exp2.py

   1
   2 import numpy as np
   3 from analysis.analysis import error
   4 from graph.graph import Random
   5 from Optimizers import DOPTIMIZER as dopt
   6 from utilities import utilities as ut
   7 from Problems.logistic_regression import LR_L2
   8 from Optimizers import COPTIMIZER as copt
   9 import os
  10 from utilities.plot_utils import plot_exp2
  11
  12
  13 seed       = np.random.randint(12345)
  14 seed       = 8075
  15 np.random.seed(seed)
  16
  17
  18 #### create asynchronous setup
  19 num_nodes  = 20
  20 mean_comp  = 1/np.array([5,10,15,20,25,30,35,40,45,50,55,60,65,70,75,80,85,90,95,100])
  21
  22 T_active_exp2, Tv_nodes_exp2, node_comp_time_exp2 = \
  23     ut.create_computation_time(num_nodes, max_iter=int(12000), comp_time_dist='exp', mean_comp=mean_comp,\
  24                                 min_comp = None, max_comp=None, variance_comp=None, make_integer=True)
  25
  26
  27 #### Logistic Regression Problem ==> p: dimension of the model,   L: L-smooth constant,  N: total number of training samples,  b: average number of local samples,  l: regularization factor
  28 lr_0 = LR_L2(num_nodes, limited_labels=False, balanced=True, train=1000 , regularization=True, lamda=None)
  29
  30 #### Create gossip matrices
  31 zero_row_sum, zero_column_sum, row_stochastic, col_stochastic, N_out, neighbors  = Random(num_nodes, prob=0.6, Laplacian_dividing_factor= 2).directed()
  32
  33 #### some parameters of the algorithms
  34 depoch    = 12500
  35 rho       = 0.1
  36 alpha     = 0.7
  37 gamma     = 0.5
  38 eta       = 1.0
  39 expScale  = 1/50
  40
  41 step_asy_dagp   = 1.5/lr_0.L
  42 step_asy_spa    = 1/lr_0.L/12
  43 step_center     = 1/lr_0.L/2
  44 step_size_appg  = 1.5/lr_0.L
  45
  46
  47 ## find the optimal solution of Logistic regression
  48 theta_c0    = np.random.normal(0,1,lr_0.p)
  49 cepoch      = 60000
  50 _, theta_opt, F_opt = copt.CGD(lr_0,step_center, cepoch, theta_c0)
  51 error_lr_0 = error(lr_0,theta_opt,F_opt)
  52
  53
  54 #### Run the optimization algorithms and compute the performance metrics
  55 x_asy_dagp, _, _, _, _, _ = \
  56     dopt.Asy_DAGP(T_active_exp2, Tv_nodes_exp2,  lr_0, zero_row_sum, zero_column_sum, step_asy_dagp, depoch, num_nodes, lr_0.p, rho, alpha, gamma, eta, neighbors, \
  57                     cons=False, delay_type='exp', min_delay=None, max_delay=None, expScale_delay=expScale, drop_msg=False, drop_prob=0.)
  58
  59 x_asyspa, _, _, _ = \
  60     dopt.Asy_SPA(T_active_exp2, Tv_nodes_exp2, lr_0, step_asy_spa, depoch, num_nodes, lr_0.p, N_out, neighbors, delay_type='exp', min_delay=None, max_delay=None,\
  61                    expScale_delay=expScale, decreasing_step_size=False, correct_step_size=True, drop_msg=False, drop_prob=0.)
  62
  63 x_appg, _, _, _ = \
  64     dopt.APPG(T_active_exp2, Tv_nodes_exp2, lr_0, step_size_appg, depoch, num_nodes, lr_0.p,  N_out, neighbors, \
  65                 delay_type='exp', min_delay=None, max_delay=None, expScale_delay=expScale, drop_msg=False, drop_prob=0.)
  66
  67
  68 res_F_asy_dagp = error_lr_0.cost_gap_path( np.sum(x_asy_dagp, axis = 1)/num_nodes)
  69 res_F_asyspa   = error_lr_0.cost_gap_path( np.sum(x_asyspa,   axis = 1)/num_nodes)
  70 res_F_appg     = error_lr_0.cost_gap_path( np.sum(x_appg,     axis = 1)/num_nodes)
  71
  72
  73 #### save data and plot results
  74 plot_exp2(T_active_exp2, res_F_asy_dagp, res_F_asyspa, res_F_appg, \
  75            current_dir=os.path.dirname(os.path.abspath(__file__)), save_results_folder='exp2')