def get_marked_imgs(gtn, imgs_test, num_f, sample, param, order_way):
if order_way is 'SequencedMeasure':
order_file = os.path.join(param['data_path'], 'Order_' + param['save_exp'])
if os.path.isfile(order_file):
order = load_pr(order_file, 'order')
else:
order = gtn.mps.markov_measurement(if_restore=True)[0]
save_pr(para['data_path'], 'Order_' + para['save_exp'], [order], ['order'])
order_now = copy.copy(order.reshape(-1,)[:num_f])
elif order_way is 'MaxSEE':
ent = gtn.mps.calculate_onsite_reduced_density_matrix()[0]
order = np.argsort(ent)[::-1]
order_now = copy.copy(order.reshape(-1,)[:num_f])
elif order_way is 'Variance':
tmp = TNmachineLearning.MachineLearningFeatureMap(param['d'], param['dataset'])
tmp.load_data()
tmp.select_samples([param['class']])
variance = tmp.variance_pixels()
order = np.argsort(variance)[::-1]
order_now = copy.copy(order.reshape(-1,)[:num_f])
elif order_way is 'RandomMeasure':
order = np.random.permutation(gtn.length)
order_now = copy.copy(order.reshape(-1,)[:num_f])
else:
order = None
order_now = None
img_part = imgs_test.images.copy()[order_now, sample]
img_new = gtn.generate_features(img_part, pos=order_now, f_max=1, f_min=0,
is_display=False, way=generate_way)
img_new = img_new.reshape(imgs_test.img_size)
marked_img = imgs_test.mark_pixels_on_full_image(img_new, order_now)
return imgs_test.images[:, sample].reshape(imgs_test.img_size), img_new, marked_img, order
def get_sequence(gtn, num_f, param, order_way):
if order_way is 'SequencedMeasure':
order_file = os.path.join(param['data_path'],
'Order_' + param['save_exp'])
if os.path.isfile(order_file):
order = load_pr(order_file, 'order')
else:
order = gtn.mps.markov_measurement(if_restore=True)[0]
save_pr(param['data_path'], 'Order_' + param['save_exp'], [order],
['order'])
order_now = copy.copy(order[:num_f])
elif order_way is 'MaxSEE':
ent = gtn.mps.calculate_onsite_reduced_density_matrix()[0]
order = np.argsort(ent)[::-1]
order_now = copy.copy(order[:num_f])
elif order_way is 'Variance':
tmp = TNmachineLearning.MachineLearningFeatureMap(
param['d'], param['dataset'])
tmp.load_data()
tmp.select_samples([param['class']])
variance = tmp.variance_pixels()
order = np.argsort(variance)[::-1]
order_now = copy.copy(order[:num_f])
elif order_way is 'RandomMeasure':
order = np.random.permutation(gtn.length)
order_now = copy.copy(order[:num_f])
else:
order = None
order_now = None
return order, order_now
def decision_mps(para=None):
if para is None:
para = pm.parameters_decision_mps()
a = TNmachineLearning.DecisionTensorNetwork(
para['dataset'],
2,
para['chi'],
'mps',
para['classes'],
para['numbers'],
if_reducing_samples=para['if_reducing_samples'])
a.images2vecs_test_samples(para['classes'])
print(a.vLabel)
for n in range(0, a.length):
bf.print_sep()
print('Calculating the %i-th tensor' % n)
a.update_tensor_decision_mps_svd(n)
# a.update_tensor_decision_mps_svd_threshold_algo(n)
# a.update_tensor_decision_mps_gradient_algo(n)
a.update_v_ctr_train(n)
a.calculate_intermediate_accuracy_train(n)
if a.remaining_samples_train.__len__() == 0:
print('All samples are classified correctly. Training stopped.')
break
print('The current accuracy = %g' % a.intermediate_accuracy_train[n])
print('Entanglement: ' + str(a.lm[n].reshape(1, -1)))
if para['if_reducing_samples']:
print('Number of remaining samples: ' +
str(a.remaining_samples_train.__len__()))
def gcmpm_one_class(para=None):
if para is None:
para = pm.parameters_gcmpm_one_class()
para['save_exp'] = save_exp_gcmpm_one_class(para)
if para['parallel'] is True:
par_pool = para['n_nodes']
else:
par_pool = None
if para['if_load'] and os.path.isfile(para['save_exp']):
a = bf.load_pr(os.path.join(para['data_path'], para['save_exp']), 'a')
else:
a = TNmachineLearning.MachineLearningMPS(para['d'], para['chi'], para['dataset'],
par_pool=par_pool)
a.images2vecs([para['class']], [100])
a.initialize_virtual_vecs_train()
a.update_virtual_vecs_train('all', 'all', 'both')
a.mps.correct_orthogonal_center(0, normalize=True)
a.mps.mps[0] /= np.linalg.norm(a.mps.mps[0].reshape(-1, ))
mps0 = a.mps.mps.copy()
for t in range(0, para['sweep_time']):
# from left to right
if para['if_print_detail']:
print('At the ' + str(t) + '-th sweep, from left to right')
for nt in range(0, a.length):
a.update_tensor_gradient(nt, para['step'])
if nt != a.length-1:
a.update_virtual_vecs_train('all', nt, 'left')
# from left to right
print('At the ' + str(t) + '-th sweep, from right to left')
for nt in range(a.length-1, -1, -1):
a.update_tensor_gradient(nt, para['step'])
if nt != 0:
a.update_virtual_vecs_train('all', nt, 'right')
if t > para['check_time0'] and ((t+1) % para['check_time'] == 0
or t+1 == para['sweep_time']):
fid = ln_fidelity_per_site(mps0, a.mps.mps)
if fid < (para['step'] * para['ratio_step_tol']):
print('After ' + str(t+1) + ' sweeps: fid = %g' % fid)
para['step'] *= para['step_ratio']
elif t+1 == para['sweep_time']:
print('After all ' + str(t+1) + ' sweeps finished, fid = %g. '
'Consider to increase the sweep times.' % fid)
else:
print('After ' + str(t+1) + ' sweeps, fid = %g.' % fid)
mps0 = a.mps.mps.copy()
if para['step'] < para['step_min']:
print('Now step = ' + str(para['step']) + ' is sufficiently small. Break the loop')
break
else:
print('Now step = ' + str(para['step']))
if para['if_save']:
save_pr(para['data_path'], para['save_exp'], [a, para], ['a', 'para'])
return a, para
def gtnc(para_tot=None):
print('Preparing parameters')
if para_tot is None:
para_tot = pm.parameters_gcmpm()
n_class = len(para_tot['classes'])
paras = bf.empty_list(n_class)
for n in range(0, n_class):
paras[n] = copy.deepcopy(para_tot)
paras[n]['class'] = int(para_tot['classes'][n])
paras[n]['chi'] = para_tot['chi'][n]
paras[n]['theta'] = para_tot['theta']
paras[n]['save_exp'] = save_exp_gtn_one_class(paras[n])
classifiers = bf.empty_list(n_class)
for n in range(0, n_class):
print_dict(paras[n])
data = para_tot['data_path'] + paras[n]['save_exp']
if para_tot['if_load'] and os.path.isfile(data):
print('The classifier already exists. Load directly')
classifiers[n] = load_pr(data, 'a')
else:
print('Training the MPS of ' + str(para_tot['classes'][n]))
classifiers[n] = gtn_one_class(paras[n])[0]
# if para_tot['if_save']:
# save_pr('../data_tnml/gcmpm/', paras[n]['save_exp'],
# [classifiers[n]], ['classifier'])
# classifiers[n].mps.check_orthogonality_by_tensors(tol=1e-12)
# ==================== Testing accuracy ====================
print('Calculating the testing accuracy')
b = TNmachineLearning.MachineLearningFeatureMap(para_tot['d'])
b.load_data(data_path='..\\..\\..\\MNIST\\',
file_sample='t10k-images.idx3-ubyte',
file_label='t10k-labels.idx1-ubyte',
is_normalize=True)
b.select_samples(para_tot['classes'])
if classifiers[0].is_dct:
b.dct(shift=para_tot['shift'], factor=para_tot['factor'])
b.images2vecs(para_tot['theta'] * np.pi / 2)
fid = bf.empty_list(n_class)
for n in range(0, n_class):
fid[n] = b.compute_fidelities(classifiers[n].mps.mps)
max_fid = np.argmin(np.hstack(fid), axis=1)
predict = np.zeros(max_fid.shape, dtype=int)
for n in range(0, n_class):
predict += (max_fid == n) * int(para_tot['classes'][n])
# plot(predict)
# plot(b.labels)
accuracy = np.sum(predict == b.labels, dtype=float) / b.numVecSample
print(accuracy)
def gcmpm(para_tot=None):
print('Preparing parameters')
if para_tot is None:
para_tot = pm.parameters_gcmpm()
n_class = len(para_tot['classes'])
paras = bf.empty_list(n_class)
for n in range(0, n_class):
paras[n] = para_tot.copy()
paras[n]['class'] = para_tot['classes'][n]
paras[n]['chi'] = para_tot['chi'][n]
paras[n]['save_exp'] = save_exp_gcmpm_one_class(paras[n])
classifiers = bf.empty_list(n_class)
for n in range(0, n_class):
data = '../data_tnml/gcmpm/' + paras[n]['save_exp']
if para_tot['if_load'] and os.path.isfile(data):
print('The classifier already exists. Load directly')
classifiers[n] = load_pr(data, 'classifier')
else:
print('Training the MPS of ' + str(para_tot['classes'][n]))
classifiers[n] = gcmpm_one_class(paras[n])[0]
if para_tot['if_save']:
save_pr('../data_tnml/gcmpm/', paras[n]['save_exp'],
[classifiers[n]], ['classifier'])
# Testing accuracy
print('Calculating the testing accuracy')
labels = para_tot['classes']
b = TNmachineLearning.MachineLearningFeatureMap('MNIST', para_tot['d'],
file_sample='t10k-images.idx3-ubyte',
file_label='t10k-labels.idx1-ubyte')
b.images2vecs(para_tot['classes'], ['all', 'all'])
fid = np.zeros((n_class, ))
num_wrong = 0
for ni in range(0, b.numVecSample):
for n in range(0, n_class):
fid[n] = b.fidelity_mps_image(classifiers[n].mps.mps, ni)
n_max = int(np.argmax(fid))
if labels[n_max] != b.LabelNow[ni]:
num_wrong += 1
accuracy = num_wrong/b.numVecSample
print(accuracy)
# =============================================================
para = parameters_gtn_one_class()
para['dct'] = False
para['d'] = 2
para['step'] = 0.2 # initial gradient step
para['if_save'] = True
para['if_load'] = True
para['dataset'] = dataset
para['class'] = which_class
para['chi'] = chi
psnr_av = np.zeros(var_values.shape)
mse_av = np.zeros(var_values.shape)
ssim = np.zeros(var_values.shape)
b = TNmachineLearning.MachineLearningFeatureMap(para['d'], para['dataset'])
if var_name is not 'which_class':
b.load_data(data_path=os.path.join(b.project_path, '..\\..\\MNIST\\' + para['dataset'] + '\\'),
file_sample='t10k-images.idx3-ubyte',
file_label='t10k-labels.idx1-ubyte', is_normalize=True)
b.select_samples([para['class']])
is_order_calculated = False
for t in range(var_values.size):
# Modify parameter
print('For ' + var_name + ' = ' + str(var_values[t]))
exec(var_name + ' = ' + str(var_values[t]))
if var_name in para:
exec('para[\'' + var_name + '\'] = ' + str(var_values[t]))
print('Train the generative TN')
def gtn_one_class(para=None, images=None, labels=None):
if 'to_black_and_white' not in para:
para['to_black_and_white'] = False
if para is None:
para = pm.parameters_gtn_one_class()
para['save_exp'] = save_exp_gtn_one_class(para)
if para['if_load'] and os.path.isfile(
os.path.join(para['data_path'], para['save_exp'])):
a = bf.load_pr(os.path.join(para['data_path'], para['save_exp']), 'a')
else:
a = TNmachineLearning.MachineLearningMPS(para['d'], para['chi'],
para['dataset'])
if para['dataset'] is 'custom':
a.input_data(copy.deepcopy(images), copy.deepcopy(labels))
else:
a.load_data()
if a.is_there_labels:
a.select_samples([para['class']])
if para['to_black_and_white']:
a.to_black_and_white()
if para['dct'] is True:
a.dct(shift=para['shift'], factor=para['factor'])
a.images2vecs(theta_max=para['theta'] * np.pi / 2)
a.initial_mps(center=0, ini_way='1')
a.initialize_virtual_vecs_train()
a.mps.correct_orthogonal_center(0)
a.update_tensor_gradient(0, para['step'])
nll0 = a.compute_nll()
step = copy.deepcopy(para['step'])
print('Iniitially, NLL = ' + str(nll0))
for t in range(0, para['sweep_time']):
# from left to right
if para['if_print_detail']:
print('At the ' + str(t + 1) + '-th sweep, from left to right')
t0 = time.time()
tt0 = time.clock()
for nt in range(0, a.length):
a.mps.correct_orthogonal_center(nt)
if nt != 0:
a.update_virtual_vecs_train(nt - 1, 'left')
a.update_tensor_gradient(nt, step)
# from right to left
if para['if_print_detail']:
print('At the ' + str(t + 1) + '-th sweep, from right to left')
for nt in range(a.length - 1, -1, -1):
a.mps.correct_orthogonal_center(nt)
if nt != a.length - 1:
a.update_virtual_vecs_train(nt + 1, 'right')
a.update_tensor_gradient(nt, step)
if para['if_print_detail']:
print('Wall time cost for one loop: %s' % (time.time() - t0))
print('CPU time cost for one loop: %s' % (time.clock() - tt0))
if t > (para['check_time0'] - 2) and (
(t + 1) % para['check_time'] == 0
or t + 1 == para['sweep_time']):
nll = a.compute_nll()
print('NLL = ' + str(nll))
# fid = fidelity_per_site(mps0, a.mps.mps)
fid = abs(nll - nll0) / nll0
if fid < (step * para['step_ratio']):
print('After ' + str(t + 1) + ' sweeps: fid = %g' % fid)
step *= para['step_ratio']
# mps0 = copy.deepcopy(a.mps.mps)
nll0 = nll
elif t + 1 == para['sweep_time']:
print('After all ' + str(t + 1) +
' sweeps finished, fid = %g. '
'Consider to increase the sweep times.' % fid)
else:
print('After ' + str(t + 1) + ' sweeps, fid = %g.' % fid)
# mps0 = copy.deepcopy(a.mps.mps)
nll0 = nll
if step < para['step_min']:
print('Now step = ' + str(step) +
' is sufficiently small. Break the loop')
break
else:
print('Now step = ' + str(step))
a.clear_before_save()
if para['if_save']:
save_pr(para['data_path'], para['save_exp'], [a, para],
['a', 'para'])
return a, para
def labeled_gtn(para):
if para is None:
para = pm.parameters_labeled_gtn()
para['save_exp'] = save_exp_labeled_gtn(para)
if para['parallel'] is True:
par_pool = para['n_nodes']
else:
par_pool = None
# Preparing testing dataset
b = TNmachineLearning.MachineLearningFeatureMap(
para['d'],
file_sample='t10k-images.idx3-ubyte',
file_label='t10k-labels.idx1-ubyte')
b.load_data()
b.select_samples(para['classes'])
b.add_labels_to_images()
b.images2vecs(para['theta'])
data_file = os.path.join(para['data_path'], para['save_exp'])
if para['if_load'] and os.path.isfile(data_file):
print('Data exist. Load directly.')
a = bf.load_pr(data_file, 'a')
else:
a = TNmachineLearning.MachineLearningMPS(para['d'],
para['chi'],
para['dataset'],
par_pool=par_pool)
a.load_data()
a.select_samples(para['classes'])
a.add_labels_to_images()
a.images2vecs(para['theta'] * np.pi / 2)
a.initial_mps()
a.mps.correct_orthogonal_center(0, normalize=True)
a.initialize_virtual_vecs_train()
a.update_virtual_vecs_train_all_tensors('both')
accuracy0 = 0
for t in range(0, para['sweep_time']):
# from left to right
for nt in range(0, a.length):
a.update_tensor_gradient(nt, para['step'])
if nt != a.length - 1:
a.update_virtual_vecs_train(nt, 'left')
# from left to right
for nt in range(a.length - 1, -1, -1):
a.update_tensor_gradient(nt, para['step'])
if nt != 0:
a.update_virtual_vecs_train(nt, 'right')
if t > para['check_time0'] and ((t + 1) % para['check_time'] == 0
or t + 1 == para['sweep_time']):
b.input_mps(a.mps)
accuracy = b.calculate_accuracy()
print('After the ' + str(t) +
'-th sweep, the testing accuracy = ' + str(accuracy))
if abs(accuracy - accuracy0) < (para['step'] *
para['ratio_step_tol']):
para['step'] *= para['step_ratio']
accuracy0 = accuracy
print('Converged. Reduce the gradient step to ' +
str(para['step']))
elif t + 1 == para['sweep_time']:
print('After all ' + str(t + 1) +
' sweeps finished, not converged. '
'Consider to increase the sweep times.')
else:
accuracy0 = accuracy
if para['step'] < para['step_min']:
print('Now step = ' + str(para['step']) +
' is sufficiently small. Break the loop')
break
else:
print('Now step = ' + str(para['step']))
a.clear_before_save()
if para['if_save']:
save_pr(para['data_path'], para['save_exp'], [a, para],
['a', 'para'])
accuracy = b.calculate_accuracy()
print('The final testing accuracy = ' + str(accuracy))
return a, para
from library import TNmachineLearning
from library.Parameters import parameters_lazy_learning
import numpy as np
from library.TensorBasicModule import khatri
from library.BasicFunctions import plot
import torch
if_torch = True
num_parts = 20 # to save GPU memory
para = parameters_lazy_learning()
print_dtime = 20
a = TNmachineLearning.MachineLearningFeatureMap(para['d'], para['dataset'])
a.load_data()
a.select_samples(para['classes'], para['num_samples'])
a.images2vecs(theta_max=para['theta'] * np.pi)
b = TNmachineLearning.MachineLearningFeatureMap(para['d'], para['dataset'])
b.load_data(data_path='..\\..\\..\\MNIST\\',
file_sample='t10k-images.idx3-ubyte',
file_label='t10k-labels.idx1-ubyte',
is_normalize=True)
b.select_samples(para['classes'], para['num_samples'])
b.images2vecs(theta_max=para['theta'] * np.pi)
accuracy = list()
num = list()
if if_torch:
for c2 in para['classes']:
fid = list()