Functional Programming + Python - Pain
Phi
Phi library for functional programming in Python that intends to remove as much of the pain as possible from your functional programming experience in Python.
Import
For demonstration purposes we will import right now everything we will need for the rest of the exercises like this
from phi.api import *
but you can also import just what you need from the phi module.
Math-like Lambdas
Operators
Using the P object you can create quick lambdas using any operator. You can write things like
f = (P * 6) / (P + 2) # lambda x: (x * 6) / (x + 2)
assert f(2) == 3 # (2 * 6) / (2 + 2) == 12 / 4 == 3
where the expression for f is equivalent to
f = lambda x: (x * 6) / (x + 2)
getitem
You can also use the P object to create lambdas that access the items of a collection
f = P[0] + P[-1] # lambda x: x[0] + x[-1]
assert f([1,2,3,4]) == 5 #1 + 4 == 5
field access
If you want create lambdas that access the field of some entity you can use the Rec (for Record) object an call that field on it
from collections import namedtuple
Point = namedtuple('Point', ['x', 'y'])
f = Rec.x + Rec.y # lambda p: p.x + p.y
assert f(Point(3, 4)) == 7 # point.x + point.y == 3 + 4 == 7
method calling
If you want to create a lambda that calls the method of an object you use the Obj object and call that method on it with the parameters
f = Obj.upper() + ", " + Obj.lower() # lambda s: s.upper() + ", " + s.lower()
assert f("HEllo") == "HELLO, hello" # "HEllo".upper() + ", " + "HEllo".lower() == "HELLO" + ", " + "hello" == "HELLO, hello"
Here no parameters were needed but in general
f = Obj.some_method(arg1, arg2, ...) #lambda obj: obj.some_method(arg1, arg2, ...)
is equivalent to
f = lambda obj: obj.some_method(arg1, arg2, ...)
Composition
>> and <<
You can use the >> operator to forward compose expressions
f = P + 7 >> math.sqrt #executes left to right
assert f(2) == 3 # math.sqrt(2 + 7) == math.sqrt(9) == 3
This is preferred because it is more readable, but you can use the << to compose them backwards just like the mathematical definition of function composition
f = math.sqrt << P + 7 #executes right to left
assert f(2) == 3 # math.sqrt(2 + 7) == math.sqrt(9) == 3
Seq and Pipe
If you need to do a long or complex composition you can use Seq (for ‘Sequence’) instead of many chained >>
f = Seq(
str,
P + "00",
int,
math.sqrt
)
assert f(1) == 10 # sqrt(int("1" + "00")) == sqrt(100) == 10
If you want to create a composition and directly apply it to an initial value you can use Pipe
assert 10 == Pipe(
1, #input
str, # "1"
P + "00", # "1" + "00" == "100"
int, # 100
math.sqrt #sqrt(100) == 10
)
Combinators
List, Tuple, Set, Dict
There are a couple of combinators like List, Tuple, Set, Dict that help you create compound functions that return the container types list, tuple, set and dict respectively. For example, you can pass List a couple of expressions to get a function that returns a list with the values of these functions
f = List( P + 1, P * 10 ) #lambda x: [ x +1, x * 10 ]
assert f(3) == [ 4, 30 ] # [ 3 + 1, 3 * 10 ] == [ 4, 30 ]
The same logic applies for Tuple and Set. With Dict you have to use keyword arguments
f = Dict( x = P + 1, y = P * 10 ) #lambda x: [ x +1, x * 10 ]
d = f(3)
assert d == { 'x': 4, 'y': 30 } # { 'x': 3 + 1, 'y': 3 * 10 } == { 'x': 4, 'y': 30 }
assert d.x == 4 #access d['x'] via field access as d.x
assert d.y == 30 #access d['y'] via field access as d.y
As you see, Dict returns a custom dict that also allows field access, this is useful because you can use it in combination with Rec.
State: Read and Write
Internally all these expressions are implemented in such a way that they not only pass their computed values but also pass a state dictionary between them in a functional manner. By reading from and writing to this state dictionary the Read and Write combinators can help you “save” the state of intermediate computations to read them later
assert [70, 30] == Pipe(
3,
Write(s = P * 10), #s = 3 * 10 == 30
P + 5, #30 + 5 == 35
List(
P * 2 # 35 * 2 == 70
,
Read('s') #s == 30
)
)
If you need to perform many reads inside a list -usually for output- you can use ReadList instead
assert [2, 4, 22] == Pipe(
1,
Write(a = P + 1), #a = 1 + 1 == 2
Write(b = P * 2), #b = 2 * 2 == 4
P * 5, # 4 * 5 == 20
ReadList('a', 'b', P + 2) # [a, b, 20 + 2] == [2, 4, 22]
)
ReadList interprets string elements as Reads, so the previous is translated to
List(Read('a'), Read('b'), P + 2)
Then, Then2, …, Then5, ThenAt
To create a partial expression from a function e.g.
def repeat_word(word, times, upper=False):
if upper:
word = word.upper()
return [ word ] * times
use the Then combinator which accepts a function plus all but the 1st of its *args + **kwargs
f = P[::-1] >> Then(repeat_word, 3)
g = P[::-1] >> Then(repeat_word, 3, upper=True)
assert f("ward") == ["draw", "draw", "draw"]
assert g("ward") == ["DRAW", "DRAW", "DRAW"]
and assumes that the 1st argument of the function will be applied last, e.g. word in the case of repeat_word. If you need the 2nd argument to be applied last use Then2, and so on. In general you can use ThenAt(n, f, *args, **kwargs) where n is the position of the argument that will be applied last. For example
# since map and filter receive the iterable on their second argument, you have to use `Then2`
f = Then2(filter, P % 2 == 0) >> Then2(map, P**2) >> list #lambda x: map(lambda z: z**2, filter(lambda z: z % 2 == 0, x))
assert f([1,2,3,4,5]) == [4, 16] #[2**2, 4**2] == [4, 16]
Be aware that P already has the map and filter methods so you can write the previous more easily as
f = P.filter(P % 2 == 0) >> P.map(P**2) >> list #lambda x: map(lambda z: z**2, filter(lambda z: z % 2 == 0, x))
assert f([1,2,3,4,5]) == [4, 16] #[2**2, 4**2] == [4, 16]
Val
If you need to create a constant function with a given value use Val
f = Val(42) #lambda x: 42
assert f("whatever") == 42
Others
Check out the With, If and more, combinators on the documentation. The P object also offers some useful combinators as methods such as Not, First, Last plus almost all python built in functions as methods:
f = Obj.split(' ') >> P.map(len) >> sum >> If( (P < 15).Not(), "Great! Got {0} letters!".format).Else("Too short, need at-least 15 letters")
assert f("short frase") == "Too short, need at-least 15 letters"
assert f("some longer frase") == "Great! Got 15 letters!"
The DSL
Phi has a small omnipresent DSL that has these simple rules:
- Any element of the class
Expressionis an element of the DSL.Pand all the combinators are of theExpressionclass. - Any callable of arity 1 is an element of the DSL.
- The container types
list,tuple,set, anddictare elements of the DSL. They are translated to their counterpartsList,Tuple,SetandDict, their internal elements are forwarded. - Any value
xthat does not comply with any of the previous rules is also an element of the DSL and is translated toVal(x).
Using the DSL, the expression
f = P**2 >> List( P, Val(3), Val(4) ) #lambda x: [ x**2]
assert f(10) == [ 100, 3, 4 ] # [ 10**2, 3, 4 ] == [ 100, 3, 4 ]
can be rewritten as
f = P**2 >> [ P, 3, 4 ]
assert f(10) == [ 100, 3, 4 ] # [ 10 ** 2, 3, 4 ] == [ 100, 3, 4 ]
Here the values 3 and 4 are translated to Val(3) and Val(4) thanks to the 4th rule, and [...] is translated to List(...) thanks to the 3rd rule. Since the DSL is omnipresent you can use it inside any core function, so the previous can be rewritten using Pipe as
assert [ 100, 3, 4 ] == Pipe(
10,
P**2, # 10**2 == 100
[ P, 3, 4 ] #[ 100, 3, 4 ]
)
F
You can compile any element to an Expression using F
f = F((P + "!!!", 42, Obj.upper())) #Tuple(P + "!!!", Val(42), Obj.upper())
assert f("some tuple") == ("some tuple!!!", 42, "SOME TUPLE")
Other example
f = F([ P + n for n in range(5) ]) >> [ len, sum ] # lambda x: [ len([ x, x+1, x+2, x+3, x+4]), sum([ x, x+1, x+2, x+3, x+4]) ]
assert f(10) == [ 5, 60 ] # [ len([10, 11, 12, 13, 14]), sum([10, 11, 12, 13, 14])] == [ 5, (50 + 0 + 1 + 2 + 3 + 4) ] == [ 5, 60 ]
Fluent Programming
All the functions you’ve seen are ultimately methods of the PythonBuilder class which inherits from the Expression, therefore you can also fluently chain methods instead of using the >> operator. For example
f = Dict(
x = 2 * P,
y = P + 1
).Tuple(
Rec.x + Rec.y,
Rec.y / Rec.x
)
assert f(1) == (4, 1) # ( x + y, y / x) == ( 2 + 2, 2 / 2) == ( 4, 1 )
This more complicated previous example
f = Obj.split(' ') >> P.map(len) >> sum >> If( (P < 15).Not(), "Great! Got {0} letters!".format).Else("Too short, need at-least 15 letters")
assert f("short frase") == "Too short, need at-least 15 letters"
assert f("some longer frase") == "Great! Got 15 letters!"
can be be rewritten as
f = (
Obj.split(' ')
.map(len)
.sum()
.If( (P < 15).Not(),
"Great! Got {0} letters!".format
).Else(
"Too short, need at-least 15 letters"
)
)
assert f("short frase") == "Too short, need at-least 15 letters"
assert f("some longer frase") == "Great! Got 15 letters!"
Integrability
Register, Register2, …, Register5, RegistarAt
If you want to have custom expressions to deal with certain data types, you can create a custom class that inherits from Builder or PythonBuilder
from phi import PythonBuilder
class MyBuilder(PythonBuilder):
pass
M = MyBuilder()
and register your function in it using the Register class method
def remove_longer_than(some_list, n):
return [ elem from elem in some_list if len(elem) <= n ]
MyBuilder.Register(remove_longer_than, "my.lib.")
Or better even use Register as a decorator
@MyBuilder.Register("my.lib.")
def remove_longer_than(some_list, n):
return [ elem for elem in some_list if len(elem) <= n ]
Now the method MyBuilder.remove_longer_than exists on this class. You can then use it like this
f = Obj.lower() >> Obj.split(' ') >> M.remove_longer_than(6)
assert f("SoMe aRe LONGGGGGGGGG") == ["some", "are"]
As you see the argument n = 6 was partially applied to remove_longer_than, an expression which waits for the some_list argument to be returned. Internally the Registar* method family uses the Then* method family.
PatchAt
If you want to register a batch of functions from a module or class automatically you can use the PatchAt class method. It’s an easy way to integrate an entire module to Phi’s DSL. See PatchAt.
Libraries
Phi currently powers the following libraries that integrate with its DSL:
- PythonBuilder : helps you integrate Python’s built-in functions and keywords into the phi DSL and it also includes a bunch of useful helpers for common stuff.
phi’s globalPobject is an instance of this class. [Shipped with Phi] - TensorBuilder: a TensorFlow library enables you to easily create complex deep neural networks by leveraging the phi DSL to help define their structure.
- NumpyBuilder: Comming soon!
Documentation
Check out the complete documentation.
More Examples
The global phi.P object exposes most of the API and preferably should be imported directly. The most simple thing the DSL does is function composition:
from phi.api import *
def add1(x): return x + 1
def mul3(x): return x * 3
x = Pipe(
1.0, #input 1
add1, #1 + 1 == 2
mul3 #2 * 3 == 6
)
assert x == 6
Use phi lambdas to create the functions
from phi.api import *
x = Pipe(
1.0, #input 1
P + 1, #1 + 1 == 2
P * 3 #2 * 3 == 6
)
assert x == 6
Create a branched computation instead
from phi.api import *
[x, y] = Pipe(
1.0, #input 1
[
P + 1 #1 + 1 == 2
,
P * 3 #1 * 3 == 3
]
)
assert x == 2
assert y == 3
Compose it with a function equivalent to f(x) = (x + 3) / (x + 1)
from phi.api import *
[x, y] = Pipe(
1.0, #input 1
(P + 3) / (P + 1), #(1 + 3) / (1 + 1) == 4 / 2 == 2
[
P + 1 #2 + 1 == 3
,
P * 3 #2 * 3 == 6
]
)
assert x == 3
assert y == 6
Give names to the branches
from phi.api import *
result = Pipe(
1.0, #input 1
(P + 3) / (P + 1), #(1 + 3) / (1 + 1) == 4 / 2 == 2
dict(
x = P + 1 #2 + 1 == 3
,
y = P * 3 #2 * 3 == 6
)
)
assert result.x == 3
assert result.y == 6
Divide x by y.
from phi.api import *
result = Pipe(
1.0, #input 1
(P + 3) / (P + 1), #(1 + 3) / (1 + 1) == 4 / 2 == 2
dict(
x = P + 1 #2 + 1 == 3
,
y = P * 3 #2 * 3 == 6
),
Rec.x / Rec.y #3 / 6 == 0.5
)
assert result == 0.5
Save the value from the (P + 3) / (P + 1) computation as s and load it at the end in a branch
from phi.api import *
[result, s] = Pipe(
1.0, #input 1
Write(s = (P + 3) / (P + 1)), #s = 4 / 2 == 2
dict(
x = P + 1 #2 + 1 == 3
,
y = P * 3 #2 * 3 == 6
),
[
Rec.x / Rec.y #3 / 6 == 0.5
,
Read('s') #s == 2
]
)
assert result == 0.5
assert s == 2
Add 3 to the loaded s for fun and profit
from phi.api import *
[result, s] = Pipe(
1.0, #input 1
Write(s = (P + 3) / (P + 1)), #s = 4 / 2 == 2
dict(
x = P + 1 #2 + 1 == 3
,
y = P * 3 #2 * 3 == 6
),
[
Rec.x / Rec.y #3 / 6 == 0.5
,
Read('s') + 3 # 2 + 3 == 5
]
)
assert result == 0.5
assert s == 5
Use the Read and Write field access lambda style just because
from phi.api import *
[result, s] = Pipe(
1.0, #input 1
(P + 3) / (P + 1), #4 / 2 == 2
Write.s, #s = 2
dict(
x = P + 1 #2 + 1 == 3
,
y = P * 3 #2 * 3 == 6
),
[
Rec.x / Rec.y #3 / 6 == 0.5
,
Read.s + 3 # 2 + 3 == 5
]
)
assert result == 0.5
assert s == 5
Add an input Val of 9 on a branch and add to it 1 just for the sake of it
from phi.api import *
[result, s, val] = Pipe(
1.0, #input 1
(P + 3) / (P + 1), Write.s, #4 / 2 == 2, saved as 's'
dict(
x = P + 1 #2 + 1 == 3
,
y = P * 3 #2 * 3 == 6
),
[
Rec.x / Rec.y #3 / 6 == 0.5
,
Read.s + 3 # 2 + 3 == 5
,
Val(9) + 1 #input 9 and add 1, gives 10
]
)
assert result == 0.5
assert s == 5
assert val == 10
Do the previous only if y > 7 else return "Sorry, come back later."
from phi.api import *
[result, s, val] = Pipe(
1.0, #input 1
(P + 3) / (P + 1), Write.s, #4 / 2 == 2, saved as 's'
dict(
x = P + 1 #2 + 1 == 3
,
y = P * 3 #2 * 3 == 6
),
[
Rec.x / Rec.y #3 / 6 == 0.5
,
Read.s + 3 # 2 + 3 == 5
,
If( Rec.y > 7,
Val(9) + 1 #input 9 and add 1, gives 10
).Else(
"Sorry, come back later."
)
]
)
assert result == 0.5
assert s == 5
assert val == "Sorry, come back later."
Now, what you have to understand that everything you’ve done with these expression is to create and apply a single function. Using Seq we can get the standalone function and then use it to get the same values as before
from phi.api import *
f = Seq(
(P + 3) / (P + 1), Write.s, #4 / 2 == 2, saved as 's'
dict(
x = P + 1 #2 + 1 == 3
,
y = P * 3 #2 * 3 == 6
),
[
Rec.x / Rec.y #3 / 6 == 0.5
,
Read.s + 3 # 2 + 3 == 5
,
If( Rec.y > 7,
Val(9) + 1 #input 9 and add 1, gives 10
).Else(
"Sorry, come back later."
)
]
)
[result, s, val] = f(1.0)
assert result == 0.5
assert s == 5
assert val == "Sorry, come back later."
Even More Examples
from phi.api import *
avg_word_length = Pipe(
"1 22 333",
Obj.split(" "), # ['1', '22', '333']
P.map(len), # [1, 2, 3]
P.sum() / P.len() # sum([1,2,3]) / len([1,2,3]) == 6 / 3 == 2
)
assert 2 == avg_word_length
from phi.api import *
assert False == Pipe(
[1,2,3,4], P
.filter(P % 2 != 0) #[1, 3], keeps odds
.Contains(4) #4 in [1, 3] == False
)
from phi.api import *
assert {'a': 97, 'b': 98, 'c': 99} == Pipe(
"a b c", Obj
.split(' ').Write.keys # keys = ['a', 'b', 'c']
.map(ord), # [ord('a'), ord('b'), ord('c')] == [97, 98, 99]
lambda it: zip(Ref.keys, it), # [('a', 97), ('b', 98), ('c', 99)]
dict # {'a': 97, 'b': 98, 'c': 99}
)
Installation
pip install phi
Bleeding Edge Release
pip install git+https://github.com/cgarciae/phi.git@develop
Status
- Version:
. - Documentation coverage: 100%. Please create an issue if documentation is unclear, it is a high priority of this library.
- Milestone: reach 1.0.0 after feedback from the community.