How to save the attention weight in official Hgt model example

[EverOasis]

I’ve successfully train the hgt model on my dataset, but haven’t figured out how to save the attention weights and updated node features from the official hgt model tutorials

[dyru]

Do you mean you want to save the attention scores and node features of each intermediate layer? In that case, you can modify the HGTLayer a little bit to include attn_score in its output.

Then group the intermediate output into a list during layer-wise forward

all_layer_h, all_layer_attn_score = [], []
for i in range(self.n_layers):
    h, attn_score = self.gcs[i](G, h)
    all_layer_h.append(h)
    all_layer_attn_score.append(attn_score)

[EverOasis]

Thanks for the reply .I want to save the node features and attention attn_score in original graph，so at the end of training, it is straightforward to get the updated node weights and the attention scores on each edge.What am I supposed to do?

[minjie]

Hi @EverOasis , could you please elaborate more on what you want? I feel @dyru 's answer is already able to achieve that.