About learning rate and camera sampling

[seasonSH]

Thanks for your contribution! I found some problems when using the code:

1. The learning rate is supposed to decay to 0.01xinitial_lr at the end but it almost never changed because the scheduler is updated every epoch instead of every step (BTW, why don't you use the cosine annealing as in the paper?)
2. In the paper, the authors mentioned that sampling camera position uniformly by angle leads to oversampling on the overhead view. So by 50% chance, they sample uniformly on the sphere surface. Seems like it is not implemented.
3. As for the w(t). The author mentioned that "we absorb the constant alpha_t = delta z_t / delta x into w(t)". I think it means that since alpha(t) can be considered as part of w(t), we can simply ignore it and choose whatever w(t) we want. So that would mean w(t)=sigma_t^2. correct me if I am wrong:)

[ashawkey]

@seasonSH Huge thanks for reporting!

1. Yes this is such a silly bug! I will correct it very soon after some testings.
2. Yes, this has not been implemented. I'm not sure whether it has a significant influence. (Also the camera jitter is temporarily disabled now, I don't have enough time to test them all right now...)
3. Good insight. The authors also mention that a uniform weight has similar performance, I guess this weight will not affect results significantly. I'll have a try later.