[Doc] Scan the API docs and make many changes (#2080)

* WIP: api

* dgl.sampling, dgl.data

* dgl.sampling; dgl.dataloading

* sampling packages

* convert

* subgraph

* deprecate

* subgraph APIs

* All docstrings for convert/subgraph/transform

* almost all funcs under dgl namespace

* WIP: DGLGraph

* done graph query

* message passing functions

* lint

* fix merge error

* fix test

* lint

* fix
Co-authored-by: Quan Gan <coin2028@hotmail.com>

python/dgl/data/graph_serialize.py

 """For Graph Serialization"""
 from __future__ import absolute_import
 import os
-from ..base import dgl_warning
+from ..base import dgl_warning, DGLError
 from ..heterograph import DGLHeteroGraph
 from .._ffi.object import ObjectBase, register_object
 from .._ffi.function import _init_api
@@ -66,16 +66,23 @@ class GraphData(ObjectBase):
 
 
 def save_graphs(filename, g_list, labels=None):
-    r"""
-    Save DGLGraphs and graph labels to file
+    r"""Save graphs and optionally their labels to file.
+
+    Besides saving to local files, DGL supports writing the graphs directly
+    to S3 (by providing a ``"s3://..."`` path) or to HDFS (by providing
+    ``"hdfs://..."`` a path).
+
+    The function saves both the graph structure and node/edge features to file
+    in DGL's own binary format. For graph-level features, pass them via
+    the :attr:`labels` argument.
 
     Parameters
     ----------
     filename : str
-        File name to store graphs.
+        The file name to store the graphs and labels.
     g_list: list
-        DGLGraph or list of DGLGraph/DGLHeteroGraph
-    labels: dict[str, tensor]
+        The graphs to be saved.
+    labels: dict[str, Tensor]
         labels should be dict of tensors, with str as keys
 
     Examples
@@ -83,7 +90,7 @@ def save_graphs(filename, g_list, labels=None):
     >>> import dgl
     >>> import torch as th
 
-    Create :code:`DGLGraph`/:code:`DGLHeteroGraph` objects and initialize node
+    Create :class:`DGLGraph` objects and initialize node
     and edge features.
 
     >>> g1 = dgl.graph(([0, 1, 2], [1, 2, 3]))
@@ -96,55 +103,66 @@ def save_graphs(filename, g_list, labels=None):
     >>> graph_labels = {"glabel": th.tensor([0, 1])}
     >>> save_graphs("./data.bin", [g1, g2], graph_labels)
 
+    See Also
+    --------
+    load_graphs
     """
     # if it is local file, do some sanity check
     if filename.startswith('s3://') is False:
-        assert not os.path.isdir(filename), "filename {} is an existing directory.".format(filename)
+        if os.path.isdir(filename):
+            raise DGLError("Filename {} is an existing directory.".format(filename))
         f_path, _ = os.path.split(filename)
         if not os.path.exists(f_path):
             os.makedirs(f_path)
 
     g_sample = g_list[0] if isinstance(g_list, list) else g_list
-    if type(g_sample) == DGLHeteroGraph: # Doesn't support DGLHeteroGraph's derived class
+    if type(g_sample) == DGLHeteroGraph:  # Doesn't support DGLHeteroGraph's derived class
         save_heterographs(filename, g_list, labels)
     else:
-        raise Exception(
-            "Invalid argument g_list. Must be a DGLGraph or a list of DGLGraphs/DGLHeteroGraphs")
+        raise DGLError(
+            "Invalid argument g_list. Must be a DGLGraph or a list of DGLGraphs.")
 
 
 
 def load_graphs(filename, idx_list=None):
-    """
-    Load DGLGraphs from file
+    """Load graphs and optionally their labels from file saved by :func:`save_graphs`.
+
+    Besides loading from local files, DGL supports loading the graphs directly
+    from S3 (by providing a ``"s3://..."`` path) or from HDFS (by providing
+    ``"hdfs://..."`` a path).
 
     Parameters
     ----------
     filename: str
-        filename to load graphs
-    idx_list: list of int
-        list of index of graph to be loaded. If not specified, will
-        load all graphs from file
+        The file name to load graphs from.
+    idx_list: list[int], optional
+        The indices of the graphs to be loaded if the file contains multiple graphs.
+        Default is loading all the graphs stored in the file.
 
     Returns
     --------
-    graph_list: list of DGLGraphs / DGLHeteroGraph
+    graph_list: list[DGLGraph]
         The loaded graphs.
     labels: dict[str, Tensor]
         The graph labels stored in file. If no label is stored, the dictionary is empty.
-        Regardless of whether the ``idx_list`` argument is given or not, the returned dictionary
-        always contains labels of all the graphs.
+        Regardless of whether the ``idx_list`` argument is given or not,
+        the returned dictionary always contains the labels of all the graphs.
 
     Examples
     ----------
-    Following the example in save_graphs.
+    Following the example in :func:`save_graphs`.
 
     >>> from dgl.data.utils import load_graphs
     >>> glist, label_dict = load_graphs("./data.bin") # glist will be [g1, g2]
     >>> glist, label_dict = load_graphs("./data.bin", [0]) # glist will be [g1]
 
+    See Also
+    --------
+    save_graphs
     """
     # if it is local file, do some sanity check
-    assert filename.startswith('s3://') or os.path.exists(filename), "file {} does not exist.".format(filename)
+    if not (filename.startswith('s3://') or os.path.exists(filename)):
+        raise DGLError("File {} does not exist.".format(filename))
 
     version = _CAPI_GetFileVersion(filename)
     if version == 1:
@@ -155,7 +173,7 @@ def load_graphs(filename, idx_list=None):
     elif version == 2:
         return load_graph_v2(filename, idx_list)
     else:
-        raise Exception("Invalid DGL Version Number")
+        raise DGLError("Invalid DGL Version Number.")
 
 
 def load_graph_v2(filename, idx_list=None):

python/dgl/dataloading/__init__.py

-"""Classes that involves iterating over nodes or edges in a graph and generates
-computation dependency of necessary nodes with neighborhood sampling methods.
+"""The ``dgl.dataloading`` package contains:
 
-This includes
-
-* :py:class:`~dgl.dataloading.pytorch.NodeDataLoader` for iterating over the nodes in
-  a graph in minibatches.
-
-* :py:class:`~dgl.dataloading.pytorch.EdgeDataLoader` for iterating over the edges in
-  a graph in minibatches.
+* 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 multi-layer GNNs.
 
 * Negative samplers for link prediction.
 
-NOTE: this module is experimental and the interfaces may be subject to changes in
-future releases.
+For a holistic explanation on how different components work together.
+Read the user guide :ref:`guide-minibatch`.
+
+.. note::
+    This package is experimental and the interfaces may be subject
+    to changes in future releases. It currently only has implementations in PyTorch.
 """
 from .neighbor import *
 from .dataloader import *

python/dgl/generators.py

 """Module for various graph generator functions."""
-# pylint: disable= dangerous-default-value
 
 from . import backend as F
 from . import convert
@@ -7,13 +6,14 @@ from . import random
 
 __all__ = ['rand_graph', 'rand_bipartite']
 
-def rand_graph(num_nodes, num_edges, idtype=F.int64, device=F.cpu(),
-               formats=['coo', 'csr', 'csc']):
-    """Generate a random graph of the given number of nodes/edges.
+def rand_graph(num_nodes, num_edges, idtype=F.int64, device=F.cpu()):
+    """Generate a random graph of the given number of nodes/edges and return.
 
-    It uniformly chooses ``num_edges`` from all pairs and form a graph.
+    It uniformly chooses ``num_edges`` from all possible node pairs and form a graph.
+    The random choice is without replacement, which means there will be no multi-edge
+    in the resulting graph.
 
-    TODO(minjie): support RNG as one of the arguments.
+    To control the randomness, set the random seed via :func:`dgl.seed`.
 
     Parameters
     ----------
@@ -22,34 +22,51 @@ def rand_graph(num_nodes, num_edges, idtype=F.int64, device=F.cpu(),
     num_edges : int
         The number of edges
     idtype : int32, int64, optional
-        Integer ID type. Must be int32 or int64. Default: int64.
+        The data type for storing the structure-related graph information
+        such as node and edge IDs. It should be a framework-specific data type object
+        (e.g., torch.int32). By default, DGL uses int64.
     device : Device context, optional
-        Device on which the graph is created. Default: CPU.
-    formats : str or list of str
-        It can be ``'coo'``/``'csr'``/``'csc'`` or a sublist of them,
-        Force the storage formats.  Default: ``['coo', 'csr', 'csc']``.
+        The device of the resulting graph. It should be a framework-specific device
+        object (e.g., torch.device). By default, DGL stores the graph on CPU.
 
     Returns
     -------
-    DGLHeteroGraph
-        Generated random graph.
+    DGLGraph
+        The generated random graph.
+
+    See Also
+    --------
+    rand_bipartite
+
+    Examples
+    --------
+    >>> import dgl
+    >>> dgl.rand_graph(100, 10)
+    Graph(num_nodes=100, num_edges=10,
+          ndata_schemes={}
+          edata_schemes={})
     """
+    #TODO(minjie): support RNG as one of the arguments.
     eids = random.choice(num_nodes * num_nodes, num_edges, replace=False)
-    rows = F.copy_to(F.astype(eids / num_nodes, idtype), device)
-    cols = F.copy_to(F.astype(eids % num_nodes, idtype), device)
-    g = convert.graph((rows, cols),
-                      num_nodes=num_nodes,
-                      idtype=idtype, device=device)
-    return g.formats(formats)
+    eids = F.zerocopy_to_numpy(eids)
+    rows = F.zerocopy_from_numpy(eids // num_nodes)
+    cols = F.zerocopy_from_numpy(eids % num_nodes)
+    rows = F.copy_to(F.astype(rows, idtype), device)
+    cols = F.copy_to(F.astype(cols, idtype), device)
+    return convert.graph((rows, cols),
+                         num_nodes=num_nodes,
+                         idtype=idtype, device=device)
 
 def rand_bipartite(utype, etype, vtype,
                    num_src_nodes, num_dst_nodes, num_edges,
-                   idtype=F.int64, device=F.cpu(),
-                   formats=['csr', 'coo', 'csc']):
-    """Generate a random bipartite graph of the given number of src/dst nodes and
-    number of edges.
+                   idtype=F.int64, device=F.cpu()):
+    """Generate a random uni-directional bipartite graph and return.
+
+    It uniformly chooses ``num_edges`` from all possible node pairs and form a graph.
+    The random choice is without replacement, which means there will be no multi-edge
+    in the resulting graph.
 
-    It uniformly chooses ``num_edges`` from all pairs and form a graph.
+    To control the randomness, set the random seed via :func:`dgl.seed`.
 
     Parameters
     ----------
@@ -60,28 +77,43 @@ def rand_bipartite(utype, etype, vtype,
     vtype : str, optional
         The name of the destination node type.
     num_src_nodes : int
-        The number of source nodes, the :math:`|U|` in :math:`G=(U,V,E)`.
+        The number of source nodes.
     num_dst_nodes : int
-        The number of destination nodes, the :math:`|V|` in :math:`G=(U,V,E)`.
+        The number of destination nodes.
     num_edges : int
         The number of edges
     idtype : int32, int64, optional
-        Integer ID type. Must be int32 or int64. Default: int64.
+        The data type for storing the structure-related graph information
+        such as node and edge IDs. It should be a framework-specific data type object
+        (e.g., torch.int32). By default, DGL uses int64.
     device : Device context, optional
-        Device on which the graph is created. Default: CPU.
-    formats : str or list of str
-        It can be ``'coo'``/``'csr'``/``'csc'`` or a sublist of them,
-        Force the storage formats.  Default: ``['coo', 'csr', 'csc']``.
+        The device of the resulting graph. It should be a framework-specific device
+        object (e.g., torch.device). By default, DGL stores the graph on CPU.
 
     Returns
     -------
-    DGLHeteroGraph
-        Generated random bipartite graph.
+    DGLGraph
+        The generated random bipartite graph.
+
+    See Also
+    --------
+    rand_graph
+
+    Examples
+    --------
+    >>> import dgl
+    >>> dgl.rand_bipartite('user', 'buys', 'game', 50, 100, 10)
+    Graph(num_nodes={'game': 100, 'user': 50},
+          num_edges={('user', 'buys', 'game'): 10},
+          metagraph=[('user', 'game', 'buys')])
     """
+    #TODO(minjie): support RNG as one of the arguments.
     eids = random.choice(num_src_nodes * num_dst_nodes, num_edges, replace=False)
-    rows = F.copy_to(F.astype(eids / num_dst_nodes, idtype), device)
-    cols = F.copy_to(F.astype(eids % num_dst_nodes, idtype), device)
-    g = convert.heterograph({(utype, etype, vtype): (rows, cols)},
-                            {utype: num_src_nodes, vtype: num_dst_nodes},
-                            idtype=idtype, device=device)
-    return g.formats(formats)
+    eids = F.zerocopy_to_numpy(eids)
+    rows = F.zerocopy_from_numpy(eids // num_dst_nodes)
+    cols = F.zerocopy_from_numpy(eids % num_dst_nodes)
+    rows = F.copy_to(F.astype(rows, idtype), device)
+    cols = F.copy_to(F.astype(cols, idtype), device)
+    return convert.heterograph({(utype, etype, vtype): (rows, cols)},
+                               {utype: num_src_nodes, vtype: num_dst_nodes},
+                               idtype=idtype, device=device)

python/dgl/nn/__init__.py

-"""Package for neural network common components."""
+"""The ``dgl.nn`` package contains framework-specific implementations for
+common Graph Neural Network layers (or module in PyTorch, Block in MXNet).
+Users can directly import ``dgl.nn.<layer_name>`` (e.g., ``dgl.nn.GraphConv``),
+and the package will dispatch the layer name to the actual implementation
+according to the backend framework currently in use.
+
+Note that there are coverage differences among frameworks. If you encounter
+an ``ImportError: cannot import name 'XXX'`` error, that means the layer is
+not available to the current backend. If you wish a module to appear in DGL,
+please `create an issue <https://github.com/dmlc/dgl/issues>`_ started with
+"[Feature Request] NN Module XXXModel". If you want to contribute a NN module,
+please `create a pull request <https://github.com/dmlc/dgl/pulls>`_ started
+with "[NN] XXX module".
+"""
+
 import importlib
 import sys
 import os

python/dgl/random.py

@@ -8,14 +8,12 @@ from . import ndarray as nd
 __all__ = ['seed']
 
 def seed(val):
-    """Set the seed of randomized methods in DGL.
-
-    The randomized methods include various samplers and random walk routines.
+    """Set the random seed of DGL.
 
     Parameters
     ----------
     val : int
-        The seed
+        The seed.
     """
     _CAPI_SetSeed(val)
 
@@ -41,8 +39,6 @@ def choice(a, size, replace=True, prob=None):  # pylint: disable=invalid-name
 
     It out-performs numpy for non-uniform sampling in general cases.
 
-    TODO(minjie): support RNG as one of the arguments.
-
     Parameters
     ----------
     a : 1-D tensor or int
@@ -61,6 +57,7 @@ def choice(a, size, replace=True, prob=None):  # pylint: disable=invalid-name
     samples : 1-D tensor
         The generated random samples
     """
+    #TODO(minjie): support RNG as one of the arguments.
     if isinstance(size, tuple):
         num = np.prod(size)
     else:

python/dgl/readout.py

@@ -28,9 +28,9 @@ def readout_nodes(graph, feat, weight=None, *, op='sum', ntype=None):
     feat : str
         Node feature name.
     weight : str, optional
-        Node weight name. If None, no weighting will be performed,
-        otherwise, weight each node feature with field :attr:`feat`.
-        for aggregation. The weight feature shape must be compatible with
+        Node weight name. None means aggregating without weights.
+        Otherwise, multiply each node feature by node feature :attr:`weight`
+        before aggregation. The weight feature shape must be compatible with
         an element-wise multiplication with the feature tensor.
     op : str, optional
         Readout operator. Can be 'sum', 'max', 'min', 'mean'.
@@ -39,7 +39,7 @@ def readout_nodes(graph, feat, weight=None, *, op='sum', ntype=None):
 
     Returns
     -------
-    tensor
+    Tensor
         Result tensor.
 
     Examples
@@ -101,22 +101,28 @@ def readout_edges(graph, feat, weight=None, *, op='sum', etype=None):
     Parameters
     ----------
     graph : DGLGraph.
-        Input graph.
+        The input graph.
     feat : str
-        Edge feature name.
+        The edge feature name.
     weight : str, optional
-        Edge weight name. If None, no weighting will be performed,
+        The edge weight feature name. If None, no weighting will be performed,
         otherwise, weight each edge feature with field :attr:`feat`.
         for summation. The weight feature shape must be compatible with
         an element-wise multiplication with the feature tensor.
     op : str, optional
         Readout operator. Can be 'sum', 'max', 'min', 'mean'.
-    etype : str, tuple of str, optional
-        Edge type. Can be omitted if there is only one edge type in the graph.
+    etype : str or (str, str, str), optional
+        The type names of the edges. The allowed type name formats are:
+
+        * ``(str, str, str)`` for source node type, edge type and destination node type.
+        * or one ``str`` edge type name if the name can uniquely identify a
+          triplet format in the graph.
+
+        Can be omitted if the graph has only one type of edges.
 
     Returns
     -------
-    tensor
+    Tensor
         Result tensor.
 
     Examples
@@ -166,31 +172,55 @@ def readout_edges(graph, feat, weight=None, *, op='sum', etype=None):
 
 def sum_nodes(graph, feat, weight=None, *, ntype=None):
     """Syntax sugar for ``dgl.readout_nodes(graph, feat, weight, ntype=ntype, op='sum')``.
+
+    See Also
+    --------
+    readout_nodes
     """
     return readout_nodes(graph, feat, weight, ntype=ntype, op='sum')
 
 def sum_edges(graph, feat, weight=None, *, etype=None):
     """Syntax sugar for ``dgl.readout_edges(graph, feat, weight, etype=etype, op='sum')``.
+
+    See Also
+    --------
+    readout_edges
     """
     return readout_edges(graph, feat, weight, etype=etype, op='sum')
 
 def mean_nodes(graph, feat, weight=None, *, ntype=None):
     """Syntax sugar for ``dgl.readout_nodes(graph, feat, weight, ntype=ntype, op='mean')``.
+
+    See Also
+    --------
+    readout_nodes
     """
     return readout_nodes(graph, feat, weight, ntype=ntype, op='mean')
 
 def mean_edges(graph, feat, weight=None, *, etype=None):
     """Syntax sugar for ``dgl.readout_edges(graph, feat, weight, etype=etype, op='mean')``.
+
+    See Also
+    --------
+    readout_edges
     """
     return readout_edges(graph, feat, weight, etype=etype, op='mean')
 
 def max_nodes(graph, feat, weight=None, *, ntype=None):
     """Syntax sugar for ``dgl.readout_nodes(graph, feat, weight, ntype=ntype, op='max')``.
+
+    See Also
+    --------
+    readout_nodes
     """
     return readout_nodes(graph, feat, weight, ntype=ntype, op='max')
 
 def max_edges(graph, feat, weight=None, *, etype=None):
     """Syntax sugar for ``dgl.readout_edges(graph, feat, weight, etype=etype, op='max')``.
+
+    See Also
+    --------
+    readout_edges
     """
     return readout_edges(graph, feat, weight, etype=etype, op='max')
 
@@ -210,15 +240,15 @@ def softmax_nodes(graph, feat, *, ntype=None):
     Parameters
     ----------
     graph : DGLGraph.
-        Input graph.
+        The input graph.
     feat : str
-        Node feature name.
+        The node feature name.
     ntype : str, optional
-        Node type. Can be omitted if there is only one node type in the graph.
+        The node type name. Can be omitted if there is only one node type in the graph.
 
     Returns
     -------
-    tensor
+    Tensor
         Result tensor.
 
     Examples
@@ -269,15 +299,21 @@ def softmax_edges(graph, feat, *, etype=None):
     Parameters
     ----------
     graph : DGLGraph.
-        Input graph.
+        The input graph.
     feat : str
-        Edge feature name.
-    etype : str, typle of str, optional
-        Edge type. Can be omitted if there is only one edge type in the graph.
+        The edge feature name.
+    etype : str or (str, str, str), optional
+        The type names of the edges. The allowed type name formats are:
+
+        * ``(str, str, str)`` for source node type, edge type and destination node type.
+        * or one ``str`` edge type name if the name can uniquely identify a
+          triplet format in the graph.
+
+        Can be omitted if the graph has only one type of edges.
 
     Returns
     -------
-    tensor
+    Tensor
         Result tensor.
 
     Examples
@@ -535,9 +571,10 @@ def _topk_on(graph, typestr, feat, k, descending, sortby, ntype_or_etype):
            topk_indices
 
 def topk_nodes(graph, feat, k, *, descending=True, sortby=None, ntype=None):
-    """Perform a graph-wise top-k on node features :attr:`feat` in
-    :attr:`graph` by feature at index :attr:`sortby`. If :attr:
-    `descending` is set to False, return the k smallest elements instead.
+    """Return a graph-level representation by a graph-wise top-k on
+    node features :attr:`feat` in :attr:`graph` by feature at index :attr:`sortby`.
+
+    If :attr:`descending` is set to False, return the k smallest elements instead.
 
     If :attr:`sortby` is set to None, the function would perform top-k on
     all dimensions independently, equivalent to calling
@@ -569,6 +606,11 @@ def topk_nodes(graph, feat, k, *, descending=True, sortby=None, ntype=None):
         :math:`B` is the batch size of the input graph, :math:`D`
         is the feature size.
 
+    Notes
+    -----
+    If an example has :math:`n` nodes and :math:`n<k`, the ``sorted_feat``
+    tensor will pad the :math:`n+1` to :math:`k` th rows with zero;
+
     Examples
     --------
 
@@ -631,20 +673,16 @@ def topk_nodes(graph, feat, k, *, descending=True, sortby=None, ntype=None):
               [0.0880, 0.6379, 0.4451, 0.6893, 0.5197]]]), tensor([[[1, 0, 1, 3, 1],
               [3, 2, 0, 2, 2],
               [2, 3, 2, 1, 3]]]))
-
-    Notes
-    -----
-    If an example has :math:`n` nodes and :math:`n<k`, the ``sorted_feat``
-    tensor will pad the :math:`n+1` to :math:`k`th rows with zero;
     """
     return _topk_on(graph, 'nodes', feat, k,
                     descending=descending, sortby=sortby,
                     ntype_or_etype=ntype)
 
 def topk_edges(graph, feat, k, *, descending=True, sortby=None, etype=None):
-    """Perform a graph-wise top-k on node features :attr:`feat` in
-    :attr:`graph` by feature at index :attr:`sortby`. If :attr:
-    `descending` is set to False, return the k smallest elements instead.
+    """Return a graph-level representation by a graph-wise top-k
+    on edge features :attr:`feat` in :attr:`graph` by feature at index :attr:`sortby`.
+
+    If :attr:`descending` is set to False, return the k smallest elements instead.
 
     If :attr:`sortby` is set to None, the function would perform top-k on
     all dimensions independently, equivalent to calling
@@ -676,6 +714,11 @@ def topk_edges(graph, feat, k, *, descending=True, sortby=None, etype=None):
         :math:`B` is the batch size of the input graph, :math:`D`
         is the feature size.
 
+
+    Notes
+    -----
+    If an example has :math:`n` nodes and :math:`n<k`, the ``sorted_feat``
+    tensor will pad the :math:`n+1` to :math:`k` th rows with zero;
     Examples
     --------
 
@@ -738,11 +781,6 @@ def topk_edges(graph, feat, k, *, descending=True, sortby=None, etype=None):
               [0.0880, 0.6379, 0.4451, 0.6893, 0.5197]]]), tensor([[[1, 0, 1, 3, 1],
               [3, 2, 0, 2, 2],
               [2, 3, 2, 1, 3]]]))
-
-    Notes
-    -----
-    If an example has :math:`n` nodes and :math:`n<k`, the ``sorted_feat``
-    tensor will pad the :math:`n+1` to :math:`k`th rows with zero;
     """
     return _topk_on(graph, 'edges', feat, k,
                     descending=descending, sortby=sortby,

python/dgl/sampling/__init__.py

-"""This module contains the implementations of various sampling operators.
+"""The ``dgl.sampling`` package contains operators and utilities for
+sampling from a graph via random walks, neighbor sampling, etc. They
+are typically used together with the ``DataLoader`` s in the
+``dgl.dataloading`` package. The user guide :ref:`guide-minibatch`
+gives a holistic explanation on how different components work together.
 """
+
 from .randomwalks import *
 from .pinsage import *
 from .neighbor import *

python/dgl/utils/internal.py

@@ -891,4 +891,12 @@ def set_num_threads(num_threads):
     """
     _CAPI_DGLSetOMPThreads(num_threads)
 
+def alias_func(func):
+    """Return an alias function with proper docstring."""
+    @wraps(func)
+    def _fn(*args, **kwargs):
+        return func(*args, **kwargs)
+    _fn.__doc__ = """Alias of :func:`dgl.{}`.""".format(func.__name__)
+    return _fn
+
 _init_api("dgl.utils.internal")