When sampling frames, instead of padding it directly there, instead do it during the collate to batch. This might lower the computational cost as less frames are sometimes passed through the transformations.
Should probably verify if this is a good idea or not, as the last few inputs would be static.
When sampling frames, instead of padding it directly there, instead do it during the collate to batch. This might lower the computational cost as less frames are sometimes passed through the transformations.
Should probably verify if this is a good idea or not, as the last few inputs would be static.