ImageProjectiveTransform should be named ImageProjectiveTransformV2

[andresusanopinto]

Describe the bug

• It looks like ImageProjectiveTransform is based on ImageProjectiveTransformV2 present in tf.contrib in TF-1.x.

Describe the expected behavior

• Op names should not colide between 1.x and 2.x environments or saved models won't be shareable across them.

[andresusanopinto]

Links to ops:

• tf.contrib.image
• tensorflow_addons.image

Looking at code it really seems that the op "ImageProjectiveTransformV2" from tf.contrib.image was renamed to "ImageProjectiveTransform" (which was a name already taken by other op). I guess this was un-intentional or in an attempt to cleanup the names, but in reality it can break moving models across TF-1.x and TF-2.x.

[seanpmorgan]

Thanks @andresusanopinto. I think you are correct that this is a problem even though we don't support 1.x because of saved models.

@karmel @martinwicke Should we maintain all V2 op naming just to be safe?

[karmel]

Yes-- we continue to support v1 graphdef/ops in v2, so maintaining separate namespaces will help avoid collisions. Eventually, we will be able to prefix op names with appropriate namespaces (ie, tensorflow_addons:OP_NAME), but we don't do that yet.

[seanpmorgan]

Thanks! Adding help wanted, and this bug fix would require reviewing all Ops that we've moved over.

[jharmsen]

Looks like there are only three ops in addon so far.

Two match:

• AdjustHsvInYiq [contrib | addon]
• SkipGramGenerateCandidates [contrib | addon]

As noted, one has a name mismatch:

• ImageProjectiveTransformV2 [contrib] became ImageProjectiveTransform [addon]

[jharmsen]

Going forward is the right thing to rename addon ImageProjectiveTransform to ImageProjectiveTransformV2?

[martinwicke]

I think for now, yes -- that'll make sure that people can use the existing SavedModels, etc. For the future, we need to figure out proper namespacing.

[karmel]

cc @tomerk