We Keep Coding

Efficient Way to Repeatedly Split Large NumPy Array and Record Middle

I have a large NumPy array nodes = np.arange(100_000_000) and I need to rearrange this array by:
Recording and then removing the middle value in the array
Split the array into the left half and right half
Repeat Steps 1-2 for each half
Stop when all values are exhausted
So, for a smaller input example nodes = np.arange(10), the output would be:
[5 2 8 1 4 7 9 0 3 6]
This was accomplished by naively doing:
import numpy as np
def split(node, out):
mid = len(node) // 2
out.append(node[mid])
return node[:mid], node[mid+1:]
def reorder(a):
nodes = [a.tolist()]
out = []
while nodes:
tmp = []
for node in nodes:
for n in split(node, out):
if n:
tmp.append(n)
nodes = tmp
return np.array(out)
if __name__ == "__main__":
nodes = np.arange(10)
print(reorder(nodes))
However, this is way too slow for nodes = np.arange(100_000_000) and so I am looking for a much faster solution.

You can vectorize your function with Numpy by working on groups of slices.
Here is an implementation:
# Similar to [e for tmp in zip(a, b) for e in tmp] ,
# but on Numpy arrays and much faster
def interleave(a, b):
assert len(a) == len(b)
return np.column_stack((a, b)).reshape(len(a) * 2)
# n is the length of the input range (len(a) in your example)
def fast_reorder(n):
if n == 0:
return np.empty(0, dtype=np.int32)
startSlices = np.array([0], dtype=np.int32)
endSlices = np.array([n], dtype=np.int32)
allMidSlices = np.empty(n, dtype=np.int32) # Similar to "out" in your implementation
midInsertCount = 0 # Actual size of allMidSlices
# Generate a bunch of middle values as long as there is valid slices to split
while midInsertCount < n:
# Generate the new mid/left/right slices
midSlices = (endSlices + startSlices) // 2
# Computing the next slices is not needed for the last step
if midInsertCount + len(midSlices) < n:
# Generate the nexts slices (possibly with invalid ones)
newStartSlices = interleave(startSlices, midSlices+1)
newEndSlices = interleave(midSlices, endSlices)
# Discard invalid slices
isValidSlices = newStartSlices < newEndSlices
startSlices = newStartSlices[isValidSlices]
endSlices = newEndSlices[isValidSlices]
# Fast appending
allMidSlices[midInsertCount:midInsertCount+len(midSlices)] = midSlices
midInsertCount += len(midSlices)
return allMidSlices[0:midInsertCount]
On my machine, this is 89 times faster than your scalar implementation with the input np.arange(100_000_000) dropping from 2min35 to 1.75s. It also consume far less memory (rougthly 3~4 times less). Note that if you want a faster code, then you probably need to use a native language like C or C++.

Edit:
The question has been updated to have a much smaller input array so I leave the below for historical reasons. Basically it was likely a typo but we often get accustomed to computers working with insanely large numbers and when memory is involved they can be a real problem.
There is already a numpy based solution submitted by someone else that I think fits the bill.
Your code requires an insane amount of RAM just to hold 100 billion 64 bit integers. Do you have 800GB of RAM? Then you convert the numpy array to a list which will be substantially larger than the array (each packed 64 bit int in the numpy array will become a much less memory efficient python int object and the list will have a pointer to that object). Then you make a lot of slices of the list which will not duplicate the data but will duplicate the pointers to the data and use even more RAM. You also append all the result values to a list a single value at a time. Lists are very fast for adding items generally but with such an extreme size this will not only be slow but the way the list is allocated is likely to be extremely wasteful RAM wise and contribute to major problems (I believe they double in size when they get to a certain level of fullness so you will end up allocating more RAM than you need and doing many allocations and likely copies). What kind of machine are you running this on? There are ways to improve your code but unless you're running it on a super computer I don't know that you're going to ever finish that calculation. I only..only? have 32GB of RAM and I'm not going to even try to create a 100B int_64 numpy array as I don't want to use up ssd write life for a mass of virtual memory.
As for improving your code stick to numpy arrays don't change to a python list it will greatly increase the RAM you need. Preallocate a numpy array to put the answer in. Then you need a new algorithm. Anything recursive or recursive like (ie a loop splitting the input,) will require tracking a lot of state, your nodes list is going to be extraordinarily gigantic and again use a lot of RAM. You could use len(a) to indicate values that are removed from your list and scan through the entire array each time to figure out what to do next but that will save RAM in favour of a tremendous amount of searching a gigantic array. I feel like there is an algorithm to cut numbers from each end and place them in the output and just track the beginning and end but I haven't figured it out at least not yet.
I also think there is a simpler algorithm where you just track the number of splits you've done instead of making a giant list of slices and keeping it all in memory. Take the middle of the left half and then the middle of the right then count up one and when you take the middle of the left half's left half you know you have to jump to the right half then the count is one so you jump over to the original right half's left half and on and on... Based on the depth into the halves and the length of the input you should be able to jump around without scanning or tracking all of those slices though I haven't been able to dedicate much time to thinking this through in my head.
With a problem of this nature if you really need to push the limits you should consider using C/C++ so you can be as efficient as possible with RAM usage and because you're doing an insane number of tiny things which doesn't map well to python performance.

Justification of constants used in random.sample

I'm looking into the source code for the function sample in random.py (python standard library).
The idea is simple:
If a small sample (k) is needed from a large population (n): Just pick k random indices, since it is unlikely you'll pick the same number twice as the population is so large. And if you do, just pick again.
If a relatively large sample (k) is needed, compared to the total population (n): It is better to keep track of what you have picked.
My Question
There are a few constants involved, setsize = 21 and setsize += 4 ** _log(3*k,4). The critical ratio is roughly k : 21+3k. The comment says # size of a small set minus size of an empty list and # table size for big sets.
Where have these specific numbers come from? What is there justification?
The comments shed some light, however I find they bring as many questions as they answer.
I would kind of understand, size of a small set but find the "minus size of an empty list" confusing. Can someone shed any light on this?
what is meant specifically by "table" size, as apposed to say "set size".
Looking on the github repository, it looks like a very old version simply used the ratio k : 6*k, as the critical ratio, but I find that equally mysterious.
The code
def sample(self, population, k):
"""Chooses k unique random elements from a population sequence or set.
Returns a new list containing elements from the population while
leaving the original population unchanged. The resulting list is
in selection order so that all sub-slices will also be valid random
samples. This allows raffle winners (the sample) to be partitioned
into grand prize and second place winners (the subslices).
Members of the population need not be hashable or unique. If the
population contains repeats, then each occurrence is a possible
selection in the sample.
To choose a sample in a range of integers, use range as an argument.
This is especially fast and space efficient for sampling from a
large population: sample(range(10000000), 60)
"""
# Sampling without replacement entails tracking either potential
# selections (the pool) in a list or previous selections in a set.
# When the number of selections is small compared to the
# population, then tracking selections is efficient, requiring
# only a small set and an occasional reselection. For
# a larger number of selections, the pool tracking method is
# preferred since the list takes less space than the
# set and it doesn't suffer from frequent reselections.
if isinstance(population, _Set):
population = tuple(population)
if not isinstance(population, _Sequence):
raise TypeError("Population must be a sequence or set. For dicts, use list(d).")
randbelow = self._randbelow
n = len(population)
if not 0 <= k <= n:
raise ValueError("Sample larger than population or is negative")
result = [None] * k
setsize = 21 # size of a small set minus size of an empty list
if k > 5:
setsize += 4 ** _ceil(_log(k * 3, 4)) # table size for big sets
if n <= setsize:
# An n-length list is smaller than a k-length set
pool = list(population)
for i in range(k): # invariant: non-selected at [0,n-i)
j = randbelow(n-i)
result[i] = pool[j]
pool[j] = pool[n-i-1] # move non-selected item into vacancy
else:
selected = set()
selected_add = selected.add
for i in range(k):
j = randbelow(n)
while j in selected:
j = randbelow(n)
selected_add(j)
result[i] = population[j]
return result
(I apologise is this question would be better placed in math.stackexchange. I couldn't think of any probability/statistics-y reasons for this particular ratio, and the comments sounded as though, it was maybe something to do with the amount of space that sets and lists use - but could't find any details anywhere).

This code is attempting to determine whether using a list or a set would take more space (instead of trying to estimate the time cost, for some reason).
It looks like 21 was the difference between the size of an empty list and a small set on the Python build this constant was determined on, expressed in multiples of the size of a pointer. I don't have a build of that version of Python, but testing on my 64-bit CPython 3.6.3 gives a difference of 20 pointer sizes:
>>> sys.getsizeof(set()) - sys.getsizeof([])
160
and comparing the 3.6.3 list and set struct definitions to the list and set definitions from the change that introduced this code, 21 seems plausible.
I said "the difference between the size of an empty list and a small set" because both now and at the time, small sets used a hash table contained inside the set struct itself instead of externally allocated:
setentry smalltable[PySet_MINSIZE];
The
if k > 5:
setsize += 4 ** _ceil(_log(k * 3, 4)) # table size for big sets
check adds the size of the external table allocated for sets larger than 5 elements, with size again expressed in number of pointers. This computation assumes the set never shrinks, since the sampling algorithm never removes elements. I am not currently sure whether this computation is exact.
Finally,
if n <= setsize:
compares the base overhead of a set plus any space used by an external hash table to the n pointers required by a list of the input elements. (It doesn't seem to account for the overallocation performed by list(population), so it may be underestimating the cost of the list.)

Is there any efficient way to increment the corresponding set positions of an integer in an integer array?

Any solution consuming less than O(Bit Length) time is welcome. I need to process around 100 million large integers.
answer = [0 for i in xrange(100)]
def pluginBits(val):
global answer
for j in xrange(len(answer)):
if val <= 0:
break
answer[j] += (val & 1)
val >>= 1

A speedier way to do this would be to use '{:b}'.format(someval) to convert from integer to a string of '1's and '0's. Python still needs to do similar work to perform this conversion, but doing it at the C layer in the interpreter internals involves significantly less overhead for larger values.
For conversion to actual list of integer 1s and 0s, you could do something like:
# Done once at top level to make translation table:
import string
bitstr_to_intval = string.maketrans(b'01', b'\x00\x01')
# Done for each value to convert:
bits = bytearray('{:b}'.format(origint).translate(bitstr_to_intval))
Since bytearray is a mutable sequence of values in range(256) that iterates the actual int values, you don't need to convert to list; it should be usable in 99% of the places the list would be used, using less memory and running faster.
This does generate the values in the reverse of the order your code produces (that is, bits[-1] here is the same as your answer[0], bits[-2] is your answer[1], etc.), and it's unpadded, but since you're summing bits, the padding isn't needed, and reversing the result is a trivial reversing slice (add [::-1] to the end). Summing the bits from each input can be made much faster by making answer a numpy array (that allows a bulk element-wise addition at the C layer), and putting it all together gets:
import string
bitstr_to_intval = string.maketrans(b'01', b'\x00\x01')
answer = numpy.zeros(100, numpy.uint64)
def pluginBits(val):
bits = bytearray('{:b}'.format(val).translate(bitstr_to_intval))[::-1]
answer[:len(bits)] += bits
In local tests, this definition of pluginBits takes a little under one-seventh the time to sum the bits at each position for 10,000 random input integers of 100 bits each, and gets the same results.

Speeding up Numpy Masking

I'm still an amature when it comes to thinking about how to optimize. I have this section of code that takes in a list of found peaks and finds where these peaks,+/- some value, are located in a multidimensional array. It then adds +1 to their indices of a zeros array. The code works well, but it takes a long time to execute. For instance it is taking close to 45min to run if ind has 270 values and refVals has a shape of (3050,3130,80). I understand that its a lot of data to churn through, but is there a more efficient way of going about this?
maskData = np.zeros_like(refVals).astype(np.int16)
for peak in ind:
tmpArr = np.ma.masked_outside(refVals,x[peak]-2,x[peak]+2).astype(np.int16)
maskData[tmpArr.mask == False ] += 1
tmpArr = None
maskData = np.sum(maskData,axis=2)

Approach #1 : Memory permitting, here's a vectorized approach using broadcasting -
# Craate +,-2 limits usind ind
r = x[ind[:,None]] + [-2,2]
# Use limits to get inside matches and sum over the iterative and last dim
mask = (refVals >= r[:,None,None,None,0]) & (refVals <= r[:,None,None,None,1])
out = mask.sum(axis=(0,3))
Approach #2 : If running out of memory with the previous one, we could use a loop and use NumPy boolean arrays and that could be more efficient than masked arrays. Also, we would perform one more level of sum-reduction, so that we would be dragging less data with us when moving across iterations. Thus, the alternative implementation would look something like this -
out = np.zeros(refVals.shape[:2]).astype(np.int16)
x_ind = x[ind]
for i in x_ind:
out += ((refVals >= i-2) & (refVals <= i+2)).sum(-1)
Approach #3 : Alternatively, we could replace that limit based comparison with np.isclose in approach #2. Thus, the only step inside the loop would become -
out += np.isclose(refVals,i,atol=2).sum(-1)

an efficient way to speed up some numpy operations

I am trying to find an efficient code instead of the following piece of code (that is only one part of my code), to increase the speed:
for pr in some_list:
Tp = T[partition[pr]].sum(0)
Tpx = np.dot(Tp, xhat)
hp = h[partition[[pr]].sum(0)
up = (uk[partition[pr][:]].sum(0))/len(partition[pr])
hpu = hpu + np.dot(hp.T, up)
Tpu = Tpu + np.dot(Tp.T, up)
I have at least two more similar blocks of code. As you can see, I used fancy indexing three times (really couldn't find another way). In my algorithm, I need this part to be done very quickly, but it's not happening now. I will really appreciate any suggestion.
Thank you all.
Best,

If your partitions are few and have many elements each, you should consider swapping around the indices of your objects. Summing an array of shape (30,1000) along its second dimension should be faster than summing an array of shape (1000,30) along its first dimension, since in the former case you are always summing contiguous blocks of memory (i.e. arr[k,:] for each k) for each remaining index. So if you put the summation index last (and get rid of some trailing singleton dimension while you're at it), you might get speed-up.
As hpaulj noted in a comment, it's not clear how your loop could be vectorized. However, since it's performance-critical, you could still try vectorizing some of the work.
I suggest that you store hp, up and Tp for each partition (following pre-allocation), then perform the scalar/matrix products in a single vectorized step. Also note that Tpx is unused in your example, so I omitted it here (whatever you're doing with it, you can do it similarly to the other examples):
part_len = len(some_list) # number of partitions, N
Tpshape = (part_len,) + T.shape[1:] # (N,30,100) if T was (1000,30,100)
hpshape = (part_len,) + h.shape[1:] # (N,30,1) if h was (1000,30,1)
upshape = (part_len,) + uk.shape[1:] # (N,30,1) if uk was (1000,30,1)
Tp = np.zeros(Tpshape)
hp = np.zeros(hpshape)
up = np.zeros(upshape)
for ipr,pr in enumerate(some_list):
Tp[ipr,:,:] = T[partition[pr]].sum(0)
hp[ipr,:,:] = h[partition[[pr]].sum(0)
up[ipr,:,:] = uk[partition[pr]].sum(0)/len(partition[pr])
# compute vectorized dot products:
#Tpx unclear in original, omitted
# sum over second index (dot), sum over first index (sum in loop)
hpu = np.einsum('abc,abd->cd',hp,up) # shape (1,1)
Tpu = np.einsum('abc,abd->cd',Tp,up) # shape (100,1)
Clearly the key player is numpy.einsum. And of course if hpu and Tpu had some prior values before the loop, you have to increment those values with the results from einsum above.
As for einsum, it performs summations and contractions of arrays of arbitrary dimensions. The pattern apearing above, 'abc,abd->cd', when applied to 3d arrays A and B, will return a 2d array C, with the following definition (math pseudocode):
C(c,d) = sum_a sum_b A(a,b,c)*B(a,b,d)
For a given fix a summation index, what's inside is
sum_b A(a,b,c)*B(a,b,d)
which, if the c and d indices are kept, will be euqivalent to np.dot(A(a,:,:).T,B(a,:,:)). Since we're summing these matrices with respect to a too, we're supposed to do exactly what your loopy version does, adding up each np.dot() contribution of the total sums.