Question about the trigger training process

[panicpan2000]

Hi. I have some questions.

1. glabels (line 120-121 of gta.py) is not used in the following step. Intuitively, it looks like that the code miss a step like bkd_dr.data['labels'] = glabels. Now, it seems that the generators try to let the graph with the trigger be predicted as the original label. In this case, I am not sure if the loss of the two generators is calculated with true labels or target labels.
2. I can not find the graph attention mechanism mentioned in the paper to obtain node encoding. It looks like the generator directly uses the adjacency matrix as the input and the adjacency matrix is only combined with a random trigger.
   Thank you.