Training Hybrid NN with Torch Layer

Hi, I’m trying to implement a hybrid NN model that takes two inputs x1 and x2. The two inputs goes through a same classical layer, and the two outputs are fed into quantum layer.

def qnode(inputs):
    qml.U3(inputs[0], inputs[1], inputs[2], wires=0)
    qml.U3(inputs[3], inputs[4], inputs[5], wires=1)
    return qml.probs(wires=[0,1])

class HybridModel(torch.nn.Module):
    def __init__(self):
        self.clayer = torch.nn.Linear(2, 3)
        self.qlayer = qml.qnn.TorchLayer(qnode, weight_shapes={})
    def forward(self, x1, x2):
        x1 = self.clayer(x1)
        x2 = self.clayer(x2)
        x = torch.concat([x1, x2])
        x = self.qlayer(x)
        return x[0]
model = HybridModel()

When training the model, the parameters don’t change at all.

data_loader =
    list(zip(X1, X2, y)), batch_size=20, shuffle=True, drop_last=True

opt = torch.optim.SGD(model.parameters(), lr=0.2)
loss_fn = torch.nn.MSELoss()
epochs = 5

for epoch in range(epochs):

  running_loss = 0

  for x1s, x2s, ys in data_loader:
    preds = [model(x1, x2) for x1, x2 in zip(x1s, x2s)]
    preds = torch.Tensor(preds)
    loss_evaluated = loss_fn(preds, ys)
    loss_evaluated.requires_grad = True

    running_loss += loss_evaluated

  avg_loss = running_loss / 2000
  print("avg loss: " + str(avg_loss))

The code outputs,

Parameter containing:
tensor([[ 0.6520, -0.0508],
[ 0.5716, -0.5352],
[-0.7030, 0.2188]], requires_grad=True)
Parameter containing:
tensor([[ 0.6520, -0.0508],
[ 0.5716, -0.5352],
[-0.7030, 0.2188]], requires_grad=True)
Parameter containing:
tensor([[ 0.6520, -0.0508],
[ 0.5716, -0.5352],
[-0.7030, 0.2188]], requires_grad=True)

without parameters changing.

Here, I only have trainable parameters with classical layer. So, I declared “weight_shapes={}” when making a quantum layer. I am not sure if this is a problem or not.

What am I doing wrong here? Here is the minimal code example. Sorry I couldn’t attach the file as I am a new user.

Hi @takh2324, welcome to the Forum!

I think the problem is in how you’re modifying the data. You can take a look at the example here and modify it so that the trainable parameters are only on the classical layers.

You can download the example notebook at the end of the page, and then change:

  1. def qnode(inputs, weights) to def qnode(inputs)
  2. Comment the two lines within the qnode and add instead qml.U3(inputs[0], inputs[1], inputs[0], wires=0)
  3. Change weight_shapes = {"weights": (n_layers, n_qubits)} to weight_shapes = {}

With these changes you will notice that the code still trains but only on the classical part.

I hope this helps you find the solution to your problem! Please let me know if you have any further questions.

Thanks for the reply @CatalinaAlbornoz . I have already seen the example, but I’m afraid my hybrid NN model is bit different from the notebook. The example takes 1 input, feeds into 1 classical layer, then feeds into 2 different quantum layers . But I want 2 inputs, feed into 1 same classical layer which will give two outputs, then feed two outputs into 1 qlayer together. I have tried different approaches but none of them works for me.

Hi @takh2324, if you share a minimum example of your problem I may be able to understand it better and look deeper into a possible solution.