Vectorized Random Shear.

[sebastian-sz]

What does this PR do?

Refactors RandomShear to subclass VectorizedBaseImageAugmentationLayer.

This gives performance benefits.

Fixes #1518

I had little error in my previous benchmark. The correct performance numbers are still better but the difference between implementations is smaller.

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you read the contributor guideline,
  Pull Request section?
• Was this discussed/approved via a Github issue? Please add a link
  to it if that's the case.
• Did you write any new necessary tests?
• If this adds a new model, can you run a few training steps on TPU in Colab to ensure that no XLA incompatible OP are used?

Who can review?

@LukeWood @ianstenbit

[sebastian-sz]

The test of a different layer fails, will update tomorrow.

[LukeWood]

LGTM @sebastian-sz but can you rebase on #1534 and include results from the demo?

[sebastian-sz]

@LukeWood I'm getting errors after rebase. There seems to be an issue when calling keras_cv.bounding_box.convert_format after casting ragged boxes to dense:
Incompatible shapes: [9,5,1] vs. [9,1]

The error is thrown in: _xyxy_to_rel_xyxy:

keras-cv/keras_cv/bounding_box/converters.py

Line 230 in 1974263

left, right = left / image_width, right / image_width

Standalone code to reproduce:

import demo_utils
import keras_cv.bounding_box

dataset = demo_utils.load_voc_dataset(bounding_box_format="xyxy")

for d in dataset.take(1):
    d["bounding_boxes"] = keras_cv.bounding_box.to_dense(
        d["bounding_boxes"], default_value=0
    )
    keras_cv.bounding_box.convert_format(
        boxes=d["bounding_boxes"],
        source="xyxy",
        target="rel_xyxy",
        images=d["images"],
    )

The error goes away if I comment out the Ragged check:

keras-cv/keras_cv/bounding_box/converters.py

Lines 514 to 516 in 1974263

	if isinstance(boxes, tf.RaggedTensor):
	height = tf.expand_dims(height, axis=-1)
	width = tf.expand_dims(width, axis=-1)

#            if isinstance(boxes, tf.RaggedTensor):
            height = tf.expand_dims(height, axis=-1)
            width = tf.expand_dims(width, axis=-1)

but this might break something somewhere else. Not sure how to proceed with this.

[ianstenbit]

Thanks for the PR!
This LGTM -- @LukeWood can you PTAL at the demo issue?

[ianstenbit]

nit: let's use keras directly and add from tensorflow import keras above

[sebastian-sz]

@ianstenbit Done.

[LukeWood]

** sebastian-sz ** commented

looks like an unrelated error. I will file a bug for this converter issue, in the meantime can you change the bounding box format used in the demo in a way to avoid this issue (if not possible no worries)

[sebastian-sz]

@LukeWood commented out the Ragged check.

Old RandomShear:

Vectorized RandomShear:

Note that both layers sample different shear values.

[LukeWood]

Thanks @sebastian-sz !