Support for custom HeteroData mini-batch class in remote backends (#…

…6377)
pyg-team · Jan 9, 2023 · 3ea1d81 · 3ea1d81
1 parent 19d5cbc
commit 3ea1d81
Show file tree

Hide file tree

Showing 4 changed files with 15 additions and 3 deletions.
diff --git a/CHANGELOG.md b/CHANGELOG.md
@@ -5,6 +5,7 @@ The format is based on [Keep a Changelog](http://keepachangelog.com/en/1.0.0/).
 
 ## [2.3.0] - 2023-MM-DD
 ### Added
+- Added support for custom `HeteroData` mini-batch class in remote backends ([#6377](https://github.com/pyg-team/pytorch_geometric/pull/6377))
 - Added the `GNNFF` model ([#5866](https://github.com/pyg-team/pytorch_geometric/pull/5866))
 - Added `MLPAggregation`, `SetTransformerAggregation`, `GRUAggregation`, and `DeepSetsAggregation` as adaptive readout functions ([#6301](https://github.com/pyg-team/pytorch_geometric/pull/6301), [#6336](https://github.com/pyg-team/pytorch_geometric/pull/6336), [#6338](https://github.com/pyg-team/pytorch_geometric/pull/6338))
 - Added `Dataset.to_datapipe` for converting PyG datasets into a torchdata `DataPipe`([#6141](https://github.com/pyg-team/pytorch_geometric/pull/6141))

diff --git a/torch_geometric/loader/link_loader.py b/torch_geometric/loader/link_loader.py
@@ -104,6 +104,9 @@ class LinkLoader(torch.utils.data.DataLoader):
             (2) it may slown down data loading,
             (3) it requires operating on CPU tensors.
             (default: :obj:`False`)
+        custom_cls (HeteroData, optional): A custom
+            :class:`~torch_geometric.data.HeteroData` class to return for
+            mini-batches in case of remote backends. (default: :obj:`None`)
         **kwargs (optional): Additional arguments of
             :class:`torch.utils.data.DataLoader`, such as :obj:`batch_size`,
             :obj:`shuffle`, :obj:`drop_last` or :obj:`num_workers`.
@@ -120,6 +123,7 @@ def __init__(
         transform: Optional[Callable] = None,
         transform_sampler_output: Optional[Callable] = None,
         filter_per_worker: bool = False,
+        custom_cls: Optional[HeteroData] = None,
         **kwargs,
     ):
         # Remove for PyTorch Lightning:
@@ -142,6 +146,7 @@ def __init__(
         self.transform = transform
         self.transform_sampler_output = transform_sampler_output
         self.filter_per_worker = filter_per_worker
+        self.custom_cls = custom_cls
 
         if (self.neg_sampling is not None and self.neg_sampling.is_binary()
                 and edge_label is not None and edge_label.min() == 0):
@@ -220,7 +225,7 @@ def filter_fn(
                                           self.link_sampler.edge_permutation)
             else:  # Tuple[FeatureStore, GraphStore]
                 data = filter_custom_store(*self.data, out.node, out.row,
-                                           out.col, out.edge)
+                                           out.col, out.edge, self.custom_cls)
 
             for key, batch in (out.batch or {}).items():
                 data[key].batch = batch

diff --git a/torch_geometric/loader/node_loader.py b/torch_geometric/loader/node_loader.py
@@ -68,6 +68,9 @@ class NodeLoader(torch.utils.data.DataLoader):
             (2) it may slown down data loading,
             (3) it requires operating on CPU tensors.
             (default: :obj:`False`)
+        custom_cls (HeteroData, optional): A custom
+            :class:`~torch_geometric.data.HeteroData` class to return for
+            mini-batches in case of remote backends. (default: :obj:`None`)
         **kwargs (optional): Additional arguments of
             :class:`torch.utils.data.DataLoader`, such as :obj:`batch_size`,
             :obj:`shuffle`, :obj:`drop_last` or :obj:`num_workers`.
@@ -81,6 +84,7 @@ def __init__(
         transform: Optional[Callable] = None,
         transform_sampler_output: Optional[Callable] = None,
         filter_per_worker: bool = False,
+        custom_cls: Optional[HeteroData] = None,
         **kwargs,
     ):
         # Remove for PyTorch Lightning:
@@ -97,6 +101,7 @@ def __init__(
         self.transform = transform
         self.transform_sampler_output = transform_sampler_output
         self.filter_per_worker = filter_per_worker
+        self.custom_cls = custom_cls
 
         self.input_data = NodeSamplerInput(
             input_id=None,
@@ -153,7 +158,7 @@ def filter_fn(
                                           self.node_sampler.edge_permutation)
             else:  # Tuple[FeatureStore, GraphStore]
                 data = filter_custom_store(*self.data, out.node, out.row,
-                                           out.col, out.edge)
+                                           out.col, out.edge, self.custom_cls)
 
             for key, batch in (out.batch or {}).items():
                 data[key].batch = batch

diff --git a/torch_geometric/loader/utils.py b/torch_geometric/loader/utils.py
@@ -159,13 +159,14 @@ def filter_custom_store(
     row_dict: Dict[str, Tensor],
     col_dict: Dict[str, Tensor],
     edge_dict: Dict[str, Tensor],
+    custom_cls: Optional[HeteroData] = None,
 ) -> HeteroData:
     r"""Constructs a `HeteroData` object from a feature store that only holds
     nodes in `node` end edges in `edge` for each node and edge type,
     respectively."""
 
     # Construct a new `HeteroData` object:
-    data = HeteroData()
+    data = custom_cls() if custom_cls is not None else HeteroData()
 
     # Filter edge storage:
     # TODO support edge attributes