我有一个很大的熊猫数据框。数据框如下所示:
>> df
"a_1" "a_2" "b_1" "c_2" ...
"d_1" nan 0.2 nan nan
"d_2" 0.1 nan nan 1
"e_1" nan 1 nan 0.2
"e_2" nan 0.05 0.1 0.7
"f_2" 0.2 0.5 0.3 0.9
...
现在,我试图遍历一个元组列表,其中包含一些行名和列名:
t = [("d", "a"), ("d", "c") ...]
例如,当i = ("d", "a")
时,我想找出对应于a_1 and a_2
,d_1 and d_2
的值,并使用以下代码定位这些值:
s = df.loc[["d_1", "d_2" ], ["a_1", "a_2"]]
# print(s)
# "a_1" "a_2"
# "d_1" nan 0.2
# "d_2 0.1 nan
# convert to list and sort the values
s = s.unstack().reset_index()
s.columns = ["A","B", "Score"]
scores = s.sort_values(by="Score", ascending=False).reset_index(drop=True)
# pick the index(rank) I want and save the not-nan data to dictionary
rank = 1
try:
s = scores.loc[rank,:]
except Exception:
continue
if str(s.Score) != "nan":
d[(s.A, s.B)] = s.Score # output dictionary
现在,鉴于给定len(t) = 28350
,以上代码可以工作,但是花费的时间太长,我需要测试150多个参数集。
一次迭代(一组参数)在集群上花费3.5分钟。
我想知道是否有更好的解决方案来解决这个问题。预先感谢!
答案 0 :(得分:0)
那又怎么样呢?
d = {}
for row, col in t:
vals = df.loc[df.index.str.startswith(row),
df.columns.str.startswith(col)].stack().dropna()
if len(vals):
d[vals.idxmax()] = vals.max()