我为参数分配了不同的weight_decay，结果training loss和testing loss都变成了nan。
我打印了prediction_train,loss_train,running_loss_train,prediction_test,loss_test,and running_loss_test，它们都是nan。
我使用numpy.any(numpy.isnan(dataset))检查了数据，它返回了False。
如果我使用optimizer = torch.optim.Adam(wnn.parameters())而不是为参数分配不同的weight_decay，就不会有问题。
您能告诉我如何修复这个问题吗？这是我的代码，我自己定义了激活函数。谢谢您:)
class Morlet(nn.Module):def __init__(self): super(Morlet,self).__init__()def forward(self,x): x=(torch.cos(1.75*x))*(torch.exp(-0.5*x*x)) return xmorlet=Morlet()class WNN(nn.Module):def __init__(self): super(WNN,self).__init__() self.a1=torch.nn.Parameter(torch.randn(64,requires_grad=True)) self.b1=torch.nn.Parameter(torch.randn(64,requires_grad=True)) self.layer1=nn.Linear(30,64,bias=False) self.out=nn.Linear(64,1)def forward(self,x): x=self.layer1(x) x=(x-self.b1)/self.a1 x=morlet(x) out=self.out(x) return outwnn=WNN()optimizer = torch.optim.Adam([{'params': wnn.layer1.weight, 'weight_decay':0.01}, {'params': wnn.out.weight, 'weight_decay':0.01}, {'params': wnn.out.bias, 'weight_decay':0}, {'params': wnn.a1, 'weight_decay':0.01}, {'params': wnn.b1, 'weight_decay':0.01}])criterion = nn.MSELoss()for epoch in range(10):prediction_test_list=[]running_loss_train=0running_loss_test=0for i,(x1,y1) in enumerate(trainloader): prediction_train=wnn(x1) #print(prediction_train) loss_train=criterion(prediction_train,y1) #print(loss_train) optimizer.zero_grad() loss_train.backward() optimizer.step() running_loss_train+=loss_train.item() #print(running_loss_train)tr_loss=running_loss_train/train_set_y_array.shape[0]for i,(x2,y2) in enumerate(testloader): prediction_test=wnn(x2) #print(prediction_test) loss_test=criterion(prediction_test,y2) #print(loss_test) running_loss_test+=loss_test.item() print(running_loss_test) prediction_test_list.append(prediction_test.detach().cpu())ts_loss=running_loss_test/test_set_y_array.shape[0]print('Epoch {} Train Loss:{}, Test Loss:{}'.format(epoch+1,tr_loss,ts_loss)) test_set_y_array_plot=test_set_y_array*(dataset.max()-dataset.min())+dataset.min()prediction_test_np=torch.cat(prediction_test_list).numpy()prediction_test_plot=prediction_test_np*(dataset.max()-dataset.min())+dataset.min()plt.plot(test_set_y_array_plot.flatten(),'r-',linewidth=0.5,label='True data')plt.plot(prediction_test_plot,'b-',linewidth=0.5,label='Predicted data')plt.legend()plt.show()print('Finish training')输出结果是:
Epoch 1 Train Loss:nan, Test Loss:nan图表上只有真实数据，如图所示。
回答：
权重衰减对学习参数应用L2正则化，快速浏览您的代码，您在这里使用a1权重作为分母x=(x-self.b1)/self.a1，并且设置了0.01的权重衰减，这可能导致一些a1权重变为零，除以零会导致什么结果呢？
