In your paper, you say that you introduce Gaussian noise to the retrieved semantic distribution, but, in query_semantics.py, you introduce random noise to the retrieved semantic distribution. Maybe I don't see the Gaussian noise?
Thanks for pointing this out! I have changed torch.rand_like into torch.randn_like, and I expect that this would not affect the performance a lot with a lower sigma1.
In your paper, you say that you introduce Gaussian noise to the retrieved semantic distribution, but, in query_semantics.py, you introduce random noise to the retrieved semantic distribution. Maybe I don't see the Gaussian noise?