I have used your model on some data sets, and the performance is a bit strange, can you help provide your results?

model.py at line 155 and 165:

    es = torch.sigmoid((query[:self.k//2] * keys[:self.k//2]).sum(-1, keepdim=True) / (self.d ** 0.5))
    eso = torch.sigmoid((query[self.k//2:] * keys[self.k//2:]).sum(-1, keepdim=True) / (self.d ** 0.5))

Unlike the paper, I wonder why you put sigmoid on here.

hello，i want to know where i can get the spatial embedding file(suffix is npy)？parser.add_argument('--SE_file', default='**.npy', help='spatial emebdding file')

Hello, author,
As far as I can read, I find there may be the following problems in this code.
Firstly, I find that the ratio of training set / verification set / test set in the paper is 7 / 1 / 2, but the data distribution ratio in this code is 6 / 2 / 2.
Secondly, whether the results in the paper are the best results measured from the validation data set or the last test set, or the best results selected from the validation set and test set!
Finally, hope to get an answer!
Thanks!