If I wrote a Dataset class, say, for fashion-mnist, now for debug on cpu, i just want a small proportion of the 60k images to get code run in a quick manner, but it is requisite to have all 10 classes. Is there an easy implementation of this?
You could use a Subset or SubsetRandomSampler using a smaller set of sample indices.
Since you would like to have all classes you would need to create those indices first using the original target attribute of the dataset.