def synthetic_run_params(m, w, n, p, gamma, sigma_step, sigma, tol_cg,
MAXIT_cg, DENSITY, power):
# Data LOAD
X_train, y_train, X_test, y_test, w_true = gen_data.gen_data_SVM(
m, n, p, DENSITY)
m, N = np.shape(X_train)
n = 2 * N + 1
# form l1-SVM constraint matrix
A, b, c = svm_func.formL1SVM(X_train, y_train)
# IPM Standard
t_stan_1 = time.time()
x, y, s, iter_out_stan, iter_in_cg_vec_stan, kap_AD_vec, time_ls_stan = ipm_func.ipm_standard(
m, n, A, b, c, tol_cg)
t_stan_2 = time.time()
# IPM Ours
t_ipm_1 = time.time()
x, y, s, iter_out_ipm, iter_in_cg_vec_ipm, kap_ADW_vec, v_vec, time_ls_ipm = ipm_func.ipm(
m, n, w, A, b, c, tol_cg)
t_ipm_2 = time.time()
results = dict()
results['ipm_time'] = t_ipm_2 - t_ipm_1
results['stan_time'] = t_stan_2 - t_stan_1
results['niter_out_stan'] = iter_out_stan
results['niter__out_ipm'] = iter_out_ipm
results['niter_inner_stan'] = iter_in_cg_vec_stan
results['niter_inner_cg'] = iter_in_cg_vec_ipm
results['inner_cg_vec'] = iter_in_cg_vec_stan
results['kap_AD_vec'] = kap_AD_vec
results['kap_ADW_vec'] = kap_ADW_vec
results['v_vec'] = v_vec
return results
t_stan_1 = time.time()
x, y, s, iter_out_stan, iter_in_cg_vec_stan, kap_AD_vec, time_ls_stan = ipm_func.ipm_standard(
m, n, A, b, c, tol_cg)
t_stan_2 = time.time()
p_ipm_vec_stan = np.dot(c, x)
d_ipm_vec_stan = np.dot(b, y)
print('\niter stan out = ', iter_out_stan)
print('p = ', p_ipm_vec_stan)
print('d = ', d_ipm_vec_stan)
#------------------------------------------------------------
# IPM Ours
t_ipm_1 = time.time()
x, y, s, iter_out_ipm, iter_in_cg_vec_ipm, kap_ADW_vec, v_vec, time_ls_ipm = ipm_func.ipm(
m, n, w, A, b, c, tol_cg)
t_ipm_2 = time.time()
p_ipm_final = np.dot(c, x)
d_ipm_final = np.dot(b, y)
v_vec_mean = np.mean(v_vec)
print('\nv_vec_mean = ', v_vec_mean)
#------------------------------------------------------------
print(
'\n-------------------------- \nSVM Results: \n--------------------------'
)
# w_approx_ipm = x[0:N] - x[N:2*N]
# w_approx_cvx = np.transpose(np.squeeze(x_cvx[0:N] - x_cvx[N:2*N]))
# print 'x = \n',x[0:5]
if __name__ == '__main__':
# X_train, y_train, X_test, y_test, w_true = gen_data.gen_data_SVM(m,n,p,DENSITY)
X_train, y_train, X_test, y_test = load_real_data.load_ARCENE()
m, N = np.shape(X_train)
n = 2 * N + 1
# form l1-SVM constraint matrix
A, b, c = svm_func.formL1SVM(X_train, y_train)
MAXIT_cg_func = lambda k: 10
tol_ipm_mu = pow(10, -1 * 6)
w = 150
feasible_point = feasible_start(A, b, c)
zeta = 10000
s = zeta * np.ones(n)
y = np.zeros(m)
feasible_point = (feasible_point[0], y, s)
x,y,s,iter_out_ipm,iter_in_cg_vec_ipm,kap_ADW_vec,v_vec,time_ls_ipm, mu_vec_infe \
= ipm_func.ipm(m,n,w,A,b,c,tol_cg)
x,y,s,iter_out_ipm,iter_in_cg_vec_ipm,kap_ADW_vec,v_vec,time_ls_ipm, mu_vec_fe \
= ipm_func.ipm(m,n,w,A,b,c,tol_cg, start = feasible_point)
t_stan_1 = time.time() x,y,s,iter_out_stan,iter_in_cg_vec_stan,kap_AD_vec,time_ls_stan = ipm_func.ipm_standard(m,n,A,b,c,tol_cg) t_stan_2 = time.time() p_ipm_vec_stan = np.dot(c,x) d_ipm_vec_stan = np.dot(b,y) print '\niter stan out = ', iter_out_stan print 'p = ',p_ipm_vec_stan print 'd = ',d_ipm_vec_stan print 'test = ', iter_out_stan #------------------------------------------------------------ # IPM Sketch (Ours) w = w_vec[0] x,y,s,iter_out_ipm_2,iter_in_cg_vec_ipm_2,kap_ADW_vec_2,v_vec_2,time_ls_ipm = ipm_func.ipm(m,n,w,A,b,c,tol_cg) w = w_vec[1] x,y,s,iter_out_ipm_5,iter_in_cg_vec_ipm_5,kap_ADW_vec_5,v_vec_5,time_ls_ipm = ipm_func.ipm(m,n,w,A,b,c,tol_cg) w = w_vec[2] x,y,s,iter_out_ipm_8,iter_in_cg_vec_ipm_8,kap_ADW_vec_8,v_vec_8,time_ls_ipm = ipm_func.ipm(m,n,w,A,b,c,tol_cg) #------------------------------------------------------------ print '\n-------------------------- \nResults: \n--------------------------' #print 'time ipm = ', t_ipm_2 - t_ipm_1, ' secs' #print '\niter stan out = ', iter_out_stan #print 'iter ipm out = ', iter_out_ipm #print '\niter stan in cg = \n', iter_in_cg_vec_stan
def loop_w_tolcg(X_train,y_train,w_vec,tol_cg_vec,data_desc):
# loop over various (w, tol_cg) pairs.
# save the results as text files, then plot. (can be re-ploted later)
#------------------------------------------------------------
t_total_1 = time.time()
#------------------------------------------------------------
# form l1-SVM constraint matrix
A,b,c = svm_func.formL1SVM(X_train,y_train)
#------------------------------------------------------------
# Parameters
# get dimensions
m,N = np.shape(X_train)
n = 2*N+1
print '---------------------------', '\n'
print '(m x n) = (', m, 'x', n, ')'
print '\n', '---------------------------'
#------------------------------------------------------------
# CVX LP
#t_cvx_1 = time.time()
#x_cvx, p_cvx = svm_func.run_CVXPY_LP(A,b,c)
#t_cvx_2 = time.time()
#print'\np* cvx = ', p_cvx
#print 'time cvx = ', t_cvx_2 - t_cvx_1, ' secs'
#------------------------------------------------------------
# Loop over w and tol_cg
num_w = len(w_vec)
num_tol_cg = len(tol_cg_vec)
v_norm = np.zeros((num_w,num_tol_cg))
it_ipm = np.zeros((num_w,num_tol_cg)) # outer iterations ours
it_sta = np.zeros(num_tol_cg) # outer iterations standard
it_cg_ipm = np.zeros((num_w,num_tol_cg)) # inner CG iterations ours
it_cg_sta = np.zeros(num_tol_cg) # inner CG iterations standard
kap_ipm = np.zeros((num_w,num_tol_cg)) # condi #
kap_sta = np.zeros(num_tol_cg) # condi #
#er_rel = np.zeros((num_w,num_tol_cg))
#------------------------------------------------------------
for ind_tol_cg in range(num_tol_cg):
tol_cg = tol_cg_vec[ind_tol_cg]
#------------------------------------------------------------
# IPM Standard
t_ls_stan = 0
x,y,s,t_iter_stan,iter_in_cg_stan_vec,kap_AD_vec,time_ls = ipm_func.ipm_standard(m,n,A,b,c,tol_cg)
#x,y,s,t_iter_stan,t_ls_stan = ipm_func.ipm_standard(m,n,A,b,c,0)
#it_sta = t_iter_stan*np.ones((num_w,num_tol_cg))
it_sta[ind_tol_cg] = t_iter_stan
it_cg_sta[ind_tol_cg] = np.ndarray.max(iter_in_cg_stan_vec)
kap_sta[ind_tol_cg] = np.ndarray.max(kap_AD_vec)
p_ipm_vec_stan = np.dot(c,x)
d_ipm_vec_stan = np.dot(b,y)
print 'p stan = ',p_ipm_vec_stan
print 'd stan = ',d_ipm_vec_stan
#------------------------------------------------------------
for ind_w in range(num_w):
w = w_vec[ind_w]
#------------------------------------------------------------
# IPM Ours
x,y,s,iter_out_ipm,iter_in_cg_ipm_vec,kap_ADW_vec,v_vec,time_ls = ipm_func.ipm(m,n,w,A,b,c,tol_cg)
#x,y,s,t_iter_ipm,t_ls,iter_cg_ipm,v_vec = ipm_func.ipm(m,n,w,A,b,c,tol_cg)
p_ipm_final = np.dot(c,x)
d_ipm_final = np.dot(b,y)
print 'p ipm = ',p_ipm_final
print 'd ipm = ',d_ipm_final
v_vec_mean = np.mean(v_vec)
#------------------------------------------------------------
# post process
#diff = np.transpose(x_cvx) - x
#error_rel = 100*np.linalg.norm(diff,2)/np.linalg.norm(x_cvx,2)
#------------------------------------------------------------
# Record
v_norm[ind_w,ind_tol_cg] = v_vec_mean
it_ipm[ind_w,ind_tol_cg] = iter_out_ipm
it_cg_ipm[ind_w,ind_tol_cg] = np.ndarray.max(iter_in_cg_ipm_vec)
kap_ipm[ind_w,ind_tol_cg] = np.ndarray.max(kap_ADW_vec)
#er_rel[ind_w,ind_tol_cg] = error_rel
#------------------------------------------------------------
#------------------------------------------------------------
#------------------------------------------------------------
#------------------------------------------------------------
print '\n-------------------------- \nResults: \n--------------------------'
#print 'time ipm = ', t_ipm_2 - t_ipm_1, ' secs'
#print 'time cvx = ', t_cvx_2 - t_cvx_1, ' secs'
print 'w_vec = ',w_vec
print 'tol_cg_vec = ',tol_cg_vec
print '\nv_norm = \n', v_norm
print 'iter ipm = \n', it_ipm
print 'iter sta = \n', it_sta
print 'it_cg_ipm = \n', it_cg_ipm
print 'it_cg_sta = \n', it_cg_sta
print 'kap_ipm = \n', kap_ipm
print 'kap_sta = \n', kap_sta
t_total_2 = time.time()
print 'time total = ', t_total_2 - t_total_1, ' secs'
#------------------------------------------------------------
np.savetxt(directory+'ipm_out/w_vec_'+str(data_desc)+'_m'+str(m)+'_n'+str(n)+'.txt', w_vec) # int
np.savetxt(directory+'ipm_out/w_vec_'+str(data_desc)+'_m'+str(m)+'_n'+str(n)+'.txt', w_vec, fmt='%i') # int
np.savetxt(directory+'ipm_out/tol_cg_vec_'+str(data_desc)+'_m'+str(m)+'_n'+str(n)+'.txt', tol_cg_vec, fmt='%.10f')
np.savetxt(directory+'ipm_out/v_norm_'+str(data_desc)+'_m'+str(m)+'_n'+str(n)+'.txt', v_norm, fmt='%.10f')
np.savetxt(directory+'ipm_out/it_ipm_'+str(data_desc)+'_m'+str(m)+'_n'+str(n)+'.txt', it_ipm, fmt='%i') # int
np.savetxt(directory+'ipm_out/it_sta_'+str(data_desc)+'_m'+str(m)+'_n'+str(n)+'.txt', it_sta, fmt='%i') # int
np.savetxt(directory+'ipm_out/it_cg_ipm_'+str(data_desc)+'_m'+str(m)+'_n'+str(n)+'.txt', it_cg_ipm, fmt='%i') # int
np.savetxt(directory+'ipm_out/it_cg_sta_'+str(data_desc)+'_m'+str(m)+'_n'+str(n)+'.txt', it_cg_sta, fmt='%i') # int
np.savetxt(directory+'ipm_out/kap_ipm_'+str(data_desc)+'_m'+str(m)+'_n'+str(n)+'.txt', kap_ipm, fmt='%i') # int
np.savetxt(directory+'ipm_out/kap_sta_'+str(data_desc)+'_m'+str(m)+'_n'+str(n)+'.txt', kap_sta, fmt='%i') # int
#np.savetxt(directory+'ipm_out/er_rel_'+str(data_desc)+'_m'+str(m)+'_n'+str(n)+'.txt', er_rel, fmt='%.10f')
# take all the data; not subset of the data (might want to take a subset later)
#ind_w_subset = range(num_w)
#ind_tol_subset = range(num_tol_cg)
num_w = len(w_vec)
num_tol_cg = len(tol_cg_vec)
ind_w_subset = range(num_w)
ind_tol_subset = range(num_tol_cg)