Python

• Abstract Base Classes
• Coroutine examples
• Cheat Sheet: (pdf, html, source)

Hello Summary

CONSTANT=10

# recursive, exception, conditions, lambda
def fib(n):
  if n < 0:
    raise Exception('n={}, should be positive'.format(n))

  if n == 0:
    return 0

  check_n = lambda n: n == 1 or n == 2

  if check_n(n):
    return 1

  return fib(n-1) + fib(n-2)

# generator
def gen(n):
  yield 1
  for i in range(5, n):
    yield n

# class, attribute, method
class math:
  def __init__(self, c=CONSTANT):
    self.c = c

  @staticmethod
  def addition(a, b=2):
    return a+b

  def times_ten_and_hundred(n):
    return n*self.c, n*self.c**c

if __name__ == '__main__':
  # variables
  a, b = 1, 3

  # list/array(l), tuple(t), dict(d)
  l, t, d= [a, b], (a, b), {'a': a, 'b': b}

  # generator
  g = gen(6)
  # next(g) is equivalent to call g.__next__()
  print(next(g)) # 1 
  print(next(g)) # 5
  print(next(g)) # StopIteration

  try:
    fib(-1)
  except Exception as e:
    print("error: {}".format(e)) # error: n=-1, should be positive

  print(math.addition(a, b=b)) # 4
  print(math.addition(d['a'])) # 3

  # instance
  r1, r2 = math().times_ten_and_hundred(5)
  assert r1 == 50 # ok
  assert r2 == 500 # ok

Data Structure

List

>>> a = []
>>> a.append(1)
>>> a.append(2)
>>> a.append(3)
>>> a.pop() # stack <=> lifo
3
>>> a.pop(0) # queue <=> fifo (if implemented on linkedlist, complexity O(1) to add/del element at beginning)
1
>>> a
[2]

Note: a list might have different type of elements (int, dict, ...), while import array cannot.

Linked List

See linked list

Sorting algorithms

See sorting

Exceptions

class Custom(Exception):
  pass

try:
  print(does_not_exists)
except Exception as e:
  raise Custom("Your variable does not exists") from e 

Typing

from __future__ import annotations

def concat(a: list | str, b: list | str):
  return a + b 

>>> concat("hello", " world")
'hello world'
>>> concat([1, 2], [3, 4])
[1, 2, 3, 4]

See pep-0604

Concurrency Compute (thread/process)

See concurrent.futures

Example 1 (customed)

from concurrent import futures
import time

def task(seq):
    print('Start:', time.strftime('%X'))
    time.sleep(max(seq))
    print('End:', time.strftime('%X'))
    return list(reversed(seq))

def main():
    max_workers = 2
    timeout = 300

    jobs = []
    not_done = set()
    pendings = 0

    with futures.ProcessPoolExecutor(max_workers=max_workers) as executor: # you can also use futures.ThreadPoolExecutor
        for seq in [[3,2,1], [10,5,0]]:
            if pendings > max_workers:
                _u, not_done = futures.wait(not_done, timeout, return_when=futures.FIRST_COMPLETED)
                pendings = len(not_done)
                print(f"{pendings} jobs pending")
            job = executor.submit(task, seq)
            jobs.append(job)
            not_done.add(job)
            pendings += 1

    done, not_done = futures.wait(jobs, timeout, return_when=futures.ALL_COMPLETED)
    results = [d.result() for d in done]
    return results

>>> r = main()
Start: 15:08:39
End: 15:08:49
Start: 15:08:49
End: 15:08:52

>>> r
[[1, 2, 3], [0, 5, 10]]

Example 2 (shorten)

from concurrent import futures
import time

def task(seq):
  print('Start:', time.strftime('%X'))
  time.sleep(max(seq))
  print('End:', time.strftime('%X'))
  return list(reversed(seq))

def main():
  results = []
  with futures.ThreadPoolExecutor(max_workers=2) as executor:
    tasks = {executor.submit(task, seq=seq) for seq in [[3,2,1],[10,5,0]]}
    for completed_task in futures.as_completed(tasks, timeout=60):
      results.append(completed_task.result())
  return results

>>> r=main()
Start: 12:30:36
Start: 12:30:36
End: 12:30:39
End: 12:30:46

>>> r
[[1, 2, 3], [0, 5, 10]]

Concurrency IO (Async)

Useful when you write/read on disk/network devices (e.g. tcp connections)

Read also:

• official doc for asyncio
• gather(), wait(), and TaskGroup()
• How does asyncio work?
• How to speed up async requests in Python

Classic usage

import asyncio
import time

async def coroutine(seq):
	await asyncio.sleep(max(seq))
	return list(reversed(seq))

async def main():
	t = asyncio.create_task(coroutine([3, 2, 1]))
	t2 = asyncio.create_task(coroutine([10, 5, 0]))  # Python 3.7+
	print('Start:', time.strftime('%X'))
	a1 = await asyncio.gather(t, t2)
	print('End:', time.strftime('%X'))  # Should be 10 seconds
	print(f'Both tasks done: {all((t.done(), t2.done()))}')
	print('Start:', time.strftime('%X'))
	a2 = await asyncio.gather(coroutine([1,2,3]), coroutine([1,2,2]))
	print('End:', time.strftime('%X'))  # Should be 3 seconds
	return a1, a2

Note: don't use time.sleep()

>>> asyncio.run(main())
Start: 15:08:39
End: 15:08:49
Both tasks done: True
Start: 15:08:49
End: 15:08:52

>>> a1
[[1, 2, 3], [0, 5, 10]]

source

Run lot of asynchronous functions with eventloop

import asyncio
import time

loop = asyncio.new_event_loop() # or loop = asyncio.get_running_loop()

# by default it will create threads, but you can create processes by changing:
# import concurrent.futures
# loop.set_default_executor(concurrent.futures.ProcessPoolExecutor())
# you can also use: 
# with concurrent.futures.ProcessPoolExecutor() as pool_executor:
#   loop.run_in_executor(pool_executor, func, params)

async def send_request(id):
  while not loop.is_closed():
    print(time.strftime('%X'), ' Hello', id)
    await asyncio.sleep(id)  # non-blocking, so other threads can take the lead, see https://stackoverflow.com/questions/56729764/asyncio-sleep-vs-time-sleep
    print(time.strftime('%X'), ' Done', id)
    # it is possible to remove the while not loop... by doing so, but it is more costly (destroy/recreate thread), moreover it might be dangerous that a dying thread create a new one ?:
    # loop.create_task(send_request(id))  

for i in range(1,32):
    loop.create_task(send_request(i))

try:
  loop.run_forever()
except KeyboardInterrupt:
  [task.cancel() for task in asyncio.all_tasks(loop)]
finally:
  loop.close()

Result:

11:40:26  Hello 1
11:40:26  Hello 2
11:40:26  Hello 3
11:40:26  Hello 4
11:40:26  Hello 5
11:40:26  Hello 6
...
11:40:26  Hello 30
11:40:26  Hello 31
11:40:27  Done 1
11:40:27  Hello 1
11:40:28  Done 2
11:40:28  Done 1
11:40:28  Hello 2
11:40:28  Hello 1
11:40:29  Done 3
11:40:29  Hello 3
...

Run synchronous functions in different threads

In case you can't touch a function (signature, content), but you want to run it in a thread or process to continue other tasks in parallel. Careful, if something is blocking in the function, you can't touch it, it will monopolize the thread. Might be problematic if you have more calls than available threads.

import asyncio
import time
import concurrent

def synchronous(i):
  print(time.strftime('%X'), ' Hello', i)
  time.sleep(i) # simulate work that cant be predictable and/or changed
  print(time.strftime('%X'), ' Done', i)

async def sync_wrapper(i):
  loop = asyncio.get_running_loop()

  while True:  # optional: in this example we consider each sync fonction is recalled
    # NOTE it seems to work as well if executor=None, however
    # i'm not sure if it really spawn a thread with executor=None
    with concurrent.futures.ThreadPoolExecutor() as executor:
      await loop.run_in_executor(executor, synchronous, i)

loop = asyncio.new_event_loop() # or loop = asyncio.get_running_loop()

for i in range(1,4):
  loop.create_task(sync_wrapper(i))

try:
  loop.run_forever()
finally:
  loop.close()

Result:

19:01:23  Hello 1
19:01:23  Hello 2
19:01:23  Hello 3
19:01:24  Done 1
19:01:24  Hello 1
19:01:25  Done 2
19:01:25  Hello 2
19:01:25  Done 1
19:01:25  Hello 1
19:01:26  Done 3
19:01:26  Hello 3
19:01:26  Done 1
19:01:26  Hello 1
19:01:27  Done 2
19:01:27  Hello 2
19:01:27  Done 1

Alternative with python 3.9

asyncio.to_thread(blocking)

source

Coupled with generators

# generator async
async def coroutine_generator(n):
  for i in range(n):
    yield i*i

# usage of generator async
async def print_gen(n):
  # generator coroutine will use `__aiter__() and __anext__()`
  async for i in coroutine_generator(n):
    print(i)

await print_gen(5) # 0 1 4 9 16
# alternatively you can do: asyncio.run(print_gen(5))

See:

• example 1
• example 2

Libraries

Machine Learning

• tslearn: time series analysis
• scikit-learn: classical machine learning
• sktime: time series model (classification, clustering) compatible with scikit-learn interfaces
• kats: a kit to analyze time series data, key statistics and characteristics, detecting change points and anomalies, to forecasting future trends.
• tsfresh: time series feature extraction
• DESlib: easy-to-use ensemble learning library focused on the implementation of the state-of-the-art techniques for dynamic classifier and ensemble selection
• PyOD: outlier detection
• river: online machine learning with streaming data
• unionml: deploy machine learning microservices
• Flyte: workflow automation platform for complex, mission-critical data, and ML processes at scale
• ALEPython: Accumulated Local Effects (or ALE) interepretability

Statistics

• Pingouin: ANOVA, post-hocs (parametric and non-parametric), multivariate
• statsmodels (see also scikit posthocs)

Misc

• modin: scale pandas (using dask or ray)
• polars: alternative to pandas written in rust
• streamlit: create web apps for machine learning projects
• luma.lcd: display drivers for HD44780, PCD8544, ST7735, ST7789, ST7567, HT1621, UC1701X, ILI9341
• ruptures: time series change point detection in Python
• SymPy: algebra system
• Adafruit CircuitPython RFM69: CircuitPython RFM69 packet radio module. This supports basic RadioHead-compatible sending and receiving of packets with RFM69 series radios (433/915Mhz).
• aiohttp: Asynchronous HTTP client/server framework for asyncio (core dev of cython)
• aiocache: Asyncio cache manager for redis, memcached and memory

Numpy

>>> a
array([0.34399327, 0.51971385, 0.42075315, 0.65919112])
>>> b
array([0.42685801, 0.52210862, 0.52210862, 0.52210862])
>>> a*b
array([0.14683628, 0.27134708, 0.21967885, 0.34416937])
>>> np.sum(a*b)
0.9820315790772367
>>> a.dot(np.transpose(b))
0.9820315790772366
>>> a.dot(b.T)
0.9820315790772366
>>> a.dot(b)
0.9820315790772366

Pandas

Combine multiple rows of lists into one big list

lst = df['col_of_lists'].explode()

source

Make Pandas DataFrame apply() use all cores ?

import multiprocessing as mp

with mp.Pool(mp.cpu_count()) as pool:
    df['newcol'] = pool.map(f, df['col'])

source

FAQ

• Are Python coroutines stackless or stackful?
• Achieving true parallelism for pure Python functions in CPython
• Is there any way to kill a Thread?
• Writing Python like it's Rust
• Making Python 100x faster with less than 100 lines of Rust