai/noah_niu/32/2709_2.png) [@Noah\_niu](https://discuss.pennylane.ai/u/Noah_niu)\ **Post date:** [July 30, 2024, 10:30pm UTC](https://discuss.pennylane.ai/t/question-of-parameter-shift-method-in-broadcasting/4940/3 "2024-07-30T22:30:11Z")
Yes, if I remove the diff\_method, the codes work. But my question is, is there any method to use the parameter-shift rule in broadcasting since backpropagation cannot be applied to real quantum devices? If I don’t use broadcasting but update the quantum parameter using the parameter-shift rule manually, the code could work but the training time was too long even for just one epoch. Can you give me some advice on this? I appreciate your help! ---
**Author:** ![CatalinaAlbornoz](https://yyz2.discourse-cdn.com/flex012/user_avatar/discuss.pennylane.ai/catalinaalbornoz/32/1196_2.png) [@CatalinaAlbornoz](https://discuss.pennylane.ai/u/CatalinaAlbornoz)\ **Post date:** [August 1, 2024, 6:28am UTC](https://discuss.pennylane.ai/t/question-of-parameter-shift-method-in-broadcasting/4940/4 "2024-08-01T06:28:37Z")
Hi @Noah_niu, I’m not sure whether there are other options. Let me look into it and get back to you on this. ---
**Author:** ![Noah\_niu](https://yyz2.discourse-cdn.com/flex012/user_avatar/discuss.pennylane.ai/noah_niu/32/2709_2.png) [@Noah\_niu](https://discuss.pennylane.ai/u/Noah_niu)\ **Post date:** [August 2, 2024, 8:33pm UTC](https://discuss.pennylane.ai/t/question-of-parameter-shift-method-in-broadcasting/4940/5 "2024-08-02T20:33:16Z")
Many thanks in advance! ---
**Author:** ![CatalinaAlbornoz](https://yyz2.discourse-cdn.com/flex012/user_avatar/discuss.pennylane.ai/catalinaalbornoz/32/1196_2.png) [@CatalinaAlbornoz](https://discuss.pennylane.ai/u/CatalinaAlbornoz)\ **Post date:** [August 6, 2024, 12:39am UTC](https://discuss.pennylane.ai/t/question-of-parameter-shift-method-in-broadcasting/4940/6 "2024-08-06T00:39:26Z")
Hi @Noah_niu ! Sorry it took me some time to look into this. It doesn’t seem like there’s any obvious solution on how to make this work with TorchLayer. Using finite-diff also doesn’t work. You need to avoid using TorchLayer completely and make a class that inherits from nn.Module. I made a working example based on the [TorchLayers demo](https://pennylane.ai/qml/demos/tutorial_qnn_module_torch/) and the [Transfer Learning demo](https://pennylane.ai/qml/demos/tutorial_quantum_transfer_learning/). See the code below: ```auto import torch import matplotlib.pyplot as plt import numpy as np from sklearn.datasets import make_moons # Set random seeds torch.manual_seed(42) np.random.seed(42) X, y = make_moons(n_samples=200, noise=0.1) y_ = torch.unsqueeze(torch.tensor(y), 1) # used for one-hot encoded labels y_hot = torch.scatter(torch.zeros((200, 2)), 1, y_, 1) c = ["#1f77b4" if y_ == 0 else "#ff7f0e" for y_ in y] # colours for each class plt.axis("off") plt.scatter(X[:, 0], X[:, 1], c=c) plt.show() import pennylane as qml import torch.nn as nn n_qubits = 2 n_layers = 6 dev = qml.device("default.qubit", wires=n_qubits) @qml.qnode(dev,diff_method='parameter-shift') def qnode(inputs, weights): qml.AngleEmbedding(inputs, wires=range(n_qubits)) qml.BasicEntanglerLayers(weights, wires=range(n_qubits)) return [qml.expval(qml.PauliZ(wires=i)) for i in range(n_qubits)] class HybridModel(nn.Module): def __init__ (self): super(). __init__ () self.clayer_1 = nn.Linear(2, n_qubits) self.q_params = nn.Parameter(torch.randn(n_layers , n_qubits)) self.clayer_2 = nn.Linear(n_qubits, 2) self.softmax = nn.Softmax(dim=1) def forward(self, inputs): q_in = self.clayer_1(inputs) # Apply the quantum circuit to each element of the batch and append to q_out q_out = torch.Tensor(0, n_qubits) for elem in q_in: q_out_elem = torch.hstack(qnode(elem, self.q_params)).float().unsqueeze(0) q_out = torch.cat((q_out, q_out_elem)) # return the two-dimensional prediction from the postprocessing layer x = self.clayer_2(q_out) return self.softmax(x) model = HybridModel() # Data and batches X = torch.tensor(X, requires_grad=True).float() y_hot = y_hot.float() batch_size = 5 batches = 200 // batch_size data_loader = torch.utils.data.DataLoader( list(zip(X, y_hot)), batch_size=batch_size, shuffle=True, drop_last=True ) # Optimization epochs = 6 loss = nn.L1Loss() opt = torch.optim.SGD(model.parameters(), lr=0.2) for epoch in range(epochs): running_loss = 0 for xs, ys in data_loader: opt.zero_grad() loss_evaluated = loss(model(xs), ys) loss_evaluated.backward() opt.step() running_loss += loss_evaluated avg_loss = running_loss / batches print("Average loss over epoch {}: {:.4f}".format(epoch + 1, avg_loss)) y_pred = model(X) predictions = torch.argmax(y_pred, axis=1).detach().numpy() correct = [1 if p == p_true else 0 for p, p_true in zip(predictions, y)] accuracy = sum(correct) / len(correct) print(f"Accuracy: {accuracy * 100}%") ``` It took me a while to make it but I hope it can help you to see how to update your code by using nn.Parameter instead of TorchLayer and breaking up the batches with `for elem in q_in:`. Let me know if you have any questions on the code specifics!