pyg-nightly 2.7.0.dev20250825__py3-none-any.whl → 2.7.0.dev20250827__py3-none-any.whl

torch_geometric/sampler/__init__.py

@@ -3,7 +3,7 @@ r"""Graph sampler package."""
 from .base import (BaseSampler, NodeSamplerInput, EdgeSamplerInput,
                    SamplerOutput, HeteroSamplerOutput, NegativeSampling,
                    NumNeighbors)
-from .neighbor_sampler import NeighborSampler
+from .neighbor_sampler import NeighborSampler, BidirectionalNeighborSampler
 from .hgt_sampler import HGTSampler
 
 __all__ = classes = [
@@ -15,5 +15,6 @@ __all__ = classes = [
     'NumNeighbors',
     'NegativeSampling',
     'NeighborSampler',
+    'BidirectionalNeighborSampler',
     'HGTSampler',
 ]

torch_geometric/sampler/base.py

@@ -3,7 +3,7 @@ import math
 import warnings
 from abc import ABC, abstractmethod
 from collections import defaultdict
-from dataclasses import dataclass
+from dataclasses import dataclass, field
 from enum import Enum
 from typing import Any, Dict, List, Literal, Optional, Union
 
@@ -12,8 +12,10 @@ from torch import Tensor
 
 from torch_geometric.data import Data, FeatureStore, GraphStore, HeteroData
 from torch_geometric.sampler.utils import (
+    global_to_local_node_idx,
     local_to_global_node_idx,
     to_bidirectional,
+    unique_unsorted,
 )
 from torch_geometric.typing import EdgeType, EdgeTypeStr, NodeType, OptTensor
 from torch_geometric.utils.mixin import CastMixin
@@ -209,6 +211,7 @@ class SamplerOutput(CastMixin):
     # TODO(manan): refine this further; it does not currently define a proper
     # API for the expected output of a sampler.
     metadata: Optional[Any] = None
+    _seed_node: OptTensor = field(repr=False, default=None)
 
     @property
     def global_row(self) -> Tensor:
@@ -220,8 +223,17 @@ class SamplerOutput(CastMixin):
 
     @property
     def seed_node(self) -> Tensor:
-        return local_to_global_node_idx(
-            self.node, self.batch) if self.batch is not None else None
+        # can be set manually if the seed nodes are not contained in the
+        # sampled nodes
+        if self._seed_node is None:
+            self._seed_node = local_to_global_node_idx(
+                self.node, self.batch) if self.batch is not None else None
+        return self._seed_node
+
+    @seed_node.setter
+    def seed_node(self, value: Tensor):
+        assert len(value) == len(self.node)
+        self._seed_node = value
 
     @property
     def global_orig_row(self) -> Tensor:
@@ -263,6 +275,230 @@ class SamplerOutput(CastMixin):
 
         return out
 
+    @classmethod
+    def collate(cls, outputs: List['SamplerOutput'],
+                replace: bool = True) -> 'SamplerOutput':
+        r"""Collate a list of :class:`~torch_geometric.sampler.SamplerOutput`
+        objects into a single :class:`~torch_geometric.sampler.SamplerOutput`
+        object. Requires that they all have the same fields.
+        """
+        if len(outputs) == 0:
+            raise ValueError("Cannot collate an empty list of SamplerOutputs")
+        out = outputs[0]
+        has_edge = out.edge is not None
+        has_orig_row = out.orig_row is not None
+        has_orig_col = out.orig_col is not None
+        has_batch = out.batch is not None
+        has_num_sampled_nodes = out.num_sampled_nodes is not None
+        has_num_sampled_edges = out.num_sampled_edges is not None
+
+        try:
+            for i, sample_output in enumerate(outputs):  # noqa
+                assert not has_edge == (sample_output.edge is None)
+                assert not has_orig_row == (sample_output.orig_row is None)
+                assert not has_orig_col == (sample_output.orig_col is None)
+                assert not has_batch == (sample_output.batch is None)
+                assert not has_num_sampled_nodes == (
+                    sample_output.num_sampled_nodes is None)
+                assert not has_num_sampled_edges == (
+                    sample_output.num_sampled_edges is None)
+        except AssertionError:
+            error_str = f"Output {i+1} has a different field than the first output"  # noqa
+            raise ValueError(error_str)  # noqa
+
+        for other in outputs[1:]:
+            out = out.merge_with(other, replace=replace)
+        return out
+
+    def merge_with(self, other: 'SamplerOutput',
+                   replace: bool = True) -> 'SamplerOutput':
+        """Merges two SamplerOutputs.
+        If replace is True, self's nodes and edges take precedence.
+        """
+        if not replace:
+            return SamplerOutput(
+                node=torch.cat([self.node, other.node], dim=0),
+                row=torch.cat([self.row, len(self.node) + other.row], dim=0),
+                col=torch.cat([self.col, len(self.node) + other.col], dim=0),
+                edge=torch.cat([self.edge, other.edge], dim=0)
+                if self.edge is not None and other.edge is not None else None,
+                batch=torch.cat(
+                    [self.batch, len(self.node) + other.batch], dim=0) if
+                self.batch is not None and other.batch is not None else None,
+                num_sampled_nodes=self.num_sampled_nodes +
+                other.num_sampled_nodes if self.num_sampled_nodes is not None
+                and other.num_sampled_nodes is not None else None,
+                num_sampled_edges=self.num_sampled_edges +
+                other.num_sampled_edges if self.num_sampled_edges is not None
+                and other.num_sampled_edges is not None else None,
+                orig_row=torch.cat(
+                    [self.orig_row,
+                     len(self.node) +
+                     other.orig_row], dim=0) if self.orig_row is not None
+                and other.orig_row is not None else None,
+                orig_col=torch.cat(
+                    [self.orig_col,
+                     len(self.node) +
+                     other.orig_col], dim=0) if self.orig_col is not None
+                and other.orig_col is not None else None,
+                metadata=[self.metadata, other.metadata],
+            )
+        else:
+
+            # NODES
+            old_nodes, new_nodes = self.node, other.node
+            old_node_uid, new_node_uid = [old_nodes], [new_nodes]
+
+            # batch tracks disjoint subgraph samplings
+            if self.batch is not None and other.batch is not None:
+                # Transform the batch indices to be global node ids
+                old_batch_nodes = self.seed_node
+                new_batch_nodes = other.seed_node
+                old_node_uid.append(old_batch_nodes)
+                new_node_uid.append(new_batch_nodes)
+
+            # NOTE: if any new node fields are added,
+            # they need to be merged here
+
+            old_node_uid = torch.stack(old_node_uid, dim=1)
+            new_node_uid = torch.stack(new_node_uid, dim=1)
+
+            merged_node_uid = unique_unsorted(
+                torch.cat([old_node_uid, new_node_uid], dim=0))
+            num_old_nodes = old_node_uid.shape[0]
+
+            # Recompute num sampled nodes for second output,
+            # subtracting out nodes already seen in first output
+            merged_node_num_sampled_nodes = None
+            if (self.num_sampled_nodes is not None
+                    and other.num_sampled_nodes is not None):
+                merged_node_num_sampled_nodes = copy.copy(
+                    self.num_sampled_nodes)
+                curr_index = 0
+                # NOTE: There's an assumption here that no two nodes will be
+                # sampled twice in the same SampleOutput object
+                for minibatch in other.num_sampled_nodes:
+                    size_of_intersect = torch.cat([
+                        old_node_uid,
+                        new_node_uid[curr_index:curr_index + minibatch]
+                    ]).unique(dim=0, sorted=False).shape[0] - num_old_nodes
+                    merged_node_num_sampled_nodes.append(size_of_intersect)
+                    curr_index += minibatch
+
+            merged_nodes = merged_node_uid[:, 0]
+            merged_batch = None
+            if self.batch is not None and other.batch is not None:
+                # Restore the batch indices to be relative to the nodes field
+                ref_merged_batch_nodes = merged_node_uid[:, 1].unsqueeze(
+                    -1).expand(-1, 2)  # num_nodes x 2
+                merged_batch = global_to_local_node_idx(
+                    merged_node_uid, ref_merged_batch_nodes)
+
+            # EDGES
+            is_bidirectional = self.orig_row is not None \
+                and self.orig_col is not None \
+                and other.orig_row is not None \
+                and other.orig_col is not None
+            if is_bidirectional:
+                old_row, old_col = self.orig_row, self.orig_col
+                new_row, new_col = other.orig_row, other.orig_col
+            else:
+                old_row, old_col = self.row, self.col
+                new_row, new_col = other.row, other.col
+
+            # Transform the row and col indices to be global node ids
+            # instead of relative indices to nodes field
+            # Edge uids build off of node uids
+            old_row_idx, old_col_idx = local_to_global_node_idx(
+                old_node_uid,
+                old_row), local_to_global_node_idx(old_node_uid, old_col)
+            new_row_idx, new_col_idx = local_to_global_node_idx(
+                new_node_uid,
+                new_row), local_to_global_node_idx(new_node_uid, new_col)
+
+            old_edge_uid, new_edge_uid = [old_row_idx, old_col_idx
+                                          ], [new_row_idx, new_col_idx]
+
+            row_idx = 0
+            col_idx = old_row_idx.shape[1]
+            edge_idx = old_row_idx.shape[1] + old_col_idx.shape[1]
+
+            if self.edge is not None and other.edge is not None:
+                if is_bidirectional:
+                    # bidirectional duplicates edge ids
+                    old_edge_uid_ref = torch.stack([self.row, self.col],
+                                                   dim=1)  # num_edges x 2
+                    old_orig_edge_uid_ref = torch.stack(
+                        [self.orig_row, self.orig_col],
+                        dim=1)  # num_orig_edges x 2
+
+                    old_edge_idx = global_to_local_node_idx(
+                        old_edge_uid_ref, old_orig_edge_uid_ref)
+                    old_edge = self.edge[old_edge_idx]
+
+                    new_edge_uid_ref = torch.stack([other.row, other.col],
+                                                   dim=1)  # num_edges x 2
+                    new_orig_edge_uid_ref = torch.stack(
+                        [other.orig_row, other.orig_col],
+                        dim=1)  # num_orig_edges x 2
+
+                    new_edge_idx = global_to_local_node_idx(
+                        new_edge_uid_ref, new_orig_edge_uid_ref)
+                    new_edge = other.edge[new_edge_idx]
+
+                else:
+                    old_edge, new_edge = self.edge, other.edge
+
+                old_edge_uid.append(old_edge.unsqueeze(-1))
+                new_edge_uid.append(new_edge.unsqueeze(-1))
+
+            old_edge_uid = torch.cat(old_edge_uid, dim=1)
+            new_edge_uid = torch.cat(new_edge_uid, dim=1)
+
+            merged_edge_uid = unique_unsorted(
+                torch.cat([old_edge_uid, new_edge_uid], dim=0))
+            num_old_edges = old_edge_uid.shape[0]
+
+            merged_edge_num_sampled_edges = None
+            if (self.num_sampled_edges is not None
+                    and other.num_sampled_edges is not None):
+                merged_edge_num_sampled_edges = copy.copy(
+                    self.num_sampled_edges)
+                curr_index = 0
+                # NOTE: There's an assumption here that no two edges will be
+                # sampled twice in the same SampleOutput object
+                for minibatch in other.num_sampled_edges:
+                    size_of_intersect = torch.cat([
+                        old_edge_uid,
+                        new_edge_uid[curr_index:curr_index + minibatch]
+                    ]).unique(dim=0, sorted=False).shape[0] - num_old_edges
+                    merged_edge_num_sampled_edges.append(size_of_intersect)
+                    curr_index += minibatch
+
+            merged_row = merged_edge_uid[:, row_idx:col_idx]
+            merged_col = merged_edge_uid[:, col_idx:edge_idx]
+            merged_edge = merged_edge_uid[:, edge_idx:].squeeze() \
+                if self.edge is not None and other.edge is not None else None
+
+            # restore to row and col indices relative to nodes field
+            merged_row = global_to_local_node_idx(merged_node_uid, merged_row)
+            merged_col = global_to_local_node_idx(merged_node_uid, merged_col)
+
+            out = SamplerOutput(
+                node=merged_nodes,
+                row=merged_row,
+                col=merged_col,
+                edge=merged_edge,
+                batch=merged_batch,
+                num_sampled_nodes=merged_node_num_sampled_nodes,
+                num_sampled_edges=merged_edge_num_sampled_edges,
+                metadata=[self.metadata, other.metadata],
+            )
+            # Restores orig_row and orig_col if they existed before merging
+            if is_bidirectional:
+                out = out.to_bidirectional(keep_orig_edges=True)
+            return out
+
 
 @dataclass
 class HeteroSamplerOutput(CastMixin):
@@ -439,6 +675,25 @@ class HeteroSamplerOutput(CastMixin):
 
         return out
 
+    @classmethod
+    def collate(cls, outputs: List['HeteroSamplerOutput'],
+                replace: bool = True) -> 'HeteroSamplerOutput':
+        r"""Collate a list of
+        :class:`~torch_geometric.sampler.HeteroSamplerOutput`objects into a
+        single :class:`~torch_geometric.sampler.HeteroSamplerOutput` object.
+        Requires that they all have the same fields.
+        """
+        # TODO(zaristei)
+        raise NotImplementedError
+
+    def merge_with(self, other: 'HeteroSamplerOutput',
+                   replace: bool = True) -> 'HeteroSamplerOutput':
+        """Merges two HeteroSamplerOutputs.
+        If replace is True, self's nodes and edges take precedence.
+        """
+        # TODO(zaristei)
+        raise NotImplementedError
+
 
 @dataclass(frozen=True)
 class NumNeighbors: