如何使用GPU训练模型？

[lingjiajie]

您好！

我是的代码是在案例“混合量子经典神经网络模型”的基础上进行修改的。修改的代码如下：

` x_train, y_train, x_test, y_test = data_select(1000, 100) model = Net().toGPU()

optimizer = Adam(model.parameters(), lr=0.005)

分类任务使用交叉熵函数

loss_func = CategoricalCrossEntropy()

训练次数

epochs = 10 train_loss_list = [] val_loss_list = [] train_acc_list =[]

val_acc_list = []

for epoch in range(1, epochs):

total_loss = []
model.train()
batch_size = 1
correct = 0
n_train = 0

for x, y in data_generator(x_train, y_train, batch_size=1, shuffle=True):

    x = x.reshape(-1, 1, 28, 28)
    x = QTensor(x)
    y = QTensor(y)
    x = x.to_gpu(DEV_GPU_0)
    y = y.to_gpu(DEV_GPU_0)
    optimizer.zero_grad()
    output = model(x)
    loss = loss_func(y, output)
    loss_np = np.array(loss.data)
    np_output = np.array(output.data, copy=False)
    mask = (np_output.argmax(1) == y.argmax(1))
    correct += np.sum(np.array(mask))
    n_train += batch_size
    loss.backward()
    optimizer._step()
    total_loss.append(loss_np)

train_loss_list.append(np.sum(total_loss) / len(total_loss))
train_acc_list.append(np.sum(correct) / n_train)
print("{:.0f} loss is : {:.10f}".format(epoch, train_loss_list[-1]))

`

但是出现下面的问题：

terminate called after throwing an instance of 'std::invalid_argument' what(): tensor --> numpy device not supported, toCPU() first. Aborted (core dumped)

请问这是哪里出了问题？谢谢！

[kevzos]

    loss = loss_func(y, output)
    loss_np = np.array(loss.data)
    np_output = np.array(output.data, copy=False)

loss 是QTensor变量。使用loss.to_numpy()转化为numpy @lingjiajie

[lingjiajie]

并不是您提的这里有问题，我想应该是有问题

模型部分的代码如下： `

class Net(Module):

def __init__(self):

    super(Net, self).__init__()
    self.conv1 = Conv2D(input_channels=1, output_channels=6, kernel_size=(5, 5), stride=(1, 1), padding="valid")
    self.maxpool1 = MaxPool2D([2, 2], [2, 2], padding="valid")
    self.conv2 = Conv2D(input_channels=6, output_channels=16, kernel_size=(5, 5), stride=(1, 1), padding="valid")
    self.maxpool2 = MaxPool2D([2, 2], [2, 2], padding="valid")
    self.fc1 = Linear(input_channels=256, output_channels=64)
    self.fc2 = Linear(input_channels=64, output_channels=1)
    self.hybrid = Hybrid(np.pi / 2)
    self.fc3 = Linear(input_channels=1, output_channels=2)

def forward(self, x):
    x = F.ReLu()(self.conv1(x))  # 1 6 24 24
    x = self.maxpool1(x)
    x = F.ReLu()(self.conv2(x))  # 1 16 8 8
    x = self.maxpool2(x)
    x = tensor.flatten(x, 1)   # 1 256
    x = F.ReLu()(self.fc1(x))  # 1 64
    x = self.fc2(x)    # 1 1
    x = self.hybrid(x)
    x = self.fc3(x)
    return x

model = Net().toGPU()

`

[kevzos]

@lingjiajie 可以直接附上你的完整代码么，我来看下。

[kevzos]

@lingjiajie 你说的很对，那个示例不能很简单的改为GPU版本。 我这里给出那个示例使用GPU的代码，这个代码只能支持batch=1。 hybird_gpu_b1.txt 也可以使用QuantumLayerV2接口，封装的更加好一些，可以支持多种device以及batchsize： hybird_gpu_qlayer.txt

并不是您提的这里有问题，我想应该是有问题

模型部分的代码如下： `

class Net(Module):

def __init__(self):

    super(Net, self).__init__()
    self.conv1 = Conv2D(input_channels=1, output_channels=6, kernel_size=(5, 5), stride=(1, 1), padding="valid")
    self.maxpool1 = MaxPool2D([2, 2], [2, 2], padding="valid")
    self.conv2 = Conv2D(input_channels=6, output_channels=16, kernel_size=(5, 5), stride=(1, 1), padding="valid")
    self.maxpool2 = MaxPool2D([2, 2], [2, 2], padding="valid")
    self.fc1 = Linear(input_channels=256, output_channels=64)
    self.fc2 = Linear(input_channels=64, output_channels=1)
    self.hybrid = Hybrid(np.pi / 2)
    self.fc3 = Linear(input_channels=1, output_channels=2)

def forward(self, x):
    x = F.ReLu()(self.conv1(x))  # 1 6 24 24
    x = self.maxpool1(x)
    x = F.ReLu()(self.conv2(x))  # 1 16 8 8
    x = self.maxpool2(x)
    x = tensor.flatten(x, 1)   # 1 256
    x = F.ReLu()(self.fc1(x))  # 1 64
    x = self.fc2(x)    # 1 1
    x = self.hybrid(x)
    x = self.fc3(x)
    return x

model = Net().toGPU()

`

[lingjiajie]

谢谢，十分感谢。

您的回复可以解决我的问题。

再次感谢您的帮助！

[kevzos]

@lingjiajie 如果解决了您的问题，请close这个issue。