print(train_size)
train_set = np.random.randn(train_size,dim)
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')
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()
# adapt every 10 timesteps using quadratic(after inital burn)
test.train_method = 'GQ'
test.adapt = 20 # Sets number of sub-cylcing steps
# do 100 steps
while test.iter < test.maxit:
test.step()
if test.iter % 20 == 0 and test.active is not None:
print('Step', test.iter, 'Active dimension', test.active.shape[1])
print('Subspace Distance', subspace_dist(true_as, test.active))
f2_avr += test.fhist
print('ASTARS trial', trial, ' minval', test.fhist[-1])
f_avr /= ntrials
f2_avr /= ntrials
plt.semilogy(f_avr, label='Stars')
plt.semilogy(f2_avr, label='Astars')
plt.legend()
plt.show()
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
dist = None
L1_hist = None
while test.iter < test.maxit:
test.step()
test2.step()
if test2.active is not None:
temp = np.array(subspace_dist(test2.active,f.active))
if dist is None:
dist = np.copy(temp)
L1_hist = np.copy(np.array(test2.L1))
else:
dist = np.append(dist,temp)
L1_hist = np.append(L1_hist,test2.L1)
#update average of f
f_avr += test.fhist
f2_avr += test2.fhist
trial_final += [email protected]@test2.x
#final answer
#project test2 solution
# adapt every f.adapt timesteps using quadratic(after inital burn)
test.train_method = 'GQ'
test.adapt = f.adapt # Sets retraining steps
#test.subcycle = True # turn on subcycling
test.threshadapt = True
#test.debug = True
test.regul = f.regul
test.threshold = f.threshold
# do 100 steps
while test.iter < test.maxit:
test.step()
if test.active is not None and test.iter // test.adapt:
print('distance', subspace_dist(test.active, f.active))
f2_avr += test.fhist
if maxit > test.tr_stop:
FAASTARS_adim_sto = np.zeros((maxit - test.tr_stop - 1, 1))
FAASTARS_sub_dist_sto = np.zeros((maxit - test.tr_stop - 1, 1))
# data dump
FAASTARS_f_sto = np.hstack(
(FAASTARS_f_sto, np.transpose([test.fhist])))
FAASTARS_x_sto = np.vstack((FAASTARS_x_sto, np.transpose(test.xhist)))
if maxit > test.tr_stop:
FAASTARS_adim_sto = np.hstack(
(FAASTARS_adim_sto, np.transpose([test.adim_hist])))