def run_ed_msr(J2, nsite):
from qstate.classifier.rules import marshall_sign_rule
h = load_hamiltonian('J1J2', size=(nsite, ), J2=J2)
H = h.get_mat()
e0, v0 = sps.linalg.eigsh(H, which='SA', k=1)
v0 = v0.ravel()
marshall_signs = marshall_sign_rule(h.configs)
plt.ion()
scatter_vec_phase(v0[marshall_signs == 1], color='r')
scatter_vec_phase(v0[marshall_signs == -1], color='b')
plt.legend([r'$+$', r'$-$'])
pdb.set_trace()
def __call__(self, problem, optimizer):
require_new_vv(problem, optimizer.n_iter)
require_e0v0(problem)
v0 = problem.cache['v0']
num_iter = optimizer.n_iter
vvi = 'vv-%s' % num_iter
vv = problem.cache[vvi]
plt.ion()
fig = gcf()
fig.set_figwidth(10, forward=True)
fig.set_figheight(5, forward=True)
fig.clear()
plt.subplot(121)
compare_wf(vv, v0)
plt.subplot(122)
scatter_vec_phase(vv, self.vv_pre)
plt.xlim(-self.D, self.D)
plt.ylim(-self.D, self.D)
plt.pause(0.01)
self.vv_pre = vv
def run_wanglei4(J2,
size,
optimize_method='adam',
momentum=0.,
do_plot_wf=True,
compare_to_exact=True,
learning_rate=1e-2):
from models.wanglei4 import WangLei4
# definition of a problem
h = load_hamiltonian('J1J2', size=size, J2=J2)
rbm = WangLei4(input_shape=size,
NF=8,
K=3,
num_features=[8],
version='conv',
itype='complex128')
# visualize network
from poornn import viznn
viznn(rbm, filename='data/%s.png' % rbm.__class__.__name__)
problem = ModelProbDef(hamiltonian=h,
rbm=rbm,
reg_method='delta',
sr_layerwise=False)
sr, rbm, vmc = problem.sr, problem.rbm, problem.vmc
vmc.inverse_rate = 0.05
optimizer = get_optimizer(wrt=rbm.get_variables(),
fprime=problem.compute_gradient,
optimize_method=optimize_method,
step_rate=learning_rate,
momentum=momentum)
# setup canvas
if do_plot_wf:
plt.ion()
fig = plt.figure(figsize=(10, 5))
# Exact Results
if compare_to_exact or compare_wf:
H, e0, v0, configs = analyse_exact(h, do_printsign=False)
el = [] # to store energy
vv_pre = None
print('\nRunning 0-th Iteration.')
for info in optimizer:
# `sampels` and `opq_vals` are cached!
ei = problem.cache['opq_vals'][0]
if do_plot_wf:
vv = rbm.tovec(mag=h.mag)
vv = vv / np.linalg.norm(vv)
fig.clear()
plt.subplot(121)
compare_wf(vv, v0)
plt.subplot(122)
scatter_vec_phase(vv, vv_pre)
D = 0.8
plt.xlim(-D, D)
plt.ylim(-D, D)
plt.pause(0.01)
vv_pre = vv
if compare_to_exact:
err = abs(e0 - ei) / (abs(e0) + abs(ei)) * 2
print('E/site = %s (%s), Error = %.4f%%' %
(ei / h.nsite, e0 / h.nsite, err * 100))
else:
print('E/site = %s' % (ei / h.nsite, ))
el.append(ei)
num_iter = info['n_iter']
#optimizer.step_rate *= 0.995
if num_iter >= 1000:
break
print('\nRunning %s-th Iteration.' % (num_iter + 1))
np.savetxt('data/el-%s%s.dat' % (h.nsite, 'p' if h.periodic else 'o'), el)
pdb.set_trace()
def run_target_sign(J2, nsite):
'''Given Sign train amplitude, the arbituary state version.'''
from models.wanglei import WangLei
# definition of a problem
h = load_hamiltonian('J1J2', size=(nsite, ), J2=J2)
rbm = WangLei(input_shape=(h.nsite, ),
version='linear',
use_conv=True,
itype='complex128')
problem = ModelProbDef(hamiltonian=h,
rbm=rbm,
reg_method='delta',
sr_layerwise=True)
sr, rbm, vmc = problem.sr, problem.rbm, problem.vmc
vmc.inverse_rate = 0.05
optimizer = get_optimizer(wrt=rbm.get_variables(),
fprime=problem.compute_gradient,
optimize_method='gd',
step_rate=1e-1,
momentum=momentum)
do_plot_wf = True
compare_to_exact = True
# setup canvas
if do_plot_wf:
plt.ion()
fig = plt.figure(figsize=(10, 5))
# Exact Results
if compare_to_exact or compare_wf:
H, e0, v0, configs = analyse_exact(h, do_printsign=False)
el = [] # to store energy
vv_pre = None
print('\nRunning 0-th Iteration.')
for info in optimizer:
# `sampels` and `opq_vals` are cached!
ei = problem.cache['opq_vals'][0]
if do_plot_wf:
vv = rbm.tovec(mag=h.mag)
vv = vv / np.linalg.norm(vv)
fig.clear()
plt.subplot(121)
compare_wf(vv, v0)
plt.subplot(122)
scatter_vec_phase(vv, vv_pre)
plt.xlim(-0.3, 0.3)
plt.ylim(-0.3, 0.3)
plt.pause(0.01)
vv_pre = vv
num_iter = info['n_iter']
#optimizer.step_rate *= 0.995
if num_iter >= 200:
break
print('\nRunning %s-th Iteration.' % (num_iter + 1))
np.savetxt('data/el-%s%s.dat' % (h.nsite, 'p' if h.periodic else 'o'), el)
pdb.set_trace()
def run_rtheta_switch(J2, nsite, rtheta_training_ratio, switch_step, momentum=0., \
do_plot_wf=True, compare_to_exact=True, do_check_sample=False):
from models.wanglei2 import WangLei2
# definition of a problem
h = load_hamiltonian('J1J2', size=(nsite, ), J1=1., J2=J2)
rbm = WangLei2(input_shape=(h.nsite, ),
num_feature_hidden=4,
use_msr=False,
theta_period=2,
with_linear=False,
itype='float64')
#rbm.thnn = get_exact_thnn4(fixed_var=True)
problem = ModelProbDef(hamiltonian=h, rbm=rbm, reg_method='sd')
sr, rbm, vmc = problem.sr, problem.rbm, problem.vmc
vmc.inverse_rate = 0.05
optimizer = get_optimizer(wrt=rbm.get_variables(),
fprime=problem.compute_gradient,
optimize_method='gd',
step_rate=3e-3,
momentum=momentum)
# setup canvas
if do_plot_wf or do_check_sample:
plt.ion()
fig = plt.figure(figsize=(10, 5))
# Exact Results
if compare_to_exact or compare_wf:
H, e0, v0, configs = analyse_exact(h, do_printsign=False, num_eng=5)
el = [] # to store energy
vv_pre = None
print('\nRunning 0-th Iteration.')
sr.rtheta_training_ratio = [rtheta_training_ratio[0], 0]
for info in optimizer:
# `sampels` and `opq_vals` are cached!
ei = problem.cache['opq_vals'][0]
if do_plot_wf or do_check_sample:
amp = []
signs = []
for config in h.configs:
amp.append(rbm.forward(config))
signs.append(rbm.get_sign(config))
amp = np.asarray(amp)
amp = amp / np.linalg.norm(amp)
vv = amp * signs
#vv = rbm.tovec(mag=h.mag)
if do_plot_wf:
fig.clear()
plt.subplot(121)
#compare_wf(amp, v0)
compare_wf(vv, v0)
plt.subplot(122)
scatter_vec_phase(vv, vv_pre)
plt.xlim(-0.8, 0.8)
plt.ylim(-0.8, 0.8)
plt.pause(0.01)
vv_pre = vv
if do_check_sample:
fig.clear()
check_sample(rbm, h, problem.cache['samples'])
plt.pause(0.01)
if compare_to_exact:
err = abs(e0 - ei) / (abs(e0) + abs(ei)) * 2
print('E/site = %s (%s), Error = %.4f%%' %
(ei / h.nsite, e0 / h.nsite, err * 100))
else:
print('E/site = %s' % (ei / h.nsite, ))
el.append(ei)
k = info['n_iter']
if k >= 800:
break
if k % (2 * switch_step) < switch_step:
print('\nRunning %s-th Iteration (optimize amplitudes).' % (k + 1))
sr.rtheta_training_ratio = [rtheta_training_ratio[0], 0]
else:
print('\nRunning %s-th Iteration (optimize signs).' % (k + 1))
sr.rtheta_training_ratio = [0, rtheta_training_ratio[1]]
np.savetxt('data/el-%s%s.dat' % (h.nsite, 'p' if h.periodic else 'o'), el)
pdb.set_trace()
def rbm_given_sign(J2, nsite):
from models.poorrbm import RBM
do_plot_wf = True
compare_to_exact = True
# definition of a problem
h = load_hamiltonian('J1J2', size=(nsite, ), J2=J2)
# Exact Results
if compare_to_exact or compare_wf:
H, e0, v0, configs = analyse_exact(h, do_printsign=False)
rbm = RBM(input_shape=(h.nsite, ),
num_feature_hidden=4,
itype='float64',
sign_func=sign_func_from_vec(h.configs, v0))
problem = ModelProbDef(hamiltonian=h,
rbm=rbm,
reg_method='delta',
sr_layerwise=False)
sr, rbm, vmc = problem.sr, problem.rbm, problem.vmc
vmc.inverse_rate = 0.05
optimizer = get_optimizer(wrt=rbm.get_variables(),
fprime=problem.compute_gradient,
optimize_method='gd',
step_rate=3e-2,
momentum=momentum)
# setup canvas
if do_plot_wf:
plt.ion()
fig = plt.figure(figsize=(10, 5))
el = [] # to store energy
vv_pre = None
print('\nRunning 0-th Iteration.')
for info in optimizer:
# `sampels` and `opq_vals` are cached!
ei = problem.cache['opq_vals'][0]
if do_plot_wf:
vv = rbm.tovec(mag=h.mag)
vv = vv / np.linalg.norm(vv)
fig.clear()
plt.subplot(121)
compare_wf(vv, v0)
plt.subplot(122)
scatter_vec_phase(vv, vv_pre)
plt.xlim(-0.3, 0.3)
plt.ylim(-0.3, 0.3)
plt.pause(0.01)
vv_pre = vv
if compare_to_exact:
err = abs(e0 - ei) / (abs(e0) + abs(ei)) * 2
print('E/site = %s (%s), Error = %.4f%%' %
(ei / h.nsite, e0 / h.nsite, err * 100))
else:
print('E/site = %s' % (ei / h.nsite, ))
el.append(ei)
num_iter = info['n_iter']
#optimizer.step_rate *= 0.995
if num_iter >= 2000:
break
print('\nRunning %s-th Iteration.' % (num_iter + 1))
np.savetxt('data/el-%s%s.dat' % (h.nsite, 'p' if h.periodic else 'o'), el)
pdb.set_trace()
def run_rtheta_toy(J2, nsite, version, rtheta_training_ratio, momentum=0.):
from models.wanglei2 import WangLei2
from models.toythnn import ToyTHNN
h = load_hamiltonian('J1J2', size=(nsite, ), J2=J2)
if version == '2l':
from models.psnn_leo import PSNN
thnn = PSNN((nsite, ),
nf=16,
batch_wise=False,
period=2,
output_mode='theta')
elif version == '1l':
from qstate.classifier import PSNN
thnn = PSNN((nsite, ),
batch_wise=False,
period=2,
output_mode='theta',
use_msr=False)
elif version == 'toy':
thnn = ToyTHNN(h)
# definition of a problem
H = h.get_mat()
rbm = get_ground_toynn(h, thnn=thnn, train_amp=True, theta_period=2)
problem = ModelProbDef(hamiltonian=h,
rbm=rbm,
reg_method='sd',
sr_layerwise=False if version == 'toy' else True)
sr, rbm, vmc = problem.sr, problem.rbm, problem.vmc
sr.rtheta_training_ratio = rtheta_training_ratio
optimizer = get_optimizer(wrt=rbm.get_variables(),
fprime=problem.compute_gradient,
optimize_method='gd',
step_rate=3e-3,
momentum=momentum)
do_plot_wf = True
compare_to_exact = True
# setup canvas
if do_plot_wf:
plt.ion()
fig = plt.figure(figsize=(10, 5))
# Exact Results
if compare_to_exact or compare_wf:
H, e0, v0, configs = analyse_exact(h, do_printsign=False)
el = [] # to store energy
vv_pre = None
print('\nRunning 0-th Iteration.')
for info in optimizer:
# `sampels` and `opq_vals` are cached!
ei = problem.cache['opq_vals'][0]
if do_plot_wf:
vv = rbm.tovec(mag=h.mag)
vv = vv / np.linalg.norm(vv)
plt.clf()
plt.subplot(121)
compare_wf(vv, v0)
plt.subplot(122)
scatter_vec_phase(vv, vv_pre)
plt.pause(0.01)
vv_pre = vv
if compare_to_exact:
err = abs(e0 - ei) / (abs(e0) + abs(ei)) * 2
print('E/site = %s (%s), Error = %.4f%%' %
(ei / h.nsite, e0 / h.nsite, err * 100))
else:
print('E/site = %s' % (ei / h.nsite, ))
el.append(ei)
if info['n_iter'] >= 300:
plt.savefig('data/SIGN-N%s-J2%s-%s.png' % (nsite, J2, version))
break
print('\nRunning %s-th Iteration.' % (info['n_iter'] + 1))
np.savetxt('data/el-%s%s.dat' % (h.nsite, 'p' if h.periodic else 'o'), el)
pdb.set_trace()
def run_rtheta(J2, nsite, rtheta_training_ratio, momentum=0.):
from models.wanglei2 import WangLei2
# definition of a problem
h = load_hamiltonian('J1J2', size=(nsite, ), J2=J2)
rbm = WangLei2(input_shape=(h.nsite, ),
num_feature_hidden=4,
use_msr=False,
theta_period=2,
with_linear=False,
itype='float64')
problem = ModelProbDef(hamiltonian=h, rbm=rbm, reg_method='sd')
sr, rbm, vmc = problem.sr, problem.rbm, problem.vmc
sr.rtheta_training_ratio = rtheta_training_ratio
optimizer = get_optimizer(wrt=rbm.get_variables(),
fprime=problem.compute_gradient,
optimize_method='gd',
step_rate=1e-2,
momentum=momentum)
do_plot_wf = True
compare_to_exact = True
do_check_sample = False
# setup canvas
if do_plot_wf or do_check_sample:
plt.ion()
fig = plt.figure(figsize=(10, 5))
# Exact Results
if compare_to_exact or compare_wf:
H, e0, v0, configs = analyse_exact(h, do_printsign=False, num_eng=10)
el = [] # to store energy
vv_pre = None
print('\nRunning 0-th Iteration.')
for info in optimizer:
# `sampels` and `opq_vals` are cached!
ei = problem.cache['opq_vals'][0]
if do_plot_wf or do_check_sample:
amps = []
thetas = []
signs = []
for config in h.configs:
amps.append(rbm.forward(config))
thetas.append(rbm.thnn.forward(config))
signs.append(np.exp(1j * thetas[-1]))
amps = np.asarray(amps)
amps = amps / np.linalg.norm(amps)
vv = amps * signs
#vv = rbm.tovec(mag=h.mag)
if do_plot_wf:
fig.clear()
plt.subplot(121)
#compare_wf(amps, v0)
compare_wf(vv, v0)
plt.subplot(122)
scatter_vec_phase(vv,
vv_pre,
winding=np.int32(
np.floor(np.array(thetas) / 2 / np.pi)))
plt.xlim(-0.8, 0.8)
plt.ylim(-0.8, 0.8)
plt.pause(0.01)
vv_pre = vv
if do_check_sample:
fig.clear()
check_sample(rbm, h, problem.cache['samples'])
plt.pause(0.01)
if compare_to_exact:
err = abs(e0 - ei) / (abs(e0) + abs(ei)) * 2
print('E/site = %s (%s), Error = %.4f%%' %
(ei / h.nsite, e0 / h.nsite, err * 100))
else:
print('E/site = %s' % (ei / h.nsite, ))
el.append(ei)
if info['n_iter'] >= 800:
break
print('\nRunning %s-th Iteration.' % (info['n_iter'] + 1))
np.savetxt('data/el-%s%s.dat' % (h.nsite, 'p' if h.periodic else 'o'), el)
pdb.set_trace()
