def toy_f(x, var=1E-2):
return x[0]**2 + var * np.random.randn(1)
init_pt = 5 * np.random.randn(20)
ntrials = 20
maxit = 250
f_avr = np.zeros(maxit + 1) #set equal to number of iterations + 1
for trial in range(ntrials):
#sim setup
test = Stars_sim(toy_f,
init_pt,
L1=2.0,
var=1E-4,
verbose=False,
maxit=maxit)
test.STARS_only = True
test.get_mu_star()
test.get_h()
# do 100 steps
while test.iter < test.maxit:
test.step()
#update average of f
f_avr += test.fhist
f2_avr = np.zeros(maxit + 1)
for trial in range(ntrials):
#np.random.seed(9)
init_pt = np.zeros(f.dim) #prior mean
#init_pt /= np.linalg.norm(init_pt)
ntrials = 3
maxit = 500
f_avr = np.zeros(maxit+1)
f2_avr = np.zeros(maxit+1)
trial_final = np.zeros(f.dim)
# STARS no sphere
for trial in range(ntrials):
#sim setup
test = Stars_sim(f, init_pt, L1 = None, var = None, verbose = False, maxit = maxit)
test.STARS_only = True
test.get_mu_star()
test.get_h()
test.update_L1 = True
test2 = Stars_sim(f, init_pt, L1 = None, var = None, verbose = True, maxit = maxit)
test2.update_L1 = True
#test.STARS_only = True
test2.get_mu_star()
test2.get_h()
test2.train_method = 'GQ'
test2.adapt = 2*f.dim
test2.regul = test2.var
test2.pad_train = 2.0
test2.explore_weight = 2.0
#test2.regul = None
for loop in range(7):
if loop != 0: #append new data
new_pts = np.random.randn(train_size,dim)
train_set = np.vstack((train_set,new_pts))
print('training data size',train_set.shape)
#train active subspace
f_data = toy_f(train_set)
print('data size', f_data.shape)
#don't normalize
sub_sp.compute(X=train_set,f=f_data,sstype='QPHD')
#usual threshold
adim = find_active(sub_sp.eigenvals,sub_sp.eigenvecs)
print('Subspace Distance',subspace_dist(true_as,sub_sp.eigenvecs[:,0:adim]))
test = Stars_sim(toy_f, init_pt, L1 = 2.0, var = 1E-4, verbose = False, maxit = train_size*3)
test.STARS_only = True
test.get_mu_star()
test.get_h()
# do 100 steps
while test.iter < test.maxit:
test.step()
if test.iter > (dim+2)*(dim+1)//4:
#compute active subspace
train_x = np.hstack((test.xhist[:,0:test.iter+1],test.yhist[:,0:test.iter]))
train_f = np.hstack((test.fhist[0:test.iter+1],test.ghist[0:test.iter]))
train_x = train_x.T
sub_sp.compute(X=train_x,f=train_f,sstype='QPHD')
adim = find_active(sub_sp.eigenvals,sub_sp.eigenvecs)
print('Subspace Distance',subspace_dist(true_as,sub_sp.eigenvecs[:,0:adim]))
#maxit random points as training data
train_set = np.random.randn(dim, maxit)
f_data = toy_f(train_set)
print('data size', f_data.shape)
#don't normalize
sub_sp.compute(X=train_set.T, f=f_data, sstype='QPHD')
#usual threshold
adim = find_active(sub_sp.eigenvals, sub_sp.eigenvecs)
print('Subspace Distance', subspace_dist(true_as, sub_sp.eigenvecs[:, 0:adim]))
for trial in range(ntrials):
#sim setup
test = Stars_sim(toy_f,
init_pt,
L1=2.0,
var=sigma**2,
verbose=False,
maxit=maxit)
test.STARS_only = True
test.get_mu_star()
test.get_h()
# do 100 steps
while test.iter < test.maxit:
test.step()
train_x = np.hstack(
(test.xhist[:, 0:test.iter + 1], test.yhist[:, 0:test.iter]))
train_f = np.hstack((test.fhist[0:test.iter + 1], test.ghist[0:test.iter]))
train_x = train_x.T
sub_sp.compute(X=train_x, f=train_f, sstype='QPHD')
adim = find_active(sub_sp.eigenvals, sub_sp.eigenvecs)
return mag * (np.dot(weights, x))**2 + sig * np.random.randn(1)
our_L1 = 2.0 * mag * dim
init_pt = np.random.randn(dim)
ntrials = 500
maxit = 200
f_avr = np.zeros(maxit + 1) #set equal to number of iterations + 1
for trial in range(ntrials):
#sim setup
test = Stars_sim(toy_f,
init_pt,
L1=our_L1,
var=our_var,
verbose=False,
maxit=maxit)
test.STARS_only = True
test.update_L1 = True
test.get_mu_star()
test.get_h()
# do 100 steps
while test.iter < test.maxit:
test.step()
#update average of f
f_avr += test.fhist
print('STARS trial', trial, ' minval', test.fhist[-1])
f2_avr = np.zeros(maxit + 1)
init_pt = np.random.randn(dim)
print(nesterov_2_f(init_pt))
ntrials = 10
maxit = 10000
f_avr = np.zeros(maxit + 1) #set equal to number of iterations + 1
for trial in range(ntrials):
#sim setup
test = Stars_sim(nesterov_2_f,
init_pt,
L1=4,
var=1E-12,
verbose=False,
maxit=maxit)
test.STARS_only = True
test.get_mu_star()
test.get_h()
# do 100 steps
while test.iter < test.maxit:
test.step()
#if test.iter % 100 == 0:
#print('iter',test.iter,test.fhist[test.iter])
#update average of f
f_avr += test.fhist
f2_avr = np.zeros(maxit + 1)
ntrials = 20
maxit = 600
dim = f.dim
f_avr = np.zeros(maxit+1)
# Start the clock!
start = timeit.default_timer()
# STARS
for trial in range(ntrials):
test = Stars_sim(f, this_init_pt, L1 = f.L1, var = f.var, verbose = False, maxit = maxit)
test.STARS_only = True
test.get_mu_star()
test.get_h()
while test.iter < test.maxit:
test.step()
#update average of f
f_avr += test.fhist
print('STARS trial',trial,' minval',test.fhist[-1])
a_dims = [2]
n_a_dims = np.size(a_dims)
f2_avr = np.zeros((maxit+1,n_a_dims))
j=0
dim = f.dim
#np.random.seed(9)
#init_pt = f.initscl*np.random.randn(dim)
init_pt = np.ones(dim)
ntrials = f.ntrials
maxit = f.maxit
f_avr = np.zeros(maxit+1)
f2_avr = np.zeros(maxit+1)
f3_avr = np.zeros(maxit+1)
f4_avr = np.zeros(maxit+1)
# STARS
for trial in range(ntrials):
#sim setup
test = Stars_sim(f, init_pt, L1 = f.L1, var = f.var, verbose = False, maxit = maxit)
test.STARS_only = True
test.get_mu_star()
test.get_h()
# STARS steps
while test.iter < test.maxit:
test.step()
#update average of f
f_avr += test.fhist
# FAASTARS (3 scenarios: no extensions, adaptive thresholding, and active subcycling)
for trial in range(ntrials):
#sim setup
#prinp.random.seed(9)
init_pt = np.ones(f.dim)
init_pt /= np.linalg.norm(init_pt)
ntrials = 30
maxit = 900
f_avr = np.zeros(maxit+1)
f_av2 = np.copy(f_avr)
# STARS, no weights
for trial in range(ntrials):
#sim setup
test = Stars_sim(f, init_pt, L1 = f.L1, var = f.var, verbose = False, maxit = maxit)
#test.STARS_only = True
test.get_mu_star()
test.get_h()
test.train_method = 'GQ'
test.threshold = .999
# do 100 steps
while test.iter < test.maxit:
test.step()
#update average of f
f_avr += test.fhist
# data dump
#print('initial first component',noisy_pred[:,0])
#print('current first component',pred[:,0])
return rms_loss(weights, pred, data)
#stars setup
maxit = 500
init_pt = np.hstack((init_weights, noisy_pred.flatten()))
ntrials = 1
f_avr = np.zeros(maxit + 1) #set equal to number of iterations + 1
for trial in range(ntrials):
#sim setup
test = Stars_sim(stars_wrapper,
init_pt,
L1=400.0,
var=1E-4,
verbose=False,
maxit=maxit)
test.STARS_only = True
test.get_mu_star()
test.get_h()
# do 100 steps
while test.iter < test.maxit:
test.step()
#update average of f
f_avr += test.fhist
print('STARS min', test.x[0:10])
f2_avr = np.zeros(maxit + 1)
f_avr = np.zeros(maxit + 1)
#initialize storage for data dump
STARS_f_sto = np.zeros((maxit + 1, 1))
STARS_x_sto = np.zeros((1, dim))
ASTARS_f_sto = np.zeros((maxit + 1, 1))
ASTARS_x_sto = np.zeros((1, dim))
FAASTARS_f_sto = np.zeros((maxit + 1, 1))
FAASTARS_x_sto = np.zeros((1, dim))
# STARS
for trial in range(ntrials):
#sim setup
test = Stars_sim(f,
init_pt,
L1=f.L1,
var=f.var,
verbose=False,
maxit=maxit)
test.STARS_only = True
test.get_mu_star()
test.get_h()
# do 100 steps
while test.iter < test.maxit:
test.step()
#update average of f
f_avr += test.fhist
# data dump
#STARS_f_sto = np.hstack((STARS_f_sto, np.transpose([test.fhist])))
#STARS_x_sto = np.vstack((STARS_x_sto,np.transpose(test.xhist)))
#test rbf on random normal samples
lin_samp = np.random.normal(size=(15,10))
f_data = testfun(lin_samp.T)
ss,lin_surrog=train_rbf(lin_samp,f_data,noise=1E-4)
print('Active subspace from linear rbf',ss.eigenvecs)
quad_samp = np.random.normal(size=(90,10))
fq_data = testfun(quad_samp.T)
ss2,quad_surrog = train_rbf(quad_samp,fq_data,noise=1E-4)
init_pt=np.random.rand(10)
stars_test=Stars_sim(testfun,init_pt,L1=2.0,var=1E-2,maxit=80)
stars_test.get_mu_star()
stars_test.get_h()
while stars_test.iter < stars_test.maxit:
stars_test.step()
stars_test.compute_active()
print('Active Variables after STARS run',stars_test.active)
print('Active Weights',stars_test.wts)
plt.semilogy(stars_test.fhist)
plt.figure()
plt.plot(stars_test.xhist[0,:])
plt.plot(stars_test.xhist[-1,:])
"""
import numpy as np
import astars
print(dir(astars))
from astars.stars_sim import Stars_sim
def testfun(x):
return x[0]**2 + x[-1]**2 + np.random.normal(scale=.01)
test = True
for trials in range(99):
init_pt = np.random.rand(10, 1)
stars_test = Stars_sim(testfun, init_pt, L1=2.0, var=1E-2)
stars_test.get_mu_star()
stars_test.get_h()
while stars_test.iter < stars_test.maxit:
stars_test.STARS_step()
#Error Bound for additive noise
error_bound = (4 * stars_test.L1 * (stars_test.dim + 4) /
(101) * np.linalg.norm(init_pt)**2 +
3 * np.sqrt(2) / 5 * np.sqrt(stars_test.var) *
(stars_test.dim + 4))
err = np.mean(stars_test.fhist)
if err > error_bound:
print('Unit test failed', err)
test = False
#init_pt[6]+=.08
#init_pt[7]+=2.5
#init_pt[8]+=1700
#init_pt[9]+=.025
print(init_pt)
ntrials = 2
maxit = 200
f_avr = np.zeros(maxit+1) #set equal to number of iterations + 1
for trial in range(ntrials):
#sim setup
test = Stars_sim(wing_barrier, init_pt, L1 = 200, var = 1E-4, verbose = True, maxit = maxit)
test.STARS_only = True
test.debug = True
test.get_mu_star()
test.get_h()
# do 100 steps
while test.iter < test.maxit:
test.step()
#if np.isnan(test.x).any:
# print(test.xhist[:,0:test.iter+1],test.yhist[:,0:test.iter+1],test.fhist[0:test.iter+1],test.ghist[0:test.iter+1])
# print(test.x)
# raise SystemExit('nan in current iterate')
#update average of f
f_avr += test.fhist
init_pt = np.ones(dci.dim) #prior mean
#init_pt /= np.linalg.norm(init_pt)
ntrials = 1
maxit = 100
f_avr = np.zeros(maxit + 1)
f2_avr = np.zeros(maxit + 1)
f3_avr = np.zeros(maxit + 1)
trial_final = np.zeros(dci.dim)
# STARS no sphere
for trial in range(ntrials):
#sim setup
test = Stars_sim(dci2,
init_pt,
L1=dci2.L1,
var=dci2.var,
verbose=False,
maxit=maxit)
test.STARS_only = True
test.get_mu_star()
test.get_h()
test2 = Stars_sim(dci,
init_pt,
f_obj=dci,
L1=dci.L1,
var=dci.var,
verbose=True,
maxit=maxit)
#test.STARS_only = True
test2.get_mu_star()
test2.get_h()
STARS_f_sto = np.zeros((maxit + 1, ntrials))
STARS_x_sto = np.zeros((1, dim))
STARS_L1_sto = np.zeros((maxit + 1, ntrials))
STARS_var_sto = np.zeros(ntrials)
FAASTARS_f_sto = np.zeros((maxit + 1, ntrials))
FAASTARS_x_sto = np.zeros((1, dim))
FAASTARS_L1_sto = np.zeros((maxit + 1, ntrials))
FAASTARS_var_sto = np.zeros(ntrials)
for trial in range(ntrials):
#sim setup
test = Stars_sim(f,
init_pt,
L1=None,
var=None,
verbose=False,
maxit=maxit,
true_as=f.active,
train_method='GQ')
test.STARS_only = True
print('Inital L1', test.L1)
print('Inital var', test.var)
test.update_L1 = True
test.get_mu_star()
test.get_h()
# do training steps
while test.iter < test.tr_stop:
test.step()
#update average of f and save for start of astars...?
