Hi, my question is regarding this code
Later the script torch.save the dataset as dataset cache to facilitate the data loading next time.
The second time I load the dataset from cached is indeed much faster.
But when I try to remove the data in the
traindir, then load the cache, there is an error.
So the code still needs to visit the dataset path to read the video on the fly, correct?
Then how can it claim to “pre-compute” the data?
And the saved dataset looks much smaller than actual dataset.
so what’s under the hood? I only know the transform can be saved in it.