charles-viss

This still seems undesirable though because as you mention in the docs, the length of a dataset/dataloader can vary depending on your augmentation or sampling strategy

For example, one use case is training over a fixed dataset with or without category-weighted sampling. Because custom data samplers change the size of the dataloader, i've used the `records_per_epoch`...

One helpful feature could be to have the length of an epoch defined by the length of the dataloader by default, but then could be overridden if a `records_per_epoch` parameter...

We are currently using determined version 0.21.2

In one situation we use a `WeightedRandomSampler` where that could work. In other cases though, such as when using a custom data sampler for triplet loss, the data samplers have...