Why use cashedMnist?

[Meric3]

Hello vlukiyanov.

First, thanks for your kind code.
I just have a question about the dataset you implemented.
There are a lot of dataset classes for mnist using pytorch.
But I don't see this style. I mean transform byte.
Why you use this cachedMnist?

[vlukiyanov]

I think the transformations were meant to imitate the original DEC code

https://github.com/piiswrong/dec/blob/e551e01f90a3d67d2ef9c90e968c8faf5d8f857d/dec/dec.py#L373

I believe it should do something similar, but obviously a formal unit test would probably help here to make sure it matches the Caffe.

On the caching, thats just to speed it up.