Hey @saswatpp and thanks a lot for the PR! I have some comments inline that make the implementation (hopefully) a little cleaner. On top of that, you need to run our auto-formatters on your code to pass CI. Have a look at our contributing guide. In short, you can do the following:

pip install pre-commit
pre-commit install  # activates the auto-formatter for each subsequent commit
pre-commit run --all-files  # only needed once, since you already commited
git commit -am "fix code format"

Also, I couldn't understand what exactly the tests for datasets does just by looking at the code. Do we have some resources for that ?

The test for the datasets are located in test/test_datasets.py. From the contributor perspective you need to add a new test case that configures some basics and provides fake data.

class SUN397TestCase(datasets_utils.ImageDatasetTestCase):
    DATASET_CLASS = datasets.SUN397

    ADDITIONAL_CONFIGS = datasets_utils.combinations_grid(
        split=("train", "test"),
        # There is no need to test all individual partitions, since they all behave the same
        partition=(1, 10, "all"),
    )

    def inject_fake_data(tmp_dir, config):
        ...

inject_fake_data should prepare tmp_dir (in your case the same as root in the dataset) in a way that the structure of the files is equal to what would be there if I had instantiated the dataset with download=True. It needs to return the number of samples in the dataset. If you want to know more, you can read the documentation of the underlying test case.

Hey @saswatpp, I've suggested one change to the label generation. With that mypy is happy and I can't reproduce the test error locally. Let's see if CI is happy to.

Thanks a lot @saswatpp for the PR. LGTM!

Thanks a lot for the PR @saswatpp and @pmeier for the review. I made a few comments below, the download() function needs some minor fixes but other than that it looks great. LMK what you think.

Thanks a lot @saswatpp !

Running the tests locally, I'm seeing a lot of warnings about unclosed files:

test/test_datasets.py::LSUNTestCase::test_transforms
  /Users/nicolashug/dev/vision/torchvision/datasets/lsun.py:31: ResourceWarning: unclosed file <_io.BufferedWriter name='_cache_varfolderssyvkyzpyhrlqqcgnTtmpclcnsatowervallmdb'>
    pickle.dump(self.keys, open(cache_file, "wb"))

This is simlar to what happened in #5116 (review), perhaps @pmeier can advise on how he fixed it over there?

Other than that, LGTM!