浅谈pytorch grad_fn以及权重梯度不更新的问题

train_pred = Variable(train_pred.float(), requires_grad=True)`

  train_pred = model(data)
  train_pred = torch.max(train_pred, 1)[1].data.squeeze()
  train_pred = Variable(train_pred.float(), requires_grad=False)
  train_loss = F.binary_cross_entropy(validation_pred.float(), target)
  train_loss.backward()

 for batch_idx, (data, target) in enumerate(train_loader):
  # Get Samples
  label = target.view(target.size(0), 1).long()
  target_onehot = torch.zeros(data.shape[0], args.num_classes).scatter_(1, label, 1)
  data, target_onehot = Variable(data.cuda()), Variable(target_onehot.cuda().float())
  
  model.zero_grad()

  # Predict
  train_pred = model(data)
  train_loss = F.binary_cross_entropy(train_pred, target_onehot)
  train_loss.backward()
  optimizer.step()