1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
|
from typing import Any, Callable, Iterator, List, Optional, Tuple, Union
import torch
from torch import Tensor
from torch_geometric.data import Data, FeatureStore, GraphStore, HeteroData
from torch_geometric.loader.base import DataLoaderIterator
from torch_geometric.loader.mixin import (
AffinityMixin,
LogMemoryMixin,
MultithreadingMixin,
)
from torch_geometric.loader.utils import (
filter_custom_hetero_store,
filter_custom_store,
filter_data,
filter_hetero_data,
get_input_nodes,
infer_filter_per_worker,
)
from torch_geometric.sampler import (
BaseSampler,
HeteroSamplerOutput,
NodeSamplerInput,
SamplerOutput,
)
from torch_geometric.typing import InputNodes, OptTensor
class NodeLoader(
torch.utils.data.DataLoader,
AffinityMixin,
MultithreadingMixin,
LogMemoryMixin,
):
r"""A data loader that performs mini-batch sampling from node information,
using a generic :class:`~torch_geometric.sampler.BaseSampler`
implementation that defines a
:meth:`~torch_geometric.sampler.BaseSampler.sample_from_nodes` function and
is supported on the provided input :obj:`data` object.
Args:
data (Any): A :class:`~torch_geometric.data.Data`,
:class:`~torch_geometric.data.HeteroData`, or
(:class:`~torch_geometric.data.FeatureStore`,
:class:`~torch_geometric.data.GraphStore`) data object.
node_sampler (torch_geometric.sampler.BaseSampler): The sampler
implementation to be used with this loader.
Needs to implement
:meth:`~torch_geometric.sampler.BaseSampler.sample_from_nodes`.
The sampler implementation must be compatible with the input
:obj:`data` object.
input_nodes (torch.Tensor or str or Tuple[str, torch.Tensor]): The
indices of seed nodes to start sampling from.
Needs to be either given as a :obj:`torch.LongTensor` or
:obj:`torch.BoolTensor`.
If set to :obj:`None`, all nodes will be considered.
In heterogeneous graphs, needs to be passed as a tuple that holds
the node type and node indices. (default: :obj:`None`)
input_time (torch.Tensor, optional): Optional values to override the
timestamp for the input nodes given in :obj:`input_nodes`. If not
set, will use the timestamps in :obj:`time_attr` as default (if
present). The :obj:`time_attr` needs to be set for this to work.
(default: :obj:`None`)
transform (callable, optional): A function/transform that takes in
a sampled mini-batch and returns a transformed version.
(default: :obj:`None`)
transform_sampler_output (callable, optional): A function/transform
that takes in a :class:`torch_geometric.sampler.SamplerOutput` and
returns a transformed version. (default: :obj:`None`)
filter_per_worker (bool, optional): If set to :obj:`True`, will filter
the returned data in each worker's subprocess.
If set to :obj:`False`, will filter the returned data in the main
process.
If set to :obj:`None`, will automatically infer the decision based
on whether data partially lives on the GPU
(:obj:`filter_per_worker=True`) or entirely on the CPU
(:obj:`filter_per_worker=False`).
There exists different trade-offs for setting this option.
Specifically, setting this option to :obj:`True` for in-memory
datasets will move all features to shared memory, which may result
in too many open file handles. (default: :obj:`None`)
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`.
"""
def __init__(
self,
data: Union[Data, HeteroData, Tuple[FeatureStore, GraphStore]],
node_sampler: BaseSampler,
input_nodes: InputNodes = None,
input_time: OptTensor = None,
transform: Optional[Callable] = None,
transform_sampler_output: Optional[Callable] = None,
filter_per_worker: Optional[bool] = None,
custom_cls: Optional[HeteroData] = None,
input_id: OptTensor = None,
**kwargs,
):
if filter_per_worker is None:
filter_per_worker = infer_filter_per_worker(data)
self.data = data
self.node_sampler = node_sampler
self.input_nodes = input_nodes
self.input_time = input_time
self.transform = transform
self.transform_sampler_output = transform_sampler_output
self.filter_per_worker = filter_per_worker
self.custom_cls = custom_cls
self.input_id = input_id
kwargs.pop('dataset', None)
kwargs.pop('collate_fn', None)
# Get node type (or `None` for homogeneous graphs):
input_type, input_nodes, input_id = get_input_nodes(
data, input_nodes, input_id)
self.input_data = NodeSamplerInput(
input_id=input_id,
node=input_nodes,
time=input_time,
input_type=input_type,
)
iterator = range(input_nodes.size(0))
super().__init__(iterator, collate_fn=self.collate_fn, **kwargs)
def __call__(
self,
index: Union[Tensor, List[int]],
) -> Union[Data, HeteroData]:
r"""Samples a subgraph from a batch of input nodes."""
out = self.collate_fn(index)
if not self.filter_per_worker:
out = self.filter_fn(out)
return out
def collate_fn(self, index: Union[Tensor, List[int]]) -> Any:
r"""Samples a subgraph from a batch of input nodes."""
input_data: NodeSamplerInput = self.input_data[index]
out = self.node_sampler.sample_from_nodes(input_data)
if self.filter_per_worker: # Execute `filter_fn` in the worker process
out = self.filter_fn(out)
return out
def filter_fn(
self,
out: Union[SamplerOutput, HeteroSamplerOutput],
) -> Union[Data, HeteroData]:
r"""Joins the sampled nodes with their corresponding features,
returning the resulting :class:`~torch_geometric.data.Data` or
:class:`~torch_geometric.data.HeteroData` object to be used downstream.
"""
if self.transform_sampler_output:
out = self.transform_sampler_output(out)
if isinstance(out, SamplerOutput):
if isinstance(self.data, Data):
data = filter_data( #
self.data, out.node, out.row, out.col, out.edge,
self.node_sampler.edge_permutation)
else: # Tuple[FeatureStore, GraphStore]
# Hack to detect whether we are in a distributed setting.
if (self.node_sampler.__class__.__name__ ==
'DistNeighborSampler'):
edge_index = torch.stack([out.row, out.col])
data = Data(edge_index=edge_index)
# Metadata entries are populated in
# `DistributedNeighborSampler._collate_fn()`
data.x = out.metadata[-3]
data.y = out.metadata[-2]
data.edge_attr = out.metadata[-1]
else:
data = filter_custom_store( #
*self.data, out.node, out.row, out.col, out.edge,
self.custom_cls)
if 'n_id' not in data:
data.n_id = out.node
if out.edge is not None and 'e_id' not in data:
edge = out.edge.to(torch.long)
perm = self.node_sampler.edge_permutation
data.e_id = perm[edge] if perm is not None else edge
data.batch = out.batch
data.num_sampled_nodes = out.num_sampled_nodes
data.num_sampled_edges = out.num_sampled_edges
if out.orig_row is not None and out.orig_col is not None:
data._orig_edge_index = torch.stack([
out.orig_row,
out.orig_col,
], dim=0)
data.input_id = out.metadata[0]
data.seed_time = out.metadata[1]
data.batch_size = out.metadata[0].size(0)
elif isinstance(out, HeteroSamplerOutput):
if isinstance(self.data, HeteroData):
data = filter_hetero_data( #
self.data, out.node, out.row, out.col, out.edge,
self.node_sampler.edge_permutation)
else: # Tuple[FeatureStore, GraphStore]
# Hack to detect whether we are in a distributed setting.
if (self.node_sampler.__class__.__name__ ==
'DistNeighborSampler'):
import torch_geometric.distributed as dist
data = dist.utils.filter_dist_store(
*self.data, out.node, out.row, out.col, out.edge,
self.custom_cls, out.metadata,
self.input_data.input_type)
else:
data = filter_custom_hetero_store( #
*self.data, out.node, out.row, out.col, out.edge,
self.custom_cls)
for key, node in out.node.items():
if 'n_id' not in data[key]:
data[key].n_id = node
for key, edge in (out.edge or {}).items():
if edge is not None and 'e_id' not in data[key]:
edge = edge.to(torch.long)
perm = self.node_sampler.edge_permutation
if perm is not None and perm.get(key, None) is not None:
edge = perm[key][edge]
data[key].e_id = edge
data.set_value_dict('batch', out.batch)
data.set_value_dict('num_sampled_nodes', out.num_sampled_nodes)
data.set_value_dict('num_sampled_edges', out.num_sampled_edges)
if out.orig_row is not None and out.orig_col is not None:
for key in out.orig_row.keys():
data[key]._orig_edge_index = torch.stack([
out.orig_row[key],
out.orig_col[key],
], dim=0)
input_type = self.input_data.input_type
data[input_type].input_id = out.metadata[0]
data[input_type].seed_time = out.metadata[1]
data[input_type].batch_size = out.metadata[0].size(0)
else:
raise TypeError(f"'{self.__class__.__name__}'' found invalid "
f"type: '{type(out)}'")
return data if self.transform is None else self.transform(data)
def _get_iterator(self) -> Iterator:
if self.filter_per_worker:
return super()._get_iterator()
# if not self.is_cuda_available and not self.cpu_affinity_enabled:
# TODO: Add manual page for best CPU practices
# link = ...
# Warning('Dataloader CPU affinity opt is not enabled, consider '
# 'switching it on with enable_cpu_affinity() or see CPU '
# f'best practices for PyG [{link}])')
# Execute `filter_fn` in the main process:
return DataLoaderIterator(super()._get_iterator(), self.filter_fn)
def __repr__(self) -> str:
return f'{self.__class__.__name__}()'
|
