在PyTorch中训练神经网络时，损失总是“为”

Question

I assigned different weight_decay for the parameters, and the training loss and testing loss were all nan.

I printed the prediction_train,loss_train,running_loss_train,prediction_test,loss_test,and running_loss_test ,they were all nan.

And I have checked the data with numpy.any(numpy.isnan(dataset)) , it returned False .

If I use optimizer = torch.optim.Adam(wnn.parameters()) rather than assigning different weight_decay for the parameters, there would be no problem.

Could you please tell me how to fix it? Here are the codes, I defined the activation function by myself. Thank you:)

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 x

morlet=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 out
wnn=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=0
running_loss_test=0
for 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')

The output was:

Epoch 1 Train Loss:nan, Test Loss:nan

And there was only the true data on the plot, as the picture shows.

Answer 1

重量衰减将L2正则化应用于学习参数，快速浏览一下你的代码，你在这里使用a1权重作为语法x=(x-self.b1)/self.a1 ，权重衰减为.01，这可能导致消除那些a1权重中的一些为零，以及除零的结果是什么？