def analyse_weights2(configfile, bench_id, num_iter):
config = load_config(configfile)
# folder to store data, containing config.py
folder = os.path.dirname(configfile)
e0 = config['hamiltonian']['EG']
sys.path.insert(0, folder)
from config import modifyconfig_and_getnn
rbm = modifyconfig_and_getnn(config, bench_id)
optimizer, problem = pconfig(config, rbm)
h, sr, rbm, vmc = problem.hamiltonian, problem.sr, problem.rbm, problem.vmc
# load data
data = np.load(folder + '/variables-%d%d.npy' % (bench_id, num_iter))
rbm.set_variables(data)
# compute overlap
H, e0, v0, configs = analyse_exact(h, do_printsign=False)
vec = rbm.tovec(mag=h.mag)
vec = vec / np.linalg.norm(vec)
overlap = abs(vec.T.conj().dot(v0))
print('overlap = %s' % overlap)
# show weights in conv layers.
keys = [1, 2, 4, 5, 6, 7]
# configs, config_indexer = h.get_config_table()
#configs = np.array([[1,-1]*(h.nsite//2)])
#configs = np.array([[1,-1,-1,1]*(h.nsite//4)])
plt.ion()
num_try = 100
for i in range(num_try):
config = configs[np.random.choice(len(configs))]
plt.clf()
ys = rbm.forward(config, full_output=True)
for ik, k in enumerate(keys):
ax = plt.subplot('61%d' % (ik + 1),
sharex=ax if ik < 4 and ik > 0 else None)
if k == 7:
vmin, vmax = 0, 5e5
elif k == 5:
vmin, vmax = 0, 2e5
elif k == 6:
vmin, vmax = 0, 5e5
elif k == 4:
vmin, vmax = 0, 5
else:
vmin, vmax = None, None
if ik < 2:
plt.pcolor(np.atleast_2d(np.squeeze(ys[k]).real),
vmin=vmin,
vmax=vmax)
else:
plt.pcolor(np.atleast_2d(np.squeeze(abs(ys[k]))),
vmin=vmin,
vmax=vmax)
plt.colorbar()
plt.pause(0.01)
pdb.set_trace()
def analyse_psiscaling(J2, size):
h = load_hamiltonian('J1J2', size=size, J2=J2)
configs, indexer = h.get_config_table()
H, e0, v0, configs = analyse_exact(h, do_printsign=False, num_eng=1)
rand = np.random.randn(len(v0))
func_list = [
lambda x: x**3, lambda x: x**5, lambda x: np.sinh(x),
lambda x: np.cosh(x)
]
labels = ['x^3', 'x^5', 'sinh(x)', 'cosh(x)']
x = np.arange(1, len(v0) + 1)
amp = abs(v0)
amp = np.sort(amp)[::-1]
plt.ion()
#plt.plot(np.log(amp))
plt.plot(x, amp)
for func in func_list:
amp2 = abs(func(rand))
amp2 = np.sort(amp2)[::-1] / np.linalg.norm(amp2)
plt.plot(x, amp2)
#plt.xscale('log')
plt.legend(['$v$'] + labels)
pdb.set_trace()
def show_prod(J2, size):
h = load_hamiltonian('J1J2', size=size, J2=J2)
H = h.get_mat()
c0, indexer = h.get_config_table()
e0, v0 = sps.linalg.eigsh(H, which='SA', k=1)
v0 = v0.ravel()
order = np.argsort(abs(v0))[::-1]
v0 = v0[order]
c0 = c0[order]
amp = np.abs(v0) * 100
nn = np.sum(c0 * np.roll(c0, 1, axis=1), axis=1)
nnn = np.sum(c0 * np.roll(c0, 2, axis=1), axis=1)
hndim = len(c0)
#wlist = np.linspace(0,hndim,2000)
#nn = dos(np.arange(hndim), wlist=wlist, weights=nn, eta=2.)*hndim
plt.ion()
plt.plot(np.arange(hndim), amp)
plt.fill_between(np.arange(hndim), nn, 0, alpha=0.5, color='r')
plt.fill_between(np.arange(hndim), nnn, 0, alpha=0.5, color='g')
plt.legend([
'$|\psi|$', r'$\sum_iS^z_i\times S^z_{i+1}$',
r'$\sum_iS^z_i\times S^z_{i+2}$'
])
plt.axhline(0, color='#999999')
pdb.set_trace()
plt.savefig('notes/img/prod_J2%sN%s.png' % (J2, size), dpi=300)
def analyse_weights(subfolder, bench_id, num_iter):
optimizer, problem = get_opt_prob(subfolder, bench_id, num_iter=num_iter)
h, sr, rbm, vmc = problem.hamiltonian, problem.sr, problem.rbm, problem.vmc
# compute overlap
H, e0, v0, configs = analyse_exact(h, do_printsign=False)
vec = rbm.tovec(mag=h.mag)
vec = vec / np.linalg.norm(vec)
overlap = abs(vec.T.conj().dot(v0))
print('overlap = %s' % overlap)
# show weights in conv layers.
convlayer = rbm.layers[2]
data = convlayer.weight #.sum(axis=0, keepdims=True)
plt.ion()
plt.imshow(np.real(
abs(data).sum(axis=0,
keepdims=True).reshape([-1,
convlayer.weight.shape[-1]])),
vmin=-0.5,
vmax=0.5)
nfo, nfi = data.shape[:2]
poses = range(nfo * nfi)
plt.yticks(poses, [(pos // nfi, pos % nfi) for pos in poses])
plt.ylabel('out, in')
plt.colorbar()
pdb.set_trace()
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 analyse_weights2(configfile, bench_id, num_iter):
optimizer, problem = get_opt_prob(subfolder, bench_id, num_iter=num_iter)
h, sr, rbm, vmc = problem.hamiltonian, problem.sr, problem.rbm, problem.vmc
# compute overlap
H, e0, v0, configs = analyse_exact(h, do_printsign=False)
vec = rbm.tovec(mag=h.mag)
vec = vec / np.linalg.norm(vec)
overlap = abs(vec.T.conj().dot(v0))
print('overlap = %s' % overlap)
# show weights in conv layers.
keys = [1, 2, 4, 5, 6, 7]
# configs, config_indexer = h.get_config_table()
#configs = np.array([[1,-1]*(h.nsite//2)])
#configs = np.array([[1,-1,-1,1]*(h.nsite//4)])
plt.ion()
num_try = 100
for i in range(num_try):
config = configs[np.random.choice(len(configs))]
plt.clf()
ys = rbm.forward(config, full_output=True)
for ik, k in enumerate(keys):
ax = plt.subplot('61%d' % (ik + 1),
sharex=ax if ik < 4 and ik > 0 else None)
if k == 7:
vmin, vmax = 0, 5e5
elif k == 5:
vmin, vmax = 0, 2e5
elif k == 6:
vmin, vmax = 0, 5e5
elif k == 4:
vmin, vmax = 0, 5
else:
vmin, vmax = None, None
if ik < 2:
plt.pcolor(np.atleast_2d(np.squeeze(ys[k]).real),
vmin=vmin,
vmax=vmax)
else:
plt.pcolor(np.atleast_2d(np.squeeze(abs(ys[k]))),
vmin=vmin,
vmax=vmax)
plt.colorbar()
plt.pause(0.01)
pdb.set_trace()
def analyse_poly(configfile, num_iter, bench_id_list):
config = load_config(configfile)
# folder to store data, containing config.py
folder = os.path.dirname(configfile)
e0 = config['hamiltonian']['EG']
labels = [
'polynomial', 'legendre', 'hermite', 'chebyshev', 'laguerre',
'hermiteE'
]
legends = []
sys.path.insert(0, folder)
from config import modifyconfig_and_getnn
plt.ion()
for ib, (bench_id, label) in enumerate(zip(bench_id_list, labels)):
rbm = modifyconfig_and_getnn(config, bench_id)
optimizer, problem = pconfig(config, rbm)
h, sr, rbm, vmc = problem.hamiltonian, problem.sr, problem.rbm, problem.vmc
if ib == 0:
H, e0, v0, configs = analyse_exact(h, do_printsign=False)
# load data
data = np.load(folder + '/variables-%d%d.npy' % (bench_id, num_iter))
rbm.set_variables(data)
# compute overlap
vec = rbm.tovec(mag=h.mag)
vec = vec / np.linalg.norm(vec)
overlap = abs(vec.T.conj().dot(v0))
print('overlap = %s' % overlap)
# show weights in conv layers.
for layer in rbm.layers:
if hasattr(layer, 'kernel_dict'):
polylayer = layer
data = polylayer.get_variables()
plt.plot(np.real(data))
legends.append('%s (%.4f)' % (label, overlap))
plt.axhline(y=0, ls='--', color='k')
plt.legend(legends)
pdb.set_trace()
plt.savefig('notes/img/polyweight.png')
def analyse_weights(configfile, bench_id, num_iter):
config = load_config(configfile)
# folder to store data, containing config.py
folder = os.path.dirname(configfile)
e0 = config['hamiltonian']['EG']
sys.path.insert(0, folder)
from config import modifyconfig_and_getnn
rbm = modifyconfig_and_getnn(config, bench_id)
optimizer, problem = pconfig(config, rbm)
h, sr, rbm, vmc = problem.hamiltonian, problem.sr, problem.rbm, problem.vmc
# load data
data = np.load(folder + '/variables-%d%d.npy' % (bench_id, num_iter))
rbm.set_variables(data)
# compute overlap
H, e0, v0, configs = analyse_exact(h, do_printsign=False)
vec = rbm.tovec(mag=h.mag)
vec = vec / np.linalg.norm(vec)
overlap = abs(vec.T.conj().dot(v0))
print('overlap = %s' % overlap)
# show weights in conv layers.
convlayer = rbm.layers[2]
data = convlayer.weight #.sum(axis=0, keepdims=True)
plt.ion()
plt.imshow(np.real(
abs(data).sum(axis=0,
keepdims=True).reshape([-1,
convlayer.weight.shape[-1]])),
vmin=-0.5,
vmax=0.5)
nfo, nfi = data.shape[:2]
poses = range(nfo * nfi)
plt.yticks(poses, [(pos // nfi, pos % nfi) for pos in poses])
plt.ylabel('out, in')
plt.colorbar()
pdb.set_trace()
def analyse_polycurve(configfile,
num_iter,
bench_id_list,
show_var=False,
token=''):
config = load_config(configfile)
# folder to store data, containing config.py
folder = os.path.dirname(configfile)
e0 = config['hamiltonian']['EG']
legends = []
sys.path.insert(0, folder)
from config import modifyconfig_and_getnn
x = np.linspace(-2, 2, 200)
plt.ion()
for bench_id in bench_id_list:
rbm = modifyconfig_and_getnn(config, bench_id)
# show weights in conv layers.
for nit in np.atleast_1d(num_iter):
rbm_var = np.load(folder + '/variables-%d%d.npy' % (bench_id, nit))
rbm.set_variables(rbm_var)
for layer in rbm.layers:
if hasattr(layer, 'kernel_dict'):
polylayer = layer
legends.append(layer.kernel + '-%d' % nit)
if show_var:
data = polylayer.get_variables()
plt.plot(data.real)
else:
data = polylayer.forward(x)
plt.plot(x, data.real)
plt.legend(legends)
#plt.ylim(-20,20)
pdb.set_trace()
if show_var:
plt.savefig('notes/img/polyvar-%s.png' % token)
else:
plt.savefig('notes/img/polycurve-%s.png' % token)
def scale_ed_msr(size, J2MIN=0, J2MAX=1, NJ2=51, yscale='log'):
from qstate.classifier.rules import marshall_sign_rule
J2L = np.linspace(J2MIN, J2MAX, NJ2)
e0l, el = [], []
for i, J2 in enumerate(J2L):
h = load_hamiltonian('J1J2', size=size, J2=J2)
H = h.get_mat()
e0, v0 = sps.linalg.eigsh(H, which='SA', k=1)
v0 = v0.ravel()
configs, config_indexer = h.get_config_table()
marshall_signs = marshall_sign_rule(configs, size=size)
v = abs(v0) * marshall_signs
el.append(v.T.conj().dot(H.dot(v)))
e0l.append(e0.item())
print('%s' % i)
np.savetxt('notes/data/scale_msr_%s.dat' % (size, ), list(zip(el, e0l)))
plt.ion()
plt.plot(J2L, np.array(el) - e0l)
plt.xlabel(r'$J_2$')
plt.ylabel(r'$E-E_0$')
plt.yscale(yscale)
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_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()
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()