Plot runtime of map_fn, vectorized_map, and a true vectorized implementation for RandomErasing

[LukeWood]

This will help us determine which flows we need to support/how important each improvement is on the previous.

[bhack]

Do we need to add also XLA jit_compile? As nobody clarified on #146 if it was or not orthogonal to vectorized_map and to a true vectorized implementation.

[bhack]

For jit_compile we could check if we want to suggest users to enable with tf.config.optimizer.set_jit like in https://www.tensorflow.org/xla/tutorials/autoclustering_xla instead of enforcing ourself in the @tf.function decorator.

We have already some layers performance tests in Keras with tf.function vs XLA jit_compile matrix in:

https://github.com/keras-team/keras/blob/master/keras/benchmarks/layer_benchmarks/layer_benchmarks_test.py

[LukeWood]

The purpose of this experiment is to gather performance numbers so we can make a decision. I'll be doing benchmarking as thoroughly as possible to get this data.

[bhack]

Yes I just meant if we want to collect also some performance with jit_compile in the performance matrix.

[chjort]

@LukeWood
I have made some performance improvements to keras_cv.utils.fill_utils in PR #159. So I would recommend benchmarking with these improvements. I could include these performance improvements in its own PR without the GridMask changes if you would like.

You can see the new fill_utils.py here: https://github.com/keras-team/keras-cv/blob/796cdd2af4dbeef27f973f67b1c9dcc37e5f8b27/keras_cv/utils/fill_utils.py

[LukeWood]

the goal is to determine how good vectorized_map is more than anything, so I think the benchmark is fine.