TorchSharp: Version 0.99.6 training process is normal, 0.100.3 training process error

An interesting question!! Without using breakpoints, it crashes in 3 ephos, but with a large number of breakpoints, it can run at least 15 ephos.

I suspect it has something to do with GC.

@lindadamama , Can you prepare a simple example and dataset so that everyone can reproduce the problem. You currently provide too little information, it is difficult to analyze the problem.

Error message exited with code -1073741819. It should be that Torch2.0.1 occupies more memory than Torch1.13, causing a memory overflow, Another error is that some Tensors in dataloader were released due to unknown reasons during the training process of the training dataset,but there was no similar situation at 0.99.6 The main problem currently is the release of ResultMasks Var loss=criterion.call (result, item. ResultMasks); @GeorgeS2019 @NiklasGustafsson code of lost tensor ：

 foreach (var item in dataloader)
                {
                    using (var d = torch.NewDisposeScope())
                    {
                        var output = this.module.call(item.Images);
                        if (output is Tensor result)
                        {
                            var loss = criterion.call(result, item.ResultMasks);
                            optimizer.zero_grad();
                            loss.backward();
                            var opt = optimizer.step();
                            var loss_val = loss.to(CPU).item<float>();
                            lossArry.Add(loss_val);
                            result?.Dispose();
                        }
                    };
                }

Is there a problem with the code written above? Or does it need to be changed to the following notation?

   using (var d = torch.NewDisposeScope())
                {
                    using var dataset = new SegDataset(datadir, labelDir, numClass);
                    using var dataloader = new torch.utils.data.DataLoader<DataSetItem, BatchItem>(dataset, batchSize, CollateFn.Collate, doShuffle, device, num_worker: 4);
                    foreach (var item in dataloader)
                    {
                        var output = this.module.call(item.Images);
                        if (output is Tensor result)
                        {
                            var loss = criterion.call(result, item.ResultMasks);
                            optimizer.zero_grad();
                            loss.backward();
                            var opt = optimizer.step();
                            var loss_val = loss.to(CPU).item<float>();
                            lossArry.Add(loss_val);
                            result?.Dispose();
                        }

                    }
                };

I’m sorry,I have modified it