使用cugraph创建Graph返回NoneType无报错的问题求助

尝试从dask_cudf DataFrame、pandas DataFrame以及仅3条边的样本数据创建cugraph Graph时，每次都返回NoneType且无任何报错信息。但使用Karate数据集执行完全相同的步骤却能正常生成Graph，且所有数据的列类型均一致。

从dask_cudf创建Graph的代码 #

cluster = LocalCUDACluster()
client = Client(cluster)
Comms.comms.initialize(p2p=True)

edges = dask.read_csv('.csv')
edges = edges.groupby(['Source','Target'])['retweet_from'].count()
edges = edges.to_frame(name="weight").reset_index()
edges = edges.map_partitions(cudf.DataFrame.from_pandas)
G = cugraph.Graph().from_dask_cudf_edgelist(edges,
                                            source = 'Source',
                                            destination = 'Target',
                                            edge_attr = 'weight')

G.__class__
NoneType

使用Karate数据集创建Graph的代码（正常运行） #

url = 'https://raw.githubusercontent.com/rapidsai/cugraph/branch-22.10/datasets/karate.csv'
df = pd.read_csv(url,delimiter=' ', header=None, names=["0", "1", "2"],
dtype={"0": "int32", "1": "int32","2": "float32"})

G = cugraph.Graph()
G.from_pandas_edgelist(df, source='0', destination='1',edge_attr='2', renumber=False)

G.__class__
cugraph.structure.graph_classes.Graph

从pandas DataFrame创建Graph的代码 #

edges = pd.read_csv('.csv')
edges = edges.groupby(['Source','Target'])['retweet_from'].count()
edges = edges.to_frame(name="weight").reset_index()
edges['Source'] = edges['Source'].astype("int32")
edges['Target'] = edges['Target'].astype("int32")
edges['weight'] = edges['weight'].astype("float32")
edges.dtypes
Source      int32
Target      int32
weight    float32
dtype: object

G = cugraph.Graph()
G = G.from_pandas_edgelist(edges,source = 'Source',destination = 'Target',edge_attr = 'weight', renumber=False)

G.__class__
NoneType

使用3条边的样本数据创建Graph的代码 #

data = [[1, 3,3], [2, 1,1], [3, 1, 7]]
edges = pd.DataFrame(data, columns=['Source', 'Target', 'weight'])
edges['Source'] = edges['Source'].astype("int32")
edges['Target'] = edges['Target'].astype("int32")
edges['weight'] = edges['weight'].astype("float32")
G = cugraph.Graph()
G = G.from_pandas_edgelist(edges,source = 'Source',
                                            destination='Target',edge_attr = 'weight', renumber=False)
G.__class__
NoneType

问题出在方法调用的赋值方式上：cugraph的from_pandas_edgelist、from_dask_cudf_edgelist等方法是原地修改Graph实例，不会返回新的Graph对象，而是返回None。

你的错误写法：

G = G.from_pandas_edgelist(...)  # 将None赋值给G，导致G变成NoneType

正确写法：

G = cugraph.Graph()
G.from_pandas_edgelist(...)  # 直接调用方法修改G，不赋值

对应到各个场景的修正：

1. dask_cudf场景修正

cluster = LocalCUDACluster()
client = Client(cluster)
Comms.comms.initialize(p2p=True)

edges = dask.read_csv('.csv')
edges = edges.groupby(['Source','Target'])['retweet_from'].count()
edges = edges.to_frame(name="weight").reset_index()
edges = edges.map_partitions(cudf.DataFrame.from_pandas)

G = cugraph.Graph()  # 先创建实例
G.from_dask_cudf_edgelist(edges,  # 直接调用方法，不赋值给G
                          source='Source',
                          destination='Target',
                          edge_attr='weight')

G.__class__  # 现在会返回cugraph.structure.graph_classes.Graph

2. pandas DataFrame场景修正

edges = pd.read_csv('.csv')
edges = edges.groupby(['Source','Target'])['retweet_from'].count()
edges = edges.to_frame(name="weight").reset_index()
edges['Source'] = edges['Source'].astype("int32")
edges['Target'] = edges['Target'].astype("int32")
edges['weight'] = edges['weight'].astype("float32")

G = cugraph.Graph()
G.from_pandas_edgelist(edges,  # 去掉赋值操作
                       source='Source',
                       destination='Target',
                       edge_attr='weight',
                       renumber=False)

G.__class__  # 正常返回Graph类型

3. 3条边样本数据场景修正

data = [[1, 3,3], [2, 1,1], [3, 1, 7]]
edges = pd.DataFrame(data, columns=['Source', 'Target', 'weight'])
edges['Source'] = edges['Source'].astype("int32")
edges['Target'] = edges['Target'].astype("int32")
edges['weight'] = edges['weight'].astype("float32")

G = cugraph.Graph()
G.from_pandas_edgelist(edges,  # 去掉赋值操作
                       source='Source',
                       destination='Target',
                       edge_attr='weight',
                       renumber=False)

G.__class__  # 正常返回Graph类型

本质原因是你在自己的代码中错误地将方法返回的None赋值给了Graph变量G，而Karate数据集的代码中没有做这个赋值操作，所以能正常保留Graph实例。

内容的提问来源于stack exchange，提问作者padul