pyg-nightly 2.7.0.dev20250904__py3-none-any.whl → 2.7.0.dev20250905__py3-none-any.whl

{pyg_nightly-2.7.0.dev20250904.dist-info → pyg_nightly-2.7.0.dev20250905.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: pyg-nightly
-Version: 2.7.0.dev20250904
+Version: 2.7.0.dev20250905
 Summary: Graph Neural Network Library for PyTorch
 Keywords: deep-learning,pytorch,geometric-deep-learning,graph-neural-networks,graph-convolutional-networks
 Author-email: Matthias Fey <matthias@pyg.org>

{pyg_nightly-2.7.0.dev20250904.dist-info → pyg_nightly-2.7.0.dev20250905.dist-info}/RECORD

@@ -1,4 +1,4 @@
-torch_geometric/__init__.py,sha256=kAQ6fnK2P5phSQX95Xx4qdMo6r6k-i2p8cDLRERlxuw,2292
+torch_geometric/__init__.py,sha256=JaI2udwlYGKK9_OGrMEQn05CBfOSEPFFsk9OtBnPN2c,2292
 torch_geometric/_compile.py,sha256=9yqMTBKatZPr40WavJz9FjNi7pQj8YZAZOyZmmRGXgc,1351
 torch_geometric/_onnx.py,sha256=ODB_8cwFUiwBUjngXn6-K5HHb7IDul7DDXuuGX7vj_0,8178
 torch_geometric/backend.py,sha256=lVaf7aLoVaB3M-UcByUJ1G4T4FOK6LXAg0CF4W3E8jo,1575
@@ -294,7 +294,7 @@ torch_geometric/loader/temporal_dataloader.py,sha256=Z7L_rYdl6SYBQXAgtr18FVcmfMH
 torch_geometric/loader/utils.py,sha256=3hzKzIgB52QIZu7Jdn4JeXZaegIJinIQfIUP9DrUWUQ,14903
 torch_geometric/loader/zip_loader.py,sha256=3lt10fD15Rxm1WhWzypswGzCEwUz4h8OLCD1nE15yNg,3843
 torch_geometric/metrics/__init__.py,sha256=3krvDobW6vV5yHTjq2S2pmOXxNfysNG26muq7z48e94,699
-torch_geometric/metrics/link_pred.py,sha256=mRQTSYYJgLKXFCelZHMKVOSbPED11JVhbryp7ajjxDU,31137
+torch_geometric/metrics/link_pred.py,sha256=1_hE3KiRqAdZLI6QuUbjgyFC__mTyFu_RimM3bD8wRw,31678
 torch_geometric/nn/__init__.py,sha256=kQHHHUxFDht2ztD-XFQuv98TvC8MdodaFsIjAvltJBw,874
 torch_geometric/nn/data_parallel.py,sha256=YiybTWoSFyfSzlXAamZ_-y1f7B6tvDEFHOuy_AyJz9Q,4761
 torch_geometric/nn/encoding.py,sha256=3DCOCO-XFt-lMb97sHWGN-4KeGUFY5lVo9P00SzrCNk,3559
@@ -620,7 +620,7 @@ torch_geometric/utils/_trim_to_layer.py,sha256=cauOEzMJJK4w9BC-Pg1bHVncBYqG9XxQe
 torch_geometric/utils/_unbatch.py,sha256=B0vjKI96PtHvSBG8F_lqvsiJE134aVjUurPZsG6UZRI,2378
 torch_geometric/utils/augmentation.py,sha256=1F0YCuaklZ9ZbXxdFV0oOoemWvLd8p60WvFo2chzl7E,8600
 torch_geometric/utils/convert.py,sha256=RE5n5no74Xu39-QMWFE0-1RvTgykdK33ymyjF9WcuSs,21938
-torch_geometric/utils/cross_entropy.py,sha256=ZFS5bivtzv3EV9zqgKsekmuQyoZZggPSclhl_tRNHxo,3047
+torch_geometric/utils/cross_entropy.py,sha256=_6whuSCWKNzavOLf83uZbI9RFU0wGRHQict_4-8XIYs,3773
 torch_geometric/utils/dropout.py,sha256=gg0rDnD4FLvBaKSoLAkZwViAQflhLefJm6_Mju5dmQs,11416
 torch_geometric/utils/embedding.py,sha256=Ac_MPSrZGpw-e-gU6Yz-seUioC2WZxBSSzXFeclGwMk,5232
 torch_geometric/utils/functions.py,sha256=orQdS_6EpzWSmBHSok3WhxCzLy9neB-cin1aTnlXY-8,703
@@ -646,7 +646,7 @@ torch_geometric/utils/undirected.py,sha256=H_nfpI0_WluOG6VfjPyldvcjL4w5USAKWu2x5
 torch_geometric/visualization/__init__.py,sha256=b-HnVesXjyJ_L1N-DnjiRiRVf7lhwKaBQF_2i5YMVSU,208
 torch_geometric/visualization/graph.py,sha256=mfZHXYfiU-CWMtfawYc80IxVwVmtK9hbIkSKhM_j7oI,14311
 torch_geometric/visualization/influence.py,sha256=CWMvuNA_Nf1sfbJmQgn58yS4OFpeKXeZPe7kEuvkUBw,477
-pyg_nightly-2.7.0.dev20250904.dist-info/licenses/LICENSE,sha256=ic-27cMJc1kWoMEYncz3Ya3Ur2Bi3bNLWib2DT763-o,1067
-pyg_nightly-2.7.0.dev20250904.dist-info/WHEEL,sha256=G2gURzTEtmeR8nrdXUJfNiB3VYVxigPQ-bEQujpNiNs,82
-pyg_nightly-2.7.0.dev20250904.dist-info/METADATA,sha256=-vA9MGqxs1s0QEIIV_Vj3gZop1fABzvtq8fMNiso7jY,64100
-pyg_nightly-2.7.0.dev20250904.dist-info/RECORD,,
+pyg_nightly-2.7.0.dev20250905.dist-info/licenses/LICENSE,sha256=ic-27cMJc1kWoMEYncz3Ya3Ur2Bi3bNLWib2DT763-o,1067
+pyg_nightly-2.7.0.dev20250905.dist-info/WHEEL,sha256=G2gURzTEtmeR8nrdXUJfNiB3VYVxigPQ-bEQujpNiNs,82
+pyg_nightly-2.7.0.dev20250905.dist-info/METADATA,sha256=7zFs8SgMgRdkwK679QLO1W08UGscHMakHb-3NXsiEok,64100
+pyg_nightly-2.7.0.dev20250905.dist-info/RECORD,,

torch_geometric/__init__.py

@@ -31,7 +31,7 @@ from .lazy_loader import LazyLoader
 contrib = LazyLoader('contrib', globals(), 'torch_geometric.contrib')
 graphgym = LazyLoader('graphgym', globals(), 'torch_geometric.graphgym')
 
-__version__ = '2.7.0.dev20250904'
+__version__ = '2.7.0.dev20250905'
 
 __all__ = [
     'Index',

torch_geometric/metrics/link_pred.py

@@ -21,6 +21,19 @@ class LinkPredMetricData:
     edge_label_index: Union[Tensor, Tuple[Tensor, Tensor]]
     edge_label_weight: Optional[Tensor] = None
 
+    def __post_init__(self) -> None:
+        # Filter all negative weights - they should not be used as ground-truth
+        if self.edge_label_weight is not None:
+            pos_mask = self.edge_label_weight > 0
+            self.edge_label_weight = self.edge_label_weight[pos_mask]
+            if isinstance(self.edge_label_index, Tensor):
+                self.edge_label_index = self.edge_label_index[:, pos_mask]
+            else:
+                self.edge_label_index = (
+                    self.edge_label_index[0][pos_mask],
+                    self.edge_label_index[1][pos_mask],
+                )
+
     @property
     def pred_rel_mat(self) -> Tensor:
         r"""Returns a matrix indicating the relevance of the `k`-th prediction.
@@ -374,8 +387,6 @@ class LinkPredMetricCollection(torch.nn.ModuleDict):
         if self.weighted and edge_label_weight is None:
             raise ValueError(f"'edge_label_weight' is a required argument for "
                              f"weighted '{self.__class__.__name__}' metrics")
-        if not self.weighted:
-            edge_label_weight = None
 
         data = LinkPredMetricData(  # Share metric data across metrics.
             pred_index_mat=pred_index_mat,

torch_geometric/utils/cross_entropy.py

@@ -18,30 +18,51 @@ class SparseCrossEntropy(torch.autograd.Function):
     ) -> Tensor:
         assert inputs.dim() == 2
 
-        logsumexp = inputs.logsumexp(dim=-1)
-        ctx.save_for_backward(inputs, edge_label_index, edge_label_weight,
-                              logsumexp)
+        # Support for both positive and negative weights:
+        # Positive weights scale the logits *after* softmax.
+        # Negative weights scale the denominator *before* softmax:
+        pos_y = edge_label_index
+        neg_y = pos_weight = neg_weight = None
 
-        out = inputs[edge_label_index[0], edge_label_index[1]]
-        out.neg_().add_(logsumexp[edge_label_index[0]])
         if edge_label_weight is not None:
-            out *= edge_label_weight
+            pos_mask = edge_label_weight >= 0
+            pos_y = edge_label_index[:, pos_mask]
+            pos_weight = edge_label_weight[pos_mask]
+
+            if pos_y.size(1) < edge_label_index.size(1):
+                neg_mask = ~pos_mask
+                neg_y = edge_label_index[:, neg_mask]
+                neg_weight = edge_label_weight[neg_mask]
+
+            if neg_y is not None and neg_weight is not None:
+                inputs = inputs.clone()
+                inputs[
+                    neg_y[0],
+                    neg_y[1],
+                ] += neg_weight.abs().log().clamp(min=1e-12)
+
+        logsumexp = inputs.logsumexp(dim=-1)
+        ctx.save_for_backward(inputs, pos_y, pos_weight, logsumexp)
+
+        out = inputs[pos_y[0], pos_y[1]]
+        out.neg_().add_(logsumexp[pos_y[0]])
+        if pos_weight is not None:
+            out *= pos_weight
 
         return out.sum() / inputs.size(0)
 
     @staticmethod
     @torch.autograd.function.once_differentiable
     def backward(ctx: Any, grad_out: Tensor) -> Tuple[Tensor, None, None]:
-        inputs, edge_label_index, edge_label_weight, logsumexp = (
-            ctx.saved_tensors)
+        inputs, pos_y, pos_weight, logsumexp = ctx.saved_tensors
 
         grad_out = grad_out / inputs.size(0)
-        grad_out = grad_out.expand(edge_label_index.size(1))
+        grad_out = grad_out.expand(pos_y.size(1))
 
-        if edge_label_weight is not None:
-            grad_out = grad_out * edge_label_weight
+        if pos_weight is not None:
+            grad_out = grad_out * pos_weight
 
-        grad_logsumexp = scatter(grad_out, edge_label_index[0], dim=0,
+        grad_logsumexp = scatter(grad_out, pos_y[0], dim=0,
                                  dim_size=inputs.size(0), reduce='sum')
 
         # Gradient computation of `logsumexp`: `grad * (self - result).exp()`
@@ -49,7 +70,7 @@ class SparseCrossEntropy(torch.autograd.Function):
         grad_input.exp_()
         grad_input.mul_(grad_logsumexp.view(-1, 1))
 
-        grad_input[edge_label_index[0], edge_label_index[1]] -= grad_out
+        grad_input[pos_y[0], pos_y[1]] -= grad_out
 
         return grad_input, None, None