尝试第二次向后浏览该图，但缓冲区已被释放

Question

目前，我正在尝试使用MNIST示例对最不可能的类方法进行编码。但是，当num_steps> 1时，此代码将不会运行。我找不到重复多次的方法。

'''python
def ll_attack(model, criterion, images, alpha, epsilon, num_steps=2):
    outputs = model(images)
    idx = torch.min(outputs,1)[1]
    idx = idx.detach_()

    for i in range(num_steps):
        ouputs = model(images)

        model.zero_grad()
        loss = criterion(outputs, idx).to(device)
        loss.backward()
        pertubed_images = images - alpha*images.grad.sign()

        out = torch.clamp(pertubed_images, -epsilon, epsilon).detach_()

    return out

correct = 0

total = 0
criterion = torch.nn.CrossEntropyLoss()
for images, labels in test_loader:
    images=images.to(device)
    labels=labels.to(device)

    images.requires_grad =True

    pertubed_images = ll_attack(new_net, criterion, images, alpha, epsilon)
    labels = labels.to(device)
    outputs = new_net(pertubed_images)

    adv_pred = torch.max(outputs.data, 1)[1]

    total += 64
    correct += (adv_pred == labels).sum()

print('Accuracy:{}'.format(int(correct)/total))`
'''

RuntimeError：尝试第二次向后浏览图形，但缓冲区已被释放。第一次回叫时，请指定keep_graph = True。