Python DataFrames 的假设混合策略行为

Question

以下按预期工作

from datetime import datetime

from hypothesis.extra.pandas import columns, data_frames, indexes
import hypothesis.strategies as st


def boundarize(d: datetime):
    return d.replace(minute=15 * (d.minute // 15), second=0, microsecond=0)


min_date = datetime(2022, 4, 1, 22, 22, 22)
max_date = datetime(2022, 5, 1, 22, 22, 22)

dfs = data_frames(
    index=indexes(
        elements=st.datetimes(min_value=min_date, max_value=max_date).map(boundarize),
        min_size=3,
        max_size=5,
    ).map(lambda idx: idx.sort_values()),
    columns=columns("A B C".split(), dtype=int),
)
dfs.example()

输出类似于

                              A           B           C
2022-04-06 12:45:00      -11482  1588438979 -1994987295
2022-04-08 15:45:00  -833447611           3         -51
2022-04-24 06:15:00  -465371373   990274387      -14969
2022-05-01 01:15:00  1750446827  1214440777         116
2022-05-01 06:15:00      -44089       30508       58737

现在，当我尝试通过

生成具有均匀 spaced DatetimeIndex 值的类似 DataFrame 时

from datetime import datetime

from hypothesis.extra.pandas import columns, data_frames, indexes
import hypothesis.strategies as st


def boundarize(d: datetime):
    return d.replace(minute=15 * (d.minute // 15), second=0, microsecond=0)


min_date_start = datetime(2022, 4, 1, 11, 11, 11)
max_date_start = datetime(2022, 4, 2, 11, 11, 11)
min_date_end = datetime(2022, 5, 1, 22, 22, 22)
max_date_end = datetime(2022, 5, 2, 22, 22, 22)

dfs = data_frames(
    index=st.builds(pd.date_range,
                    start=st.datetimes(min_value=min_date_start, max_value=max_date_start).map(boundarize),
                    end=st.datetimes(min_value=min_date_end, max_value=max_date_end).map(boundarize),
                    freq=st.just("15T"),
                    ),
    columns=columns("A B C".split(), dtype=int),
)
dfs.example()

输出如下，请注意整数列在第一个示例中不存在时始终为零：

                     A  B  C
2022-04-01 15:45:00  0  0  0
2022-04-01 16:00:00  0  0  0
2022-04-01 16:15:00  0  0  0
2022-04-01 16:30:00  0  0  0
2022-04-01 16:45:00  0  0  0
...                 .. .. ..
2022-05-01 21:15:00  0  0  0
2022-05-01 21:30:00  0  0  0
2022-05-01 21:45:00  0  0  0
2022-05-01 22:00:00  0  0  0
2022-05-01 22:15:00  0  0  0

[2907 rows x 3 columns]

这是预期的行为还是我遗漏了什么？

编辑:

回避“随机连续子集”的方法（请参阅下面我的评论），我还尝试使用预定义索引

from datetime import datetime

from hypothesis.extra.pandas import columns, data_frames
import hypothesis.strategies as st

min_date_start = datetime(2022, 4, 1, 8, 0, 0)

dfs = data_frames(
    index=st.just(pd.date_range(start=min_date_start, periods=10, freq="15T")),
    columns=columns("A B C".split(), dtype=int),
)
dfs.example()

它也给出所有零列

                     A  B  C
2022-04-01 08:00:00  0  0  0
2022-04-01 08:15:00  0  0  0
2022-04-01 08:30:00  0  0  0
2022-04-01 08:45:00  0  0  0
2022-04-01 09:00:00  0  0  0
2022-04-01 09:15:00  0  0  0
2022-04-01 09:30:00  0  0  0
2022-04-01 09:45:00  0  0  0
2022-04-01 10:00:00  0  0  0
2022-04-01 10:15:00  0  0  0

编辑 2:

我试图想出一个连续子集的手工版本，它应该减少 space 的值，以便根据 @zac-hatfield-dodds 的回答为列值留下足够的熵，但根据经验它仍然生成几乎全为零的列值

from datetime import datetime
import math

import hypothesis.strategies as st
from hypothesis.extra.pandas import columns, data_frames
import pandas as pd

time_start = datetime(2022, 4, 1, 8, 0, 0)
time_stop = datetime(2022, 4, 2, 8, 0, 0)
r = pd.date_range(start=time_start, end=time_stop, freq="15T")

def build_indices(sequence):
    first = 0
    if len(sequence) % 2 == 0:
        mid_ceiling = len(sequence) // 2
        mid_floor = mid_ceiling - 1
    else:
        mid_floor = math.floor(len(sequence) / 2)
        mid_ceiling = mid_floor + 1
    second = len(sequence) - 1
    return first, mid_floor, mid_ceiling, second

first, mid_floor, mid_ceiling, second = build_indices(r)

a = st.integers(min_value=first, max_value=mid_floor)
b = st.integers(min_value=mid_ceiling, max_value=second)

def indexer(sequence, lower, upper):
    return sequence[lower:upper]

dfs = data_frames(
    index=st.builds(lambda lower, upper: indexer(r, lower, upper), lower=a, upper=b),
    columns=columns("A B C".split(), dtype=int),
)
dfs.example()

Answer 1

您的问题是后面的索引 大得多 ，假设是运行出于熵生成列内容。如果您将索引限制为最多几十个条目，一切都应该正常。

我们有这个 soft-cap 是为了限制其他无限的递归结构，所以整体设计按预期工作，但我承认在这种情况下它既不必要也不可取。

Python DataFrames 的假设混合策略行为

Python Hypothesis mixing strategies bahavior for DataFrames

python-hypothesis