cykhash

cython wrapper for khash-sets/maps, efficient implementation of isin and unique

About:

• Brings functionality of khash (https://github.com/attractivechaos/klib/blob/master/khash.h) to Python and Cython and can be used seamlessly in numpy or pandas.

• Numpy's world is lacking the concept of a (hash-)set. This shortcoming is fixed and efficient (memory- and speedwise compared to pandas') unique and isin are implemented.

• Python-set/dict have big memory-footprint. For some datatypes the overhead can be reduced by using khash by factor 4-8.

Installation:

The recommended way to install the library is via conda package manager using the conda-forge channel:

conda install -c conda-forge cykhash

You can also install the library using pip. To install the latest release:

pip install cykhash

To install the most recent version of the module:

pip install https://github.com/realead/cykhash/zipball/master

Dependencies:

To build the library from source, Cython>=0.28 is required as well as a c-build tool chain.

See (https://github.com/realead/cykhash/blob/master/doc/README4DEVELOPER.md) for dependencies needed for development.

Quick start

Hash set and isin

Creating a hashset and using it in isin:

# prepare data:
>>> import numpy as np 
>>> a = np.arange(42, dtype=np.int64)
>>> b = np.arange(84, dtype=np.int64)
>>> result = np.empty(b.size, dtype=np.bool)

# actually usage
>>> from cykhash import Int64Set_from_buffer, isin_int64

>>> lookup = Int64Set_from_buffer(a) # create a hashset
>>> isin_int64(b, lookup, result)    # running time O(b.size)
>>> isin_int64(b, lookup, result)    # lookup is reused and not recreated

unique

Finding unique in O(n) (compared to numpy's np.unique - O(n*logn)) and smaller memory-footprint than pandas' pd.unique:

# prepare input
>>> import numpy as np
>>> a = np.array([1,2,3,3,2,1], dtype=np.int64)

# actual usage:
>>> from cykhash import unique_int64
>>> unique_buffer = unique_int64(a) # unique element are exposed via buffer-protocol

# can be converted to a numpy-array without copying via
>>> unique_array = np.ctypeslib.as_array(unique_buffer)
>>> unique_array.shape
(3,)

Hash map

Maps and sets handle nan-correctly (try it out with Python's dict/set):

>>> from cykhash import Float64toInt64Map
>>> my_map = Float64toInt64Map() # values are 64bit integers
>>> my_map[float("nan")] = 1
>>> my_map[float("nan")]
1

Functionality overview

Hash sets

Int64Set, Int32Set, Float64Set, Float32Set ( and PyObjectSet) are implemented. They are more or less drop-in replacements for Python's set. Furthermore, given the Cython-interface, efficient extensions of functionality are easily done.

The biggest advantage of these sets is that they need about 4-8 times less memory than the usual Python-sets and are somewhat faster for integers or floats.

As PyObjectSet is somewhat slower than the usual set and needs about the same amount of memory, it should be used only if all nans should be treated as equivalent.

The most efficient way to create such sets is to use XXXXSet_from_buffer(...), e.g. Int64Set_from_buffer, if the data container at hand supports buffer protocol (e.g. numpy-arrays, array.array or ctypes-arrays). Or XXXXSet_from(...) for any iterator.

Hash maps

Int64toInt64Map, Int32toInt32Map, Float64toInt64Map, Float32toInt32Map ( and PyObjectMap) are implemented. They are more or less drop-in replacements for Python's dict (however, not every piece of dict's functionality makes sense, for example setdefault(x, default) without default-argument, because None cannot be inserted, also the khash-maps don't preserve the insertion order, so there is also no reversed). Furthermore, given the Cython-interface, efficient extensions of functionality are easily done.

Biggest advantage of these sets is that they need about 4-8 times less memory than the usual Python-dictionaries and are somewhat faster for integers or floats.

As PyObjectMap is somewhat slower than the usual dict and needs about the same amount of memory, it should be used only if all nans should be treated as equivalent.

isin

• implemented are isin_int64, isin_int32, isin_float64, isin_float32
• using hash set instead of arrays in isin function has the advantage, that the look-up data structure doesn't have to be reconstructed for every call, thus reducing the running time from O(n+m)to O(n), where n is the number of queries and m-number of elements in the look up array.
• Thus cykash's isin can be order of magnitude faster than the numpy's or pandas' versions.

all, none, any, and count_if

• siblings functions of isin_XXX are:
  • all_XXX/all_XXX_from_iterator which return True if all elements of the query array can be found in the set.
  • any_XXX/any_XXX_from_iterator which return True if at least one element of the query array can be found in the set.
  • none_XXX/none_XXX_from_iterator which return True if none of elements from the query array can be found in the set.
  • count_if_XXX/count_if_XXX_from_iterator which return the number of elements from the query array can be found in the set.
• all_XXX, any_XXX, none_XXX and count_if_XXX are faster than using isin_XXX and applying numpy's versions of these function on the resulting array.
• from_iterator version works with any iterable, but the version for buffers are more efficient.

unique

• implemented are unique_int64, unique_int32, unique_float64, unique_float32
• returns an object which implements the buffer protocol, so np.ctypeslib.as_array (recommended) or np.frombuffer (less safe, as memory can get reinterpreted) can be used to create numpy arrays.
• differently as pandas, the returned uniques aren't in the order of the appearance. If order of appearence is important use unique_stable_xxx-versions, which needs somewhat more memory.
• the signature is unique_xxx(buffer, size_hint=0.0) the initial memory-consumption of the hash-set will be len(buffer)*size_hint unless size_hint<=0.0, in this case it will be ensured, that no rehashing is needed even if all elements are unique in the buffer.

As pandas uses maps instead of sets internally for unique, it needs about 4 times more peak memory and is 1.6-3 times slower.

Floating-point numbers as keys

There is a problem with floating-point sets or maps, i.e. Float64Set, Float32Set, Float64toInt64Map and Float32toInt32Map: The standard definition of "equal" and hash-function based on the bit representation don't define a meaningful or desired behavior for the hash set:

• NAN != NAN and thus it is not equivalence relation
• -0.0 == 0.0 but hash(-0.0)!=hash(0.0), but x==y => hash(x)==hash(y) is neccessary for set to work properly.

This problem is resolved through following special case handling:

• hash(-0.0):=hash(0.0)
• hash(x):=hash(NAN) for any not a number x.
• x is equal y <=> x==y || (x!=x && y!=y)

A consequence of the above rule, that the equivalence classes of {0.0, -0.0} and e{x | x is not a number} have more than one element. In the set these classes are represented by the first seen element from the class.

The above holds also for PyObjectSet (this behavior is not the same as fro Python-set which shows a different behavior for nans).

Examples:

Hash sets

Python: Creates a set from a numpy-array and looks up whether an element is in the resulting set:

>>> import numpy as np
>>> from cykhash import Int64Set_from_buffer
>>> a =  np.arange(42, dtype=np.int64)
>>> my_set = Int64Set_from_buffer(a) # no reallocation will be needed
>>> 41 in my_set 
True
>>> 42 not in my_set
True

Python: Create a set from an iterable and looks up whether an element is in the resulting set:

>>> from cykhash import Int64Set_from
>>> my_set = Int64Set_from(range(42)) # no reallocation will be needed
>>> assert 41 in my_set and 42 not in my_set

Cython: Create a set and put some values into it:

from cykhash.khashsets cimport Int64Set
my_set = Int64Set(number_of_elements_hint=12)  # reserve place for at least 12 integers
cdef Py_ssize_t i
for i in range(12):
   my_set.add(i)
assert 11 in my_set and 12 not in my_set

Hash maps

Python: Creating int64->float64 map using Int64toFloat64Map_from_buffers:

>>> import numpy as np
>>> from cykhash import Int64toFloat64Map_from_buffers
>>> keys = np.array([1, 2, 3, 4], dtype=np.int64)
>>> vals = np.array([5, 6, 7, 8], dtype=np.float64)
>>> my_map = Int64toFloat64Map_from_buffers(keys, vals) # there will be no reallocation
>>> assert my_map[4] == 8.0

Python: Creating int64->int64 map from scratch:

>>> import numpy as np
>>> from cykhash import Int64toInt64Map

# my_map will not need reallocation for at least 12 elements
>>> my_map = Int64toInt64Map(number_of_elements_hint=12)
>>> for i in range(12):  my_map[i] = i+1
>>> assert my_map[5] == 6

isin

Python: Creating look-up data structure from a numpy-array, performing isin-query

>>> import numpy as np
>>> from cykhash import Int64Set_from_buffer, isin_int64
>>> a = np.arange(42, dtype=np.int64)
>>> lookup = Int64Set_from_buffer(a)

>>> b = np.arange(84, dtype=np.int64)
>>> result = np.empty(b.size, dtype=np.bool)

>>> isin_int64(b, lookup, result)    # running time O(b.size)
>>> assert np.sum(result.astype(np.int)) == 42

unique

Python: using unique_int64:

>>> import numpy as np
>>> from cykhash import unique_int64
>>> a = np.array([1,2,3,3,2,1], dtype=np.int64)
>>> u = np.ctypeslib.as_array(unique_int64(a)) # there will be no reallocation
>>> assert set(u) == {1,2,3}

Python: using unique_stable_int64:

>>> import numpy as np
>>> from cykhash import unique_stable_int64
>>> a = np.array([3,2,1,1,2,3], dtype=np.int64)
>>> u = np.ctypeslib.as_array(unique_stable_int64(a)) # there will be no reallocation
>>> assert list(u) == [3,2,1] 

API

See (https://github.com/realead/cykhash/blob/master/doc/README_API.md) for a more detailed API description.

Performance

See (https://github.com/realead/cykhash/blob/master/doc/README_PERFORMANCE.md) for results of performance tests.

Trivia

• This project was inspired by the following stackoverflow question: https://stackoverflow.com/questions/50779617/pandas-pd-series-isin-performance-with-set-versus-array.

• pandas also uses khash (and thus was a source of inspiration), but wraps only maps and doesn't wrap sets. Thus, pandas' unique needs more memory as it should. Those maps are also never exposed, so there is no way to reuse the look-up structure for multiple calls to isin.

• khash is a good choice, but there are other alternatives, e.g. https://github.com/sparsehash/sparsehash. See also https://stackoverflow.com/questions/48129713/fastest-way-to-find-all-unique-elements-in-an-array-with-cython/48142655#48142655 for a comparison for different unique implementations.

• A similar approach for sets/maps in pure Cython: https://github.com/realead/tighthash, which is quite slower than khash.

• There is no dependency on numpy: this library uses buffer protocol, thus it works for array.array, numpy.ndarray, ctypes-arrays and anything else. However, some interfaces are somewhat cumbersome (which type should be created as answer?) and for convenient usage it might be a good idea to wrap the functionality so objects of right types are created.

History:

Release 2.0.0 (09.11.2021):

• Implementation of any, all, none and count_if
• Hash-sets are now (almost) drop-in replacements of Python's sets
• Breaking change: iterator from maps doesn't no longer returns items but only keys. However there are following new methods keys(), values() and items()which return so called mapvies, which correspond more or less to dictviews (but for mapsview doesn't hold that "Dictionary order is guaranteed to be insertion order.").
• Hash-Maps are now (almost) drop-in replacements of Python's dicts. Differences: insertion order isn't preserved, thus there is also no reversed()-method, setdefault(key, default) isn't possible without default because None cannot be inserted in the map
• Better hash-functions for float64, float32, int64 and int32 (gh-issue #4).
• Breaking change: different names/signatures for maps
• supports tracemalloc for Py3.6+
• supports Python 3.10

Release 1.0.2 (30.05.2020):

• can be installed via conda-forge to all operating systems
• can be installed via pip in a clean environment (Cython>=0.28 is now fetched automatically)

Release 1.0.1 (27.05.2020):

• released on PyPi

Older:

• 0.4.0: uniques_stable, preparing for release
• 0.3.0: PyObjectSet, Maps for Int64/32 and also Float64/32, unique-versions
• 0.2.0: Int32Set, Float64Set, Float32Set
• 0.1.0: Int64Set