data_loading_tutorial.py iterators

[patricknaughton01]

I like this tutorial but I think it would be better if it included an example of how to define next() and iter() methods so that the dataset can be used with enumerate.

[noqqaqq]

/assigntome

[noqqaqq]

Some thoughts looking at this issue:

1. iter and next are not needed to use enumerate - it works out of the box with len and getitem. Tutorial can be then easily improved to use enumerate in relevant places without reimplementing dataset class.
2. There is IterableDataset. I guess, implementing it would be most sufficient for this example.
3. Using IterableDataset raise some issues:

• random access is lost, which then requires looping or skipping iterator for this case (need to make sure to implement it in a way that minimize perf penalty)
• IterableDataset does not allow for shuffling, which is needed when passing dataset to Dataloader. There are ways to implement this like described here or here or here, but I don't know yet how to do it in the simplest correct way for the sake of tutorial.

To sum up, I'll follow with pushing patch for 1. and try to copy and modify dataset tutorial to cover iterable case (having both in single file in my opinion would obfuscate the examples and make them hard to track / understand).

If anyone has comments on this or hints on shuffling please let me know.

[NicolasHug]

I agree with @noqqaqq's thoughts that there's no need for next() and iter() to be present for enunerate() to work. The original point of the issue may have been to include an example for iterable datasets whoc have a __next__() method instead of __getitem__(), but like @noqqaqq also correctly pointed out, there are other resources out there to do that and that's not the original point of this tutorial.

For the simple purpose of exposing how to use enumerate(), #2407 LGTM