Code to fetch the SVHN dataset

[rcalland]

SVHN is very similar to MNIST. I added code that will download and create datasets for train and test samples the same way get_mnist() does.

This code requires scipy in order to read the matlab format that files are officially provided in.

[mattya]

Thank you for the PR!
I have two comments.

• Please add get_svhn in __init__.py (https://github.com/rcalland/chainer/blob/b1cb12d726078589bf99b31ab66f553a1ab717a9/chainer/datasets/__init__.py)
• It seems label i corresponds to the number i+1 in the image and it might be confusing.

[rcalland]

The SVHN website says:
10 classes, 1 for each digit. Digit '1' has label 1, '9' has label 9 and '0' has label 10.
which I guess is less confusing, how about switching to that system?

[rcalland]

I just realised I misunderstood slightly the label definitions on the website, and so simply setting digit "0" from label 10 to 0 makes everything sensible!

[mattya]

Thank you for the update.
It seems there are some flake8 errors in __init__.py and svhn.py. Could you fix them?

[mattya]

There are still following flake8 errors

.\chainer\datasets_init_.py:6:1: H306 imports not in alphabetical order (chainer.datasets.svhn, chainer.datasets.ptb)
.\chainer\datasets_init_.py:21:1: H306 imports not in alphabetical order (chainer.datasets.svhn.get_svhn, chainer.datasets.ptb.get_ptb_words)

[rcalland]

My version of flake8 didn't show those errors. Fixed.

[mattya]

Sorry for late reply, current CI is failing because scipy is not in the default dependency.
Could you add scipy availability check like this? https://github.com/cupy/cupy/blob/d7ad0d34f77de5b111807c0c1998176e9ab7e59a/cupy/sparse/csc.py#L1

[mattya]

Thank you for fixing and sorry for late reply.
LGTM!