3

I implemented a breadth-first search for a board game. Here I use a dict to reflect duplicate board configurations on each level. Currently, the overall search takes nearly all the RAM I have (16GB) for one start configuration. I plan to integrate intersection checks for different start configurations. So I need read access to the configurations I found, and the dict for one level doesn't change if the level is finished.

That's why I plan to convert the dict into a flat data structure (list or tuple) with keys at [2n] position and values at [2n+1] position before I evaluate the next level.

The problem is to find a fast conversion from {1: 2, 3: 4} to [1, 2, 3, 4] for dictwith more than 10**8 items.

I found sum(dict.items(), ()) from a comment by Natim to another question, which worked, but which is too slow (it seemingly stops working for dicts with more than 10**6 items).

mkrieger1
  • 19,194
  • 5
  • 54
  • 65
Wolf
  • 9,679
  • 7
  • 62
  • 108

4 Answers4

2

You can try this:

dct = {1:2, 3:4, 5:6, 7:8}

out = [None] * 2*len(dct)

for idx, (out[2*idx],out[2*idx+1]) in enumerate(dct.items()):
    pass

print(out)

Output:

[1, 2, 3, 4, 5, 6, 7, 8]

Check Runtime with dictionary that size is 50_000_000: (on colab)

from timeit import timeit
import operator, functools
from itertools import chain

def approach1(dct):
    li = []
    for k, v in dct.items():
        li.extend([k,v])
    return li

def approach2(dct):
    out = [None] * 2*len(dct)
    for idx, (out[2*idx],out[2*idx+1]) in enumerate(dct.items()):
        pass
    return (out)

def approach3(dct):
    return functools.reduce(operator.iconcat, dct.items(), [])

def approach4(dct):
    return list(chain.from_iterable(dct.items()))

def approach5(dct):
    return [i for t in dct.items() for i in t]
    
funcs = approach1, approach2, approach3, approach4, approach5
dct = {i:i for i in range(50_000_000)}

for _ in range(3):
    for func in funcs:
        t = timeit(lambda: func(dct), number=1)
        print('%.3f s ' % t, func.__name__)
    print()

Output:

8.825 s  approach1
13.243 s  approach2
4.506 s  approach3
3.809 s  approach4
7.881 s  approach5

8.391 s  approach1
13.159 s  approach2
4.487 s  approach3
3.854 s  approach4
7.946 s  approach5

8.391 s  approach1
13.197 s  approach2
4.448 s  approach3
3.681 s  approach4
7.904 s  approach5

Check Runtime with different size of dictionary: (on colab)

from timeit import timeit
import operator, functools
from itertools import chain
import pandas as pd
import seaborn as sns
import matplotlib.pyplot as plt
def app_extend(dct):
    li = []
    for k, v in dct.items():
        li.extend([k,v])
    return li
def app_enumerate(dct):
    out = [None] * 2*len(dct)
    for idx, (out[2*idx],out[2*idx+1]) in enumerate(dct.items()):
        pass
    return (out)
def app_functools(dct):
    return functools.reduce(operator.iconcat, dct.items(), [])
def app_chain(dct):
    return list(chain.from_iterable(dct.items()))
def app_for(dct):
    return [i for t in dct.items() for i in t]
funcs = app_extend, app_enumerate, app_functools, app_chain, app_for
dct_rslt = {}
for dct_size in [100_000, 250_000, 500_000, 1_000_000, 2_500_000, 5_000_000, 10_000_000, 25_000_000, 50_000_000]:
    dct = {i:i for i in range(dct_size)}
    dct_rslt[str(dct_size)] = {func.__name__ : timeit(lambda: func(dct), number=1) for func in funcs}
df = pd.DataFrame(dct_rslt).T
fig, ax = plt.subplots()
fig.set_size_inches(12, 9)
sns.lineplot(data=df)
plt.xlabel('Dictionary Size')
plt.ylabel('Time(sec)')
plt.show()

enter image description here

I'mahdi
  • 23,382
  • 5
  • 22
  • 30
0

Appending and extending a list is quite efficient in python:

def dict_to_list(d):
    li = []
    for k, v in d.items():
        li.extend([k,v])
    return li

Therefore, the above apparently naive function beats the very compact and somehow also elegant expression list(sum(d.items(), ())) in terms of performance.

Wolf
  • 9,679
  • 7
  • 62
  • 108
  • 2
    ``list(chain.from_iterable(a.items()))`` – jizhihaoSAMA Sep 24 '21 at 10:11
  • It's a bit sad that this first reaction didn't start as an answer. I think it's high likely that I keep the mantra *chain from iterable* in mind instead of the slightly faster functional recipe with more ingredients. – Wolf Sep 24 '21 at 11:48
0

You can use a list comprehension over the items of the dict:

d = {1: 2, 3: 4}
print([i for t in d.items() for i in t])

This outputs:

[1, 2, 3, 4]
blhsing
  • 91,368
  • 6
  • 71
  • 106
0

Use itertools function chain and classmethod alternative constructor from_iterable:

>>> from itertools import chain
>>> list(chain.from_iterable(dct.items()))
[1, 2, 3, 4]
>>> 

Or with operator.iconcat and functools.reduce:

>>> import operator, functools
>>> functools.reduce(operator.iconcat, dct.items(), [])
[1, 2, 3, 4]
>>> 
U13-Forward
  • 69,221
  • 14
  • 89
  • 114
  • Let us [continue this discussion in chat](https://chat.stackoverflow.com/rooms/237450/discussion-between-wolf-and-u12-forward). – Wolf Sep 24 '21 at 11:02