PyMonad is a small library implementing monads and related data abstractions -- functors, applicative functors, and monoids -- for use in implementing functional style programs. For those familiar with monads in Haskell, PyMonad aims to implement many of the features you're used to so you can use monads in python quickly and easily. For those who have never used monads but are interested, PyMonad is an easy way to learn about them in, perhaps, a slightly more forgiving environment without needing to learn Haskell.
- Easily define curried functions with the
@currydecorator. - Straight-forward partial application: just pass a curried function the number of arguments you want.
- Composition of curried functions using
<<. - Functor, Applicative Functor, and Monad operators:
<<,&, and>>. - Monoids - int, float, str, list, List, Maybe, First, and Last:
+. - Six predefined monad types.
- Maybe - For when a calculation might fail.
- Either - Similar to Maybe but with additional error reporting.
- List - For non-deterministic calculations.
- Reader - For sequencing calculations which all access the same data.
- Writer - For keeping logs of program execution.
- State - Simulating mutable state in a purely functional way.
Using pip:
pip install PyMonad
Or download the package or clone the git repository from bitbucket <https://bitbucket.org/jason_delaat/pymonad> and run:
python setup.py install
from the project directory.
Import the entire package:
from pymonad import *
Or just a single monad type:
from pymonad.Maybe import *
If you're not importing everything but want to use curried functions:
from pymonad.Reader import curry
To define a curried function use the @curry decorator:
@curry
def add(x, y):
return x + y
@curry
def func(x, y, z):
# Do something with x, y and z.
...
The above functions can be partially applied by passing them less than their full set of arguments:
add(7, 8) # Calling 'add' normally returns 15 as expected. add7 = add(7) # Partial application: 'add7' is a function taking one argument. add7(8) # Applying the final argument retruns 15... add7(400) # ... or 407, or whatever. # 'func' can be applied in any of the following ways. func(1, 2, 3) # Call func normally. func(1, 2)(3) # Partially applying two, then applying the last argument. func(1)(2, 3) # Partially applying one, then applying the last two arguments. func(1)(2)(3) # Partially applying one, partially applying again, then applying the last argument.
Curried functions can be composed with the << operator.
Functions are applied from right to left:
# Returns the first element of a list.
@curry
def head(aList):
return aList[0]
# Returns everything except the first element of the list.
@curry
def tail(aList):
return aList[1:]
second = head << tail # 'tail' will be applied first, then its result passed to 'head'
second([1, 2, 3, 4]) # returns 2
You can also compose partially applied functions:
@curry
def add(x, y):
return x + y
@curry
def mul(x, y):
return x * y
comp = add(7) << mul(2) # 'mul(2)' is evaluated first, and it's result passed to 'add(7)'
comp(4) # returns 15
# Composition order matters!
comp = mul(2) << add(7)
comp(4) # returns 22
All Monads are also Applicative Functors, and all Applicative Functors are also Functors, though the same is not necessarily true in reverse. All the types included with PyMonad are defined as all three but you can define new types however you want.
All of these types ultimately derive from Container which simply holds a value and provides some basic value equality checking.
The method getValue() will allow you to "extract" the value of monadic computations if/when necessary.
All functors define the fmap method which can be invoked via the fmap operator <<.
fmap takes functions which operate on simple types -- integers, strings, etc. -- and allows them to operate of functor types:
from pymonad.Maybe import *
from pymonad.List import *
# 'neg' knows nothing about functor types...
def neg(x):
return -x
# ... but that doesn't stop us from using it anyway.
neg << Just(9) # returns Just(-9)
neg << Nothing # returns Nothing
neg << List(1, 2, 3, 4) # returns List(-1, -2, -3, -4)
Notice that the function is on the left and the functor type is on the right.
If you think of << as a sort of fancy opening parenthesis, then normal calls and fmap calls have basically the same structure:
| Type | Function | Opener | Argument | Sep | Argument | Closer |
|---|---|---|---|---|---|---|
| Normal | neg |
( |
9 |
) |
||
| Normal | add |
( |
7 |
, |
8 |
) |
fmap |
neg |
<< |
Just(9) |
|||
amap |
add |
<< |
Just(7) |
& |
Just(8) |
Notice that << is also the function composition operator.
In fact, curried functions are instances of the Reader monad, and fmap -ing a function over another function is the same thing as function composition.
As shown next, functions that take multiple arguments can be called by chaining ("anding", amap -ing) the functors.
Functors allow you to use normal functions of a single argument -- like neg above -- with functor types.
Applicative Functors extend that capability -- via amap and its operator & -- allowing you to use normal functions of multiple arguments with functor types:
# 'add' operates on simple types, not functors or applicatives...
def add(x, y):
return x + y
# ... but we're going to use it on those types anyway.
# Note that we're still using '<<' but now in conjunction with '&'
add << Just(7) & Just(8) # returns Just(15)
add << Nothing & Just(8) # returns Nothing
add << Just(7) & Nothing # returns Nothing
add << List(1, 2, 3) & List(4, 5, 6) # returns List(5, 6, 7, 6, 7, 8, 7, 8, 9)
If << is a fancy parenthesis, & is the fancy comma used to separate arguments.
Monads allow you to sequence a series of calculations within than monad using the bind operator >>.
The first argument to >> is a monad type.
The second argument is a function which takes a single, non-monad argument and returns an instance of the same monad:
from pymonad.List import *
from pymonad.Reader import curry
# Takes a simple number type and returns a 'List' containing that value and it's negative.
def positive_and_negative(x):
return List(x, -x)
# You can call 'positive_and_negative' normally.
positive_and_negative(9) # returns List(9, -9)
# Or you can create a List...
x = List(9)
# ... and then use '>>' to apply positive_and_negative'
x >> positive_and_negative # also returns List(9, -9)
# But 'x' could also have more than one value...
x = List(1, 2)
x >> positive_and_negative # returns List(1, -1, 2, -2)
# And of course you can sequence partially applied functions.
@curry
def add_and_sub(x, y):
return List(y + x, y - x)
List(2) >> positive_and_negative >> add_and_sub(3) # creates List(2)
# applies positive_and_negative: List(2, -2)
# then add_and_sub(3): List(5, -1, 1, -5)
# final result: List(5, -1, 1, -5)
The second argument to >> is a function which takes a single, non-monad argument.
Because of that, you can use lambda to assign values to a variable withing monadic code, like this:
from pymonad.Maybe import * Just(9) >> (lambda x: # Here, 'x' takes the value '9' Just(8) >> (lambda y: # And 'y' takes the value '8' Just(x + y))) # The final returned value is 'Just(9 + 8)', or 'Just(17)'
You can also simply ignore values if you wish:
Just(9) >> Just(8) # The '9' is thrown away and the result of this computation is 'Just(8)'
Implementing other functors, applicatives, or monads is fairly straight-forward. There are three classes, serving as interfaces:
Monad --> Applicative --> Functor
To implement a new functor, create a new class which derives from Functor and override the fmap method.
To implement a new applicative functor, create a new class which derives from Applicative and override the amap and fmap methods.
To implement a new monad, create a new class which derives from Monad and override at least the bind method, and preferably the amap and fmap methods as well.
The operators, <<, &, and >> are pre-defined to call the above methods so you shouldn't need to touch them directly.
The previous version of PyMonad didn't include the method unit (called return in Haskell). unit takes a bare value, such as 8, and places it in a default context for that monad.
Haskell allows polymorphism on return types as well as supporting type inference, so you (mostly) don't have to tell return what types to expect, it just figures it out.
We can't do that in Python, so you always need to tell unit what type you're expecting.
The unit method is implemented as a class method in Functor.py, so it can be used with any functor, applicative or monad.
There is also a unit function which expects a functor type (though you can also give it an instance) and a value and invokes the corresponding unit method.
It is provided to give a more "functional look" to code, but use whichever method you prefer.
With the Maybe monad for example:
- Maybe.unit(8) # returns Just(8)
- unit(Maybe, 8) # also returns Just(8)
In either case all functors (and applicatives and monads) should implement the unit class method.
Monoids are a data type which consists of some operation for combining values of that type, and an identity value for that operation.
The operation is called mplus and the identity value is called mzero.
Despite the names, they are not necessarily addition and zero.
They can be addition and zero though, numbers are sort of the typical monoid.
In the case of numbers, zero is the identity element and addition is the operation. Monoids adhere to the following laws:
- Left and right identity:
x + 0 = 0 + x = x - Associativity:
(x + y) + z = x + (y + z) = x + y + z
Stings are also monoids with the identity element mzero equal to the empty string, and the operation mplus concatenation.
To create a new monoids type create a class deriving from Monoid and override the mzero static method which takes no arguments and should return an instance of the class containing the identity value for the monoid.
Also, override the mplus method.
For instance, numbers can be a monoid in two ways, one way with zero and addition as discussed above and the other way with one and multiplication.
We could implement that like this:
class ProductMonoid(Monoid):
@staticmethod
def mzero():
return ProductMonoid(1)
def mplus(self, other):
return ProductMonoid(self.getValue() * other.getValue())
The + operator (aka __add__()) is defined to call mplus on monoid instances, so you can simply "add" monoid values together rather than having to call mplus directly.
Similar to unit for monads, there is an mzero function which expects a type and can be used instead of the mzero method.
Unlike unit however, the mzero function serves another purpose.
Numbers, strings and lists can all be used as monoids and all already have an appropriate definition for +.
What they don't have is an mzero method.
To allow numbers, strings and lists to be used as monoids without any extra work, the mzero function will return the appropriate value for these types and will attempt to call the mzero method on anything else.
For instance:
mzero(int) # returns 0, also works with float
mzero(str) # returns ""
mzero(list) # returns []
mzero(ProductMonoid) # return ProductMonoid(1)
# etc...
If you write code involving monoids, and you're not sure what type of monoid you might be handed, you should use the mzero function and not the mzero method.
The Writer monad performs calculations and keeps a log. The log can be any monoid type -- strings being a typical example.
The Writer class doesn't have a default log type, so to use Writer you need to inherit from it.
It is extremely simple as the only thing you need to do is define the log type.
For instance:
class StringWriter(Writer):
logType = str
That's it.
Everything else is already defined by Writer.
StringWriter, NumberWriter, and ListWriter are already defined in the Writer module for you to use.
Calling unit with a Writer class packages whatever value you give it with the mzero of the log type:
unit(StringWriter, 8) # Returns Writer(8, "")
Writer constructors take two values, the first being the result of whatever calculation you've just performed, the second being the log message -- or value, or whatever -- to add to the log.
### Other Methods ###
getValue(): Returns the result and log as a two-tuple.getResult(): Returns only the result.getLog(): Returns only the log.
A quick example:
@curry
def add(x, y):
return StringWriter(x + y, "Adding " + str(x) + " and " + str(y) + ". ")
x = unit(StringWriter, 8) >> add(4) >> add(5)
print(x.getResult()) # prints 17
print(x.getLog()) # prints "Adding 8 and 4. Adding 12 and 5. "
In the definition of add`, ``StringWriter could have also been just Writer.
It's really only necessary to use subclasses when using unit, because unit checks for the logType variable.
Otherwise simply giving plain old Writer a string -- or other monoid type argument -- accomplishes the same thing.
Both unit and bind (or >>) convert *Writer types to plain Writer but using StringWriter -- or whatever -- makes your intentions more clear.
Unlike most of the other monad types, the state monad doesn't wrap values it wraps functions. Specifically, it wraps functions which accept a single 'state' argument and produce a result and a new 'state' as a 2-tuple. The 'state' can be anything: simple types like integers, lists, dictionaries, custom objects/data types, whatever. The important thing is that any given chain of stateful computations all use the same type of state.
The State constructor should only be used to create stateful computations.
Trying to use State to inject values, or even non-stateful functions, into the monad will cause it to function incorrectly.
To inject values, use the unit function.
Here's an example of using State.
We'll create a little system which can perform addition and subtraction.
Our total will never be allowed to drop below zero.
The state that we'll be keeping track of is a simple count of the total number of operations performed.
Every time we perform an addition or subtraction the count will go up by one:
@curry
def add(x, y):
return State(lambda old_state: (x + y, old_state + 1))
@curry
def subtract(y, x):
@State
def state_computation(old_state):
if x - y < 0:
return (0, old_state + 1)
else:
return (x - y, old_state + 1)
return state_computation
As mentioned, the State constructor takes a function which accepts a 'state', in this case simply an integer, and produces a result and a new state as a tuple.
Although we could have done subtract as a one-liner, I wanted to show that, if your computation is more complex than can easily be contained in a lambda expression, you can use State as a decorator to define the stateful computation.
Using these functions is now simple:
x = unit(State, 1) >> add(2) >> add(3) >> subtract(40) >> add(5)
x now contains a stateful computation but that computation hasn't been executed yet.
Since State values contain functions, you can call them like functions by supplying an initial state value:
y = x(0) # Since we're counting the total number of operations, we start at zero. print(y) # Prints (5, 4), '5' is the result and '4' is the total number of operations performed.
Calling a State function in this way will always return the (result, state) tuple.
If you're only interested in the result:
x.getResult(0) # Returns the value 5, the result of the computataion.
Or if you only care about the final state:
x.getState(0) # Returns the value 4, the final state of the computation.