Fix warnings when converting tensors with requires_grad=True to scalars.

CHANGELOG.md

@@ -21,6 +21,7 @@ The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/).
 
 ### Fixed
 
+- Fixed Pytorch warnings in several examples ([#10571](https://github.com/pyg-team/pytorch_geometric/pull/10571)))
 - Fixed `ogbn_train_cugraph` example for distributed cuGraph ([#10439](https://github.com/pyg-team/pytorch_geometric/pull/10439))
 - Added `safe_onnx_export` function with workarounds for `onnx_ir.serde.SerdeError` issues in ONNX export ([#10422](https://github.com/pyg-team/pytorch_geometric/pull/10422))
 - Fixed importing PyTorch Lightning in `torch_geometric.graphgym` and `torch_geometric.data.lightning` when using `lightning` instead of `pytorch-lightning` ([#10404](https://github.com/pyg-team/pytorch_geometric/pull/10404), [#10417](https://github.com/pyg-team/pytorch_geometric/pull/10417)))

examples/attentive_fp.py

@@ -125,7 +125,7 @@ def train():
         loss = F.mse_loss(out, data.y)
         loss.backward()
         optimizer.step()
-        total_loss += float(loss) * data.num_graphs
+        total_loss += float(loss.detach()) * data.num_graphs
         total_examples += data.num_graphs
     return sqrt(total_loss / total_examples)
 

examples/autoencoder.py

@@ -100,7 +100,7 @@ def train():
         loss = loss + (1 / train_data.num_nodes) * model.kl_loss()
     loss.backward()
     optimizer.step()
-    return float(loss)
+    return float(loss.detach())
 
 
 @torch.no_grad()

examples/dir_gnn.py

@@ -71,7 +71,7 @@ def train():
     loss = F.cross_entropy(out[data.train_mask], data.y[data.train_mask])
     loss.backward()
     optimizer.step()
-    return float(loss)
+    return float(loss.detach())
 
 
 @torch.no_grad()

examples/gcn2_cora.py

@@ -65,7 +65,7 @@ def train():
     loss = F.nll_loss(out[data.train_mask], data.y[data.train_mask])
     loss.backward()
     optimizer.step()
-    return float(loss)
+    return float(loss.detach())
 
 
 @torch.no_grad()

examples/glnn.py

@@ -44,7 +44,7 @@ def train_teacher():
     loss = F.cross_entropy(out[data.train_mask], data.y[data.train_mask])
     loss.backward()
     gnn_optimizer.step()
-    return float(loss)
+    return float(loss.detach())
 
 
 @torch.no_grad()
@@ -84,7 +84,7 @@ def train_student():
     loss = args.lamb * loss1 + (1 - args.lamb) * loss2
     loss.backward()
     mlp_optimizer.step()
-    return float(loss)
+    return float(loss.detach())
 
 
 @torch.no_grad()

examples/mixhop.py

@@ -64,7 +64,7 @@ def train():
     loss.backward()
     optimizer.step()
     scheduler.step()
-    return float(loss)
+    return float(loss.detach())
 
 
 @torch.no_grad()

examples/ogbn_proteins_deepgcn.py

@@ -86,7 +86,7 @@ def train(epoch):
         loss.backward()
         optimizer.step()
 
-        total_loss += float(loss) * int(data.train_mask.sum())
+        total_loss += float(loss.detach()) * int(data.train_mask.sum())
         total_examples += int(data.train_mask.sum())
 
         pbar.update(1)

examples/ogbn_train.py

@@ -147,7 +147,7 @@ def train(epoch: int) -> tuple[Tensor, float]:
         loss.backward()
         optimizer.step()
 
-        total_loss += float(loss)
+        total_loss += float(loss.detach())
         total_correct += int(out.argmax(dim=-1).eq(y).sum())
         pbar.update(batch.batch_size)
 

examples/pmlp.py

@@ -32,7 +32,7 @@ def train():
     loss = F.cross_entropy(out[data.train_mask], data.y[data.train_mask])
     loss.backward()
     optimizer.step()
-    return float(loss)
+    return float(loss.detach())
 
 
 @torch.no_grad()

examples/proteins_mincut_pool.py

@@ -77,7 +77,7 @@ def train(epoch):
         out, mc_loss, o_loss = model(data.x, data.edge_index, data.batch)
         loss = F.nll_loss(out, data.y.view(-1)) + mc_loss + o_loss
         loss.backward()
-        loss_all += data.y.size(0) * float(loss)
+        loss_all += data.y.size(0) * float(loss.detach())
         optimizer.step()
     return loss_all / len(train_dataset)
 
@@ -92,7 +92,7 @@ def test(loader):
         data = data.to(device)
         pred, mc_loss, o_loss = model(data.x, data.edge_index, data.batch)
         loss = F.nll_loss(pred, data.y.view(-1)) + mc_loss + o_loss
-        loss_all += data.y.size(0) * float(loss)
+        loss_all += data.y.size(0) * float(loss.detach())
         correct += int(pred.max(dim=1)[1].eq(data.y.view(-1)).sum())
 
     return loss_all / len(loader.dataset), correct / len(loader.dataset)

examples/rgcn.py

@@ -72,7 +72,7 @@ def train():
     loss = F.nll_loss(out[data.train_idx], data.train_y)
     loss.backward()
     optimizer.step()
-    return float(loss)
+    return float(loss.detach())
 
 
 @torch.no_grad()

examples/seal_link_pred.py

@@ -198,7 +198,7 @@ def train():
         loss = criterion(out.view(-1), data.y.to(torch.float))
         loss.backward()
         optimizer.step()
-        total_loss += float(loss) * data.num_graphs
+        total_loss += float(loss.detach()) * data.num_graphs
 
     return total_loss / len(train_dataset)
 

examples/tgn.py

@@ -147,7 +147,7 @@ def train():
         loss.backward()
         optimizer.step()
         memory.detach()
-        total_loss += float(loss) * batch.num_events
+        total_loss += float(loss.detach()) * batch.num_events
 
     return total_loss / train_data.num_events