A Python utility belt containing simple tools, a stdlib like feel, and extra batteries. Hashing, Caching, Timing, Progress, and more made easy!

Ubelt is a small library of robust, tested, documented, and simple functions that extend the Python standard library. It has a flat API that all behaves similarly on Windows, Mac, and Linux (up to some small unavoidable differences). Almost every function in ubelt was written with a doctest. This provides helpful documentation and example usage as well as helping achieve 100% test coverage (with minor exceptions for Python2, Windows, etc...).

  • Goal: provide simple functions that accomplish common tasks not yet addressed by the python standard library.
  • Constraints: Must be low-impact pure python; it should be easy to install and use.
  • Method: All functions are written with docstrings and doctests to ensure that a baseline level of documentation and testing always exists (even if functions are copy/pasted into other libraries)
  • Motto: Good utilities lift all codes.

Read the docs here: http://ubelt.readthedocs.io/en/latest/

These are some of the tasks that ubelt's API enables:

  • hash common data structures like list, dict, int, str, etc. (hash_data)
  • hash files (hash_file)
  • cache a block of code (Cacher, CacheStamp)
  • time a block of code (Timerit, Timer)
  • show loop progress (ProgIter)
  • download a file with optional caching and hash verification (download, grabdata)
  • run shell commands (cmd)
  • find a file or directory in candidate locations (find_path, find_exe)
  • string-format nested data structures (repr2)
  • color text with ANSI tags (color_text)
  • horizontally concatenate multiline strings (hzcat)
  • make a directory if it doesn't exist (ensuredir)
  • delete a file, link, or entire directory (delete)
  • create cross platform symlinks (symlink)
  • expand environment variables and tildes in path strings (expandpath)
  • import a module using the path to that module (import_module_from_path)
  • check if a particular flag or value is on the command line (argflag, argval)
  • get paths to cross platform data/cache/config directories (ensure_app_cache_dir, ...)
  • memoize functions (memoize, memoize_method, memoize_property)
  • build ordered sets (oset)
  • short defaultdict and OrderedDict aliases (ddict and odict)
  • map a function over the keys or values of a dictionary (map_keys, map_vals)
  • perform set operations on dictionaries (dict_union, dict_isect, dict_diff, dict_subset, ...)
  • perform dictionary operations like histogram, inversion, and sorting (dict_hist, invert_dict, sorted_keys, sorted_vals)
  • argmax/min/sort on lists and dictionaries (argmin, argsort,)
  • find duplicates in a list (find_duplicates)
  • group a sequence of items by some criterion (group_items)

Ubelt is small. Its top-level API is defined using roughly 40 lines:

from ubelt.util_arg import (argflag, argval,)
from ubelt.util_cache import (CacheStamp, Cacher,)
from ubelt.util_colors import (NO_COLOR, color_text, highlight_code,)
from ubelt.util_const import (NoParam,)
from ubelt.util_cmd import (cmd,)
from ubelt.util_dict import (AutoDict, AutoOrderedDict, ddict, dict_diff,
                             dict_hist, dict_isect, dict_subset, dict_union,
                             dzip, find_duplicates, group_items, invert_dict,
                             map_keys, map_vals, named_product, odict,
                             sorted_keys, sorted_vals, varied_values,)
from ubelt.util_download import (download, grabdata,)
from ubelt.util_download_manager import (DownloadManager,)
from ubelt.util_func import (compatible, identity, inject_method,)
from ubelt.util_format import (FormatterExtensions, repr2,)
from ubelt.util_futures import (Executor, JobPool,)
from ubelt.util_io import (delete, readfrom, touch, writeto,)
from ubelt.util_links import (symlink,)
from ubelt.util_list import (allsame, argmax, argmin, argsort, argunique,
                             boolmask, chunks, compress, flatten, iter_window,
                             iterable, peek, take, unique, unique_flags,)
from ubelt.util_hash import (hash_data, hash_file,)
from ubelt.util_import import (import_module_from_name,
                               import_module_from_path, modname_to_modpath,
                               modpath_to_modname, split_modpath,)
from ubelt.util_indexable import (IndexableWalker, indexable_allclose,)
from ubelt.util_memoize import (memoize, memoize_method, memoize_property,)
from ubelt.util_mixins import (NiceRepr,)
from ubelt.util_path import (TempDir, augpath, ensuredir, expandpath,
                             shrinkuser, userhome,)
from ubelt.util_platform import (DARWIN, LINUX, POSIX, WIN32,
                                 ensure_app_cache_dir, ensure_app_config_dir,
                                 ensure_app_data_dir, find_exe, find_path,
                                 get_app_cache_dir, get_app_config_dir,
                                 get_app_data_dir, platform_cache_dir,
                                 platform_config_dir, platform_data_dir,)
from ubelt.util_str import (codeblock, ensure_unicode, hzcat, indent,
from ubelt.util_stream import (CaptureStdout, CaptureStream, TeeStringIO,)
from ubelt.util_time import (timestamp,)
from ubelt.orderedset import (OrderedSet, oset,)
from ubelt.progiter import (ProgIter,)
from ubelt.timerit import (Timer, Timerit,)


Ubelt is distributed on pypi as a universal wheel and can be pip installed on Python 2.7, Python 3.5+. Installations are tested on CPython and PyPy implementations.

pip install ubelt

Note that our distributions on pypi are signed with GPG. The signing public key is D297D757; this should agree with the value in dev/public_gpg_key.

It is also possible to simply install it from source.

pip install git+https://github.com/Erotemic/ubelt.git


Ubelt is a migration of the most useful parts of utool(https://github.com/Erotemic/utool) into a standalone module with minimal dependencies.

The utool library contains a number of useful utility functions, but it also contained non-useful functions, as well as the kitchen sink. A number of the functions were too specific or not well documented. The ubelt is a port of the simplest and most useful parts of utool.

Note that there are other cool things in utool that are not in ubelt. Notably, the doctest harness ultimately became xdoctest. Code introspection and dynamic analysis tools were ported to xinspect. The more IPython-y tools were ported to xdev. Parts of it made their way into scriptconfig. The init-file generation was moved to mkinit. Some vim and system-y things can be found in vimtk.

Function Usefulness

When I had to hand pick a set of functions that I thought were the most useful I chose these and provided some comment on why:

import ubelt as ub

ub.ensuredir  # os.makedirs(exist_ok=True) is 3 only and too verbose
ub.Timerit  # powerful multiline alternative to timeit
ub.Cacher  # configuration based on-disk cachine
ub.cmd  # combines the best of subprocess.Popen and os.system
ub.hash_data  # hash mutable python containers, useful with Cacher to config strings
ub.repr2  # readable representations of nested data structures
ub.download  # why is this not a one liner --- also see grabdata for the same thing, but builtin caching.
ub.AutoDict  # one of the most useful tools in Perl, recursive default dicts of dicts
ub.JobPool   # easy multi-threading / multi-procesing / or single-threaded processing
ub.modname_to_modpath  # (works via static analysis)
ub.modpath_to_modname  # (works via static analysis)
ub.import_module_from_path  # (Unlike importlib, this does not break pytest)
ub.import_module_from_name  # (Unlike importlib, this does not break pytest)

But a better way might to objectively measure the frequency of usage and built a histogram of usefulness. I generated this histogram using python dev/gen_api_for_docs.py, which roughly counts the number of times I've used a ubelt function in another project. Note: this measure is biased towards older functions.

Function name Usefulness
ubelt.repr2 2001
ubelt.expandpath 728
ubelt.ProgIter 694
ubelt.ensuredir 539
ubelt.take 383
ubelt.odict 338
ubelt.map_vals 305
ubelt.dzip 272
ubelt.augpath 250
ubelt.ddict 227
ubelt.NiceRepr 220
ubelt.cmd 209
ubelt.flatten 206
ubelt.peek 189
ubelt.NoParam 184
ubelt.argval 179
ubelt.group_items 173
ubelt.argflag 173
ubelt.codeblock 171
ubelt.Timerit 164
ubelt.dict_hist 161
ubelt.iterable 144
ubelt.hash_data 124
ubelt.grabdata 106
ubelt.oset 103
ubelt.paragraph 100
ubelt.delete 97
ubelt.allsame 90
ubelt.compress 87
ubelt.color_text 84
ubelt.dict_subset 76
ubelt.dict_isect 75
ubelt.Cacher 70
ubelt.dict_diff 65
ubelt.memoize 54
ubelt.indent 54
ubelt.argsort 53
ubelt.Timer 52
ubelt.dict_union 50
ubelt.invert_dict 49
ubelt.identity 49
ubelt.find_duplicates 43
ubelt.map_keys 43
ubelt.timestamp 42
ubelt.unique 40
ubelt.chunks 38
ubelt.hzcat 37
ubelt.argmax 35
ubelt.import_module_from_path 35
ubelt.memoize_property 34
ubelt.iter_window 33
ubelt.readfrom 31
ubelt.sorted_vals 30
ubelt.hash_file 30
ubelt.writeto 30
ubelt.memoize_method 29
ubelt.symlink 28
ubelt.ensure_unicode 25
ubelt.CacheStamp 23
ubelt.touch 22
ubelt.modname_to_modpath 21
ubelt.find_exe 17
ubelt.import_module_from_name 17
ubelt.highlight_code 17
ubelt.AutoDict 13
ubelt.inject_method 11
ubelt.argmin 8
ubelt.shrinkuser 8
ubelt.split_modpath 6
ubelt.find_path 5
ubelt.download 5
ubelt.sorted_keys 5
ubelt.CaptureStdout 4
ubelt.modpath_to_modname 4
ubelt.userhome 3
ubelt.argunique 2
ubelt.AutoOrderedDict 1
ubelt.unique_flags 1
ubelt.varied_values 0
ubelt.platform_data_dir 0
ubelt.platform_config_dir 0
ubelt.named_product 0
ubelt.indexable_allclose 0
ubelt.get_app_data_dir 0
ubelt.get_app_config_dir 0
ubelt.ensure_app_data_dir 0
ubelt.ensure_app_config_dir 0
ubelt.compatible 0
ubelt.boolmask 0
ubelt.TempDir 0
ubelt.TeeStringIO 0
ubelt.POSIX 0
ubelt.OrderedSet 0
ubelt.NO_COLOR 0
ubelt.JobPool 0
ubelt.IndexableWalker 0
ubelt.FormatterExtensions 0
ubelt.Executor 0
ubelt.DownloadManager 0
ubelt.CaptureStream 0


Be sure to checkout the new Jupyter notebook: https://github.com/Erotemic/ubelt/blob/master/docs/notebooks/Ubelt%20Demo.ipynb

Here are some examples of some features inside ubelt


Quickly time a single line.

>>> import math
>>> import ubelt as ub
>>> timer = ub.Timer('Timer demo!', verbose=1)
>>> with timer:
>>>     math.factorial(100000)
tic('Timer demo!')
...toc('Timer demo!')=0.1453s

Robust Timing and Benchmarking

Easily do robust timings on existing blocks of code by simply indenting them. There is no need to refactor into a string representation or convert to a single line. With ub.Timerit there is no need to resort to the timeit module!

The quick and dirty way just requires one indent.

Note: Timerit is also defined in a standalone module: pip install timerit)

>>> import math
>>> import ubelt as ub
>>> for _ in ub.Timerit(num=200, verbose=3):
>>>     math.factorial(10000)
Timing for 200 loops
Timed for: 200 loops, best of 3
    time per loop: best=2.055 ms, mean=2.145 ± 0.083 ms

Use the loop variable as a context manager for more accurate timings or to incorporate an setup phase that is not timed. You can also access properties of the ub.Timerit class to programmatically use results.

>>> import math
>>> import ubelt as ub
>>> t1 = ub.Timerit(num=200, verbose=2)
>>> for timer in t1:
>>>     setup_vars = 10000
>>>     with timer:
>>>         math.factorial(setup_vars)
>>> print('t1.total_time = %r' % (t1.total_time,))
Timing for 200 loops
Timed for: 200 loops, best of 3
    time per loop: best=2.064 ms, mean=2.115 ± 0.05 ms
t1.total_time = 0.4427177629695507

Loop Progress

ProgIter is a no-threads attached Progress meter that writes to stdout. It is a mostly drop-in alternative to tqdm. The advantage of ``ProgIter`` is that it does not use any python threading, and therefore can be safer with code that makes heavy use of multiprocessing.

Note: ProgIter is also defined in a standalone module: pip install progiter)

>>> import ubelt as ub
>>> def is_prime(n):
...     return n >= 2 and not any(n % i == 0 for i in range(2, n))
>>> for n in ub.ProgIter(range(1000), verbose=2):
>>>     # do some work
>>>     is_prime(n)
    0/1000... rate=0.00 Hz, eta=?, total=0:00:00, wall=14:05 EST
    1/1000... rate=82241.25 Hz, eta=0:00:00, total=0:00:00, wall=14:05 EST
  257/1000... rate=177204.69 Hz, eta=0:00:00, total=0:00:00, wall=14:05 EST
  642/1000... rate=94099.22 Hz, eta=0:00:00, total=0:00:00, wall=14:05 EST
 1000/1000... rate=71886.74 Hz, eta=0:00:00, total=0:00:00, wall=14:05 EST


Cache intermediate results in a script with minimal boilerplate. It looks like 4 lines of boilerplate is the best you can do with Python 3.8 syntax. See <https://raw.githubusercontent.com/Erotemic/ubelt/master/ubelt/util_cache.py>`__ for details.

>>> import ubelt as ub
>>> cfgstr = 'repr-of-params-that-uniquely-determine-the-process'
>>> cacher = ub.Cacher('test_process', cfgstr)
>>> data = cacher.tryload()
>>> if data is None:
>>>     myvar1 = 'result of expensive process'
>>>     myvar2 = 'another result'
>>>     data = myvar1, myvar2
>>>     cacher.save(data)
>>> myvar1, myvar2 = data


The ub.hash_data constructs a hash corresponding to a (mostly) arbitrary ordered python object. A common use case for this function is to construct the cfgstr mentioned in the example for ub.Cacher. Instead of returning a hex, string, ub.hash_data encodes the hash digest using the 26 lowercase letters in the roman alphabet. This makes the result easy to use as a filename suffix.

>>> import ubelt as ub
>>> data = [('arg1', 5), ('lr', .01), ('augmenters', ['flip', 'translate'])]
>>> ub.hash_data(data)[0:8]

There exists an undocumented plugin architecture to extend this function to arbitrary types. See ubelt/util_hash.py for details.

Command Line Interaction

The builtin Python subprocess.Popen module is great, but it can be a bit clunky at times. The os.system command is easy to use, but it doesn't have much flexibility. The ub.cmd function aims to fix this. It is as simple to run as os.system, but it returns a dictionary containing the return code, standard out, standard error, and the Popen object used under the hood.

>>> import ubelt as ub
>>> info = ub.cmd('gcc --version')
>>> print(ub.repr2(info))
    'command': 'gcc --version',
    'err': '',
    'out': 'gcc (Ubuntu 5.4.0-6ubuntu1~16.04.9) 5.4.0 20160609\nCopyright (C) 2015 Free Software Foundation, Inc.\nThis is free software; see the source for copying conditions.  There is NO\nwarranty; not even for MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.\n\n',
    'proc': <subprocess.Popen object at 0x7ff98b310390>,
    'ret': 0,

Also note the use of ub.repr2 to nicely format the output dictionary.

Additionally, if you specify verbose=True, ub.cmd will simultaneously capture the standard output and display it in real time.

>>> import ubelt as ub
>>> info = ub.cmd('gcc --version', verbose=True)
gcc (Ubuntu 5.4.0-6ubuntu1~16.04.9) 5.4.0 20160609
Copyright (C) 2015 Free Software Foundation, Inc.
This is free software; see the source for copying conditions.  There is NO

A common use case for ub.cmd is parsing version numbers of programs

>>> import ubelt as ub
>>> cmake_version = ub.cmd('cmake --version')['out'].splitlines()[0].split()[-1]
>>> print('cmake_version = {!r}'.format(cmake_version))
cmake_version = 3.11.0-rc2

This allows you to easily run a command line executable as part of a python process, see what it is doing, and then do something based on its output, just as you would if you were interacting with the command line itself.

Lastly, ub.cmd removes the need to think about if you need to pass a list of args, or a string. Both will work. This utility has been tested on both Windows and Linux.

Cross-Platform Resource and Cache Directories

If you have an application which writes configuration or cache files, the standard place to dump those files differs depending if you are on Windows, Linux, or Mac. Ubelt offers a unified functions for determining what these paths are.

The ub.ensure_app_cache_dir and ub.ensure_app_resource_dir functions find the correct platform-specific location for these files and ensures that the directories exist. (Note: replacing "ensure" with "get" will simply return the path, but not ensure that it exists)

The resource root directory is ~/AppData/Roaming on Windows, ~/.config on Linux and ~/Library/Application Support on Mac. The cache root directory is ~/AppData/Local on Windows, ~/.config on Linux and ~/Library/Caches on Mac.

Example usage on Linux might look like this:

>>> import ubelt as ub
>>> print(ub.compressuser(ub.ensure_app_cache_dir('my_app')))
>>> print(ub.compressuser(ub.ensure_app_resource_dir('my_app')))

The ub.symlink function will create a symlink similar to os.symlink. The main differences are that 1) it will not error if the symlink exists and already points to the correct location. 2) it works* on Windows (*hard links and junctions are used if real symlinks are not available)

>>> import ubelt as ub
>>> dpath = ub.ensure_app_cache_dir('ubelt', 'demo_symlink')
>>> real_path = join(dpath, 'real_file.txt')
>>> link_path = join(dpath, 'link_file.txt')
>>> ub.writeto(real_path, 'foo')
>>> ub.symlink(real_path, link_path)

Downloading Files

The function ub.download provides a simple interface to download a URL and save its data to a file.

>>> import ubelt as ub
>>> url = ''
>>> fpath = ub.download(url, verbose=0)
>>> print(ub.compressuser(fpath))

The function ub.grabdata works similarly to ub.download, but whereas ub.download will always re-download the file, ub.grabdata will check if the file exists and only re-download it if it needs to.

>>> import ubelt as ub
>>> url = ''
>>> fpath = ub.grabdata(url, verbose=0, hash_prefix='944389a39')
>>> print(ub.compressuser(fpath))

New in version 0.4.0: both functions now accepts the hash_prefix keyword argument, which if specified will check that the hash of the file matches the provided value. The hasher keyword argument can be used to change which hashing algorithm is used (it defaults to "sha512").


Group items in a sequence into a dictionary by a second id list

>>> import ubelt as ub
>>> item_list    = ['ham',     'jam',   'spam',     'eggs',    'cheese', 'bannana']
>>> groupid_list = ['protein', 'fruit', 'protein',  'protein', 'dairy',  'fruit']
>>> ub.group_items(item_list, groupid_list)
{'dairy': ['cheese'], 'fruit': ['jam', 'bannana'], 'protein': ['ham', 'spam', 'eggs']}

Dictionary Histogram

Find the frequency of items in a sequence

>>> import ubelt as ub
>>> item_list = [1, 2, 39, 900, 1232, 900, 1232, 2, 2, 2, 900]
>>> ub.dict_hist(item_list)
{1232: 2, 1: 1, 2: 4, 900: 3, 39: 1}

Find Duplicates

Find all duplicate items in a list. More specifically, ub.find_duplicates searches for items that appear more than k times, and returns a mapping from each duplicate item to the positions it appeared in.

>>> import ubelt as ub
>>> items = [0, 0, 1, 2, 3, 3, 0, 12, 2, 9]
>>> ub.find_duplicates(items, k=2)
{0: [0, 1, 6], 2: [3, 8], 3: [4, 5]}

Dictionary Manipulation

Take a subset of a dictionary.

>>> import ubelt as ub
>>> dict_ = {'K': 3, 'dcvs_clip_max': 0.2, 'p': 0.1}
>>> subdict_ = ub.dict_subset(dict_, ['K', 'dcvs_clip_max'])
>>> print(subdict_)
{'K': 3, 'dcvs_clip_max': 0.2}

Take only the values, optionally specify a default value.

>>> import ubelt as ub
>>> dict_ = {1: 'a', 2: 'b', 3: 'c'}
>>> print(list(ub.take(dict_, [1, 2, 3, 4, 5], default=None)))
['a', 'b', 'c', None, None]

Apply a function to each value in the dictionary (see also ub.map_keys).

>>> import ubelt as ub
>>> dict_ = {'a': [1, 2, 3], 'b': []}
>>> newdict = ub.map_vals(len, dict_)
>>> print(newdict)
{'a': 3, 'b': 0}

Invert the mapping defined by a dictionary. By default invert_dict assumes that all dictionary values are distinct (i.e. the mapping is one-to-one / injective).

>>> import ubelt as ub
>>> mapping = {0: 'a', 1: 'b', 2: 'c', 3: 'd'}
>>> ub.invert_dict(mapping)
{'a': 0, 'b': 1, 'c': 2, 'd': 3}

However, by specifying unique_vals=False the inverted dictionary builds a set of keys that were associated with each value.

>>> import ubelt as ub
>>> mapping = {'a': 0, 'A': 0, 'b': 1, 'c': 2, 'C': 2, 'd': 3}
>>> ub.invert_dict(mapping, unique_vals=False)
{0: {'A', 'a'}, 1: {'b'}, 2: {'C', 'c'}, 3: {'d'}}

AutoDict - Autovivification

While the collections.defaultdict is nice, it is sometimes more convenient to have an infinitely nested dictionary of dictionaries.

>>> import ubelt as ub
>>> auto = ub.AutoDict()
>>> print('auto = {!r}'.format(auto))
auto = {}
>>> auto[0][10][100] = None
>>> print('auto = {!r}'.format(auto))
auto = {0: {10: {100: None}}}
>>> auto[0][1] = 'hello'
>>> print('auto = {!r}'.format(auto))
auto = {0: {1: 'hello', 10: {100: None}}}

String-based imports

Ubelt contains functions to import modules dynamically without using the python import statement. While importlib exists, the ubelt implementation is simpler to user and does not have the disadvantage of breaking pytest.

Note ubelt simply provides an interface to this functionality, the core implementation is in xdoctest (over as of version 0.7.0, the code is statically copied into an autogenerated file such that ubelt does not actually depend on xdoctest during runtime).

>>> import ubelt as ub
>>> module = ub.import_module_from_path(ub.expandpath('~/code/ubelt/ubelt'))
>>> print('module = {!r}'.format(module))
module = <module 'ubelt' from '/home/joncrall/code/ubelt/ubelt/__init__.py'>
>>> module = ub.import_module_from_name('ubelt')
>>> print('module = {!r}'.format(module))
module = <module 'ubelt' from '/home/joncrall/code/ubelt/ubelt/__init__.py'>

Related to this functionality are the functions ub.modpath_to_modname and ub.modname_to_modpath, which statically transform (i.e. no code in the target modules is imported or executed) between module names (e.g. ubelt.util_import) and module paths (e.g. ~/.local/conda/envs/cenv3/lib/python3.5/site-packages/ubelt/util_import.py).

>>> import ubelt as ub
>>> modpath = ub.util_import.__file__
>>> print(ub.modpath_to_modname(modpath))
>>> modname = ub.util_import.__name__
>>> assert ub.modname_to_modpath(modname) == modpath

Horizontal String Concatenation

Sometimes its just prettier to horizontally concatenate two blocks of text.

>>> import ubelt as ub
>>> B = ub.repr2([[1, 2], [3, 4]], nl=1, cbr=True, trailsep=False)
>>> C = ub.repr2([[5, 6], [7, 8]], nl=1, cbr=True, trailsep=False)
>>> print(ub.hzcat(['A = ', B, ' * ', C]))
A = [[1, 2], * [[5, 6],
     [3, 4]]    [7, 8]]

External tools

Some of the tools in ubelt also exist as standalone modules. I haven't decided if its best to statically copy them into ubelt or require on pypi to satisfy the dependency. There are some tools that are not used by default unless you explicitly allow for them.

Code that is currently statically included:

Code that is currently linked via pypi:

Code that is completely optional, and only used in specific cases:

  • Numpy - ub.repr2 will format a numpy array nicely by default
  • xxhash - this can be specified as a hasher to ub.hash_data
  • Pygments - used by the util_color module.

Similar Tools

UBelt is one of many Python utility libraries. Other similar libraries are listed here:


Ubelt will support Python 2.7 and 3.5 until the 1.0 release.

PRs are welcome.

Also check out my other projects (many of which are powered by ubelt):