Joblib doesn't call a custom function when n_jobs > 1

Question

I have an example with data.

As you can see from the code, every call of a function fit_by_idx() has to print 'here', but it doesn't. All is ok when n_jobs=1, but if n_jobs is more, than joblib does not call the function.

Code:

import statsmodels.tsa.holtwinters as holtwinters
import pandas as pd
import numpy as np
from joblib import Parallel, delayed

train = pd.read_csv('train.csv').drop(columns=['id'])


def iter_predict(data, model, steps, fit_args=[],  fit_kwargs={}): # steps - кол. предсказываемых точек
    def fit_by_idx(idx):
        print('here')
        endog = data.iloc[idx]
        fitted = model(endog).fit(*fit_args, optimized=False, **fit_kwargs)\
        res[idx, :] = fitted.forecast(steps)

    res = np.zeros((data.shape[0], steps))
    Parallel(n_jobs=2)(delayed(fit_by_idx)(idx) for idx in range(data.shape[0]))
    return res

iter_predict(train, holtwinters.SimpleExpSmoothing, 2, fit_kwargs={'smoothing_level': 0.5})

And here's the link to the dataset.

user3666197 · Answer 1 · 2020-04-30T04:08:53.897

Q : "if n_jobs is more, than joblib does not call the function."

Well, it does ( you may check the PID & PPID numbers ),
it just does not present the result of the print( "here" )

Using the definition from API documentation :

print( *objects, sep = ' ', end = '\n', file = sys.stdout, flush = False ) start with enforcing the flush = True

Yet, going to face more troubles down the road. joblib-spawns ( unless forced otherwise, at a cost of adverse effects on performance, if returning back to a pure-[SERIAL], GIL-controlled re-[SERIAL]-ised code execution for any n_jobs running again one-step-after-another-after-another, which does not make sense, as you pay all the costs of instantiations and other overheads, yet not receiving any speed-up benefit from doing so, does it? ). Using:

def iter_preDEMO( data,            # Pandas DF-alike data
                  #other args removed for MCVE-clarity
                  ):

    def fit_by_idx( idx ): #-------------------------------------[FUNCTION]-def-<start> To be transferred to each remote-joblib-initiated process(es)

        print( 'here[{0:_>4d}(PPID:PID={1:_>7d}:{2::>7d})]'.format( idx,
                                                                    os.getppid(), # test joblib-[FUNCTION]-def-transfer here with: lambda x = "_{0:}_" : x.format( os.getppid() )
                                                                    os.getpid()   # test joblib-[FUNCTION]-def-transfer here with: lambda x = "_{0:}_" : x.format( os.getpid()  )
                                                                    ),
                end   = "\t",
                flush = True
                )
    #------------------------------------------------------------[FUNCTION]-def-<end>

    res = np.zeros( ( data.shape[0], 3 ) )
    for aBackEND in ( 'threading', 'loky', 'multiprocessing' ):
        try:
             print( "\n____________________________Going into ['{0:}']-backend".format( aBackEND ) )
             with parallel_backend( aBackEND, n_jobs = N_JOBS ):
                  Parallel( n_jobs = N_JOBS )( delayed( fit_by_idx )( pickled_SER_DES_copy_of_idx )
                                               for                    pickled_SER_DES_copy_of_idx in range( data.shape[0] )
                                               )
        finally:
             print( "\n_____________________________Exit from ['{0:}']-backend".format( aBackEND ) )
    return res

you shall see how the things work, using a bit more detailed, print()-ed results

START: PID=_____22528

____________________________Going into ['threading']-backend
here[___0(PPID:PID=__22527:::22528)]    here[___1(PPID:PID=__22527:::22528)]    here[___2(PPID:PID=__22527:::22528)]    here[___3(PPID:PID=__22527:::22528)]    here[___4(PPID:PID=__22527:::22528)]    here[___5(PPID:PID=__22527:::22528)]    here[___6(PPID:PID=__22527:::22528)]    here[___7(PPID:PID=__22527:::22528)]    here[___8(PPID:PID=__22527:::22528)]    here[___9(PPID:PID=__22527:::22528)]    here[__10(PPID:PID=__22527:::22528)]    here[__11(PPID:PID=__22527:::22528)]    here[__12(PPID:PID=__22527:::22528)]    here[__13(PPID:PID=__22527:::22528)]    here[__14(PPID:PID=__22527:::22528)]    here[__15(PPID:PID=__22527:::22528)]    here[__16(PPID:PID=__22527:::22528)]    
_____________________________Exit from ['threading']-backend

____________________________Going into ['loky']-backend
here[___0(PPID:PID=__22527:::22528)]    here[___1(PPID:PID=__22527:::22528)]    here[___2(PPID:PID=__22527:::22528)]    here[___3(PPID:PID=__22527:::22528)]    here[___4(PPID:PID=__22527:::22528)]    here[___5(PPID:PID=__22527:::22528)]    here[___6(PPID:PID=__22527:::22528)]    here[___7(PPID:PID=__22527:::22528)]    here[___8(PPID:PID=__22527:::22528)]    here[___9(PPID:PID=__22527:::22528)]    here[__10(PPID:PID=__22527:::22528)]    here[__11(PPID:PID=__22527:::22528)]    here[__12(PPID:PID=__22527:::22528)]    here[__13(PPID:PID=__22527:::22528)]    here[__14(PPID:PID=__22527:::22528)]    here[__15(PPID:PID=__22527:::22528)]    here[__16(PPID:PID=__22527:::22528)]    
_____________________________Exit from ['loky']-backend

____________________________Going into ['multiprocessing']-backend
here[___0(PPID:PID=__22527:::22528)]    here[___1(PPID:PID=__22527:::22528)]    here[___2(PPID:PID=__22527:::22528)]    here[___3(PPID:PID=__22527:::22528)]    here[___4(PPID:PID=__22527:::22528)]    here[___5(PPID:PID=__22527:::22528)]    here[___6(PPID:PID=__22527:::22528)]    here[___7(PPID:PID=__22527:::22528)]    here[___8(PPID:PID=__22527:::22528)]    here[___9(PPID:PID=__22527:::22528)]    here[__10(PPID:PID=__22527:::22528)]    here[__11(PPID:PID=__22527:::22528)]    here[__12(PPID:PID=__22527:::22528)]    here[__13(PPID:PID=__22527:::22528)]    here[__14(PPID:PID=__22527:::22528)]    here[__15(PPID:PID=__22527:::22528)]    here[__16(PPID:PID=__22527:::22528)]    
_____________________________Exit from ['multiprocessing']-backend

 [[0. 0. 0.]
  [0. 0. 0.]
  ...
  ]

Also check this and this on your O/S, resp. your actual joblib and (hidden) pickling-SER/DES-tools versions.

Joblib doesn't call a custom function when n_jobs > 1

1 Answers1