dgl.dataloading¶
The dgl.dataloading
package contains:
Data loader classes for iterating over a set of nodes or edges in a graph and generates computation dependency via neighborhood sampling methods.
Various sampler classes that perform neighborhood sampling for multilayer GNNs.
Negative samplers for link prediction.
For a holistic explanation on how different components work together. Read the user guide Chapter 6: Stochastic Training on Large Graphs.
Note
This package is experimental and the interfaces may be subject to changes in future releases. It currently only has implementations in PyTorch.
DataLoaders¶
DGL DataLoader for minibatch training works similarly to PyTorch’s DataLoader.
It has a generator interface that returns minibatches sampled from some given graphs.
DGL provides two DataLoaders: a NodeDataLoader
for node classification task
and an EdgeDataLoader
for edge/link prediction task.

class
dgl.dataloading.pytorch.
NodeDataLoader
(g, nids, block_sampler, device='cpu', **kwargs)[source]¶ PyTorch dataloader for batchiterating over a set of nodes, generating the list of message flow graphs (MFGs) as computation dependency of the said minibatch.
 Parameters
g (DGLGraph) – The graph.
nids (Tensor or dict[ntype, Tensor]) – The node set to compute outputs.
block_sampler (dgl.dataloading.BlockSampler) – The neighborhood sampler.
device (device context, optional) – The device of the generated MFGs in each iteration, which should be a PyTorch device object (e.g.,
torch.device
).kwargs (dict) – Arguments being passed to
torch.utils.data.DataLoader
.
Examples
To train a 3layer GNN for node classification on a set of nodes
train_nid
on a homogeneous graph where each node takes messages from all neighbors (assume the backend is PyTorch):>>> sampler = dgl.dataloading.MultiLayerNeighborSampler([15, 10, 5]) >>> dataloader = dgl.dataloading.NodeDataLoader( ... g, train_nid, sampler, ... batch_size=1024, shuffle=True, drop_last=False, num_workers=4) >>> for input_nodes, output_nodes, blocks in dataloader: ... train_on(input_nodes, output_nodes, blocks)
Notes
Please refer to Minibatch Training Tutorials and User Guide Section 6 for usage.

class
dgl.dataloading.pytorch.
EdgeDataLoader
(g, eids, block_sampler, device='cpu', **kwargs)[source]¶ PyTorch dataloader for batchiterating over a set of edges, generating the list of message flow graphs (MFGs) as computation dependency of the said minibatch for edge classification, edge regression, and link prediction.
For each iteration, the object will yield
A tensor of input nodes necessary for computing the representation on edges, or a dictionary of node type names and such tensors.
A subgraph that contains only the edges in the minibatch and their incident nodes. Note that the graph has an identical metagraph with the original graph.
If a negative sampler is given, another graph that contains the “negative edges”, connecting the source and destination nodes yielded from the given negative sampler.
A list of MFGs necessary for computing the representation of the incident nodes of the edges in the minibatch.
For more details, please refer to 6.2 Training GNN for Edge Classification with Neighborhood Sampling and 6.3 Training GNN for Link Prediction with Neighborhood Sampling.
 Parameters
g (DGLGraph) – The graph.
eids (Tensor or dict[etype, Tensor]) – The edge set in graph
g
to compute outputs.block_sampler (dgl.dataloading.BlockSampler) – The neighborhood sampler.
device (device context, optional) – The device of the generated MFGs and graphs in each iteration, which should be a PyTorch device object (e.g.,
torch.device
).g_sampling (DGLGraph, optional) –
The graph where neighborhood sampling is performed.
One may wish to iterate over the edges in one graph while perform sampling in another graph. This may be the case for iterating over validation and test edge set while perform neighborhood sampling on the graph formed by only the training edge set.
If None, assume to be the same as
g
.exclude (str, optional) –
Whether and how to exclude dependencies related to the sampled edges in the minibatch. Possible values are
None,
reverse_id
,reverse_types
See the description of the argument with the same name in the docstring of
EdgeCollator
for more details.reverse_eids (Tensor or dict[etype, Tensor], optional) –
A tensor of reverse edge ID mapping. The ith element indicates the ID of the ith edge’s reverse edge.
If the graph is heterogeneous, this argument requires a dictionary of edge types and the reverse edge ID mapping tensors.
See the description of the argument with the same name in the docstring of
EdgeCollator
for more details.reverse_etypes (dict[etype, etype], optional) –
The mapping from the original edge types to their reverse edge types.
See the description of the argument with the same name in the docstring of
EdgeCollator
for more details.negative_sampler (callable, optional) –
The negative sampler.
See the description of the argument with the same name in the docstring of
EdgeCollator
for more details.kwargs (dict) – Arguments being passed to
torch.utils.data.DataLoader
.
Examples
The following example shows how to train a 3layer GNN for edge classification on a set of edges
train_eid
on a homogeneous undirected graph. Each node takes messages from all neighbors.Say that you have an array of source node IDs
src
and another array of destination node IDsdst
. One can make it bidirectional by adding another set of edges that connects fromdst
tosrc
:>>> g = dgl.graph((torch.cat([src, dst]), torch.cat([dst, src])))
One can then know that the ID difference of an edge and its reverse edge is
E
, whereE
is the length of your source/destination array. The reverse edge mapping can be obtained by>>> E = len(src) >>> reverse_eids = torch.cat([torch.arange(E, 2 * E), torch.arange(0, E)])
Note that the sampled edges as well as their reverse edges are removed from computation dependencies of the incident nodes. That is, the edge will not involve in neighbor sampling and message aggregation. This is a common trick to avoid information leakage.
>>> sampler = dgl.dataloading.MultiLayerNeighborSampler([15, 10, 5]) >>> dataloader = dgl.dataloading.EdgeDataLoader( ... g, train_eid, sampler, exclude='reverse_id', ... reverse_eids=reverse_eids, ... batch_size=1024, shuffle=True, drop_last=False, num_workers=4) >>> for input_nodes, pair_graph, blocks in dataloader: ... train_on(input_nodes, pair_graph, blocks)
To train a 3layer GNN for link prediction on a set of edges
train_eid
on a homogeneous graph where each node takes messages from all neighbors (assume the backend is PyTorch), with 5 uniformly chosen negative samples per edge:>>> sampler = dgl.dataloading.MultiLayerNeighborSampler([15, 10, 5]) >>> neg_sampler = dgl.dataloading.negative_sampler.Uniform(5) >>> dataloader = dgl.dataloading.EdgeDataLoader( ... g, train_eid, sampler, exclude='reverse_id', ... reverse_eids=reverse_eids, negative_sampler=neg_sampler, ... batch_size=1024, shuffle=True, drop_last=False, num_workers=4) >>> for input_nodes, pos_pair_graph, neg_pair_graph, blocks in dataloader: ... train_on(input_nodse, pair_graph, neg_pair_graph, blocks)
For heterogeneous graphs, the reverse of an edge may have a different edge type from the original edge. For instance, consider that you have an array of useritem clicks, representated by a user array
user
and an item arrayitem
. You may want to build a heterogeneous graph with a userclickitem relation and an itemclickedbyuser relation.>>> g = dgl.heterograph({ ... ('user', 'click', 'item'): (user, item), ... ('item', 'clickedby', 'user'): (item, user)})
To train a 3layer GNN for edge classification on a set of edges
train_eid
with typeclick
, you can write>>> sampler = dgl.dataloading.MultiLayerNeighborSampler([15, 10, 5]) >>> dataloader = dgl.dataloading.EdgeDataLoader( ... g, {'click': train_eid}, sampler, exclude='reverse_types', ... reverse_etypes={'click': 'clickedby', 'clickedby': 'click'}, ... batch_size=1024, shuffle=True, drop_last=False, num_workers=4) >>> for input_nodes, pair_graph, blocks in dataloader: ... train_on(input_nodes, pair_graph, blocks)
To train a 3layer GNN for link prediction on a set of edges
train_eid
with typeclick
, you can write>>> sampler = dgl.dataloading.MultiLayerNeighborSampler([15, 10, 5]) >>> neg_sampler = dgl.dataloading.negative_sampler.Uniform(5) >>> dataloader = dgl.dataloading.EdgeDataLoader( ... g, train_eid, sampler, exclude='reverse_types', ... reverse_etypes={'click': 'clickedby', 'clickedby': 'click'}, ... negative_sampler=neg_sampler, ... batch_size=1024, shuffle=True, drop_last=False, num_workers=4) >>> for input_nodes, pos_pair_graph, neg_pair_graph, blocks in dataloader: ... train_on(input_nodes, pair_graph, neg_pair_graph, blocks)
See also
dgl.dataloading.dataloader.EdgeCollator
Notes
Please refer to Minibatch Training Tutorials and User Guide Section 6 for usage.
For endtoend usages, please refer to the following tutorial/examples:
Edge classification on heterogeneous graph: GCMC
Link prediction on homogeneous graph: GraphSAGE for unsupervised learning
Link prediction on heterogeneous graph: RGCN for link prediction.

class
dgl.dataloading.pytorch.
GraphDataLoader
(dataset, collate_fn=None, **kwargs)[source]¶ PyTorch dataloader for batchiterating over a set of graphs, generating the batched graph and corresponding label tensor (if provided) of the said minibatch.
 Parameters
collate_fn (Function, default is None) – The customized collate function. Will use the default collate function if not given.
kwargs (dict) – Arguments being passed to
torch.utils.data.DataLoader
.
Examples
To train a GNN for graph classification on a set of graphs in
dataset
(assume the backend is PyTorch):>>> dataloader = dgl.dataloading.GraphDataLoader( ... dataset, batch_size=1024, shuffle=True, drop_last=False, num_workers=4) >>> for batched_graph, labels in dataloader: ... train_on(batched_graph, labels)
Neighbor Sampler¶
Neighbor samplers are classes that control the behavior of DataLoader
s
to sample neighbors. All of them inherit the base BlockSampler
class, but implement
different neighbor sampling strategies by overriding the sample_frontier
or
the sample_blocks
methods.

class
dgl.dataloading.neighbor.
BlockSampler
(num_layers, return_eids=False)[source]¶ Abstract class specifying the neighborhood sampling strategy for DGL data loaders.
The main method for BlockSampler is
sample_blocks()
, which generates a list of message flow graphs (MFGs) for a multilayer GNN given a set of seed nodes to have their outputs computed.The default implementation of
sample_blocks()
is to repeatnum_layers
times the following procedure from the last layer to the first layer:Obtain a frontier. The frontier is defined as a graph with the same nodes as the original graph but only the edges involved in message passing on the current layer. Customizable via
sample_frontier()
.Optionally, if the task is link prediction or edge classfication, remove edges connecting training node pairs. If the graph is undirected, also remove the reverse edges. This is controlled by the argument
exclude_eids
insample_blocks()
method.Convert the frontier into a MFG.
Optionally assign the IDs of the edges in the original graph selected in the first step to the MFG, controlled by the argument
return_eids
insample_blocks()
method.Prepend the MFG to the MFG list to be returned.
All subclasses should override
sample_frontier()
method while specifying the number of layers to sample innum_layers
argument. Parameters
Notes
For the concept of frontiers and MFGs, please refer to User Guide Section 6 and Minibatch Training Tutorials.

sample_blocks
(g, seed_nodes, exclude_eids=None)[source]¶ Generate the a list of MFGs given the destination nodes.
 Parameters
 Returns
The MFGs generated for computing the multilayer GNN output.
 Return type
Notes
For the concept of frontiers and MFGs, please refer to User Guide Section 6 and Minibatch Training Tutorials.

sample_frontier
(block_id, g, seed_nodes)[source]¶ Generate the frontier given the destination nodes.
The subclasses should override this function.
 Parameters
 Returns
The frontier generated for the current layer.
 Return type
Notes
For the concept of frontiers and MFGs, please refer to User Guide Section 6 and Minibatch Training Tutorials.

class
dgl.dataloading.neighbor.
MultiLayerNeighborSampler
(fanouts, replace=False, return_eids=False)[source]¶ Bases:
dgl.dataloading.dataloader.BlockSampler
Sampler that builds computational dependency of node representations via neighbor sampling for multilayer GNN.
This sampler will make every node gather messages from a fixed number of neighbors per edge type. The neighbors are picked uniformly.
 Parameters
fanouts (list[int] or list[dict[etype, int] or None]) –
List of neighbors to sample per edge type for each GNN layer, starting from the first layer.
If the graph is homogeneous, only an integer is needed for each layer.
If None is provided for one layer, all neighbors will be included regardless of edge types.
If 1 is provided for one edge type on one layer, then all inbound edges of that edge type will be included.
replace (bool, default True) – Whether to sample with replacement
return_eids (bool, default False) – Whether to return the edge IDs involved in message passing in the MFG. If True, the edge IDs will be stored as an edge feature named
dgl.EID
.
Examples
To train a 3layer GNN for node classification on a set of nodes
train_nid
on a homogeneous graph where each node takes messages from 5, 10, 15 neighbors for the first, second, and third layer respectively (assuming the backend is PyTorch):>>> sampler = dgl.dataloading.MultiLayerNeighborSampler([5, 10, 15]) >>> collator = dgl.dataloading.NodeCollator(g, train_nid, sampler) >>> dataloader = torch.utils.data.DataLoader( ... collator.dataset, collate_fn=collator.collate, ... batch_size=1024, shuffle=True, drop_last=False, num_workers=4) >>> for blocks in dataloader: ... train_on(blocks)
If training on a heterogeneous graph and you want different number of neighbors for each edge type, one should instead provide a list of dicts. Each dict would specify the number of neighbors to pick per edge type.
>>> sampler = dgl.dataloading.MultiLayerNeighborSampler([ ... {('user', 'follows', 'user'): 5, ... ('user', 'plays', 'game'): 4, ... ('game', 'playedby', 'user'): 3}] * 3)
Notes
For the concept of MFGs, please refer to User Guide Section 6 and Minibatch Training Tutorials.

sample_frontier
(block_id, g, seed_nodes)[source]¶ Generate the frontier given the destination nodes.
The subclasses should override this function.
 Parameters
 Returns
The frontier generated for the current layer.
 Return type
Notes
For the concept of frontiers and MFGs, please refer to User Guide Section 6 and Minibatch Training Tutorials.

class
dgl.dataloading.neighbor.
MultiLayerFullNeighborSampler
(n_layers, return_eids=False)[source]¶ Bases:
dgl.dataloading.neighbor.MultiLayerNeighborSampler
Sampler that builds computational dependency of node representations by taking messages from all neighbors for multilayer GNN.
This sampler will make every node gather messages from every single neighbor per edge type.
 Parameters
Examples
To train a 3layer GNN for node classification on a set of nodes
train_nid
on a homogeneous graph where each node takes messages from all neighbors for the first, second, and third layer respectively (assuming the backend is PyTorch):>>> sampler = dgl.dataloading.MultiLayerFullNeighborSampler(3) >>> collator = dgl.dataloading.NodeCollator(g, train_nid, sampler) >>> dataloader = torch.utils.data.DataLoader( ... collator.dataset, collate_fn=collator.collate, ... batch_size=1024, shuffle=True, drop_last=False, num_workers=4) >>> for blocks in dataloader: ... train_on(blocks)
Notes
For the concept of MFGs, please refer to User Guide Section 6 and Minibatch Training Tutorials.
Collators¶
Collators are platformagnostic classes that generates the minibatches given the graphs and indices to sample from.

class
dgl.dataloading.
NodeCollator
(g, nids, block_sampler)[source]¶ DGL collator to combine nodes and their computation dependencies within a minibatch for training node classification or regression on a single graph with neighborhood sampling.
 Parameters
Examples
To train a 3layer GNN for node classification on a set of nodes
train_nid
on a homogeneous graph where each node takes messages from all neighbors (assume the backend is PyTorch):>>> sampler = dgl.dataloading.MultiLayerNeighborSampler([15, 10, 5]) >>> collator = dgl.dataloading.NodeCollator(g, train_nid, sampler) >>> dataloader = torch.utils.data.DataLoader( ... collator.dataset, collate_fn=collator.collate, ... batch_size=1024, shuffle=True, drop_last=False, num_workers=4) >>> for input_nodes, output_nodes, blocks in dataloader: ... train_on(input_nodes, output_nodes, blocks)
Notes
For the concept of MFGs, please refer to User Guide Section 6 and Minibatch Training Tutorials.

class
dgl.dataloading.
EdgeCollator
(g, eids, block_sampler, g_sampling=None, exclude=None, reverse_eids=None, reverse_etypes=None, negative_sampler=None)[source]¶ DGL collator to combine edges and their computation dependencies within a minibatch for training edge classification, edge regression, or link prediction on a single graph with neighborhood sampling.
Given a set of edges, the collate function will yield
A tensor of input nodes necessary for computing the representation on edges, or a dictionary of node type names and such tensors.
A subgraph that contains only the edges in the minibatch and their incident nodes. Note that the graph has an identical metagraph with the original graph.
If a negative sampler is given, another graph that contains the “negative edges”, connecting the source and destination nodes yielded from the given negative sampler.
A list of MFGs necessary for computing the representation of the incident nodes of the edges in the minibatch.
 Parameters
g (DGLGraph) – The graph from which the edges are iterated in minibatches and the subgraphs are generated.
eids (Tensor or dict[etype, Tensor]) – The edge set in graph
g
to compute outputs.block_sampler (dgl.dataloading.BlockSampler) – The neighborhood sampler.
g_sampling (DGLGraph, optional) –
The graph where neighborhood sampling and message passing is performed.
Note that this is not necessarily the same as
g
.If None, assume to be the same as
g
.exclude (str, optional) –
Whether and how to exclude dependencies related to the sampled edges in the minibatch. Possible values are
None, which excludes nothing.
'reverse_id'
, which excludes the reverse edges of the sampled edges. The said reverse edges have the same edge type as the sampled edges. Only works on edge types whose source node type is the same as its destination node type.'reverse_types'
, which excludes the reverse edges of the sampled edges. The said reverse edges have different edge types from the sampled edges.
If
g_sampling
is given,exclude
is ignored and will be alwaysNone
.reverse_eids (Tensor or dict[etype, Tensor], optional) –
A tensor of reverse edge ID mapping. The ith element indicates the ID of the ith edge’s reverse edge.
If the graph is heterogeneous, this argument requires a dictionary of edge types and the reverse edge ID mapping tensors.
Required and only used when
exclude
is set toreverse_id
.For heterogeneous graph this will be a dict of edge type and edge IDs. Note that only the edge types whose source node type is the same as destination node type are needed.
reverse_etypes (dict[etype, etype], optional) –
The mapping from the edge type to its reverse edge type.
Required and only used when
exclude
is set toreverse_types
.negative_sampler (callable, optional) –
The negative sampler. Can be omitted if no negative sampling is needed.
The negative sampler must be a callable that takes in the following arguments:
The original (heterogeneous) graph.
The ID array of sampled edges in the minibatch, or the dictionary of edge types and ID array of sampled edges in the minibatch if the graph is heterogeneous.
It should return
A pair of source and destination node ID arrays as negative samples, or a dictionary of edge types and such pairs if the graph is heterogenenous.
A set of builtin negative samplers are provided in the negative sampling module.
Examples
The following example shows how to train a 3layer GNN for edge classification on a set of edges
train_eid
on a homogeneous undirected graph. Each node takes messages from all neighbors.Say that you have an array of source node IDs
src
and another array of destination node IDsdst
. One can make it bidirectional by adding another set of edges that connects fromdst
tosrc
:>>> g = dgl.graph((torch.cat([src, dst]), torch.cat([dst, src])))
One can then know that the ID difference of an edge and its reverse edge is
E
, whereE
is the length of your source/destination array. The reverse edge mapping can be obtained by>>> E = len(src) >>> reverse_eids = torch.cat([torch.arange(E, 2 * E), torch.arange(0, E)])
Note that the sampled edges as well as their reverse edges are removed from computation dependencies of the incident nodes. This is a common trick to avoid information leakage.
>>> sampler = dgl.dataloading.MultiLayerNeighborSampler([15, 10, 5]) >>> collator = dgl.dataloading.EdgeCollator( ... g, train_eid, sampler, exclude='reverse_id', ... reverse_eids=reverse_eids) >>> dataloader = torch.utils.data.DataLoader( ... collator.dataset, collate_fn=collator.collate, ... batch_size=1024, shuffle=True, drop_last=False, num_workers=4) >>> for input_nodes, pair_graph, blocks in dataloader: ... train_on(input_nodes, pair_graph, blocks)
To train a 3layer GNN for link prediction on a set of edges
train_eid
on a homogeneous graph where each node takes messages from all neighbors (assume the backend is PyTorch), with 5 uniformly chosen negative samples per edge:>>> sampler = dgl.dataloading.MultiLayerNeighborSampler([15, 10, 5]) >>> neg_sampler = dgl.dataloading.negative_sampler.Uniform(5) >>> collator = dgl.dataloading.EdgeCollator( ... g, train_eid, sampler, exclude='reverse_id', ... reverse_eids=reverse_eids, negative_sampler=neg_sampler, >>> dataloader = torch.utils.data.DataLoader( ... collator.dataset, collate_fn=collator.collate, ... batch_size=1024, shuffle=True, drop_last=False, num_workers=4) >>> for input_nodes, pos_pair_graph, neg_pair_graph, blocks in dataloader: ... train_on(input_nodse, pair_graph, neg_pair_graph, blocks)
For heterogeneous graphs, the reverse of an edge may have a different edge type from the original edge. For instance, consider that you have an array of useritem clicks, representated by a user array
user
and an item arrayitem
. You may want to build a heterogeneous graph with a userclickitem relation and an itemclickedbyuser relation.>>> g = dgl.heterograph({ ... ('user', 'click', 'item'): (user, item), ... ('item', 'clickedby', 'user'): (item, user)})
To train a 3layer GNN for edge classification on a set of edges
train_eid
with typeclick
, you can write>>> sampler = dgl.dataloading.MultiLayerNeighborSampler([15, 10, 5]) >>> collator = dgl.dataloading.EdgeCollator( ... g, {'click': train_eid}, sampler, exclude='reverse_types', ... reverse_etypes={'click': 'clickedby', 'clickedby': 'click'}) >>> dataloader = torch.utils.data.DataLoader( ... collator.dataset, collate_fn=collator.collate, ... batch_size=1024, shuffle=True, drop_last=False, num_workers=4) >>> for input_nodes, pair_graph, blocks in dataloader: ... train_on(input_nodes, pair_graph, blocks)
To train a 3layer GNN for link prediction on a set of edges
train_eid
with typeclick
, you can write>>> sampler = dgl.dataloading.MultiLayerNeighborSampler([15, 10, 5]) >>> neg_sampler = dgl.dataloading.negative_sampler.Uniform(5) >>> collator = dgl.dataloading.EdgeCollator( ... g, train_eid, sampler, exclude='reverse_types', ... reverse_etypes={'click': 'clickedby', 'clickedby': 'click'}, ... negative_sampler=neg_sampler) >>> dataloader = torch.utils.data.DataLoader( ... collator.dataset, collate_fn=collator.collate, ... batch_size=1024, shuffle=True, drop_last=False, num_workers=4) >>> for input_nodes, pos_pair_graph, neg_pair_graph, blocks in dataloader: ... train_on(input_nodes, pair_graph, neg_pair_graph, blocks)
Notes
For the concept of MFGs, please refer to User Guide Section 6 and Minibatch Training Tutorials.

class
dgl.dataloading.
GraphCollator
[source]¶ Given a set of graphs as well as their graphlevel data, the collate function will batch the graphs into a batched graph, and stack the tensors into a single bigger tensor. If the example is a container (such as sequences or mapping), the collate function preserves the structure and collates each of the elements recursively.
If the set of graphs has no graphlevel data, the collate function will yield a batched graph.
Examples
To train a GNN for graph classification on a set of graphs in
dataset
(assume the backend is PyTorch):>>> dataloader = dgl.dataloading.GraphDataLoader( ... dataset, batch_size=1024, shuffle=True, drop_last=False, num_workers=4) >>> for batched_graph, labels in dataloader: ... train_on(batched_graph, labels)
Negative Samplers for Link Prediction¶
Negative samplers are classes that control the behavior of the EdgeDataLoader
to generate negative edges.

class
dgl.dataloading.negative_sampler.
Uniform
(k)[source]¶ Negative sampler that randomly chooses negative destination nodes for each source node according to a uniform distribution.
For each edge
(u, v)
of type(srctype, etype, dsttype)
, DGL generatesk
pairs of negative edges(u, v')
, wherev'
is chosen uniformly from all the nodes of typedsttype
. The resulting edges will also have type(srctype, etype, dsttype)
. Parameters
k (int) – The number of negative examples per edge.
Examples
>>> g = dgl.graph(([0, 1, 2], [1, 2, 3])) >>> neg_sampler = dgl.dataloading.negative_sampler.Uniform(2) >>> neg_sampler(g, [0, 1]) (tensor([0, 0, 1, 1]), tensor([1, 0, 2, 3]))
Async Copying to/from GPUs¶
Data can be copied from the CPU to the GPU
while the GPU is being used for
computation, using the AsyncTransferer
.
For the transfer to be fully asynchronous, the context the
AsyncTranserer
is created with must be a GPU context, and the input tensor must be in
pinned memory.

class
dgl.dataloading.
AsyncTransferer
(device)[source]¶ Class for initiating asynchronous copies to the GPU on a second GPU stream.
To initiate a transfer to a GPU:
>>> tensor_cpu = torch.ones(100000).pin_memory() >>> transferer = dgl.dataloading.AsyncTransferer(torch.device(0)) >>> future = transferer.async_copy(tensor_cpu, torch.device(0))
And then to wait for the transfer to finish and get a copy of the tensor on the GPU.
>>> tensor_gpu = future.wait()

__init__
(device)[source]¶ Create a new AsyncTransferer object.
 Parameters
device (Device or context object.) – The context in which the second stream will be created. Must be a GPU context for the copy to be asynchronous.

async_copy
(tensor, device)[source]¶ Initiate an asynchronous copy on the internal stream. For this call to be asynchronous, the context the AsyncTranserer is created with must be a GPU context, and the input tensor must be in pinned memory.
Currently, only transfers to the GPU are supported.
 Parameters
tensor (Tensor) – The tensor to transfer.
device (Device or context object.) – The context to transfer to.
 Returns
A Transfer object that can be waited on to get the tensor in the new context.
 Return type
Transfer
