User facing split does not match numpy

[ricardoV94]

Description

Our implementation of Split expects the sizes of the subarrays (instead of the truncation points like numpy) and these must match to the full size of the input.

This is arguably a more sensible implementation as this is valid in numpy:

import numpy as np
np.split(np.zeros(10), [5, 2])

However users wanting to use split may be tripped by the fact it is different than the numpy one. We should offer a helper for the numpy one and keep ours internally. We can canonicalize to ours when we know all split sizes are non-decreasing.

[Hashcode-Ankit]

Hi @ricardoV94 ,
I am thinking to extend the following function:

def split(x, splits_size, n_splits, axis=0):
    the_split = Split(n_splits)
    return the_split(x, axis, splits_size)


class Split(COp):
    """Partition a `TensorVariable` along some axis.

with a extra parameter numpy_split which will be bool, based on that we can create split according to numpy split function Or you want to create split in such as way it uses pytensor split for non decreasing else it do modifications accordingly (may be not use pytensor split).

[ricardoV94]

Not sure I understand your suggestion completely. Can you explain in more detail?

[Hashcode-Ankit]

Not sure I understand your suggestion completely. Can you explain in more detail?

I am thinking this way :

1. modify the split methond not Split
2. split method use pytensor Split if it is a non decreasing array
3. else it will do some modifications for expected result.

[ricardoV94]

We should probably still use the PyTensor Split under the hood, just adapt the user facing split to convert the behavior from numpy to pytensor.

We need a transition, perhaps add a kwarg mode: Literal["numpy", "pytensor"] | None = None, and have a FutureWarning when the default None is used (the user didn't specify), saying the behavior will change to match that of numpy in the future. To suppress the warning explicitly set mode="numpy" or mode="pytensor". After a while we remove the None option and the default becomes "numpy"

[Hashcode-Ankit]

@ricardoV94 will raise a pr for it
Thanks