0

I define a function using numba @jit(nopython =True) which will call a normal cdf function via scipy.special, and I got this error message: 

TypingError: Untyped global name 'norm'

The following is my code: 

import numpy as np
from numba import * 
from scipy.stats import *
timelst = np.random.randn(2083)
noncovlst = np.random.randn(2083)
lst = np.random.randint(30,size=127)
lst_num = 0
lst_num = np.append(lst_num,lst)
lst_num = np.cumsum(lst_num)
n = 127
agelst = np.random.randint(16,80,127)
edulst = np.random.randint(1,28,127)
incomelst = np.random.randn(127)



@jit(nopython=True)
def lamb(now,params):
    b = params[0]
    mu = params[1]
    s = params[2]
    beta1 = params[3]
    beta3 = params[4]
    beta4 = params[5]
    beta5 = params[6]
    beta6 = params[7]
    L = 0.0

    for i in range(n):
        timez = timelst[lst_num[i]:lst_num[i+1]]
        noncov = noncovlst[lst_num[i]:lst_num[i+1]]
        timeq = timez[timez < now]
        noncovq = noncov[timez < now]
        L += (b*now + np.exp(beta3*agelst[i]+beta5*edulst[i]
                             +beta6*incomelst[i])*(norm.cdf((now-timeq-mu)/s)*np.exp(beta1*noncovq)).sum())/n
    return L

res = [2.422855,16.320364,28.984707,-9.942004,-3.118748,-0.248391,1.903728,4.126649]

Type 

lamb(0.2,res)

The error message:

TypingError: Failed at nopython (nopython frontend)
Untyped global name 'norm'

I know drop nopython keyword will probably fix the issue, however, it will covert to object mode, which can have a huge performance penalty.

Is any better way to include the normal cdf into numba nopython mode ?

lz10086lz
  • 43
  • 5
  • Doesn't `nopython` mean don't use the imported function and objects, like `norm`? To get the most speed out of `numba` you have to write out the calculations in all their gory detail; no imported 'black boxes'. – hpaulj Apr 23 '16 at 16:29
  • http://numba.pydata.org/numba-doc/0.17.0/reference/pysupported.html#pysupported has a list of features that the `nopython` mode can handle. – hpaulj Apr 23 '16 at 16:35
  • `stats.norm.cdf` massages its inputs and ends up calling `special.ndtr`, which is a compiled function (`scipy/special/cephes/ndtr.c`). So even it had access to the code I doubt if `numba` could improve on it. – hpaulj Apr 24 '16 at 01:18

0 Answers0