|
def create_combined_dataloader( |
This just roundrobins, forever, across all dataloaders.
A better solution would involve some kind of proportional sampling. Maybe weighted by some custom weights.
Related: it would also be nice to have a way to be more explicit about which datasets are being used. Instead of iterating over all modules in a package that have a particular set of functions that identify them as a "dataset", let people craft their own list of datasets that they want to use. You can always provide a helpful wrapper around that to provide a reasonable default set of datasets.
mugato/mugato/data/utils.py
Line 84 in 1deec80
This just roundrobins, forever, across all dataloaders.
A better solution would involve some kind of proportional sampling. Maybe weighted by some custom weights.
Related: it would also be nice to have a way to be more explicit about which datasets are being used. Instead of iterating over all modules in a package that have a particular set of functions that identify them as a "dataset", let people craft their own list of datasets that they want to use. You can always provide a helpful wrapper around that to provide a reasonable default set of datasets.