Change interpolation in apply_transform() from nearest to bilinear.

[ozabluda]

Before #2446, interpolation was inconsistently sometimes nearest, sometimes bicubic. AFAIK, in other libraries and tools default interpolation is never 'nearest'. Sometimes it's bicubic, sometimes it's bilinear. For Keras bilinear is better because it's much faster, and is something a single (de)convolutional layer can learn (i.e. it is, basically a single hardcoded non-trainable (de)convolutional layer. Which is not true of nearest, even with pooling.

Also see load_img() #8435

[fchollet]

This is also a potentially breaking change.

[ozabluda]

Yes, but it doesn't affect any pretrained models. Current default is highly unexpected, since we have fractional shifts, rotations, zooms and whatnot, and probably makes things worse vast majority of the time. I actually spent some time looking at 'nearest' interpolation when I was dealing with binarized images, which had to stay binarized after data augmentation. It sucks (augmented data doesn't correspond to the original). For that reason, I am not even proposing adding order parameter, since it's so exotic. A slightly less exotic case is order=3 for, say, superresolution autoencoder, but even than moved to order=1 AFAIK, probably for the reasons in my description above.

But we could introduce order parameter with default None and print warning that default changed, but I don't think anybody would care (but it would be useful mostly for debugging). Then revert to default order=1 when we have a an API-breaking release. Or we can keep None so that sometimes we do indeed default to order=0 when, say, we have integer shifts only (something like random_crop(), for which I am still designing of a good API, and, maybe, it would work just fine with order=1.

[Dref360]

This would break existing code for segmentation DA.

[ozabluda]

what is "segmentation DA"?

[Dref360]

Sorry,
Data augmentation on segmentation map.
If you do a bilinear interpolation, you would corrupt the data. For example, a pixel with class "4" could become "4.5" if its neighbors are not of the same class.

[ozabluda]

I thought segmentation DA currently doesn't really work anyway: search for "segmentation" in #3338. Also see #6538. Also 2D softmax was merged only a couple of weeks ago. But it is a valid use case for nearest, thank you.

BTW, for binary labels, and binary crossentropy, it may make things better. Automatic soft labels.

[ahundt]

I thought segmentation DA currently doesn't really work anyway

That's what I found last time I tried things out

[fchollet]

Ok, let's go with this change. Hopefully it won't break any users...

[ozabluda]

I knew it! Use random linear interpolation for data & labels to improve stuff

mixup: Beyond Empirical Risk Minimization (1017)
https://arxiv.org/abs/1710.09412v1