Input shape for the GraphSAGE layer and sparse tensor construction

[zzzy0828]

Hi!
Recently I want to use the GraphSAGE layer to extract spatial information between different sites, each site has relevant features. Problems I encountered are as follows.
The main code is here.
''''''''''''''''''
def x_reshape(u):
return backend.reshape(u, (-1, n_site, fea)) # 自定义Reshape函数1---收time_step(x)
def a_reshape(v):
return backend.reshape(v, (-1, n_site, n_site)) # 自定义Reshape函数2---收time_step(a)

model construction

x_in = Input(shape=(time_step, n_site, fea))
a_in = Input(shape=(time_step, n_site, n_site))
x_in_v = Lambda(x_reshape, output_shape=(n_site, fea))(x_in)
a_in_v = Lambda(a_reshape, output_shape=(n_site, n_site))(a_in)

spatial information passing

output = GraphSageConv(64, activation='relu')([x_in_v, a_in_v])
output = GraphSageConv(64, activation='relu')([output, a_in_v]) # two layers for spatial feature extraction
''''''''''''''''''
1.How can I construct the 'a_in_v' into a sparse tensor? I tried many ways, but failed.
2.The input of the GraphSAGE layer is defined as:
Input
- Node features of shape (n_nodes, n_node_features);
- Binary adjacency matrix of shape (n_nodes, n_nodes).
But when I try a Input(shape = (n_site, n_site), sparse=True), it remainds me "AssertionError: A must have rank 2". Is this a contradict? How can I handle it.

The reshape function is to convert the 3D input into 2D input, which is designed for graph convolution operation.
Questions above don't exist when I use a GCN or GAT layer for feature extraction.
Really need your help.

Expect for your reply.
Thank you very much！