pipda 0.7.0

A framework for data piping in python

pipda

A framework for data piping in python

Inspired by siuba, dfply, plydata and dplython, but with simple yet powerful APIs to mimic the dplyr and tidyr packages in python

API | Change Log | Documentation

Installation

pip install -U pipda

Usage

Verbs

Verbs are functions next to the piping sign (>>) receiving the data directly.

import pandas as pd
from pipda import (
    register_verb,
    register_func,
    register_operator,
    evaluate_expr,
    Operator,
    Symbolic,
    Context
)

f = Symbolic()

df = pd.DataFrame({
    'x': [0, 1, 2, 3],
    'y': ['zero', 'one', 'two', 'three']
})

df

#      x    y
# 0    0    zero
# 1    1    one
# 2    2    two
# 3    3    three

@register_verb(pd.DataFrame)
def head(data, n=5):
    return data.head(n)

df >> head(2)
#      x    y
# 0    0    zero
# 1    1    one

@register_verb(pd.DataFrame, context=Context.EVAL)
def mutate(data, **kwargs):
    data = data.copy()
    for key, val in kwargs.items():
        data[key] = val
    return data

df >> mutate(z=1)
#    x      y  z
# 0  0   zero  1
# 1  1    one  1
# 2  2    two  1
# 3  3  three  1

df >> mutate(z=f.x)
#    x      y  z
# 0  0   zero  0
# 1  1    one  1
# 2  2    two  2
# 3  3  three  3

Functions used as verb arguments

# verb can be used as an argument passed to another verb
# dep=True make `data` argument invisible while calling
@register_verb(pd.DataFrame, context=Context.EVAL, dep=True)
def if_else(data, cond, true, false):
    cond.loc[cond.isin([True]), ] = true
    cond.loc[cond.isin([False]), ] = false
    return cond

# The function is then also a singledispatch generic function

df >> mutate(z=if_else(f.x>1, 20, 10))
#    x      y   z
# 0  0   zero  10
# 1  1    one  10
# 2  2    two  20
# 3  3  three  20

# function without data argument
@register_func
def length(strings):
    return [len(s) for s in strings]

df >> mutate(z=length(f.y))

#    x     y    z
# 0  0  zero    4
# 1  1   one    3
# 2  2   two    3
# 3  3 three    5

Context

The context defines how a reference (f.A, f['A'], f.A.B is evaluated)

@register_verb(pd.DataFrame, context=Context.SELECT)
def select(df, *columns):
    return df[list(columns)]

df >> select(f.x, f.y)
#    x     y
# 0  0  zero
# 1  1   one
# 2  2   two
# 3  3 three

How it works

data %>% verb(arg1, ..., key1=kwarg1, ...)

The above is a typical dplyr/tidyr data piping syntax.

The counterpart python syntax we expect is:

data >> verb(arg1, ..., key1=kwarg1, ...)

To implement that, we need to defer the execution of the verb by turning it into a Verb object, which holds all information of the function to be executed later. The Verb object won't be executed until the data is piped in. It all thanks to the executing package to let us determine the ast nodes where the function is called. So that we are able to determine whether the function is called in a piping mode.

If an argument is referring to a column of the data and the column will be involved in the later computation, the it also needs to be deferred. For example, with dplyr in R:

data %>% mutate(z=a)

is trying add a column named z with the data from column a.

In python, we want to do the same with:

data >> mutate(z=f.a)

where f.a is a Reference object that carries the column information without fetching the data while python sees it immmediately.

Here the trick is f. Like other packages, we introduced the Symbolic object, which will connect the parts in the argument and make the whole argument an Expression object. This object is holding the execution information, which we could use later when the piping is detected.

Documentation

https://pwwang.github.io/pipda/

See also datar for realcase usages.