Python：从数据帧获取随机数据

具有具有以下值的df：

name     algo      accuracy
tom       1         88
tommy     2         87
mark      1         88
stuart    3         100
alex      2         99
lincoln   1         88

如何从df中随机选取4条记录，条件是至少应从每个唯一的algo列值中选取一条记录。这里，algo列只有3个唯一值（1、2、3）

样本输出：

name     algo      accuracy
tom       1         88
tommy     2         87
stuart    3         100
lincoln   1         88

样本输出2：

name     algo      accuracy
mark      1         88
stuart    3         100
alex      2         99
lincoln   1         88

单程

num_sample, num_algo = 4, 3

# sample one for each algo
out = df.groupby('algo').sample(n=num_sample//num_algo)

# append one more sample from those that didn't get selected.
out = out.append(df.drop(out.index).sample(n=num_sample-num_algo) )

---

另一种方法是洗牌整个数据，枚举每个algo中的行，按该枚举排序，并获取所需数量的样本。这比第一种方法的代码略多，但更便宜，并产生更平衡的算法计数：

# shuffle data
df_random = df['algo'].sample(frac=1)

# enumerations of rows with the same algo
enums = df_random.groupby(df_random).cumcount()

# sort with `np.argsort`:
enums = enums.sort_values()

# pick the first num_sample indices
# these will be indices of the samples
# so we can use `loc`
out = df.loc[enums.iloc[:num_sample].index]

单程

num_sample, num_algo = 4, 3

# sample one for each algo
out = df.groupby('algo').sample(n=num_sample//num_algo)

# append one more sample from those that didn't get selected.
out = out.append(df.drop(out.index).sample(n=num_sample-num_algo) )

---

另一种方法是洗牌整个数据，枚举每个algo中的行，按该枚举排序，并获取所需数量的样本。这比第一种方法的代码略多，但更便宜，并产生更平衡的算法计数：

# shuffle data
df_random = df['algo'].sample(frac=1)

# enumerations of rows with the same algo
enums = df_random.groupby(df_random).cumcount()

# sort with `np.argsort`:
enums = enums.sort_values()

# pick the first num_sample indices
# these will be indices of the samples
# so we can use `loc`
out = df.loc[enums.iloc[:num_sample].index]