Implement superset behavior

[nassimberrada]

Overview

Your task is to assign yourself a function from the list below, add the additional functionality to the corresponding ivy function and ensure the new feature is being tested. For the backend framework offering the additional functionality, it will likely only require adding the additional argument in the call. For other backends, it might require more work/composition.

Task details

The Contributor guide was updated with the Superset section providing the details for adding superset of functionality amongst the supported backends. The page linked above includes some of the instructions and implementation examples, taking efficiency into account.

Please ensure that the following steps are carried out:

1. Arguments are added in functional/ivy
2. Arguments are added in all backend functions
3. Type Hints are added accordingly
4. The arguments are added in container(both static and instance) and array methods
5. Explained the parameters in docstrings
6. Add/edit examples in the docstrings (functional/ivy, container, array) to reflect the additional feature. See examples page for more details.
7. Update the tests and make sure they pass with the new parameters

Please refer to the contributor guide or post a question in the internal discussion thread. Feel free to create a PR to be reviewed by your POC (point of contact) and reference the PR/commit once completed in the PR/Commit column.

You will find in the corresponding issues, the following items:

• The framework with the additional functionality(ies)
• The features to be added

Functions list

activations

• mish #11658
• prelu #11659
• triu #11660

set

• unique_all #14041

creation

• one_hot #11661
• meshgrid

linalg

• transpose #11662
• diag #6407
• matrix_norm #6408
• matrix_rank #6409
• pinv #6410
• qr #6412
• svd #6413
• trace #6414
• matmul #6415
• inv #11638
• solve #11640
• det #11641
• eig #11642
• eigvals #11643
• eigh #11644
• eigvalsh #11645
• eigh_tridiagonal #11646
• eigvalsh_tridiagonal #11647
• lu_factor #11648
• svd_vals #11649
• cholesky #11650

manipulations

• split #6416
• unstack #6417
• constant_pad #6418
• pad #14400
• flatten #15795

searching

• argmax #6419
• argmin #6420

random

• random_uniform #6421
• random_normal #6422
• multinomial #6423
• randint #6424
• shuffle #6425

norms

• layer_norm #6426
• batch_norm #11651

layers

• conv_1d #11663
• dropout #11664
• dct #11653
• convolve #11654
• interpolate #11655

losses

• cross_entropy #11656

Suggestions

Please add any suggestions that you think we might have missed in the comment, in the following format:
“Ivy.some_function: Explanation of the change and why this is relevant superset behaviour, which backend has this extra behaviour etc.”
We will then review them as a team, and possibly add them to the list above.

[nassimberrada]

(As found in the original doc)
Ivy.logspace: The Torch does not have the endpoint argument unlike others
Ivy.unique: Numpy includes an “axis” argument, which determines which axis to operate on. Torch also has a “dim” argument that might also do the same thing.
Tensorflow.argmax does not support keepdims parameter
Ivy.fmax: Numpy includes a “where” argument that allows a condition to be broadcasted over the result.

[AnnaTz]

ivy.split: numpy and jax can receive indices to split at, not just number or size of sections.