pyg-nightly 2.7.0.dev20250221__tar.gz → 2.7.0.dev20250222__tar.gz

{pyg_nightly-2.7.0.dev20250221 → pyg_nightly-2.7.0.dev20250222}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: pyg-nightly
-Version: 2.7.0.dev20250221
+Version: 2.7.0.dev20250222
 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.dev20250221 → pyg_nightly-2.7.0.dev20250222}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend="flit_core.buildapi"
 
 [project]
 name="pyg-nightly"
-version="2.7.0.dev20250221"
+version="2.7.0.dev20250222"
 authors=[
     {name="Matthias Fey", email="matthias@pyg.org"},
 ]

{pyg_nightly-2.7.0.dev20250221 → pyg_nightly-2.7.0.dev20250222}/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.dev20250221'
+__version__ = '2.7.0.dev20250222'
 
 __all__ = [
     'Index',

{pyg_nightly-2.7.0.dev20250221 → pyg_nightly-2.7.0.dev20250222}/torch_geometric/hash_tensor.py

@@ -142,12 +142,12 @@ class HashTensor(Tensor):
         if (key.dtype in {torch.uint8, torch.int16} or _range <= 1_000_000
                 or _range <= 2 * key.numel()):
             _map = torch.full(
-                size=(_range + 2, ),
+                size=(_range + 3, ),
                 fill_value=-1,
                 dtype=torch.int64,
-                device=device,
+                device=key.device,
             )
-            _map[(key - (min_key - 1)).long()] = torch.arange(
+            _map[key.long() - (min_key.long() - 1)] = torch.arange(
                 key.numel(),
                 dtype=_map.dtype,
                 device=_map.device,
@@ -164,6 +164,8 @@ class HashTensor(Tensor):
             dtype=dtype,
         )
 
+    # Private Methods #########################################################
+
     @classmethod
     def _from_data(
         cls,
@@ -217,6 +219,33 @@ class HashTensor(Tensor):
             dtype=self.dtype,
         )
 
+    def _get(self, query: Tensor) -> Tensor:
+        if isinstance(self._map, Tensor):
+            index = query.long() - (self._min_key.long() - 1)
+            index = self._map[index.clamp_(min=0, max=self._map.numel() - 1)]
+        elif torch_geometric.typing.WITH_CUDA_HASH_MAP and query.is_cuda:
+            index = self._map.get(query)
+        elif torch_geometric.typing.WITH_CPU_HASH_MAP:
+            index = self._map.get(query.cpu())
+        else:
+            import pandas as pd
+
+            ser = pd.Series(query.cpu().numpy(), dtype=self._map)
+            index = torch.from_numpy(ser.cat.codes.to_numpy()).to(torch.long)
+
+        index = index.to(self.device)
+
+        if self._value is None:
+            return index.to(self.dtype)
+
+        out = self._value[index]
+        mask = index != -1
+        mask = mask.view([-1] + [1] * (out.dim() - 1))
+        if out.is_floating_point():
+            return out.where(mask, float('NaN'))
+        else:
+            return out.where(mask, -1)
+
     # Methods #################################################################
 
     def as_tensor(self) -> Tensor:
@@ -252,6 +281,9 @@ class HashTensor(Tensor):
                                           kwargs)
         return func(*args, **(kwargs or {}))
 
+    def index_select(self, dim: int, index: Any) -> Tensor:  # type: ignore
+        return torch.index_select(self, dim, index)
+
 
 @implements(aten.alias.default)
 def _alias(tensor: HashTensor) -> HashTensor:
@@ -387,3 +419,51 @@ def _slice(
         num_keys=tensor.size(0),
         dtype=tensor.dtype,
     )
+
+
+# Since PyTorch does only allow PyTorch tensors as indices in `index_select`,
+# we need to create a wrapper function and monkey patch `index_select` :(
+_old_index_select = torch.index_select
+
+
+def _new_index_select(
+    input: Tensor,
+    dim: int,
+    index: Any,
+    *,
+    out: Optional[Tensor] = None,
+) -> Tensor:
+
+    if dim < -input.dim() or dim >= input.dim():
+        raise IndexError(f"Dimension out of range (expected to be in range of "
+                         f"[{-input.dim()}, {input.dim()-1}], but got {dim})")
+
+    # We convert any index tensor in the first dimension into a tensor. This
+    # means that downstream handling (i.e. in `aten.index_select.default`)
+    # needs to take this pre-conversion into account.
+    if isinstance(input, HashTensor) and (dim == 0 or dim == -input.dim()):
+        index = as_key_tensor(index, device=input.device)
+    return _old_index_select(input, dim, index, out=out)
+
+
+torch.index_select = _new_index_select  # type: ignore
+
+
+@implements(aten.index_select.default)
+def _index_select(
+    tensor: HashTensor,
+    dim: int,
+    index: Tensor,
+) -> Union[HashTensor, Tensor]:
+
+    if dim == 0 or dim == -tensor.dim():
+        return tensor._get(index)
+
+    return tensor._from_data(
+        tensor._map,
+        aten.index_select.default(tensor.as_tensor(), dim, index),
+        tensor._min_key,
+        tensor._max_key,
+        num_keys=tensor.size(0),
+        dtype=tensor.dtype,
+    )

{pyg_nightly-2.7.0.dev20250221 → pyg_nightly-2.7.0.dev20250222}/torch_geometric/metrics/link_pred.py

@@ -715,22 +715,32 @@ class LinkPredPersonalization(_LinkPredMetric):
 
     Args:
         k (int): The number of top-:math:`k` predictions to evaluate against.
+        max_src_nodes (int, optional): The maximum source nodes to consider to
+            compute pair-wise dissimilarity. If specified,
+            Personalization @ :math:`k` is approximated to avoid computation
+            blowup due to quadratic complexity. (default: :obj:`2**12`)
         batch_size (int, optional): The batch size to determine how many pairs
             of user recommendations should be processed at once.
             (default: :obj:`2**16`)
     """
     higher_is_better: bool = True
 
-    def __init__(self, k: int, batch_size: int = 2**16) -> None:
+    def __init__(
+        self,
+        k: int,
+        max_src_nodes: Optional[int] = 2**12,
+        batch_size: int = 2**16,
+    ) -> None:
         super().__init__(k)
+        self.max_src_nodes = max_src_nodes
         self.batch_size = batch_size
 
         if WITH_TORCHMETRICS:
             self.add_state('preds', default=[], dist_reduce_fx='cat')
-            self.add_state('dev_tensor', torch.empty(0), dist_reduce_fx='sum')
+            self.add_state('total', torch.tensor(0), dist_reduce_fx='sum')
         else:
             self.preds: List[Tensor] = []
-            self.register_buffer('dev_tensor', torch.empty(0))
+            self.register_buffer('total', torch.tensor(0))
 
     def update(
         self,
@@ -738,11 +748,21 @@ class LinkPredPersonalization(_LinkPredMetric):
         edge_label_index: Union[Tensor, Tuple[Tensor, Tensor]],
         edge_label_weight: Optional[Tensor] = None,
     ) -> None:
+
         # NOTE Move to CPU to avoid memory blowup.
-        self.preds.append(pred_index_mat[:, :self.k].cpu())
+        pred_index_mat = pred_index_mat[:, :self.k].cpu()
+
+        if self.max_src_nodes is None:
+            self.preds.append(pred_index_mat)
+            self.total += pred_index_mat.size(0)
+        elif self.total < self.max_src_nodes:
+            remaining = int(self.max_src_nodes - self.total)
+            pred_index_mat = pred_index_mat[:remaining]
+            self.preds.append(pred_index_mat)
+            self.total += pred_index_mat.size(0)
 
     def compute(self) -> Tensor:
-        device = self.dev_tensor.device
+        device = self.total.device
         score = torch.tensor(0.0, device=device)
         total = torch.tensor(0, device=device)
 
@@ -786,6 +806,7 @@ class LinkPredPersonalization(_LinkPredMetric):
 
     def _reset(self) -> None:
         self.preds = []
+        self.total.zero_()
 
 
 class LinkPredAveragePopularity(_LinkPredMetric):