slices versus batch

[mdanb]

I usually use the batch attribute to indicate which graph a node belongs to in a Dataset. But I'm reading someone else's code and I noticed that they do it using slices so I browsed the documentation and noticed that the example for creating my own dataset also uses slices. Are these just two different ways of doing the same thing or is there a more fundamental difference between them?

[rusty1s]

They only slightly differ. I'm not really sure how you implement it via batch, but slices lets you access individual examples in O(1). Furthermore, since data is only sliced, it will not need to duplicate the output data in memory:

x = dataset.data.x[100:150]  # x holds the same underlying memory as dataset.data.x