# Python Weighted Object Picker

I've designed a class ObjectPicker that can be given objects and weights. Objects can be picked randomly out of the ObjectPicker, with heavier weighted objects having a higher chance of being selected.

The pick() method is designed to be recursive and provide $O(\log n)$ speed. What I'm not sure about is the defaulting that's currently in place. I'm wondering if there is a better way to handle it. However, I would appreciate any feedback you are willing to provide beyond the specific.

class ObjectPicker:
"""
An object that stores other objects with weights, and is able to
pick out one stored object, with a higher chance for higher weighted
objects.
"""

def __init__(self):
"""
Initialize the Object Picker object. Create the instance variable
bucket for storing objects.
"""
self.bucket = []

"""
Add an item to the bucket, with given weight.
:param item: Object to be stored.
:param weight: Weight given to the object to determine odds.
"""
if self.bucket:  # If there is anything in the bucket.
total = self.bucket[-1][-1]  # Get the end number of the last item.
else:  # If the bucket is empty.
total = 0  # End number is 0.

# Object stored in Tuple.
# (Object, Weight, Starting Num, Ending Num)
self.bucket.append((item, weight, total, total + weight))

def pick(self, choice=-1, storage=None):
"""
Pick an object from the bucket recursively,
taking weight into account.
:param choice: Number of choice.
:param storage: Storage Object to choose from.
"""
if not storage:  # If no storage is given.
if not self.bucket:  # If bucket is empty.
return None  # Return None.
else:
storage = self.bucket  # Storage is bucket
total = storage[-1][-1]  # Get final weight.

if choice < 0:  # If choice is < 0,
# Randomly choose a number to represent the choice.
choice = random.random() * total

# Start binary search in middle of storage object.
index = len(storage) // 2
start, end = storage[index][2:]

# If the choice is lower, recursively search left half.
if choice < start:
return self.pick(choice, storage[:index])
# If the choice is higher, recursively search right half.
elif choice > end:
return self.pick(choice, storage[index + 1:])
# Otherwise, choice is in number spread and return object.
else:
return storage[index][0]


# doc strings

Good. I like seeing doc strings. But, let's take a closer look.

class ObjectPicker:
"""
An object that stores other objects with weights, and is able to
pick out one stored object, with a higher chance for higher weighted
objects.
"""


Ok, that is a lot of text. Can we separate this out into two sections with a short title?

class ObjectPicker:
"""
Choose a random element taking weights into account.

Elements with a higher weight have a higher change of being chosen.
"""


On to the init.

    def __init__(self):
"""
Initialize the Object Picker object. Create the instance variable
bucket for storing objects.
"""
...


There are two things wrong with it. It is telling me the following:

• It initialises the Object Picker object: That's what __init__ is supposed to do. Good. But that does not really warrant mentioning. So you can leave that portion out.
• It creates an instance variable bucket for storing objects: That's an implementation detail. Where and when the bucket variable is initialized does not really matter. That it creates a variable bucket instead of storage or something else does not matter. Again, leave that out.

So we move to

    def __init__(self):
...


(That is, remove the doc string.)

add and pick can be given similar treatments. Importantly, add an extra empty line between the documentation, and the parameter description.

Look at the following code

if self.bucket:  # If there is anything in the bucket.
total = self.bucket[-1][-1]  # Get the end number of the last item.
else:  # If the bucket is empty.
total = 0  # End number is 0.


The comments for the if and else are really trivial. You can easily drop them. Please do.

Consider what your comments add to the code. If they detail how it's implemented, try very hard to find a way to remove them without reducing the clarity. The following is just as clear.

if self.bucket:
total = self.bucket[-1][-1]  # Get the end number of the last item.
else:
total = 0


You notice I left 1 comment in. That's because of the end number. If you follow the namedtuple advice below, you can replace that line with

total = self.bucket[-1].end


(without the comment). The code is just as clear.

## Excessive commenting in general

Another comment I'd just like to point out. Somewhere in the code I see

return None  # Return None.


As an outsider, comments like this make me believe that you are a beginner in Python, or even programming in general. Probably even following a class where the lecturer dictates comments like that.

Commenting like this screams 'I know nothing'. However, when I read the code without the comments, it's very readable, and only a few comments remain that are actually necessary. Actually, just two:

# Get the end number of the last item.


and

# Start binary search in middle of storage object.


To remove them, you'd need to refactor your code. And later on I'll give you the ingredients needed in a bit more detail.

You might want to leave in some more comments, depending on how comfortable you are, but make sure that the comments add value. Commonly, that's done by explaining why a piece of code is written, not what it does.

Let me state the following a bit more clearly. It is the excessive commenting that makes me doubt your experience. Reading your code (and explanation) tells me a different story: You know how to handle a binary search (and from what I can see, correctly), you know something about complexity theory. This is somebody to be reckoned with.

# Algorithm

## Storing of elements.

You use a tuple to store the elements, leading to code such as

total = self.bucket[-1][-1]


and

start, end = storage[index][2:]


The tuple is always the same size, and the same form. So, why not make it a namedtuple? Put the following import at the top of your file:

from collections import namedtuple

RangedElement = namedtuple('RangedElement', ['value', 'length', 'start', 'end'])


(I cheated a bit, I moved from weights to lengths, because it somewhat makes more sense when talking about start and end.)

self.bucket.append((item, weight, total, total + weight))


write

self.bucket.append(RangedElement(item, weight, total, total + weight))


Yes, it's more typing. But, instead of

total = storage[-1][-1]


you can now write

total = storage[-1].end


Which is much better.

# Getting the (random) element.

## Defaulting

You already mentioned you were not sure about the defaulting. Important should be the question 'why the defaulting'?

From what I can see, the only reason for the defaulting is the recursive algorithm you use, which is a recursive bisection algorithm (or binary search algorithm). It would be good to separate the bisecting from the choosing.

def _bisect(self, choice, storage):
"""
Bisect to find the stored object.
"""
# Bisection is currently done recursively

# Start binary search in middle of storage object.
index = len(storage) // 2
start, end = storage[index][2:]

# If the choice is lower, recursively search left half.
if choice < start:
return self._bisect(choice, storage[:index])
# If the choice is higher, recursively search right half.
elif choice > end:
return self._bisect(choice, storage[index + 1:])
# Otherwise, choice is in number spread and return object.
else:
return storage[index][0]

def pick(self):
"""
Pick an object from the bucket recursively,
taking weight into account.

:param choice: Number of choice.
:param storage: Storage Object to choose from.
"""
if not self.bucket:  # If bucket is empty.
return None  # Return None.
else:
storage = self.bucket  # Storage is bucket
total = storage[-1][-1]  # Get final weight.

# Randomly choose a number to represent the choice.
choice = random.random() * total

return self._bisect(choice, storage)


(Why I chose this: I saw that some arguments were recursion-specific, and were basically saying "Don't do this when you're already recursing". A clear pointer that we were mixing two concerns, which are now better separated).

Note that splitting like this can be somewhat dangerous when this method is called somewhere else as well with the extra parameters. Sufficient unit tests would detect it.

## Bisecting

You probably saw I named the helper method _bisect. That is because it uses a fairly trivial bisection algorithm. The algorithm works recursively. Recursive algorithms are really nice, but Python has a recursion-limit, no tail-call optimisations, and you are also building subslices of bucket.

In fact, look at the following lines:

index = len(storage) // 2
....
return self.pick(choice, storage[:index])
....
return self.pick(choice, storage[index + 1:])
....


In the first call, you're copying exactly half the list. Slicing a list is linear in the length of the resulting slice. So this alone is O(n/2) = O(n). Due to recursion, you get O(n/2 + n/4 + n/8 + ...) (still O(n), though).

So your algorithm is actually linear due to copying. However, this is fixable. Instead of slicing storage, pass in offsets (hi/lo), calculate mid = lo + (hi - lo) // 2. Then recurse using hi = mid or lo = mid + 1.

However, instead of re-inventing the wheel, why not look at the Python bisect module?

https://hg.python.org/cpython/file/3.5/Lib/bisect.py

(I'm intentionally pointing at the source, because I'm going to copy the code, not call the module)

def _bisect(self, choice, storage):
lo, hi = 0, len(storage)
while True:
mid = (lo + hi) // 2
if storage[mid].end < choice:
lo = mid + 1
elif storage[mid].start > choice:
hi = mid
else:
return storage[mid].value


Now I just hope I got my boundary conditions right (that's always a problem with a bisection algorithm). Just write plenty of unit tests for this specific part, please!

## Choosing from an empty storage?

I'm specifically talking about the following lines:

if not self.bucket:  # If bucket is empty.
return None  # Return None.


Is None also an allowed value in your buckets? If so, this could cause confusing bugs. Better:

if not self.bucket:
raise EmptyBucketError("Can't pick from an empty bucket.")


or something similar. But this is something you need to decide yourself.

# Conclusion

Great code, but it could use some (very) minor improvements with great benefits:

• namedtuple
• extract binary search from picking an element.
• Thanks for the very detailed feedback. I never thought that I could have too many comments, but once you point it out it's very clear. As a self-taught programmer, being seen as someone very new is a big concern of mine. Would you have any resource recommendations on commenting best practices? Mar 7 '16 at 15:21
• I've noticed that in my previous code, I've used defaults as basically flags, is that in general seen as bad practice, and I should split into separate concerns? Or is there a general rule of thumb? Mar 7 '16 at 15:24
• @user2004245: I don't really have any resources on 'best practices' regarding comments, but things I keep hearing often are: "Explain the why, not the how.". Try to check if there is anything the comment covers that the code does not. If not, the comment adds nothing. Mar 7 '16 at 21:08
• @user2004245: In general, flags are not that bad an idea. But (I think this explains it best): do check if it makes sense for others to supply the default values, or if you're only using them for internal purposes. As for the rule of thumb: see if a method/function/... does more than 1 thing. In this case, it did: It did something with random numbers, and it did a (complicated/recursive) lookup. Mar 7 '16 at 21:11

In Python 3.6, the random.choices function was introduced, which supports weights. Thus your class could simply become:

import random

class ObjectPicker:

def __init__(self):
self.bucket = []
self.weights = []

self.bucket.append(item)
self.weights.append(weight)

def random_pick(self):
return random.choices(self.bucket, weights=self.weights, k=1)[0]


(docstrings omitted for brevity's sake.)

This even allows getting multiple

This can be sped-up slightly if needed by saving cumulative weights, instead of the weights list, as noted in the documentation:

class ObjectPicker:

def __init__(self):
self.bucket = []
self.cum_weights = []


If you want to you can even expose k to get multiple random elements (with replacement).