如何设置智能oneliner以计算非常大的数据?

时间:2018-05-17 13:00:19

标签: python arrays performance loops

我使用维度( Xpos )屏蔽了纬度(Ypos)和经度(125,800,000)的数组。

我想计算数组中纬度和经度的差异。这是数组XposYpos类似)。

 masked_array(  data       = [ [-2.0551843643188477, -2.637551784515381, -2.720881223678589,  ..., 2.2812530994415283, 2.281250476837158, 2.281254768371582 ],
                               [-2.3242127895355225, -2.804257869720459, -2.8825504779815674, ..., 2.2812530994415283, 2.281250476837158, 2.281254768371582 ],
                               [-2.073770523071289, -2.6198980808258057, -2.708889961242676,  ..., 2.2812530994415283, 2.281250476837158, 2.281254768371582 ],
                               ...,
                               [-3.517531633377075, -2.908338785171509, -2.9069409370422363,  ..., 2.2812530994415283, 2.281250476837158, 2.281254768371582 ],
                               [-3.688690662384033, -3.0086288452148438, -3.010164260864258,  ..., 2.2812530994415283, 2.281250476837158, 2.281254768371582 ],
                               [-3.520817518234253, -2.943941116333008, -2.941941738128662,   ..., 2.2812530994415283, 2.281250476837158, 2.281254768371582 ]
                               ],
                mask       = [ [ False, False, False, ..., False, False, False ],
                               [ False, False, False, ..., False, False, False],
                               [ False, False, False, ..., False, False, False],
                               ...,
                               [ False, False, False, ..., False, False, False],
                               [ False, False, False, ..., False, False, False],
                               [ False, False, False, ..., False, False, False]
                               ],
                fill_value = 1e+20,
                dtype      = float32
                )

这是我的代码,它可以运行,但需要超长时间来计算。

Dist= np.zeros((len(XposApr),len(XposApr[0])))
DiffLon=np.zeros((len(XposApr),len(XposApr[0])))
DiffLat=np.zeros((len(XposApr),len(XposApr[0])))
for i in range (1,len(XposApr),12):
    for j in range (0,len(XposApr[0])):
        DiffLon[i][j]=(XposApr[i][j]-XposApr[i-1][j])
        DiffLat[i][j]=(YposApr[i][j]-YposApr[i-1][j])

我真的不知道如何制作着名的单行,这是我尝试的,但不起作用:

DistLon = [XposApr[i][j]-XposApr[i-1][j] for i in enumerate (XposApr) and j in enumerate (XposApr[0])]

是否可以制作一个oneliner或其他代码,使计算更快? 提前谢谢!

2 个答案:

答案 0 :(得分:0)

  

是否可以制作oneliner或其他代码,使计算更快

语法格式不是实现速度的核心,
智能矢量化或重新配方是:

直到光荣的 numba -team决定将支持掩码阵列的操作转移到前置刻录机上,来自numba.jit( ... )的JIT加速速度不足(Masked array support #2103

让我们试试numpy - 原生步骤:

内环可能首先退出慢速GIL驱动的运动:

for     i in range ( 1, len( XposApr    ), 12 ):
    #or j in range ( 0, len( XposApr[0] )     ):
        DiffLon[i][:] = ( XposApr[i][:] - XposApr[i-1][:] )
        DiffLat[i][:] = ( YposApr[i][:] - YposApr[i-1][:] )

然而,如果我正确地阅读了布局,可以通过这样的方式解决整个问题:

 DiffLon[1:len( XposApr ):12] = ( XposApr[1:len( XposApr   ):12]
                                - XposApr[0:len( XposApr )-1:12]
                                  )
 DiffLat[1:len( XposApr ):12] = ( YposApr[1:len( XposApr   ):12]
                                - YposApr[0:len( XposApr )-1:12]
                                  )

我相信@Divakar@cᴏʟᴅsᴘᴇᴇᴅ,真正的numpy - 这里的大师会准备好告诉你更聪明,更有效的切片/矢量化技巧来做同样的事,请耐心等待:o)

答案 1 :(得分:0)

感谢您的回答和@patrick的有用评论,我相信这也可以解决我的问题:

Xapr=np.array(XposApr)
Yapr=np.array(YposApr)
for i in range (1,len(XposApr)):
    DiffLon[i]=(Xapr[i]-Xapr[i-1])
    DiffLat[i]=(Yapr[i]-Yapr[i-1])