[API RFC] Support `IColumn.if_else` for Boolean Column

cond.if_else(a, b) -> IColumn, the output is defined as:

• out_i = a_i if cond_i is True
• out_i = b_i otherwise

API Demo

>>> import torcharrow as ta
>>> cond = ta.Column([True, False, True, True])
>>> a = ta.Column([1, 2, 3, 4])
>>> b = ta.Column([10, 20, 30, 40])
>>> cond.if_else(a, b)
0   1
1  20
2   3
3   4
dtype: int64, length: 4, null_count: 0

Current API

It’s currently called ite in Boolean Column: https://github.com/facebookresearch/torcharrow/blob/0dbe14d399b766a2dfd596e35cb7843e7514ad59/torcharrow/icolumn.py#L848-L853

API in other frameworks

PyArrow

pyarrow.compute has if_else. Note instead of being a member function to BooleanArray, it’s a standalone function in pyarrow.compute package that takes 3 arguments (cond, left, right)

Pandas

Doesn’t seem to have same method?

NumPy/PyTorch/TensorFlow

NumPy provides np.where: https://numpy.org/doc/stable/reference/generated/numpy.where.html PyTorch provides torch.where: https://pytorch.org/docs/stable/generated/torch.where.html TensorFlow provides tf.where: https://www.tensorflow.org/api_docs/python/tf/where

R

ifelse as a standalone function: https://www.rdocumentation.org/packages/base/versions/3.6.2/topics/ifelse

But in general, R seems to prefer standalone functions instead of fluent-style API.

Discussions

Shall we make ta.if_else as standalone function that takes 3 arguments in torcharrow package (rather than a member methods in BooleanColumn`)? – This is similar to Arrow Compute, PyTorch and TensorFlow.

• Standalone function (if_else(cond, x, y) ):

ta.if_else(a < 1, a, 1)

ta.if_else(
    tf.contains([1, 2, 3], a), 
    a, b
)

• cond.if_else(x, y)

(a < 1).if_else(a, 1)

tf.contains([1, 2, 3], a).if_else(a, b)

[OswinC]

I see. Yeah looks like readability here is more or less a taste thing :p I do agree that we should change it to the 3-arguments form for following the convention, but I can share what I see when reading (a < 1).if_else(a, 1) and ta.if_else

explicit parenthesis needs to be used in the first case (a < 1).if_else(a, 1)

The parenthesis actually separates the condition and makes it stand out more clearly when I read this expression. (isn’t parenthesis used widely for condition in many programming languages?)

The way the 2-arguments form separating condition from the candidate values also helps making the condition stand out clearly in my eyes (or brain) when the condition expression is long.

How the if_else symbol breaks condition expression and candidate expressions into 2 parts: <condition>.if_else(<then-value>, <else-value>)

e.g.

tf.contains([1, 2, 3], a).if_else(
    a, b
)

[wenleix]

Resolved by https://github.com/facebookresearch/torcharrow/pull/56