question about ref_img_sampler parameters (num_ref_imgs, frame_range, filter_key_img )

[fcakyon]

@OceanPang @GT9505 Thanks a lot for maintaining this library!

I read the related docs on these params but couldnt understand their meaning, would you tell their meaning and what they change during training:

num_ref_imgs, frame_range, filter_key_img from this config.

What is the difference between this and this as given:
num_ref_imgs=9, frame_range=9, filter_key_img=True
num_ref_imgs=2, frame_range=0, filter_key_img=True

[OceanPang]

Hi @fcakyon, num_ref_imgs means the number of reference images you want to sample (output).
frame_range means the frame intervals to sample the images and filter_key_img means whether you will put the keyframe into the sampling candidates.
For example, a key image locates at frame t. If frame_range == 2, the sampling candidates will be the images in t-2 -> t+2. filter_key_img decides whether t exists in sampling candidates.

[fcakyon]

@OceanPang thanks for the fast response.

So when num_ref_imgs=2, frame_range=0, filter_key_img=True as in here, the number of reference images you want to sample (output) is 2 and sampling candidates will be the images in t -> t. How does that work out?

[OceanPang]

@fcakyon Sorry for the late reply.
This seems like an implementation bug. @GT9505, can you take a look at this config?

[GT9505]

The code for sampling reference images is in here.
This is a bug, which may introduce some misunderstanding, in the vid config. The filter_key_img should be False, when the training data is images rather than videos. But it don't change the output of the function, since the function directly copy the input image with num_ref_imgs times, as long as frame_range=0.
We will fix it in next PR.

[fcakyon]

@GT9505 thanks for the feedback.