Hard gumbel softmax during training AIM?

[tedhuang96]

I am confused with this line of code. If hard is set as True even for training, how can the gradient flow back through AIM module in backprop?

AS-GCN/net/utils/adj_learn.py

Line 264 in ab6951c

self.edges = gumbel_softmax(self.logits, tau=0.5, hard=True)