Denormalization layer

[markub3327]

@gadagashwini

Solved #16284

[fchollet]

Thanks for the PR. I see we have an invert argument in the preprocessing layers StringLookup and
IntegerLookup. Perhaps this use case could be covered in a simpler fashion by also adding invert to the Normalization layer?

@mattdangerw what do you think?

[markub3327]

@fchollet
I changed it to version with invert=False. I think it's better solution too.

Please make a review.

[mattdangerw]

Yeah, I think invert as an argument make sense. Could you add a unit test seeing if this would also be workable for an adapted layer? I think the following snippet should work, but it would be good to check.

norm = keras.layers.Normalization(axis=None)
norm.adapt(some_data)
inv_norm = keras.layers.Normalization(
    mean=norm.mean,
    variance=norm.variance,
    axis=norm.axis,
    invert=True)

We should probably also show an example of this in the docstring, taking some data, first normalizing, then inverting. For both the adapted and non-adapted cases.

[markub3327]

@mattdangerw

I create both unit tests.

[fchollet]

LGTM, thank you

[mattdangerw]

@markub3327 thanks. Could you add an example of invert in the docstring as well?

[markub3327]

@mattdangerw

Yes, sorry I forgotten. It's done.

[fchollet]

LGTM, thanks. I'll do a round of copyedits in post-processing.

[markub3327]

@fchollet

Ok. Thanks.