def get_opt_prob(subfolder, bench_id, num_iter=None):
'''
get optimier and problem.
Args:
subfolder (str): subfolder with configuration files.
bench_id (int): benchmark id.
num_iter (None|int): load the data for specific iteration, None for random data.
Return:
climin.Minizer, Probdef: optimizer and problem.
'''
configfile = 'benchmarks/%s/config-sample.ini' % subfolder
config = load_config(configfile)
# modification to parameters
folder = os.path.dirname(configfile)
sys.path.insert(0, folder)
from config import modifyconfig_and_getnn
rbm = modifyconfig_and_getnn(config, bench_id)
optimizer, problem = pconfig(config, rbm)
# load saved data
if num_iter is not None:
data = np.load(folder + '/variables-%d%d.npy' % (bench_id, num_iter))
problem.rbm.set_variables(data)
return optimizer, problem
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 bverr(self, subfolder, bench_id, extension='png'):
'''show energy error function.'''
# get configuration and foler
configfile = 'benchmarks/%s/config-sample.ini' % subfolder
folder = os.path.dirname(configfile)
config = load_config(configfile)
# modification to parameters
sys.path.insert(0, folder)
from config import modifyconfig_and_getnn
rbm = modifyconfig_and_getnn(config, bench_id)
EG = config['hamiltonian']['EG']
if EG is None:
h = load_hamiltonian(model='J1J2',
J1=1,
J2=config['hamiltonian']['J2'],
size=config['hamiltonian']['size'])
H, EG, v, configs = analyse_exact(h, do_printsign=False, num_eng=1)
context = DShow(
(5, 3.5),
filename="notes/img/ERRL-%s-%s.%s" %
(subfolder, bench_id,
extension)) if extension is not '-' else contextlib.ExitStack()
with context:
show_el(datafiles=['%s/el-%i.dat' % (folder, bench_id)],
nsite=np.prod(config['hamiltonian']['size']),
EG=EG,
legends=['id = %s' % bench_id],
show_err=True,
xlogscale=False)
def bve(self, subfolder, bench_id, extension='png'):
'''show energy function.'''
# get configuration and foler
configfile = 'benchmarks/%s/config-sample.ini' % subfolder
folder = os.path.dirname(configfile)
config = load_config(configfile)
# modification to parameters
sys.path.insert(0, folder)
from config import modifyconfig_and_getnn
rbm = modifyconfig_and_getnn(config, bench_id)
EG = config['hamiltonian']['EG']
context = DShow(
(5, 3.5),
filename="notes/img/EL-%s-%s.%s" %
(subfolder, bench_id,
extension)) if extension is not '-' else contextlib.ExitStack()
with context:
show_el(datafiles=['%s/el-%i.dat' % (folder, bench_id)],
nsite=np.prod(config['hamiltonian']['size']),
EG=config['hamiltonian']['EG'],
legends=['id = %s' % bench_id],
show_err=False,
xlogscale=True)
def show_kernel_bentch(configfile, bentch_id):
config = load_config(configfile)
# folder to store data, containing config.py
folder = os.path.dirname(configfile)
# modification to parameters
sys.path.insert(0, folder)
from config import modifyconfig_and_getnn
rbm = modifyconfig_and_getnn(config, bentch_id)
show_kernel(rbm)
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 benchmark(self, configfile, bench_id, interactive=False):
from problems import load_config
config = load_config(configfile)
folder = os.path.dirname(configfile)
# modification to parameters
sys.path.insert(0, folder)
from config import modifyconfig_and_getnn
rbm = modifyconfig_and_getnn(config, bench_id)
n = config['mpi']['num_core']
logfile = '%s/log-%s.log' % (folder, bench_id)
options = ' '.join(['%s' % item for item in [configfile, bench_id]])
exec_code = 'python benchmark.py %s' % (options, )
if not interactive:
exec_code = 'nohup mpirun -n %s %s > %s &' % (n, exec_code,
logfile)
print('Excuting: %s' % exec_code)
os.system(exec_code)
def main():
from monitors import Print_eng_with_exact, print_eng, show_wf, DumpNetwork
configfile, bench_id = sys.argv[1:]
np.random.seed(3)
dconfig = load_config(configfile)
# folder to store data, containing config.py
folder = os.path.dirname(configfile)
# modification to parameters
sys.path.insert(0, folder)
from config import modifyconfig_and_getnn
rbm = modifyconfig_and_getnn(dconfig, int(bench_id))
e0 = dconfig['hamiltonian']['EG']
import datetime
print(datetime.datetime.now())
run_benchmark(dconfig, rbm, int(bench_id),\
monitors = [
Print_eng_with_exact(e0),\
DumpNetwork(folder=os.path.dirname(configfile),token=bench_id,step=1000)\
], folder=folder, viz=False)
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 show_bench44dncr():
'''Compare complex and real.'''
token = 'dn'
ids = [1, 6, 7]
show_err = True
plt.ion()
from problems import load_config
folder = '../benchmarks/wanglei4dn'
config = load_config(folder + '/config-sample.ini')
legends = []
for id in ids:
for version in ['c', 'r']:
show_el(datafiles=[
folder + '/el-%i%s.dat' % (id, 'c' if version == 'c' else '')
],
EG=config['hamiltonian']['EG'],
show_err=show_err,
xlogscale=not show_err,
smooth_step=20)
legends.append('%s-%s' % (version, id + 1))
plt.legend(legends)
plt.ylim(1e-5, 1)
pdb.set_trace()
plt.savefig('img/errcr-%s.png' % token)