whether `temporal_interp` is used at training time ? about nvp HOT 2 CLOSED

Hi, thank you for the amazing work.

I'm interested in the temporal interpolation (frame interpolation) function of NVP. But I found that the quality of interpolated frames is very low.

I noted that if temporal interpolation is required (i.e., opt.t_interp != -1), then temporal_interp is set as True at test time, and the model will call self.sparse_grid.forward_inter. But at training time, the temporal_interp is set as Flase as default.

So, do I need to turn on temporal_interp at training time if I want temporal interpolation ? or the sparse_grid.forward_inter is only used for inference time ?

Update:

but when I set temporal_interp=True at training time, the loss and psnr are nan

from nvp.

The sparse_grid.forward_inter is only used for inference time.
You can get the temporally interpolated video as shown on the homepage

from nvp.