mmckerns
commented
1 year ago I've slightly modified your code, and it seems to work:
import time
import math
import os
import sys
import numpy as np
import sympy as sym
from sympy.stats import Normal, density, E, std, cdf, skewness
from sympy import lambdify
from scipy.optimize import minimize
from pathos.pools import ProcessPool as Pool
import pathos
#The following creats the log-likelihood function for 10 gaussian observations.
CS1_Onsets = np.array([11.65239205, 23.67586247, 31.96652543, 40.59708088, 48.44196047, 57.50153782, 68.16787918, 76.90944557, 88.71767575, 99.527949 ])
CS_Wait_Times = np.array(CS1_Onsets[0]) # This takes care of the fact that the first observation cannot be obtained by differencing.
CS_Wait_Times = np.append(CS_Wait_Times,np.diff(CS1_Onsets))
del CS1_Onsets
u_CS = sym.Symbol('u_CS',positive=True)
var_CS = sym.Symbol('var_CS',positive=True)
x_CS = sym.Symbol('x_CS',positive=True)
Gaussian_CS = density(Normal("x_CS", u_CS, var_CS**(1/2)))
log_likelihood = 0
#Compute log-likelihood:
for i in range(0,len(CS_Wait_Times),1):
log_likelihood = log_likelihood + sym.log(Gaussian_CS(CS_Wait_Times[i]))
######################################### Optimising the log-likelihood function #####################################
Neg_log_likelihood = -1*log_likelihood
del log_likelihood
Neg_Score = [sym.diff(Neg_log_likelihood,u_CS),sym.diff(Neg_log_likelihood,var_CS)]
##The following implements the optimisation
bounds = [(0.0000000001,86400),(0.0000000001,86400)] #u_CS, var_CS,u_CS_US_Int,var_CS_US_Int,p_US
options = {"maxiter":400}
Results = [] #Store the results for different starting points.
Initial_Parameter_Estimates = [(1,1),(10,10),(100,100),(1000,1000),(10000,10000),(8,1),(80,10),(800,100),(8000,1000),(80000,10000)]
start_time = time.perf_counter()
Objective = lambdify([(u_CS, var_CS)], Neg_log_likelihood,'mpmath') #This gives the negative log-likelihood in the desired form:
Gradient = lambdify([(u_CS, var_CS)], Neg_Score,'mpmath')
##Parallel Version - The following doesnt work.
def MINI(start_loc):
import math
import numpy as np
import sympy as sym
from sympy.stats import Normal, density, E, std, cdf, skewness
from sympy import lambdify
from scipy.optimize import minimize
from sympy.stats import Normal, density, E, std, cdf, skewness
fun = Objective
x0 = start_loc
method = 'SLSQP'
jac = Gradient
bounds = [(0.0000000001, 86400), (0.0000000001, 86400)]
options = {"maxiter": 400}
result = minimize(fun=fun, x0=x0, method='SLSQP', jac=jac, bounds=bounds, options=options)
return result
if __name__ == '__main__':
pool = Pool()
from pathos.helpers import freeze_support
freeze_support()
start_time = time.perf_counter()
RESULTS = pool.map(MINI,Initial_Parameter_Estimates)
r = list(map(MINI, Initial_Parameter_Estimates))
finish = time.perf_counter()
Running_parallel = finish - start_time
print('Running time of parallel execution is:',Running_parallel)
# cleanup
pool.close()
pool.join()
pool.clear()The primary change is to use ProcessPool instead of ParallelPool:
from pathos.pools import ProcessPool as Pool
ProcessPool uses object serialization (using dill), while ParallelPool uses serialization by source extraction (using dill.source). The former is faster, and more robust. The issue is that the sympy-generated log function can't be found by ppft as it's a dynamically-generated function. ppft checks the namespace of the function MINI... but fails to find log in some cases (as you noted). For functions that aren't dynamically-generated, you can include the importing module, and it does generally find the function. ProcessPool is using multiprocess, which I imagine is what you were wanting to use in the first place.
I'm going to close the issue, but if it doesn't work for you, then please reopen.
FOXP16
Hi, I have an issue using Pathos which I have been unable to solve myself. The following is the code I am trying to run in python 3.10:
Whenever I try to run this code, the following error message is displayed (it always repeats itself 10 times - I have 14 processors in my computer and utilisation jumps to near 100% - so the processes are getting submitted):
What do you think the issue is here? I've looked through all the documentation and tried a million and one hacks to try and get this working but I just cannot do it. The following are some potentially useful notes:
if__name_==’__main__’:of my script within the definition of ‘MINI’, then everything works. The problem is that this is a toy example for a much bigger problem that I am dealing this, and using this naïve approach is cannot be used.if __name__ == '__main__':in another script, then importing that script and using the MINI function that way (I've just described the basic idea). This hack is also too inflexible to use however.Would really appreciate your help with this, pathos is a great library and I am looking forward to using it!