SvenDS9

Using the testcode/testoutput block format the output does look slightly different but imo this it's actually an improvement 

I also noticed that many datapipes in `util` also handle loading so should probably be moved to `load`. Imho the division into `load`, `transform` and `util` is fine but is...

Some tests in torchvision fail. Apparently a stream isn't being closed and my implemenation of `itertomap` `__getitem__()` causes this. The stream in question I think is being created here: https://github.com/pytorch/vision/blob/01ef0a68b6ec00452391251fc16c38e58b92bf07/test/builtin_dataset_mocks.py#L1356...

I think so as well. @pmeier Thank you for linking https://github.com/pytorch/vision/pull/6997 I think I know what's happening now! Forcefully depleting the dp (e.g by adding a `Prefetcher` in cub200 with...

> Probably in the test not every value is retrieved from the map, therefore the map is never fully loaded and the `iter` of the previous dp never "finishes". I...

128 was still to low - at least with a timeout of 5 seconds. With this new value I am mostly limited by my internet connection. I have also compared...

This can be closed.

A similar thing happens when you do: ```python dp = SequenceWrapper({'a': 100, 'b': 200, 'c': 300, 'd': 400}) print(list(dp)) > File "\lib\site-packages\torch\utils\data\datapipes\map\utils.py", line 46, in __getitem__ > return self.sequence[index] >...

Personally I would expect the data.

Dear @ejguan for some reason this PR seems to have not been merged yet. Could you please have a quick look? Thanks a lot in advance