Motif guidance scale factors seem to be missing

[amorehead]

In

protein-frame-flow/motif_scaffolding/twisting.py

Line 406 in f50d8db

var_rot = rt_sq_rot[:, None, None]

the value of rt_sq_rot during inference is always None. Also, it doesn't seem like rt_sq_rot is defined anywhere else in the code. Am I simply missing something in the codebase that defines this value?

Referencing the repo's TMLR paper, do you simply define it as follows?

rt_sq_rot = ((1 - batch['so3_t']) ** 2) / (batch['so3_t'] ** 2 + (1 - batch['so3_t']) ** 2)

If so, how does the extra / 2 term fit into the following calculation?

log_p_rot = -((R_obs - R_pred)**2 / 2 / var_rot).sum(dim=[-3, -2, -1])

Perhaps this is based on the following definition of Gaussian negative log likelihood (reference: https://stats.stackexchange.com/questions/521091/optimizing-gaussian-negative-log-likelihood)?

[Layne-Huang]

I have the same question about it