A proper way of using embedding as feature matrix

[JihoChoi]

What is the proper example of using the output (embedding) of Transformer (or LSTM) as the input feature matrix?

Each node in the graph consists of a sequence of elements and the representation of the sequence trained in RNNs.
I would like to build an end-to-end model in which the output of the RNNs is the input the to GNNs as node representations.
However, I believe, all the examples consist of static feature matrix X.
What would be the proper example of passing the embedding to the input?

from torch_geometric.data import Data

data = Data(
    x                   = x_embedding,    # output tensor from the transfomer
    y                   = y_label,
    edge_index = edge_index,
    train_mask = mask,
)

[rusty1s]

It depends on whether you want to fine-tune your transformer/LSTM or simply want to use their embeddings (detached from the computation graph). In case of the latter, you can simply input your embeddings into data.x (as you have done). In case of the former, you would need to input the sequence into the transformer as part of your model. Does that answer your question? :)

[JihoChoi]

Thank you very much!