Modifying dataFrames inside a list is not working

问题:

I have two DataFrames and I want to perform the same list of cleaning ops. I realized I can merge into one, and to everything in one pass, but I am still curios why this method is not working

test_1 = pd.DataFrame({
    "A": [1, 8, 5, 6, 0],
    "B": [15, 49, 34, 44, 63]
})
test_2 = pd.DataFrame({
    "A": [np.nan, 3, 6, 4, 9, 0],
    "B": [-100, 100, 200, 300, 400, 500]
})

Let's assume I want to only take the raws without NaNs: I tried

for df in [test_1, test_2]:
    df = df[pd.notnull(df["A"])]

but test_2 is left untouched. On the other hand if I do:

test_2 = test_2[pd.notnull(test_2["A"])]

Now I the first raw went away.

回答1:

All these slicing/indexing operations create views/copies of the original dataframe and you then reassign df to these views/copies, meaning the originals are not touched at all.

Option 1
dropna(...inplace=True)
Try an in-place dropna call, this should modify the original object in-place

df_list = [test_1, test_2]
for df in df_list:
    df.dropna(subset=['A'], inplace=True)  

Note, this is one of the few times that I will ever recommend an in-place modification, because of this use case in particular.

---

Option 2
enumerate with reassignment
Alternatively, you may re-assign to the list -

for i, df in enumerate(df_list):
    df_list[i] = df.dropna(subset=['A'])  # df_list[i] = df[df.A.notnull()]

回答2:

You are modifying copies of the dataframes rather than the original dataframes.

One way to deal with this issue is to use a dictionary. As a convenience, you can use pd.DataFrame.pipe together with dictionary comprehensions to modify your dictionaries.

def remove_nulls(df):
    return df[df['A'].notnull()]

dfs = dict(enumerate([test_1, test_2]))
dfs = {k: v.pipe(remove_nulls) for k, v in dfs.items()}

print(dfs)

# {0:    A   B
#     0  1  15
#     1  8  49
#     2  5  34
#     3  6  44
#     4  0  63,
#  1:      A    B
#     1  3.0  100
#     2  6.0  200
#     3  4.0  300
#     4  9.0  400
#     5  0.0  500}

Note: In your result dfs[1]['A'] remains float: this is because np.nan is considered float and we have not triggered a conversion to int.

回答3:

By using pd.concat

[x.reset_index(level=0,drop=True) for _, x in pd.concat([test_1,test_2],keys=[0,1]).dropna().groupby(level=0)]
Out[376]: 
[     A   B
 0  1.0  15
 1  8.0  49
 2  5.0  34
 3  6.0  44
 4  0.0  63,      A    B
 1  3.0  100
 2  6.0  200
 3  4.0  300
 4  9.0  400
 5  0.0  500]