[MXNET-141] Add tutorial Gluon Datasets and DataLoaders

[thomelane]

Description

Intro to Datasets and Dataloaders.
Using own data with Included Dataset objects (including RecordIO format).
Using own data with Custom Dataset objects.
Wrappers for converting between DataLoader and DataIters.

Checklist

N/A. Added single markdown file.

Comments

https://issues.apache.org/jira/browse/MXNET-141

[Ishitori]

Looks fine to me, @thomelane

[piiswrong]

1. DataLoader has num_workers to allow parallelization. This is very important. We should introduce it to users.
2. recordio is pretty hard to use. We are considering phasing it out. I don't think we should recommend it to users here

[thomelane]

@piiswrong agree with num_workers. Will add commentary on that, and use it in the examples.
Question about RecordIO though... doesn't it give better performance? And if so, is there a replacement planned/implemented?

[piiswrong]

It shouldn't matter that much, espesially when you are randomly accessing with shuffle

[thomelane]

@piiswrong Okay, good to know, I've removed the section on RecordIO. We should mention on the website/docs more clearly that it's not recommended, this could confuse new users otherwise. And even add a deprecation warning to the im2rec.py file.

Also realized that I had used num_workers in the tutorial already, just on the second usage of DataLoader. So I've moved the discussion to the first time DataLoader is used, so there's no chance people can miss it!

And lastly, I've removed the wrapper to convert from DataLoader to DataIter based on your email feedback. I think keeping DataIter to DataLoader would still be useful for the case where people have already implemented a lot of augmentation, etc, using DataIters.