问题：使用熊猫查找最多两列或更多列
我有一个列的数据帧A，B。我需要创建一个列C，以便为每个记录/行：
C = max(A, B)。
我应该怎么做呢？
回答 0
您可以这样获得最大值：
>>> import pandas as pd
>>> df = pd.DataFrame({"A": [1,2,3], "B": [-2, 8, 1]})
>>> df
A B
0 1 -2
1 2 8
2 3 1
>>> df[["A", "B"]]
A B
0 1 -2
1 2 8
2 3 1
>>> df[["A", "B"]].max(axis=1)
0 1
1 8
2 3
所以：
>>> df["C"] = df[["A", "B"]].max(axis=1)
>>> df
A B C
0 1 -2 1
1 2 8 8
2 3 1 3
如果您知道“ A”和“ B”是唯一的列，那么您甚至可以逃脱
>>> df["C"] = df.max(axis=1)
.apply(max, axis=1)我猜你也可以使用。
回答 1
在几乎所有正常情况下，@ DSM的答案都很好。但是，如果您是想比表面层次更深入一点的程序员类型，那么您可能想知道，在基础数组.values对于<0.24）上调用numpy函数要快一些，而不是直接调用调用在DataFrame / Series对象上定义的（cythonized）函数。
例如，您可以ndarray.max()沿第一个轴使用。
# Data borrowed from @DSM's post.
df = pd.DataFrame({"A": [1,2,3], "B": [-2, 8, 1]})
df
A B
0 1 -2
1 2 8
2 3 1
df['C'] = df[['A', 'B']].values.max(1)
# Or, assuming "A" and "B" are the only columns,
# df['C'] = df.values.max(1)
df
A B C
0 1 -2 1
1 2 8 8
2 3 1 3
如果您的数据包含NaN，则将需要numpy.nanmax：
df['C'] = np.nanmax(df.values, axis=1)
df
A B C
0 1 -2 1
1 2 8 8
2 3 1 3
您也可以使用numpy.maximum.reduce。numpy.maximum是ufunc（通用函数），每个ufunc都有reduce：
df['C'] = np.maximum.reduce(df['A', 'B']].values, axis=1)
# df['C'] = np.maximum.reduce(df[['A', 'B']], axis=1)
# df['C'] = np.maximum.reduce(df, axis=1)
df
A B C
0 1 -2 1
1 2 8 8
2 3 1 3
np.maximum.reduce并且np.max看起来或多或少是相同的（对于大多数标准尺寸的DataFrame），并且阴影的速度比快DataFrame.max。我认为这种差异大致保持不变，并且是由于内部开销（索引对齐，处理NaN等）引起的。
该图是使用perfplot生成的。基准测试代码，以供参考：
import pandas as pd
import perfplot
np.random.seed(0)
df_ = pd.DataFrame(np.random.randn(5, 1000))
perfplot.show(
setup=lambda n: pd.concat([df_] * n, ignore_index=True),
kernels=[
lambda df: df.assign(new=df.max(axis=1)),
lambda df: df.assign(new=df.values.max(1)),
lambda df: df.assign(new=np.nanmax(df.values, axis=1)),
lambda df: df.assign(new=np.maximum.reduce(df.values, axis=1)),
],
labels=['df.max', 'np.max', 'np.maximum.reduce', 'np.nanmax'],
n_range=[2**k for k in range(0, 15)],
xlabel='N (* len(df))',
logx=True,
logy=True)
@DSM’s answer is perfectly fine in almost any normal scenario. But if you’re the type of programmer who wants to go a little deeper than the surface level, you might be interested to know that it is a little faster to call numpy functions on the underlying .values for <0.24) array instead of directly calling the (cythonized) functions defined on the DataFrame/Series objects.
For example, you can use ndarray.max() along the first axis.
# Data borrowed from @DSM's post.
df = pd.DataFrame({"A": [1,2,3], "B": [-2, 8, 1]})
df
A B
0 1 -2
1 2 8
2 3 1
df['C'] = df[['A', 'B']].values.max(1)
# Or, assuming "A" and "B" are the only columns,
# df['C'] = df.values.max(1)
df
A B C
0 1 -2 1
1 2 8 8
2 3 1 3
If your data has NaNs, you will need numpy.nanmax:
df['C'] = np.nanmax(df.values, axis=1)
df
A B C
0 1 -2 1
1 2 8 8
2 3 1 3
You can also use numpy.maximum.reduce. numpy.maximum is a ufunc (Universal Function), and every ufunc has a reduce:
df['C'] = np.maximum.reduce(df['A', 'B']].values, axis=1)
# df['C'] = np.maximum.reduce(df[['A', 'B']], axis=1)
# df['C'] = np.maximum.reduce(df, axis=1)
df
A B C
0 1 -2 1
1 2 8 8
2 3 1 3
np.maximum.reduce and np.max appear to be more or less the same (for most normal sized DataFrames)—and happen to be a shade faster than DataFrame.max. I imagine this difference roughly remains constant, and is due to internal overhead (indexing alignment, handling NaNs, etc).
The graph was generated using perfplot. Benchmarking code, for reference:
import pandas as pd
import perfplot
np.random.seed(0)
df_ = pd.DataFrame(np.random.randn(5, 1000))
perfplot.show(
setup=lambda n: pd.concat([df_] * n, ignore_index=True),
kernels=[
lambda df: df.assign(new=df.max(axis=1)),
lambda df: df.assign(new=df.values.max(1)),
lambda df: df.assign(new=np.nanmax(df.values, axis=1)),
lambda df: df.assign(new=np.maximum.reduce(df.values, axis=1)),
],
labels=['df.max', 'np.max', 'np.maximum.reduce', 'np.nanmax'],
n_range=[2**k for k in range(0, 15)],
xlabel='N (* len(df))',
logx=True,
logy=True)
