这是输入:
df = pd.DataFrame({'keys': [('K0', 'K1'), ('K1', 'K2')],
'A': ['A0', 'A1']})
lookup_df = pd.DataFrame({'val': ['V1', 'V2', 'V3']},
index = ['K0', 'K1', 'K2'])
在进行一些“加入”操作后,我希望在df
中添加一个新列,将keys
中df
的密钥映射到val
lookup_df
。
输出应为:
pd.DataFrame({'keys': [('K0', 'K1'), ('K1', 'K2')],
'val': [('V0', 'V1'), ('V1', 'V2')],
'A': ['A0', 'A1']})
我能想到的一种方式是:
df['val'] = df['keys'].apply(lambda ks:
list(map(lambda k: lookup_df.loc[k].val, ks)))
还有其他更好的方法来实现这一目标吗?
答案 0 :(得分:2)
你可以这样做:
In [83]: df['val'] = df['keys'].str.join(',').str.split(',', expand=True).stack().map(lookup_df.val).unstack().apply(tuple)
In [84]: df
Out[84]:
A keys val
0 A0 (K0, K1) (V1, V2)
1 A1 (K1, K2) (V2, V3)
In [85]: lookup_df
Out[85]:
val
K0 V1
K1 V2
K2 V3
或更好一点,但速度较慢的方法(感谢@Boud):
In [5]: df['val'] = df['keys'].apply(pd.Series).stack().map(lookup_df.val).unstack().apply(tuple)
In [6]: df
Out[6]:
A keys val
0 A0 (K0, K1) (V1, V2)
1 A1 (K1, K2) (V2, V3)
针对10K行的计时DF:
In [18]: big = pd.concat([df] * 10**5, ignore_index=True)
In [19]: x = big.head(10**4)
In [20]: x.shape
Out[20]: (10000, 2)
In [21]: %timeit x['keys'].str.join(',').str.split(',', expand=True).stack().map(lookup_df.val).unstack().apply(tuple)
10 loops, best of 3: 75.1 ms per loop
In [22]: %timeit x['keys'].apply(pd.Series).stack().map(lookup_df.val).unstack().apply(tuple)
1 loop, best of 3: 5.5 s per loop
In [23]: %timeit x['keys'].apply(pd.Series).replace(lookup_df.val).apply(tuple)
1 loop, best of 3: 5.52 s per loop
In [24]: %%timeit
....: dk = pd.DataFrame(x['keys'].tolist()).applymap(lambda x: lookup_df.val[x])
....: x['val'] = zip(dk[0], dk[1])
....:
1 loop, best of 3: 1.66 s per loop
结论:最丑陋的方法目前是最快的方法
答案 1 :(得分:2)
缩短并避免字符串操作:
private func touchPointToScenePoint(recognizer: UIGestureRecognizer) -> SCNVector3 {
// Get touch point
let touchPoint = recognizer.locationInView(sceneView)
// Compute near & far points
let nearVector = SCNVector3(x: Float(touchPoint.x), y: Float(touchPoint.y), z: 0)
let nearScenePoint = sceneView.unprojectPoint(nearVector)
let farVector = SCNVector3(x: Float(touchPoint.x), y: Float(touchPoint.y), z: 1)
let farScenePoint = sceneView.unprojectPoint(farVector)
// Compute view vector
let viewVector = SCNVector3(x: Float(farScenePoint.x - nearScenePoint.x), y: Float(farScenePoint.y - nearScenePoint.y), z: Float(farScenePoint.z - nearScenePoint.z))
// Normalize view vector
let vectorLength = sqrt(viewVector.x*viewVector.x + viewVector.y*viewVector.y + viewVector.z*viewVector.z)
let normalizedViewVector = SCNVector3(x: viewVector.x/vectorLength, y: viewVector.y/vectorLength, z: viewVector.z/vectorLength)
// Scale normalized vector to find scene point
let scale = Float(15)
let scenePoint = SCNVector3(x: normalizedViewVector.x*scale, y: normalizedViewVector.y*scale, z: normalizedViewVector.z*scale)
print("2D point: \(touchPoint). 3D point: \(nearScenePoint). Far point: \(farScenePoint). scene point: \(scenePoint)")
// Return <scenePoint>
return scenePoint
}
答案 2 :(得分:1)
这并不意味着漂亮。
dk = pd.DataFrame(df['keys'].tolist()).applymap(lambda x: lookup_df.val[x])
df['val'] = zip(dk[0], dk[1])
df