0

我有一个稍微复杂的函数,它通过预定义的逐步逻辑(取决于固定边界以及基于实际值的相对边界)为给定数据分配质量级别。下面的函数 'get_quality()' 对每一行执行此操作,并且使用 pandas DataFrame.apply 对于大型数据集来说非常慢。所以我想把这个计算向量化。显然,我可以为内部 if-logic 做类似的事情df.groupby(pd.cut(df.ground_truth, [-np.inf, 10.0, 20.0, 50.0, np.inf])),然后在每个组内应用类似的子分组(基于每个组的边界),但是对于取决于给定实数/每行中的ground_truth值?

使用df['quality'] = np.vectorize(get_quality)(df['measured'], df['ground_truth'])速度已经快了很多,但是是否有一种真正的矢量化方法来计算相同的“质量”列?

import pandas as pd
import numpy as np
from bisect import bisect

quality_levels = ['WayTooLow', 'TooLow', 'OK',  'TooHigh', 'WayTooHigh']

# Note: to make the vertical borders always lead towards the 'better' score we use a small epsilon around them
eps = 0.000001

def get_quality(measured_value, real_value):
    diff = measured_value - real_value
    if real_value <= 10.0:
        i = bisect([-4.0-eps, -2.0-eps, 2.0+eps, 4.0+eps], diff)
        return quality_levels[i]
    elif real_value <= 20.0:
        i = bisect([-14.0-eps, -6.0-eps, 6.0+eps, 14.0+eps], diff)
        return quality_levels[i]
    elif real_value <= 50.0:
        i = bisect([-45.0-eps, -20.0-eps, 20.0+eps, 45.0+eps], diff)
        return quality_levels[i]
    else:
        i = bisect([-0.5*real_value-eps, -0.25*real_value-eps,
                    0.25*real_value+eps, 0.5*real_value+eps], diff)
        return quality_levels[i]

N = 100000
df = pd.DataFrame({'ground_truth': np.random.randint(0, 100, N),
                   'measured': np.random.randint(0, 100, N)})


df['quality'] = df.apply(lambda row: get_quality((row['measured']), (row['ground_truth'])), axis=1)
print(df.head())
print(df.quality2.value_counts())

#   ground_truth  measured     quality
#0            51         1   WayTooLow
#1             7        25  WayTooHigh
#2            38        95  WayTooHigh
#3            76        32   WayTooLow
#4             0        18  WayTooHigh

#OK            30035
#WayTooHigh    24257
#WayTooLow     18998
#TooLow        14593
#TooHigh       12117

4

1 回答 1

0

这可以通过np.select 实现

import numpy as np

quality_levels = ['WayTooLow', 'TooLow', 'OK',  'TooHigh', 'WayTooHigh']

def get_quality_vectorized(df):
    # Prepare the first 4 conditions, to match the 4 sets of boundaries.
    gt = df['ground_truth']
    conds = [gt <= 10, gt <= 20, gt <= 50, True]
    lo = np.select(conds, [2, 6, 20, 0.25 * gt])
    hi = np.select(conds, [4, 14, 45, 0.5 * gt])

    # Prepare inner 5 conditions, to match the 5 quality levels.
    diff = df['measured'] - df['ground_truth']
    quality_conds = [diff < -hi-eps, diff < -lo-eps, diff < lo+eps, diff < hi+eps, True]
    df['quality'] = np.select(quality_conds, quality_levels)
    return df
于 2020-04-17T09:06:09.850 回答