RIFE 4.25 Fine Tunning

[Farah-mohamed]

I am fine tunning RIFE on MODD dataset (water surface movement videos) in order for rife to be better in interpolating videos that contain waves in sea.
I need to know:
a) the fps of the dataset it was trained on, number of images in each dataset and number of epochs it was trained for?
b)when fine tunning which batch size should i choose?
c)i will use the model for sea view 4x interpolation so is it worthy to update the model in each training loop for t=0.25, 0.5 and 0.75 with their gt images, or keep training on a random t selected? if so then what should be the minimum fps of the training dataset to do so? and how much frames per each video sequence?
d)for the learning rate scheduler, can you conclude and equation to adjust its parameters based on the number of images in the dataset and epochs the model will be trained on?
Thanks in advance.

[hzwer]

a) A suitable training frame rate for RIFE is between 30-60 fps. It is recommended to have over 50,000 images, and generally, more epochs are better.

b) Choose the maximum batch size that your GPU can handle.

c) You can train solely on fixed t values of 0.25, 0.5, and 0.75.

d) Use a cosine decay for the learning rate, meaning a larger learning rate is used for most of the training period, with a rapid decay only at the end.

[netExtra]

I am fine tunning RIFE on MODD dataset (water surface movement videos) in order for rife to be better in interpolating videos that contain waves in sea. I need to know: a) the fps of the dataset it was trained on, number of images in each dataset and number of epochs it was trained for? b)when fine tunning which batch size should i choose? c)i will use the model for sea view 4x interpolation so is it worthy to update the model in each training loop for t=0.25, 0.5 and 0.75 with their gt images, or keep training on a random t selected? if so then what should be the minimum fps of the training dataset to do so? and how much frames per each video sequence? d)for the learning rate scheduler, can you conclude and equation to adjust its parameters based on the number of images in the dataset and epochs the model will be trained on? Thanks in advance.

Did you get to fine tune Rife v4.25. I would be interested in trying it if you have.

@hzwer Just a quick FYI. I did some recent testing of v4.25 vs v4.26 and the differences are very small but v4.26 is an improvement for certain certain strobing artifacts. But v4.25 is a little smoother for fast motion. In one scene between two frames there is no double image but there is in v4.26. That double image becomes a nasty artifact for some scenes. So v4.25 is still the best overall because the bad v4.26 artifacts are worse than the bad v4.25 artifacts.