Logo Questions Linux Laravel Mysql Ubuntu Git Menu
 

Difference of solution between sequential and parallel programming

I have created a python code that solves a group lasso penalized linear model. For those of you not used to work with these models, the basic idea is that you give as input a dataset (x) and a response variable (y), as well as the value of a parameter (lambda1), varying the value of this parameter changes the solution of the model. So I decided to use the multiprocessing library and solve different models (associated to different parameter values). I created a python file called "model.py" in which there are the following functions:

# -*- coding: utf-8 -*-
from __future__ import division
import functools
import multiprocessing as mp
import numpy as np
from cvxpy import *

def lm_gl_preprocessing(x, y, index, lambda1=None):
    lambda_vector = [lambda1]
    m = x.shape[1]
    n = x.shape[0]
    lambda_param = Parameter(sign="positive")
    m = m+1
    index = np.append(0, index)
    x = np.c_[np.ones(n), x]
    group_sizes = []
    beta_var = []
    unique_index = np.unique(index)
    for idx in unique_index:
        group_sizes.append(len(np.where(index == idx)[0]))
        beta_var.append(Variable(len(np.where(index == idx)[0])))
    num_groups = len(group_sizes)
    group_lasso_penalization = 0
    model_prediction = x[:, np.where(index == unique_index[0])[0]] * beta_var[0]
    for i in range(1, num_groups):
        model_prediction += x[:, np.where(index == unique_index[i])[0]] * beta_var[i]
        group_lasso_penalization += sqrt(group_sizes[i]) * norm(beta_var[i], 2)
    lm_penalization = (1.0/n) * sum_squares(y - model_prediction)
    objective = Minimize(lm_penalization + (lambda_param * group_lasso_penalization))
    problem = Problem(objective)
    response = {'problem': problem, 'beta_var': beta_var, 'lambda_param': lambda_param, 'lambda_vector': lambda_vector}
    return response

def solver(problem, beta_var, lambda_param, lambda_vector):
    beta_sol_list = []
    for i in range(len(lambda_vector)):
        lambda_param.value = lambda_vector[i]
        problem.solve(solver=ECOS)
        beta_sol = np.asarray(np.row_stack([b.value for b in beta_var])).flatten()
        beta_sol_list.append(beta_sol)
    return beta_sol_list

def parallel_solver(problem, beta_var, lambda_param, lambda_vector):
    # Divide parameter vector into chunks to be executed in parallel
    num_chunks = mp.cpu_count()
    chunks = np.array_split(lambda_vector, num_chunks)
    # Solve problem in parallel
    pool = mp.Pool(num_chunks)
    global_results = pool.map(functools.partial(solver, problem, beta_var, lambda_param), chunks)
    pool.close()
    pool.join()
    return global_results
  • The function lm_gl_preprocessing basically defines the model to be solved using the cvxpy module.
  • The function solver takes the model details from the previus function, and solves an optimization problem that leads to the final solution of the model.
  • The function parallel_solver parallellizes the solver function using multiprocessing.

If, in the python console, I start runnig the parallel solver, it gives a solution. This solution is different than the one provided by the sequential solver. If I restart the python console and start runnig the sequential solver, and then I run the parallel solver, the parallel solver gives the same solution as the sequential solver. I will show:

from __future__ import division
from sklearn.datasets import load_boston
import numpy as np
import model as t

boston = load_boston()
x = boston.data
y = boston.target
index = np.array([1, 1, 1, 2, 2, 2, 3, 3, 3, 4, 4, 4, 5])

lambda1 = 1e-3

r1 = t.lm_gl_preprocessing(x=x, y=y, index=index, lambda1=lambda1)
s_parallel_1 = t.parallel_solver(problem=r1['problem'], beta_var=r1['beta_var'], lambda_param=r1['lambda_param'], lambda_vector=r1['lambda_vector'])
print(s_parallel_1)
[[array([  4.61648376e+01,  -1.22394832e-04,   0.00000000e+00,
       0.00000000e+00,   1.37065733e-04,   1.51910696e-03,
       0.00000000e+00,   1.51910696e-03,   0.00000000e+00,
       7.00079603e-03,   1.52776114e-03,  -8.67357376e-01,
       7.16429750e-03,  -8.67357376e-01])], [], [], []]
s_1 = t.solver(problem=r1['problem'], beta_var=r1['beta_var'], lambda_param=r1['lambda_param'], lambda_vector=r1['lambda_vector'])
print(s_1)
[array([  3.62813738e+01,  -1.06995338e-01,   4.64210526e-02,
      1.97112192e-02,   2.68475527e+00,  -1.75142155e+01,
      3.80741843e+00,   5.14842823e-04,  -1.47105323e+00,
      3.04949407e-01,  -1.23508259e-02,  -9.50143293e-01,
      9.40708993e-03,  -5.25758097e-01])]
#####################################################
r1 = t.lm_gl_preprocessing(x=x, y=y, index=index, lambda1=lambda1)
s_1 = t.solver(problem=r1['problem'], beta_var=r1['beta_var'], lambda_param=r1['lambda_param'], lambda_vector=r1['lambda_vector'])
print(s_1)
[array([  3.62813738e+01,  -1.06995338e-01,   4.64210526e-02,
      1.97112192e-02,   2.68475527e+00,  -1.75142155e+01,
      3.80741843e+00,   5.14842823e-04,  -1.47105323e+00,
      3.04949407e-01,  -1.23508259e-02,  -9.50143293e-01,
      9.40708993e-03,  -5.25758097e-01])]
s_parallel_1 = t.parallel_solver(problem=r1['problem'], beta_var=r1['beta_var'], lambda_param=r1['lambda_param'], lambda_vector=r1['lambda_vector'])
print(s_parallel_1)
[[array([  3.62813738e+01,  -1.06995338e-01,   4.64210526e-02,
       1.97112192e-02,   2.68475527e+00,  -1.75142155e+01,
       3.80741843e+00,   5.14842823e-04,  -1.47105323e+00,
       3.04949407e-01,  -1.23508259e-02,  -9.50143293e-01,
       9.40708993e-03,  -5.25758097e-01])], [], [], []]

PS: I know that in this example I am using parallel programming just to solve one model with one possible parameter value, but this is just a little example designed to show the difference of solutions provided by sequential and parallel programming here. I would thank any hint since I am completely lost here.

like image 373
Álvaro Méndez Civieta Avatar asked Nov 08 '22 02:11

Álvaro Méndez Civieta


1 Answers

If I execute your code I get the same result in all the cases. This is the code that I am running (I merged the 2 files):

from __future__ import division
import functools
import multiprocessing as mp
import numpy as np
from cvxpy import *
from sklearn.datasets import load_boston

def lm_gl_preprocessing(x, y, index, lambda1=None):
    lambda_vector = [lambda1]
    m = x.shape[1]
    n = x.shape[0]
    lambda_param = Parameter(sign="positive")
    m = m+1
    index = np.append(0, index)
    x = np.c_[np.ones(n), x]
    group_sizes = []
    beta_var = []
    unique_index = np.unique(index)
    for idx in unique_index:
        group_sizes.append(len(np.where(index == idx)[0]))
        beta_var.append(Variable(len(np.where(index == idx)[0])))
    num_groups = len(group_sizes)
    group_lasso_penalization = 0
    model_prediction = x[:, np.where(index == unique_index[0])[0]] * beta_var[0]
    for i in range(1, num_groups):
        model_prediction += x[:, np.where(index == unique_index[i])[0]] * beta_var[i]
        group_lasso_penalization += sqrt(group_sizes[i]) * norm(beta_var[i], 2)
    lm_penalization = (1.0/n) * sum_squares(y - model_prediction)
    objective = Minimize(lm_penalization + (lambda_param * group_lasso_penalization))
    problem = Problem(objective)
    response = {'problem': problem, 'beta_var': beta_var, 'lambda_param': lambda_param, 'lambda_vector': lambda_vector}
    return response

def solver(problem, beta_var, lambda_param, lambda_vector):
    beta_sol_list = []
    for i in range(len(lambda_vector)):
        lambda_param.value = lambda_vector[i]
        problem.solve(solver=ECOS)
        beta_sol = np.asarray(np.row_stack([b.value for b in beta_var])).flatten()
        beta_sol_list.append(beta_sol)
    return beta_sol_list

def parallel_solver(problem, beta_var, lambda_param, lambda_vector):
    # Divide parameter vector into chunks to be executed in parallel
    num_chunks = mp.cpu_count()
    chunks = np.array_split(lambda_vector, num_chunks)
    # Solve problem in parallel
    pool = mp.Pool(num_chunks)
    global_results = pool.map(functools.partial(solver, problem, beta_var, lambda_param), chunks)
    pool.close()
    pool.join()
    return global_results

if __name__ == "__main__":
     boston = load_boston()
     x = boston.data
     y = boston.target
     index = np.array([1, 1, 1, 2, 2, 2, 3, 3, 3, 4, 4, 4, 5])

     lambda1 = 1e-3

     r1 = lm_gl_preprocessing(x=x, y=y, index=index, lambda1=lambda1)
     s_parallel_1 = parallel_solver(problem=r1['problem'], beta_var=r1['beta_var'], lambda_param=r1['lambda_param'], lambda_vector=r1['lambda_vector'])
     print(s_parallel_1)
     r1 = lm_gl_preprocessing(x=x, y=y, index=index, lambda1=lambda1)
     s_1 = solver(problem=r1['problem'], beta_var=r1['beta_var'], lambda_param=r1['lambda_param'], lambda_vector=r1['lambda_vector'])
     print(s_1)
     print ("#####################################################")
     r1 = lm_gl_preprocessing(x=x, y=y, index=index, lambda1=lambda1)
     s_1 = solver(problem=r1['problem'], beta_var=r1['beta_var'], lambda_param=r1['lambda_param'], lambda_vector=r1['lambda_vector'])
     print(s_1)
     r1 = lm_gl_preprocessing(x=x, y=y, index=index, lambda1=lambda1)
     s_parallel_1 = parallel_solver(problem=r1['problem'], beta_var=r1['beta_var'], lambda_param=r1['lambda_param'], lambda_vector=r1['lambda_vector'])
     print(s_parallel_1)

and output:

[[array([ 3.62813738e+01, -1.06995338e-01,  4.64210526e-02,  1.97112192e-02,
        2.68475527e+00, -1.75142155e+01,  3.80741843e+00,  5.14842823e-04,
       -1.47105323e+00,  3.04949407e-01, -1.23508259e-02, -9.50143293e-01,
        9.40708993e-03, -5.25758097e-01])], [], [], []]
[array([ 3.62813738e+01, -1.06995338e-01,  4.64210526e-02,  1.97112192e-02,
        2.68475527e+00, -1.75142155e+01,  3.80741843e+00,  5.14842823e-04,
       -1.47105323e+00,  3.04949407e-01, -1.23508259e-02, -9.50143293e-01,
        9.40708993e-03, -5.25758097e-01])]
#####################################################
[array([ 3.62813738e+01, -1.06995338e-01,  4.64210526e-02,  1.97112192e-02,
        2.68475527e+00, -1.75142155e+01,  3.80741843e+00,  5.14842823e-04,
       -1.47105323e+00,  3.04949407e-01, -1.23508259e-02, -9.50143293e-01,
        9.40708993e-03, -5.25758097e-01])]
[[array([ 3.62813738e+01, -1.06995338e-01,  4.64210526e-02,  1.97112192e-02,
        2.68475527e+00, -1.75142155e+01,  3.80741843e+00,  5.14842823e-04,
       -1.47105323e+00,  3.04949407e-01, -1.23508259e-02, -9.50143293e-01,
        9.40708993e-03, -5.25758097e-01])], [], [], []]

As you can see, I have the same number of CPUs (4).

My environment is Python2.7 on Linux and these are the versions of the relevant packages:

>>> import sklearn
>>> sklearn.__version__
'0.19.2'
>>> import scipy
>>> scipy.__version__
'1.1.0'
>>> import numpy 
>>> numpy.__version__
'1.15.2'
>>> import cvxpy
>>> cvxpy.__version__
'0.4.0'
>>> import multiprocessing
>>> multiprocessing.__version__
'0.70a1'
like image 181
Amedeo Avatar answered Nov 14 '22 22:11

Amedeo