Hi, I am implementing QCNN based on pennylane. For flexibility, I choose the torch interface. The forward process seems going well. However, when I check the gradient of circuit parameters, I find some of them is None, which results in that these parameters will not get updated during optimization.
Here is my code example.
def qconv_kernel(phi, params):
'''
Implementation of quantum convolution kernel circuit
------------------------------------------
:param phi: image pixels, [n_quibits]
:param params: learnable convolutional weights, [n_layers * n_qubits]
'''
# quantum circuit implemention here
return qml.expval(qml.PauliZ(wires=0))
def qconv_torch(inputs, params, qnode, n_in_channels=1, kernel_size=[2, 2], stride=None, padding=False):
"""
Convolves the input image with many applications of the same quantum circuit.
------------------------------------------------------------------------
:param inputs: input image, [C, H, W]
:param params: learnable convolutional weights, [n_out_channels, n_layers * n_qubits]
:param qnode: QNode object
:param n_in_channels: input channels of previous layer
:param kernel_size: size of qconv kernel, [kernel_h, kernel_w]
:param stride: step of qconv, [stride_h, stride_w]
:param padding: whether padding, bool
return a new image, [C_out, H_out, W_out]
"""
for j in range(0, in_h-kernel_size[0], stride[0]):
for k in range(0, in_w-kernel_size[1], stride[1]):
channel_results = []
for c_in in range(inputs.shape[0]):
img_patch = []
for h in range(kernel_size[0]):
for w in range(kernel_size[1]):
img_patch.append(inputs[c_in, j+h, k+w])
out_pixel = qnode(torch.tensor(img_patch), params[i])
out[:, j//stride[0], k//stride[1]] = torch.tensor(out_pixel)
return out
class QCNN_torch(torch.nn.Module):
def __init__(self, cfg):
super().__init__()
n_quibit = sum(cfg.MODEL.QCNN.KERNEL_SIZE)
self.DIM = cfg.MODEL.QCNN.DIM
dev = qml.device(cfg.CIRCUIT.BACKEND, wires=n_quibit)
self.qnode = qml.QNode(qconv_kernel, dev, interface='torch')
self.qconv_param = []
for i in range(len(self.DIM)-1):
param_weight = np.random.uniform(0, 2*math.pi, shape1*shape2)
param_weight = np.reshape(param_weight, (shape1, shape2))
param_weight = torch.nn.Parameter(torch.tensor(param_weight, requires_grad=True))
self.register_parameter('layer'+str(i+1), param_weight)
self.qconv_param.append(param_weight)
# FC
n_feat = ((28//(2**(len(self.DIM)-1)))**2) * self.DIM[-1]
stdv = 1./n_feat
self.fc = np.random.uniform(-stdv, stdv, n_feat*10)
self.fc = np.reshape(self.fc, (n_feat, 10))
self.fc = torch.nn.Parameter(torch.FloatTensor(self.fc))
def forward(self, x):
out = []
for b in range(len(x)):
bx = x[b]
for i in range(len(self.DIM)-1):
bx = qconv_torch(bx, self.qconv_param[i], self.qnode, n_in_channels=self.DIM[i])
out.append(bx.flatten().unsqueeze(0) @ self.fc)
return torch.cat(out, dim=0)
if __name__=='__main__':
opt = get_opt()
cfg = get_config(opt.config_file)
loss_fn = torch.nn.CrossEntropyLoss()
qcnn = QCNN_torch(cfg)
optimizer = torch.optim.Adam(params=qcnn.parameters(), lr=0.1)
optimizer.zero_grad()
inputs = torch.rand(4, 1, 28, 28)
out = qcnn(inputs)
loss = loss_fn(out, torch.LongTensor([1, 2, 3, 4]))
optimizer.param_groups[0]['params'][0].retain_grad()
loss.backward()
optimizer.step()
The problem is that I can only get the gradient of qcnn.fc but not other parameters.
Any suggestions are helpful. Thank you in advance.
