🐛 Bug

torch::nll_loss2d memory leak

When I use classification, there is no memory leak using nll_loss,But when I train with semantics segmentation, memory leaks occur when I use nll_loss2d.

To Reproduce

Steps to reproduce the behavior:

1.semantics segmentation
2.torch::nll_loss2d
3.cpp frontend

model_->train();				
auto data = batch.data.to(d), targets = batch.target.to(d);	
optimizer_->zero_grad();
auto prediction = model_->forward(data);	
loss = torch::nll_loss2d(torch::log_softmax(prediction, 1), targets);
loss.backward();//If I comment on this line, there will be no memory leak
optimizer_->step();
float lossIn = loss.item<float>();
lr = optimizer_->options.learning_rate();

• PyTorch Version (e.g., 1.0):1.1
• OS (e.g., Linux):windows 10
• How you installed PyTorch (conda, pip, source):libtorch
• Build command you used (if compiling from source):
• Python version:
• CUDA/cuDNN version:cu100
• GPU models and configuration:cuda cudnn
• Any other relevant information: