您正在尝试编写一个用户定义的聚合函数,这在 pyspark 中无法完成,请参阅https://stackoverflow.com/a/40030740 https://stackoverflow.com/a/40030740.
你可以写的是UDF
以列表形式收集的每组内的数据:
首先进行设置:
import pandas as pd
import numpy as np
from scipy.optimize import minimize
import pyspark.sql.functions as psf
from pyspark.sql.types import *
df = pd.DataFrame({
'y0': np.random.randn(20),
'y1': np.random.randn(20),
'x0': np.random.randn(20),
'x1': np.random.randn(20),
'grpVar': ['a', 'b'] * 10})
sdf = sqlContext.createDataFrame(df)
# Starting values
startVal = np.ones(2)*(1/2)
#Constraint Sum of coefficients = 0
cons = ({'type':'eq', 'fun': lambda x: 1 - sum(x)})
# Bounds on coefficients
bnds = tuple([0,1] for x in startVal)
我们将广播这些变量,因为我们需要在聚合数据帧的每一行上调用它们,它将把值复制到每个节点,这样它们就不必在驱动程序上获取它们:
sc.broadcast(startVal)
sc.broadcast(bnds)
让我们使用以下方法聚合数据collect_list
,我们将更改周围数据的结构,以便我们只有一列(您可以将每一列收集到不同的列中,但随后您必须修改将数据传递给函数的方式):
Sgrp_grpVar = sdf\
.groupby('grpVar')\
.agg(psf.collect_list(psf.struct("y0", "y1", "x0", "x1")).alias("data"))
Sgrp_grpVar.printSchema()
root
|-- grpVar: string (nullable = true)
|-- data: array (nullable = true)
| |-- element: struct (containsNull = true)
| | |-- y0: double (nullable = true)
| | |-- y1: double (nullable = true)
| | |-- x0: double (nullable = true)
| | |-- x1: double (nullable = true)
我们现在可以创建我们的UDF
,返回的数据类型对于pyspark来说太复杂,numpy arrays
pyspark 不支持,所以我们需要稍微改变一下:
def sSumSqDif(a, data):
return np.sum(
(data['y0'] - a[0]*data['x0'])**2 \
+ (data['y1'] - a[1]*data['x1'])**2)
def sRunMinimize(data, startVal=startVal, bnds=bnds, cons=cons):
data = pd.DataFrame({k:v for k,v in zip(["y0", "y1", "x0", "x1"], data)})
ResultByGrp = minimize(sSumSqDif, startVal, method='SLSQP',
bounds=bnds, constraints = cons, args=(data))
return ResultByGrp.x.tolist()
sRunMinimize_udf = lambda startVal, bnds, cons: psf.udf(
lambda data: sRunMinimize(data, startVal, bnds, cons),
ArrayType(DoubleType())
)
我们现在可以将此函数应用于每组中收集的数据:
Results = Sgrp_grpVar.select(
"grpVar",
sRunMinimize_udf(startVal, bnds, cons)("data").alias("res")
)
Results.show(truncate=False)
+------+-----------------------------------------+
|grpVar|res |
+------+-----------------------------------------+
|b |[0.4073139282953772, 0.5926860717046227] |
|a |[0.8275186444565927, 0.17248135554340727]|
+------+-----------------------------------------+
但我不认为 pyspark 是合适的工具。