def test_slc_crime():
ind = np.load('SLC_60k_inds.npy')
y = np.load('SLC_60k_y_in_day.npy').reshape(-1, 1).astype(np.float32)
TOTAL_LEN = len(ind)
NUM_TRAIN = 40000
train_ind = ind[:NUM_TRAIN]
train_y = y[:NUM_TRAIN]
test_ind = ind[NUM_TRAIN:]
test_y = y[NUM_TRAIN:]
print("Total len = %d, num train = %d, num test = %d" %
(TOTAL_LEN, NUM_TRAIN, TOTAL_LEN - NUM_TRAIN))
print(test_y[-1][0] - test_y[0][0])
nmod = ind.shape[1]
nvec = np.max(ind, axis=0) + 1
R = 8
U = [np.random.rand(nvec[k], R) for k in range(nmod)]
init_config = {}
init_config['U'] = U
init_config['batch_size_event'] = 128
init_config['batch_size_entry'] = 128
init_config['learning_rate'] = 1e-3
len_B = 128 #event
print('lauching Kmeans')
B = utils_funcs.init_base_gp_pseudo_inputs(U, train_ind, len_B)
print('Kmeans end')
# VI Sparse GP
init_config['inducing_B'] = B # init with k-means, [len_B, rank]
model = NTF_HP(train_ind, train_y, init_config)
#model.create_all_relevant_test_graph(test_ind, test_y)
model.create_standAlone_test_graph(test_ind, test_y)
#model.create_standAlone_test_graph(train_ind[-19300:], train_y[-19300:])
steps_per_epoch = int(len(train_ind) / init_config['batch_size_event'])
num_epoch = 50
test_llk = []
for epoch in range(1, num_epoch + 1):
print('epoch %d' % epoch)
model.train(steps_per_epoch, int(steps_per_epoch))
test_log_p, int_term, eventSum_term = model.test(verbose=False)
print("test_log_llk = %g, int_term = %g, eventsum_term = %g\n" %
(test_log_p, int_term, eventSum_term))
test_llk.append(test_log_p)
jb.dump(test_llk, "slc_bg_only.pkl")
def test_data_set():
(ind, y), (train_ind, train_y), (test_ind,
test_y) = utils_funcs.load_dataSet(
arg_data_name, '../Data')
nmod = ind.shape[1]
nvec = np.max(ind, axis=0) + 1
R = arg_rank
U = [np.random.rand(nvec[k], R) for k in range(nmod)]
init_config = {}
init_config['U'] = U
init_config['batch_size_event'] = 64
init_config['batch_size_entry'] = 64
init_config['learning_rate'] = arg_lr
len_B = 128 # event
print('lauching Kmeans')
B = utils_funcs.init_base_gp_pseudo_inputs(U, train_ind, len_B)
print('Kmeans end')
# VI Sparse GP
init_config['inducing_B'] = B # init with k-means, [len_B, rank]
model = NTF_HP(train_ind, train_y, init_config)
# model.create_all_relevant_test_graph(test_ind, test_y)
model.create_standAlone_test_graph(test_ind, test_y)
# model.create_standAlone_test_graph(train_ind[-19300:], train_y[-19300:])
steps_per_epoch = int(len(train_ind) / init_config['batch_size_event'])
num_epoch = 50
test_llk = []
for epoch in range(1, num_epoch + 1):
print('epoch %d' % epoch)
model.train(steps_per_epoch, int(steps_per_epoch))
test_log_p, int_term, eventSum_term = model.test(verbose=False)
print("test_log_llk = %g, int_term = %g, eventsum_term = %g\n" %
(test_log_p, int_term, eventSum_term))
test_llk.append(test_log_p)
#jb.dump(test_llk, "../Result_Logs/%s_GP_Poisson_rank_%d.jb" % ( arg_data_name, arg_rank))
#final_U = model.check_vars( model.tf_U)
#jb.dump( final_U, "../Result_Logs/%s_GP_Poisson_final_U_rank_%d.jb" %( arg_data_name, arg_rank))
utils_funcs.log_results('../Result_Logs/GP_Rayleigh.txt', arg_data_name,
arg_rank, arg_lr, test_llk)
model.sess.close()
def test_data_set():
(ind, y), (train_ind, train_y), (test_ind, test_y) = utils_funcs.load_dataSet(arg_data_name, '../Data')
nmod = ind.shape[1]
nvec = np.max(ind, axis=0) + 1
T = y[-1][0] - y[0][0]
print('T = %g' % T)
R = arg_rank
U = [np.random.rand(nvec[k], R) for k in range(nmod)]
init_config = {}
init_config['U'] = U
init_config['batch_size_event'] = 128
init_config['batch_size_entry'] = 128
init_config['learning_rate'] = arg_lr
len_B = 128 # event
print('lauching Kmeans')
B = utils_funcs.init_base_gp_pseudo_inputs(U, train_ind, len_B)
print('Kmeans end')
# VI Sparse GP
init_config['inducing_B'] = B # init with k-means, [len_B, rank]
model = NTF_HP(train_ind, train_y, init_config)
model.create_standAlone_test_graph(test_ind, test_y)
steps_per_epoch = int(len(train_ind) / init_config['batch_size_event'])
num_epoch = 50
test_llk = []
for epoch in range(1, num_epoch + 1):
print('epoch %d' % epoch)
model.train(steps_per_epoch, int(steps_per_epoch))
test_log_p, int_term, eventSum_term = model.test(verbose=False)
print("test_log_llk = %g, int_term = %g, eventsum_term = %g\n" % (test_log_p, int_term, eventSum_term))
test_llk.append(test_log_p)
utils_funcs.log_results('GP_PTF.txt', arg_data_name, arg_rank, arg_lr, test_llk)
model.sess.close()
nvec = np.max(ind, axis=0) + 1
R_list = [1, 2, 5, 8, 10]
res = []
for R in R_list:
R = 8
U = [np.random.rand(nvec[k], R) for k in range(nmod)]
init_config = {}
init_config['U'] = U
init_config['batch_size_event'] = 128
init_config['batch_size_entry'] = 128
init_config['learning_rate'] = 1e-3
len_B = 128 #event
print('lauching Kmeans')
B = utils_funcs.init_base_gp_pseudo_inputs(U, train_ind, len_B)
print('Kmeans end')
# VI Sparse GP
init_config['inducing_B'] = B # init with k-means, [len_B, rank]
model = NTF_HP(train_ind, train_y, init_config)
#model.create_all_relevant_test_graph(test_ind, test_y)
model.create_standAlone_test_graph(test_ind, test_y)
#model.create_standAlone_test_graph(train_ind[-19300:], train_y[-19300:])
steps_per_epoch = int(len(train_ind) / init_config['batch_size_event'])
num_epoch = 50
test_llk = []
def test_data_set():
(ind, y), (train_ind, train_y), (test_ind,
test_y) = utils_funcs.load_dataSet(
arg_data_name, '../Data')
nmod = ind.shape[1]
nvec = np.max(ind, axis=0) + 1
R = arg_rank
U = [np.random.rand(nvec[k], R) * 1.0 for k in range(nmod)]
init_config = {}
init_config['U'] = U
init_config['batch_size_event'] = 64
init_config['batch_size_entry'] = 64
init_config['batch_size_inner_event'] = 4096
init_config['batch_size_entryEvent'] = 4096
init_config['learning_rate'] = arg_lr
init_config['log_amp_alpha'] = 0.0
init_config['log_amp_delta'] = 0.0
init_config['log_amp_trig'] = -3
len_B = 128 # Base Rate
model_config = {
'log_amp_alpha': init_config['log_amp_alpha'],
'log_amp_delta': init_config['log_amp_delta'],
'log_amp_trig': init_config['log_amp_trig'],
'rank': arg_rank,
'MATRIX_JITTER': MATRIX_JITTER,
'DELTA_JITTER': DELTA_JITTER,
'lr': arg_lr,
'batch_size_event': init_config['batch_size_event'],
'batch_size_entry': init_config['batch_size_entry'],
'batch_size_inner_event': init_config['batch_size_inner_event'],
'batch_size_entryEvent': init_config['batch_size_entryEvent'],
'num_psd_points': len_B
}
print('launching Kmeans')
B = utils_funcs.init_base_gp_pseudo_inputs(U, train_ind, len_B)
print(B.shape)
print('Kmeans end')
# VI Sparse GP
init_config['inducing_B'] = B # init with k-means, [len_B, rank]
model = NTF_HP(train_ind, train_y, init_config)
model.create_standAlone_test_graph(test_ind, test_y)
steps_per_epoch = int(len(train_ind) / init_config['batch_size_event'])
num_epoch = 50
log_file = utils_funcs.init_log_file('./BNF.txt', arg_data_name,
model_config)
for epoch in range(1, num_epoch + 1):
print('epoch %d\n' % epoch)
model.train(steps_per_epoch, int(steps_per_epoch / 5))
test_log_p, int_term, int_term1, int_term3, eventSum_term = model.test(
128, 16, verbose=False)
log_file.write('%g\n' % test_log_p)
log_file.flush()
os.fsync(log_file.fileno())
log_file.close()
model.sess.close()