What does I * (1 - valid_inds) mean？

[JayChan-USTC]

hi,author!
During training, when calculating A, the following formula is used,
A = torch.matmul(prev_sim_map, A) * valid_inds + I * (1 - valid_inds)
A = A * valid_inds + I * (1 - valid_inds).

https://github.com/TRI-ML/RAM/blob/master/src/lib/model/networks/base_model.py#:~:text=if%20prev_sim_map%20is%20not%20None,A%20*%20valid_inds%20%2B%20I%20*%20(1%20-%20valid_inds)
(line 334-line 342)

I have two questions,
(1) I don't know what I * (1 - valid_inds) means,
(2) Why need to add I * (1 - valid_inds) after calculating (torch.matmul(prev_sim_map, A) * valid_inds) and (A * valid_inds)

Hope to get your reply, thank you!

[Wuziyi616]

I think if they don't add I * (1 - valid_inds), then some rows in A will be all-zeros, which will lead to divide-by-error in later lines. I don't think that's important tho. After all, the invalid tracks should never be used in loss computation.