Places365 dataset

[pmeier]

This adds the Places365 dataset. As the name implies it comprises 365 categories of "places". It is structured in 4 different splits:

• training-standard (~ 1.8e6 images)
• training-challenge (~ 8e6 images)
• val (~ 18e3 images)
• test (~ 330e3 images)

With the flag small the user can switch between high resolution and small (256 x 256) images.

[fmassa]

Very nice work, thanks a lot @pmeier !

I think this looks good to be merged, I have only a few questions / comments.

This PR is missing an entry in the documentation in https://github.com/pytorch/vision/blob/master/docs/source/datasets.rst , can you add one there (we have just sorted it alphabetically)

For a future PR, I think it might make sense to split the tests for Places365 in its own Tester class (still in the same file is ok), as we have quite a few test functions there.

[pmeier]

For a future PR, I think it might make sense to split the tests for Places365 in its own Tester class (still in the same file is ok), as we have quite a few test functions there.

So basically

class Places365Tester(TestCase):
    pass

?

[fmassa]

So basically

class Places365Tester(TestCase):
    pass

Yes, something like

class TesterBase(TestCase):
    # base implementation for tester
    def generic_classification_dataset_test(...)
    ...

class Tester(TesterBase):  # find better name
    # all datasets tests that are implemented as a single function
    def test_imagefolder(...):
    ...

class Places365Tester(TesterBase):
    ...

[fmassa]

Thanks a lot!