dgl.function¶
In DGL, message passing is expressed by two APIs:
send(edges, message_func)
for computing the messages along the given edges.recv(nodes, reduce_func)
for collecting the incoming messages, perform aggregation and so on.
Although the two-stage abstraction can cover all the models that are defined in the message passing paradigm, it is inefficient because it requires storing explicit messages. See the DGL blog post for more details and performance results.
Our solution, also explained in the blog post, is to fuse the two stages into one kernel so no explicit messages are generated and stored. To achieve this, we recommend using our built-in message and reduce functions so that DGL can analyze and map them to fused dedicated kernels. Here are some examples (in PyTorch syntax).
import dgl
import dgl.function as fn
import torch as th
g = ... # create a DGLGraph
g.ndata['h'] = th.randn((g.number_of_nodes(), 10)) # each node has feature size 10
g.edata['w'] = th.randn((g.number_of_edges(), 1)) # each edge has feature size 1
# collect features from source nodes and aggregate them in destination nodes
g.update_all(fn.copy_u('h', 'm'), fn.sum('m', 'h_sum'))
# multiply source node features with edge weights and aggregate them in destination nodes
g.update_all(fn.u_mul_e('h', 'w', 'm'), fn.max('m', 'h_max'))
# compute edge embedding by multiplying source and destination node embeddings
g.apply_edges(fn.u_mul_v('h', 'h', 'w_new'))
fn.copy_u
, fn.u_mul_e
, fn.u_mul_v
are built-in message functions, while fn.sum
and fn.max
are built-in reduce functions. We use u
, v
and e
to represent
source nodes, destination nodes, and edges among them, respectively. Hence, copy_u
copies the source
node data as the messages, u_mul_e
multiplies source node features with edge features, for example.
To define a unary message function (e.g. copy_u
) specify one input feature name and one output
message name. To define a binary message function (e.g. u_mul_e
) specify
two input feature names and one output message name. During the computation,
the message function will read the data under the given names, perform computation, and return
the output using the output name. For example, the above fn.u_mul_e('h', 'w', 'm')
is
the same as the following user-defined function:
def udf_u_mul_e(edges):
return {'m' : edges.src['h'] * edges.data['w']}
To define a reduce function, one input message name and one output node feature name
need to be specified. For example, the above fn.max('m', 'h_max')
is the same as the
following user-defined function:
def udf_max(nodes):
return {'h_max' : th.max(nodes.mailbox['m'], 1)[0]}
Broadcasting is supported for binary message function, which means the tensor arguments
can be automatically expanded to be of equal sizes. The supported broadcasting semantic
is standard and matches NumPy
and PyTorch. If you are not familiar
with broadcasting, see the linked topics to learn more. In the
above example, fn.u_mul_e
will perform broadcasted multiplication automatically because
the node feature 'h'
and the edge feature 'w'
are of different shapes, but they can be broadcast.
All DGL’s built-in functions support both CPU and GPU and backward computation so they
can be used in any autograd system. Also, built-in functions can be used not only in update_all
or apply_edges
as shown in the example, but wherever message and reduce functions are
required (e.g. pull
, push
, send_and_recv
).
DGL Built-in Function¶
Here is a cheatsheet of all the DGL built-in functions.
Category |
Functions |
Memo |
---|---|---|
Unary message function |
|
|
|
||
|
alias of |
|
|
alias of |
|
Binary message function |
|
|
|
||
|
||
|
||
|
||
|
||
|
alias of |
|
Reduce function |
|
|
|
||
|
||
|
Message functions¶
|
Builtin message function that computes message using source node feature. |
|
Builtin message function that computes message using edge feature. |
|
Builtin message function that computes message by performing binary operation mul between src feature and edge feature. |
|
Builtin message function that computes message using source node feature. |
|
Builtin message function that computes message using edge feature. |
|
Builtin message function that computes a message on an edge by performing element-wise add between features of u and v if the features have the same shape; otherwise, it first broadcasts the features to a new shape and performs the element-wise operation. |
|
Builtin message function that computes a message on an edge by performing element-wise sub between features of u and v if the features have the same shape; otherwise, it first broadcasts the features to a new shape and performs the element-wise operation. |
|
Builtin message function that computes a message on an edge by performing element-wise mul between features of u and v if the features have the same shape; otherwise, it first broadcasts the features to a new shape and performs the element-wise operation. |
|
Builtin message function that computes a message on an edge by performing element-wise div between features of u and v if the features have the same shape; otherwise, it first broadcasts the features to a new shape and performs the element-wise operation. |
|
Builtin message function that computes a message on an edge by performing element-wise add between features of u and e if the features have the same shape; otherwise, it first broadcasts the features to a new shape and performs the element-wise operation. |
|
Builtin message function that computes a message on an edge by performing element-wise sub between features of u and e if the features have the same shape; otherwise, it first broadcasts the features to a new shape and performs the element-wise operation. |
|
Builtin message function that computes a message on an edge by performing element-wise mul between features of u and e if the features have the same shape; otherwise, it first broadcasts the features to a new shape and performs the element-wise operation. |
|
Builtin message function that computes a message on an edge by performing element-wise div between features of u and e if the features have the same shape; otherwise, it first broadcasts the features to a new shape and performs the element-wise operation. |
|
Builtin message function that computes a message on an edge by performing element-wise add between features of v and u if the features have the same shape; otherwise, it first broadcasts the features to a new shape and performs the element-wise operation. |
|
Builtin message function that computes a message on an edge by performing element-wise sub between features of v and u if the features have the same shape; otherwise, it first broadcasts the features to a new shape and performs the element-wise operation. |
|
Builtin message function that computes a message on an edge by performing element-wise mul between features of v and u if the features have the same shape; otherwise, it first broadcasts the features to a new shape and performs the element-wise operation. |
|
Builtin message function that computes a message on an edge by performing element-wise div between features of v and u if the features have the same shape; otherwise, it first broadcasts the features to a new shape and performs the element-wise operation. |
|
Builtin message function that computes a message on an edge by performing element-wise add between features of v and e if the features have the same shape; otherwise, it first broadcasts the features to a new shape and performs the element-wise operation. |
|
Builtin message function that computes a message on an edge by performing element-wise sub between features of v and e if the features have the same shape; otherwise, it first broadcasts the features to a new shape and performs the element-wise operation. |
|
Builtin message function that computes a message on an edge by performing element-wise mul between features of v and e if the features have the same shape; otherwise, it first broadcasts the features to a new shape and performs the element-wise operation. |
|
Builtin message function that computes a message on an edge by performing element-wise div between features of v and e if the features have the same shape; otherwise, it first broadcasts the features to a new shape and performs the element-wise operation. |
|
Builtin message function that computes a message on an edge by performing element-wise add between features of e and u if the features have the same shape; otherwise, it first broadcasts the features to a new shape and performs the element-wise operation. |
|
Builtin message function that computes a message on an edge by performing element-wise sub between features of e and u if the features have the same shape; otherwise, it first broadcasts the features to a new shape and performs the element-wise operation. |
|
Builtin message function that computes a message on an edge by performing element-wise mul between features of e and u if the features have the same shape; otherwise, it first broadcasts the features to a new shape and performs the element-wise operation. |
|
Builtin message function that computes a message on an edge by performing element-wise div between features of e and u if the features have the same shape; otherwise, it first broadcasts the features to a new shape and performs the element-wise operation. |
|
Builtin message function that computes a message on an edge by performing element-wise add between features of e and v if the features have the same shape; otherwise, it first broadcasts the features to a new shape and performs the element-wise operation. |
|
Builtin message function that computes a message on an edge by performing element-wise sub between features of e and v if the features have the same shape; otherwise, it first broadcasts the features to a new shape and performs the element-wise operation. |
|
Builtin message function that computes a message on an edge by performing element-wise mul between features of e and v if the features have the same shape; otherwise, it first broadcasts the features to a new shape and performs the element-wise operation. |
|
Builtin message function that computes a message on an edge by performing element-wise div between features of e and v if the features have the same shape; otherwise, it first broadcasts the features to a new shape and performs the element-wise operation. |
|
Builtin message function that computes a message on an edge by performing element-wise dot between features of u and v if the features have the same shape; otherwise, it first broadcasts the features to a new shape and performs the element-wise operation. |
|
Builtin message function that computes a message on an edge by performing element-wise dot between features of u and e if the features have the same shape; otherwise, it first broadcasts the features to a new shape and performs the element-wise operation. |
|
Builtin message function that computes a message on an edge by performing element-wise dot between features of v and e if the features have the same shape; otherwise, it first broadcasts the features to a new shape and performs the element-wise operation. |
|
Builtin message function that computes a message on an edge by performing element-wise dot between features of v and u if the features have the same shape; otherwise, it first broadcasts the features to a new shape and performs the element-wise operation. |
|
Builtin message function that computes a message on an edge by performing element-wise dot between features of e and u if the features have the same shape; otherwise, it first broadcasts the features to a new shape and performs the element-wise operation. |
|
Builtin message function that computes a message on an edge by performing element-wise dot between features of e and v if the features have the same shape; otherwise, it first broadcasts the features to a new shape and performs the element-wise operation. |
Reduce functions¶
|
Builtin reduce function that aggregates messages by sum. |
|
Builtin reduce function that aggregates messages by max. |
|
Builtin reduce function that aggregates messages by min. |
|
Builtin reduce function that aggregates messages by mean. |