I'm trying to specify the return type for a namedtuple in Numba and I am not able to do so. Could someone help? Consider the following minimal code:
import numba as nb
from collections import namedtuple
NT = namedtuple('NT',['sum','sum2'])
@nb.njit((nb.types.NamedTuple([nb.float64,nb.float64],NT))(nb.int64,nb.float64[:,:]),fastmath=True)
def arrsum_njit(nn,xx):
arraysum = 0.0
out = NT(sum=arraysum,sum2=arraysum)
return out
I get the error
No conversion from NT(float64 x 2) to NT(float64, float64) for '$20return_value.7', defined at None
File "numbanamedtuple.py", line 10:
def arrsum_njit(nn,xx):
<source elided>
out = NT(sum=arraysum,sum2=arraysum)
return out
^
During: typing of assignment at numbanamedtuple.py (10)
File "numbanamedtuple.py", line 10:
def arrsum_njit(nn,xx):
<source elided>
out = NT(sum=arraysum,sum2=arraysum)
return out
The problem is "overoptimized" numba compiler (bug). Add a variable of a different type to the tuple to tell the compiler to use a heterogeneous tuple (internal class).
import numba as nb
from collections import namedtuple
NT = namedtuple('NT',['sum','sum2','dummy'])
@nb.njit((nb.types.NamedTuple([nb.float64,nb.float64,nb.int64],NT))(nb.int64,nb.float64[:,:]),fastmath=True)
def arrsum_njit(nn,xx):
arraysum = 0.0
out = NT(sum=arraysum,sum2=arraysum,dummy=1)
return out
Upd: Tested: