Data augmentation for SVHN dataset

[QiushiYang]

Many thanks for your clean codes.

(1) It seems there is no code of dataloader for SVHN dataset. Would you like to introduce what kinds of data pre-processing and augmenation methods (and mean&std values used for normalization) you used for SVHN dataset in the training phase?

(2) Also, as for the results reported in Table 2 of your paper, I wonder which version of SVHN dataset (it has two versions, one including 73257 training images and 26032 test images, the other one named extra version containing 631131 images) you used to train FixMatch models?

[carlini]

The SVHN dataloader can be found here:

https://github.com/google-research/fixmatch/blob/d4985a158065947dba803e626ee9a6721709c570/libml/data.py#L297-L300

We run experiments with the smaller svhn_noextra dataset here.

[QiushiYang]

Thanks a lot! Do you mean that you used smaller version of SVHN dataset to train and to test, without the svhn_extra?

[carlini]

Yeah, we used the smaller dataset to train. This dataset, in the code, is called svhn_noextra. If you look at the ReMixMatch paper I think we also have results for how much gain you et by moving to the full svhn dataset, and it was rather small.

[QiushiYang]

Many thanks! I will have a try :-)