引言

由于图数据结构问题，直接使用Tensorflow的一些层是无法直接处理图数据的，需要借用TF-GNN框架下的MapFeatures对图数据中的节点特征或是边特征进行变换。

题外话(MapFeatures使用)

节点特征变换

from tensorflow.keras.layers import BatchNormalization
from tensorflow_gnn.keras.layers import MapFeatures

# map node features
def node_sets_fn(node_set, *, node_set_name):
    features = node_set.features
    return BatchNormalization()(features["hidden_state"])

graph = MapFeatures(node_sets_fn=node_sets_fn)(graph)

边特征变换

from tensorflow_gnn.keras.layers import MapFeatures

# Hashes edge features called "id", leaves others unchanged:
def edge_sets_fn(edge_set, *, edge_set_name):
    features = edge_set.get_features_dict()
    ids = features.pop("id")
    num_bins = 100_000 if edge_set_name == "views" else 20_000
    hashed_ids = tf.keras.layers.Hashing(num_bins=num_bins)(ids)
    features["hashed_id"] = hashed_ids
    return features
graph = MapFeatures(edge_sets_fn=edge_sets_fn)(graph)

传入额外参数

from functools import partial
from tensorflow.keras.layers import Dense
from tensorflow_gnn.keras.layers import MapFeatures

# map node features
def node_sets_fn(node_set, *, node_set_name, dim):
    features = node_set.features
    return Dense(dim)(features["hidden_state"])

graph = MapFeatures(node_sets_fn=partial(node_sets_fn, dim=64))(graph)

问题

就是在使用MapFeatures时，如果循环使用则会在存储模型的时候报错：ValueError: Unable to create dataset (name already exists)

问题demo


from functools import partial
from tensorflow.keras.layers import Dense
from tensorflow_gnn.keras.layers import MapFeatures

# map node features
def node_sets_fn(node_set, *, node_set_name, dim):
    features = node_set.features
    return Dense(dim)(features["hidden_state"])

for ln in range(layer_num):
    graph = MapFeatures(node_sets_fn=partial(node_sets_fn, dim=64))(graph)

解决方案

最后发现是在使用MapFeatures时，使用层时如Dense需要区分每一次变换时的层名


from functools import partial
from tensorflow.keras.layers import Dense
from tensorflow_gnn.keras.layers import MapFeatures

# map node features
def node_sets_fn(node_set, *, node_set_name, dim，name):
    features = node_set.features
    return Dense(dim, name=f'Dense_{name}')(features["hidden_state"])

for ln in range(layer_num):
    graph = MapFeatures(node_sets_fn=partial(node_sets_fn, dim=64，name=ln))(graph)