Read the docs |
|
Gitlab (main) |
|
Github (mirror) |
|
Pypi |
The main webpage for this project is: https://gitlab.kitware.com/computer-vision/kwarray
The kwarray module implements a small set of pure-python extensions to numpy and torch.
The kwarray module started as extensions for numpy + a simplified pandas-like DataFrame object with much faster item row and column access. But it also include an ArrayAPI, which is a wrapper that allows 100% interoperability between torch and numpy. It also contains a few algorithms like setcover and mincost_assignment.
The top-level API is:
from kwarray.arrayapi import ArrayAPI, dtype_info
from .algo_assignment import (maxvalue_assignment, mincost_assignment,
mindist_assignment,)
from .algo_setcover import (setcover,)
from .dataframe_light import (DataFrameArray, DataFrameLight, LocLight,)
from .fast_rand import (standard_normal, standard_normal32, standard_normal64,
uniform, uniform32,)
from .util_averages import (RunningStats, stats_dict,)
from .util_groups import (apply_grouping, group_consecutive,
group_consecutive_indices, group_indices,
group_items,)
from .util_misc import (FlatIndexer,)
from .util_numpy import (arglexmax, argmaxima, argminima, atleast_nd, boolmask,
isect_flags, iter_reduce_ufunc, normalize,)
from .util_random import (ensure_rng, random_combinations, random_product,
seed_global, shuffle,)
from .util_slices import (embed_slice, padded_slice,)
from .util_slider import (SlidingWindow, Stitcher,)
from .util_torch import (one_hot_embedding, one_hot_lookup,)The ArrayAPI
On of the most useful features in kwarray is the kwarray.ArrayAPI — a class that helps bridge between numpy and torch. This class consists of static methods that implement part of the numpy API and operate equivalently on either torch.Tensor or numpy.ndarray objects.
This works because every function call checks if the input is a torch tensor or a numpy array and then takes the appropriate action.
As you can imagine, it can be slow to validate your inputs on each function call. Therefore the recommended way of using the array API is via the kwarray.ArrayAPI.impl function. This function does the check once and then returns another object that directly performs the correct operations on subsequent data items of the same type.
The following example demonstrates both modes of usage.
import torch
import numpy as np
data1 = torch.rand(10, 10)
data2 = data1.numpy()
# Method 1: grab the appropriate sub-impl
impl1 = ArrayAPI.impl(data1)
impl2 = ArrayAPI.impl(data2)
result1 = impl1.sum(data1, axis=0)
result2 = impl2.sum(data2, axis=0)
assert np.all(impl1.numpy(result1) == impl2.numpy(result2))
# Method 2: choose the impl on the fly
result1 = ArrayAPI.sum(data1, axis=0)
result2 = ArrayAPI.sum(data2, axis=0)
assert np.all(ArrayAPI.numpy(result1) == ArrayAPI.numpy(result2))Other Notes:
The kwarray.ensure_rng function helps you properly maintain and control local seeded random number generation. This means that you wont clobber the random state of another library / get your random state clobbered.
DataFrameArray and DataFrameLight implement a subset of the pandas API. They are less powerful, but orders of magnitude faster. The main drawback is that you lose loc, but iloc is available.
uniform32 and standard_normal32 are faster 32-bit random number generators (compared to their 64-bit numpy counterparts).
mincost_assignment is the Munkres / Hungarian algorithm. It solves the assignment problem.
setcover - solves the minimum weighted set cover problem using either an approximate or an exact solution.
one_hot_embedding is a fast numpy / torch way to perform the often needed OHE deep-learning trick.
group_items is a fast way to group a numpy array by another numpy array. For fine grained control we also expose group_indices, which groups the indices of a numpy array, and apply_grouping, which partitions a numpy array by those indices.
boolmask effectively inverts np.where.
Usefulness:
This is the frequency that I’ve used various components of this library with in my projects:
Function name |
Usefulness |
|---|---|
239 |
|
148 |
|
50 |
|
43 |
|
40 |
|
34 |
|
28 |
|
21 |
|
17 |
|
14 |
|
14 |
|
12 |
|
12 |
|
10 |
|
8 |
|
7 |
|
7 |
|
7 |
|
6 |
|
6 |
|
6 |
|
6 |
|
5 |
|
4 |
|
4 |
|
3 |
|
3 |
|
2 |
|
2 |
|
1 |
|
1 |
|
1 |
|
0 |
|
0 |
|
0 |
|
0 |
|
0 |
|
0 |
|
0 |
|
0 |
|
0 |
|
0 |
|
0 |
|
0 |
|
0 |
|
0 |
|
0 |
|
0 |
Release files for kwarray 0.7.2
For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.
Source distribution (sdist)
| File | Size | Uploaded | |
|---|---|---|---|
| kwarray-0.7.2.tar.gz | 120.6 kB | Details |
Built distribution (wheel)
| File | Interpreter | ABI | Platform | Reset |
|---|---|---|---|---|
| kwarray-0.7.2-py3-none-any.whl | Python 3 | none | any | Details |
Total release size: 237.8 kB
Release files / kwarray-0.7.2.tar.gz
| Download URL | kwarray-0.7.2.tar.gz |
|---|---|
| Size | 120.6 kB |
| Tags | Source |
|
SHA-256 checksum How to use checksums |
370c6830141a35114c899fa6f5ef2a3272ea704f679a83177b74c6d9fe8bf162
|
|
BLAKE2b-256 checksum How to use checksums |
9dfca269954028b42841da26b248c10f8868c067ddc9fa75aaa7a5b61166ddb2
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
No |
| Uploaded via |
twine/7.0.0 CPython/3.14.7
|
Release files / kwarray-0.7.2-py3-none-any.whl
| Download URL | kwarray-0.7.2-py3-none-any.whl |
|---|---|
| Size | 117.2 kB |
| Tags | Python 3 |
|
SHA-256 checksum How to use checksums |
d34a8eb86c8c3a29ddc4cb626f4645eb928093e5c651baa2a65645db2a2be00b
|
|
BLAKE2b-256 checksum How to use checksums |
721d04fa3f608051f08b9edcaf15176659c1e7987d8f8cbe763fdd5dba73ec5a
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
No |
| Uploaded via |
twine/7.0.0 CPython/3.14.7
|
