Port TensorFlow implementation of the RandomCrop layer

[soumik12345]

This PR ports the TensorFlow implementation of the RandomCrop layer to keras_core and aims to end the dependency of this layer on tf.keras.

Linked Issue: keras-team/keras#18442

[soumik12345]

@fchollet
I ran the corresponding tests locally, and two of them failed.

keras_core/layers/preprocessing/random_crop_test.py::RandomCropTest::test_predicting_with_longer_height FAILED                [ 16%]
keras_core/layers/preprocessing/random_crop_test.py::RandomCropTest::test_predicting_with_longer_width PASSED                 [ 33%]
keras_core/layers/preprocessing/random_crop_test.py::RandomCropTest::test_random_crop PASSED                                  [ 50%]
keras_core/layers/preprocessing/random_crop_test.py::RandomCropTest::test_random_crop_full PASSED                             [ 66%]
keras_core/layers/preprocessing/random_crop_test.py::RandomCropTest::test_random_crop_partial PASSED                          [ 83%]
keras_core/layers/preprocessing/random_crop_test.py::RandomCropTest::test_tf_data_compatibility FAILED                        [100%]

[fchollet]

Thanks for the PR! 👍

[fchollet]

Looks great! The last thing you're going to need is compatibility with tf.data. To achieve this, you just need to:

• Inherit from TFDataLayer (look at other preprocessing layers, like CenterCrop for example)
• Use self.backend when accessing ops
• Pass backend=self.backend in smart_resize
• Handle seed generators for the tf.data case -- take a look at RandomBrightness for an example. Basically in the tf.data case we need to make a new seed generator.

[soumik12345]

Use self.backend when accessing ops

@fchollet
Should all ops be replace by self.backend.ops?

[fchollet]

Yes -- this is the only way to make it TF Data compatible.

[soumik12345]

@fchollet
I'm getting this error in the CI for TF tests

AttributeError: module 'keras_core.backend.tensorflow' has no attribute 'ops'

[fchollet]

Thanks for the update -- the code is looking good!

• What about the test failures?
• Do we have sufficient test coverage here? You can use pytest-cov to check for coverage.

[fchollet]

Note: you could just leave name to kwargs

[fchollet]

To avoid an eager performance overhead, please move this to a method, _randomly_crop_inputs (otherwise redefining the function at each execution is slow)

[fchollet]

Hi @soumik12345 ! Are you still working on this? Can we help? Should we take it over?

[soumik12345]

Hi @soumik12345 ! Are you still working on this? Can we help? Should we take it over?

Hi @fchollet I am stuck with the unit tests. Would really appreciate some help.

[fchollet]

I pulled it and fixed it up -- but unfortunately I'm hitting intractable issues with JAX (it works fine with the other backends): JAX does not support dynamic slicing. More specifically, slice sizes must be static in https://jax.readthedocs.io/en/latest/_autosummary/jax.lax.dynamic_slice.html

As far as I can tell this seems to prohibit random cropping in JAX. Will investigate further.

[fchollet]

Finally found a fix for JAX. This is now merged. Thank you! You can check my changes FYI.