def main(args):
mc_small = read_file(L=args.L // 2, n_samples=args.nTE)
mc_large = read_file(L=args.L, n_samples=args.nTE)
set_GPU_memory(fraction=args.GPU)
model = ModelLoader(list_ind=args.Mind)
if args.TEST > args.nTE:
args.TEST = args.nTE
obs = np.zeros([len(T_list), 5, 7])
## First upsampling
T_ren = 2.0 / np.arccosh(np.exp(2.0 / T_list))
for (iT, T) in enumerate(T_ren):
## Update model temperature ##
T_model = T_list[np.abs(T - T_list).argmin()]
model.update_temperature(T=T_model)
## Make predictions ##
data_in = temp_partition(mc_small, iT, n_samples=args.nTE)
pred = model.graph.predict(add_index(data_in))
## Calculate observables ##
obs[iT] = calculate_observables_real(temp_partition(
mc_large, iT, n_samples=args.nTE),
data_in,
pred[:, :, 0],
T=T_list[iT],
Tr=T)
print('Temperature %d / %d done!' % (iT + 1, len(T_list)))
## Save observables ##
create_directory(quantities_real_dir)
np.save(quantities_real_dir + '/%s.npy' % (model.name), obs)
def main(args):
## Renormalized temperature (inverse)
T_ren_inv = np.array([0.01, 0.01, 0.01, 0.01, 0.01,
1.21835191, 1.22976684, 1.39674347, 1.51484435, 1.65761354,
1.75902208, 1.85837041, 1.95260925, 2.07132396, 2.13716533,
2.25437054, 2.29606717, 2.38018868, 2.44845189, 2.51316151,
2.58725426, 2.6448879 , 2.7110948 , 2.74426717, 2.81525268,
2.87031377, 2.90806294, 2.98742994, 3.03780331, 3.10501399,
3.17323991, 3.19663683])
## Read data ##
L_list = [args.L, args.L//2, args.L]
data_or = temp_partition(add_index(read_file(L=args.L, n_samples=args.nTE)), args.iT, n_samples=args.nTE)
data_in = temp_partition(add_index(read_file(L=args.L//2, n_samples=args.nTE)), args.iT, n_samples=args.nTE)
## Set model ##
set_GPU_memory(fraction=args.GPU)
model = ModelLoader(list_ind=args.Mind, critical=args.CR)
print('\nModel %s loaded.'%model.name)
print('Temperature = %.4f\n'%T_list[args.iT])
if args.TEST > args.nTE:
args.TEST = args.nTE
## Find transformed temperatures ##
Tr = T_ren_inv[args.iT]
## Find closer value from T_list to update model temperature ##
iT_closer = np.abs(T_list - Tr).argmin()
model.update_temperature(T=T_list[iT_closer])
extrapolated_model = duplicate(model.graph, data_in.shape,
hid_filters=args.HF, kernels=args.K, hid_act=args.ACT)
print('\nModel %s loaded.'%model.name)
print('Temperature = %.4f'%T_list[args.iT])
print('Renormalized temperatute = %.4f\n'%T_list[iT_closer])
## Make predictions ##
pred_cont = extrapolated_model.predict(data_in)
## Calculate observables ##
data_list = [temp_partition(read_file(L=args.L, n_samples=args.nTE), iT_closer),
temp_partition(read_file(L=args.L//2, n_samples=args.nTE), iT_closer),
pred_cont[:,:,:,0] > np.random.random(pred_cont.shape[:-1])]
obj_list = [Ising(2 * x - 1) for x in data_list]
obs = np.zeros([3, 2, args.nTE])
for (i, obj) in enumerate(obj_list):
obj._calculate_magnetization()
obj._calculate_energy()
obs[i, 0] = np.abs(obj.sample_mag) * 1.0 / L_list[i]**2
obs[i, 1] = obj.sample_energy * 1.0 / L_list[i]**2
create_directory(seaborn_dir)
np.save(seaborn_dir + '/%s_extr_iT%d.npy'%(model.name, args.iT), np.array(obs))
def main(args):
## Decide if deep or not
assert len(args.K) == len(args.Nw)
if len(args.K) > 1:
from networks.trainer_deep import Trainer
args.WAVEP = None ##!!! we have not fixed weight averaging for deep model yet!
else:
from networks.trainer import Trainer
# Make Nw and K integers
args.Nw, args.K = args.Nw[0], args.K[0]
## Load data
train_data = dl.add_index(
dl.Z2_zero_majority(dl.read_file(L=args.L, n_samples=10000,
train=True)))
val_data = dl.add_index(
dl.Z2_zero_majority(
dl.read_file(L=args.L, n_samples=10000, train=False)))
for (iT, T) in enumerate(T_list):
## Prepare RBM
args.iT = iT
rbm = Trainer(args)
print('\nTemperature: %.4f - Critical: %s' % (T, str(args.CR)))
if isinstance(args.K, int):
print('RBM with %d visible units and %d hidden units.' %
(rbm.Nv**2, rbm.Nh**2 * rbm.K))
print('Number of weight parameters: %d.\n' % (rbm.Nw**2 * rbm.K))
else:
print('Created RBM with %d visible units.' % (rbm.Nv**2))
print(args.K)
print('\n')
rbm.prepare_training()
## Set up directory for saving
save_dir = 'Trained_Models/MultipleTemps/%s' % (rbm.name)
dl.create_directory(save_dir)
save_dir += '/T%.4f' % T
dl.create_directory(save_dir)
## Train and save
rbm.fit_and_save(train_data=dl.temp_partition(train_data, iT),
val_data=dl.temp_partition(val_data, iT)[:args.nVAL],
directory=save_dir)
print('Temperature %d / %d done!' % (iT + 1, len(T_list)))
return
def main(args):
if args.CR:
critical_n_samples = {8: 40000, 16: 100000}
T = 2 / np.log(1 + np.sqrt(2))
data = dl.read_file_critical(L=args.L,
n_samples=critical_n_samples[args.L],
train=False)
else:
from data.directories import T_list
T = T_list[args.iT]
data = dl.temp_partition(
dl.read_file(L=args.L, n_samples=10000, train=False), args.iT)
## Initialize
rbm = Tester(args, T=T)
print('\nInitiating testing with %s.\n' % rbm.name)
## Test
rbm.prepare_testing(data)
obs_list = rbm.test()
## Save
save_dir = 'Observables/%s' % ['T%.4f' % T, 'Critical'][int(args.CR)]
dl.create_directory(save_dir)
save_dir += '/%s' % rbm.name
dl.create_directory(save_dir)
np.save('%s/k%d_%d.npy' % (save_dir, args.kI, args.kF), obs_list)
## Saved .npy format: Array (different k (0 = correct), 12 observables)
return
def main(args):
data = dl.read_file(L=args.L, n_samples=10000, train=False)
obs_list = []
for (iT, T) in enumerate(T_list):
## Initialize
args.iT = iT
rbm = TesterTemps(args, T=T)
print('\nInitiating testing with %s.\n' % rbm.name)
## Test
rbm.prepare_testing(dl.temp_partition(data, iT))
obs_list.append(rbm.test())
print('Temperature %d / %d done!' % (iT + 1, len(T_list)))
obs_list = np.array(obs_list)
## Save
save_dir = 'Observables/%s' % rbm.name
dl.create_directory(save_dir)
np.save('%s/k%d_%d.npy' % (save_dir, args.kI, args.kF), obs_list)
## Saved .npy format: Array (temps, different k (0 = correct), 12 observables)
# Save only the last converged k
np.save('%s/k%d_converged.npy' % (save_dir, args.kF),
np.concatenate((obs_list[:, 0, :], obs_list[:, -1, :])))
## Saved .npy format: Array ((0=correct, 1=converged),temps, 12 observables)
return
def main(args):
if args.PBC:
from networks.architectures import duplicate_simple2D_pbc as duplicate
else:
from networks.architectures import duplicate_simple2D as duplicate
## Renormalized temperature (inverse)
T_ren_inv = np.array([
0.01, 0.01, 0.01, 0.01, 0.01, 1.21835191, 1.22976684, 1.39674347,
1.51484435, 1.65761354, 1.75902208, 1.85837041, 1.95260925, 2.07132396,
2.13716533, 2.25437054, 2.29606717, 2.38018868, 2.44845189, 2.51316151,
2.58725426, 2.6448879, 2.7110948, 2.74426717, 2.81525268, 2.87031377,
2.90806294, 2.98742994, 3.03780331, 3.10501399, 3.17323991, 3.19663683
])
## Read data ##
data_or = read_file(L=args.L, n_samples=args.nTE)
data_in = add_index(read_file(L=args.L // 2, n_samples=args.nTE))
## Set model ##
set_GPU_memory(fraction=args.GPU)
model = ModelLoader(list_ind=args.Mind, critical=args.CR)
if args.TEST > args.nTE:
args.TEST = args.nTE
if args.OUT:
from data.directories import output_dir
create_directory(output_dir + '/' + model.name)
if args.Tind == None:
args.Tind = np.arange(len(T_list))
obs = np.zeros([len(args.Tind), 5, 12])
for (iT, T) in enumerate(T_list[args.Tind]):
## Find transformed temperatures ##
Tr = T_ren_inv[iT]
## Find closer value from T_list to update model temperature ##
T_closer = T_list[np.abs(T_list - Tr).argmin()]
model.update_temperature(T=T_closer)
extrapolated_model = duplicate(model.graph,
data_in.shape,
hid_filters=args.HF,
kernels=args.K,
hid_act=args.ACT)
## Make predictions ##
data_in_T = temp_partition(data_in, iT, n_samples=args.nTE)
pred_cont = extrapolated_model.predict(data_in_T)
## Calculate observables ##
if iT > 5:
Tr_calc = Tr
else:
Tr_calc = T
obs[iT] = calculate_observables_real(temp_partition(
data_or, iT, n_samples=args.nTE),
data_in_T[:, :, :, 0],
pred_cont[:, :, :, 0],
T=T,
Tr=Tr_calc)
## Save network output ##
if args.OUT:
np.save(output_dir + '/%s/T%.4f.npy' % (model.name, T), pred_cont)
print('Temperature %d / %d done!' % (iT + 1, len(args.Tind)))
## Save observables ##
create_directory(quantities_real_dir)
np.save(quantities_real_dir + '/%s_extr.npy' % model.name, np.array(obs))
def main(args):
if args.CR:
T = 2 / np.log(1 + np.sqrt(2))
rbm, rbm_name = get_model(args.Mind, Nv=args.L, critical=True)
data = dl.read_file_critical(L=args.L, n_samples=100000, train=False)
else:
from data.directories import T_list
T = T_list[args.iT]
rbm, rbm_name = get_model(args.Mind, Nv=args.L, critical=False, T=T)
data = dl.temp_partition(
dl.read_file(L=args.L, n_samples=10000, train=False), args.iT)
## Calculate MC observables
obs_correct = get_observables_with_corr_and_tpf(data, T)
n_batches = args.nTE // args.BSC
## Create RBM ops
if args.REC:
data = dl.add_index(data)[:args.nTE]
ops = [
rbm.create_gibbs_sampler_random(args.BSC, k=args.GBk),
rbm.create_gibbs_sampler(k=args.GBk)
]
obs_prediction = np.zeros([n_batches, 2, len(obs_correct)])
else:
ops = rbm.create_gibbs_sampler_random(args.BSC, k=args.GBk)
obs_prediction = np.zeros([n_batches, len(obs_correct)])
print('\nInitiating testing with %s.\n' % rbm_name)
## Define Session
gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=args.GPU)
sess = tf.Session(config=tf.ConfigProto(gpu_options=gpu_options))
sess.run(tf.global_variables_initializer())
if args.REC:
for iB in range(n_batches):
sampled, recon = sess.run(ops,
feed_dict={
rbm.visible:
data[iB * args.BSC:(iB + 1) *
args.BSC]
})
obs_prediction[iB] = np.array([
get_observables_with_corr_and_tpf(sampled[:, :, :, 0], T),
get_observables_with_corr_and_tpf(recon[:, :, :, 0], T)
])
sampling_error = np.abs(obs_prediction.mean(axis=0)[0] -
obs_correct) * 100 / obs_correct
reconstr_error = np.abs(obs_prediction.mean(axis=0)[1] -
obs_correct) * 100 / obs_correct
else:
for iB in range(n_batches):
sampled = sess.run(ops)
obs_prediction[iB] = get_observables_with_corr_and_tpf(
sampled[:, :, :, 0], T)
## Debugging
print('\n')
print(sampled[:, :, :, 0])
print('\n')
obs_pred = obs_prediction.mean(axis=0)
sampling_error = np.abs(obs_pred - obs_correct) * 100 / obs_correct
print('\nGibbs k=%d' % args.GBk)
print('\nSampling:')
print(sampling_error)
if args.REC:
print('\nReconstruction:')
print(reconstr_error)
## Print observables
print('\nCorrect vs Predicted Observables:')
for (cor, pred) in zip(obs_correct, obs_pred):
print('%.6f - %.6f' % (cor, pred))
return
def main(args):
critical_n_samples = {8: 40000, 16: 100000}
## Decide if deep or not
assert len(args.K) == len(args.Nw)
if len(args.K) > 1:
from networks.trainer_deep import Trainer
args.WAVEP = None ##!!! we have not fixed weight averaging for deep model yet!
else:
from networks.trainer import Trainer
# Make Nw and K integers
args.Nw, args.K = args.Nw[0], args.K[0]
## Load data
if args.CR:
T = 2 / np.log(1 + np.sqrt(2))
save_dir = 'Trained_Models/Critical'
data = dl.add_index(
dl.read_file_critical(L=args.L,
n_samples=critical_n_samples[args.L]))
train_data = data[:args.nTR]
val_data = data[args.nTR:args.nTR + args.nVAL]
else:
from data.directories import T_list
T = T_list[args.iT]
save_dir = 'Trained_Models/T%.4f' % T
train_data = dl.add_index(
dl.temp_partition(
dl.read_file(L=args.L, n_samples=10000, train=True), args.iT))
val_data = dl.add_index(
dl.temp_partition(
dl.read_file(L=args.L, n_samples=10000, train=False), args.iT))
## Prepare RBM
rbm = Trainer(args)
print('\nTemperature: %.4f - Critical: %s' % (T, str(args.CR)))
if isinstance(args.K, int):
print('RBM with %d visible units and %d hidden units.' %
(rbm.Nv**2, rbm.Nh**2 * rbm.K))
print('Number of weight parameters: %d.\n' % (rbm.Nw**2 * rbm.K))
else:
print('Created RBM with %d visible units.' % (rbm.Nv**2))
print(args.K)
print('\n')
rbm.prepare_training()
## Set up directory for saving
dl.create_directory(save_dir)
save_dir += '/%s' % (rbm.name)
dl.create_directory(save_dir)
## Train and save
rbm.fit_and_save(train_data, val_data, directory=save_dir)
return
def main(args):
data_mc = read_file(L=args.L, n_samples=args.nTE)
set_GPU_memory(fraction=args.GPU)
model = ModelLoader(list_ind=args.Mind)
if args.TEST > args.nTE:
args.TEST = args.nTE
tpf = np.zeros([args.UP, len(T_list)])
obs = np.zeros([args.UP, len(T_list), 3, 7])
## WARNING: Does not contain original MC observables
pred_cont = []
## First upsampling
T_ren = 2.0 / np.arccosh(np.exp(2.0 / T_list))
for (iT, T) in enumerate(T_ren):
## Update model temperature ##
T_model = T_list[np.abs(T - T_list).argmin()]
model.update_temperature(T=T_model)
## Make predictions ##
pred_cont.append(make_prediction(
data_in=add_index(temp_partition(data_mc, iT, n_samples=args.nTE)),
graph=model.graph, hid_filters=args.HF, kernels=args.K, hid_act=args.ACT))
if iT < args.TS:
pred_cont[iT] = np.round(pred_cont[iT])
else:
pred_cont[iT] = (pred_cont[iT] > np.random.random(pred_cont[iT].shape)).astype(np.int)
## Calculate observables ##
obs[0, iT] = calculate_observables_rep(pred_cont[iT][:,:,0], Tr=T)
tpf[0, iT] = two_point_function(2 * pred_cont[iT][:,:,0] - 1, k=int(pred_cont[iT].shape[1]**0.8/5))
print('Temperature %d / %d done!'%(iT+1, len(T_list)))
print('\nUpsampling 1 / %d completed!\n'%args.UP)
for iUP in range(1, args.UP):
T_ren = 2.0 / np.arccosh(np.exp(2.0 / T_ren))
for (iT, T) in enumerate(T_ren):
T_model = T_list[np.abs(T - T_list).argmin()]
model.update_temperature(T=T_model)
## Make predictions ##
if iT < args.TS:
pred_cont[iT] = np.round(pred_cont[iT])
else:
pred_cont[iT] = (pred_cont[iT] > np.random.random(pred_cont[iT].shape)).astype(np.int)
pred_cont[iT] = make_prediction(data_in=pred_cont[iT], graph=model.graph,
hid_filters=args.HF, kernels=args.K, hid_act=args.ACT)
## Calculate observables ##
obs[iUP, iT] = calculate_observables_rep(pred_cont[iT][:,:,0], Tr=T)
tpf[iUP, iT] = two_point_function(2 * pred_cont[iT][:,:,0] - 1, k=int(pred_cont[iT].shape[1]**0.8/5))
print('Temperature %d / %d done!'%(iT+1, len(T_list)))
print('\nUpsampling %d / %d completed!\n'%(iUP+1, args.UP))
## Save observables ##
create_directory(quantities_rep_dir)
np.save(quantities_rep_dir + '/%s_TS%d_UP%d_VER%d.npy'%(model.name, args.TS, args.UP, args.VER), obs)
np.save(quantities_rep_dir + '/%s_TS%d_TPF_N085.npy'%(model.name, args.TS), tpf)
