我想应用高斯核θ（z_i，z_j）= exp（γ|| z_i−z_j ||²）来接收PyTorch中的嵌入。该函数接收adj（2485x2485割炬张量）和emb（2485x64割炬.nn.parameter.Parameter）。目标是使用给定的内核在adj和emb之间进行二进制交叉熵损失。

def compute_loss_ber_exp1(adj,emb,b=0.1):
    #Init
    N,d=emb.shape
    gamma=0.001

    #get indices of upper triangular matrix
    ind=torch.triu_indices(N,N,offset=1)
    labels = adj[ind[0],ind[1]]

    #compute f(z_i,z_j) = exp(-gamma||z_i-z_j||^2)
    dist=F.pdist(emb,p=2)

    logits=torch.exp(-gamma * dist**2)

    #compute loss
    loss = F.binary_cross_entropy_with_logits(logits,labels,reduction='mean')
    return loss

代码应用于

max_epochs = 1000
display_step = 250
compute_loss = compute_loss_ber_exp1

for epoch in range(max_epochs):
    opt.zero_grad()
    loss = compute_loss(adj,b)
    loss.backward()
    opt.step()
    # Training loss is printed every display_step epochs
    if epoch == 0 or (epoch + 1) % display_step == 0:
        print(f'Epoch {epoch+1:4d},loss = {loss.item():.5f}')

返回错误

---------------------------------------------------------------------------
RuntimeError                              Traceback (most recent call last)
<ipython-input-10-1f961abb1fac> in <module>
      6     opt.zero_grad()
      7     loss = compute_loss(adj,b)
----> 8     loss.backward()
      9     opt.step()
     10     # Training loss is printed every display_step epochs

.../site-packages/torch/tensor.py in backward(self,gradient,retain_graph,create_graph)
    164                 products. Defaults to ``False``.
    165         """
--> 166         torch.autograd.backward(self,create_graph)
    167 
    168     def register_hook(self,hook):

.../site-packages/torch/autograd/__init__.py in backward(tensors,grad_tensors,create_graph,grad_variables)
     97     Variable._execution_engine.run_backward(
     98         tensors,---> 99         allow_unreachable=True)  # allow_unreachable flag
    100 
    101 

RuntimeError: CUDA error: invalid configuration argument

有人知道如何解决此问题吗？我们在此错误上花费了相当多的时间，并尝试了不同的方法来进行基于距离的相似性度量，但结果却很慢。 谢谢！