def main():
try:
if not ('.pkl' in args.dataset): raise Exception('input_format')
# Get data
with open('{}'.format(args.dataset), 'r') as inputfile:
data = pkl.load(inputfile)
xn = data['xn']
N, D = xn.shape
if VERBOSE: init_time = time()
# Priors
alpha_o = np.array([1.0] * K)
nu_o = np.array([float(D)])
if nu_o[0] < D: raise Exception('degrees_of_freedom')
w_o = generate_random_positive_matrix(D)
m_o = np.array([0.0] * D)
beta_o = np.array([0.7])
# Variational parameters intialization
lambda_phi = np.random.dirichlet(alpha_o, N) \
if args.randomInit else init_kmeans(xn, N, K)
lambda_pi = np.zeros(shape=K)
lambda_beta = np.zeros(shape=K)
lambda_nu = np.zeros(shape=K)
lambda_m = np.random.rand(K, D)
lambda_w = np.array([np.copy(w_o) for _ in range(K)])
# Plot configs
if VERBOSE and D == 2:
plt.ion()
plt.style.use('seaborn-darkgrid')
fig = plt.figure(figsize=(10, 10))
ax_spatial = fig.add_subplot(1, 1, 1)
circs = []
sctZ = None
# Inference
lbs = []
aux_lbs = []
n_iters = 0
for i in range(args.maxIter * (N / BATCH_SIZE)):
# Sample xn
idx = np.random.randint(N, size=BATCH_SIZE)
x_batch = xn[idx, :]
# Variational parameter updates
lambda_pi = update_lambda_pi(lambda_pi, lambda_phi[idx, :],
alpha_o)
Nks = np.sum(lambda_phi[idx, :], axis=0)
lambda_beta = update_lambda_beta(lambda_beta, beta_o, Nks)
lambda_nu = update_lambda_nu(lambda_nu, nu_o, Nks)
lambda_m = update_lambda_m(lambda_m, lambda_phi[idx, :],
lambda_beta, m_o, beta_o, x_batch,
BATCH_SIZE, D)
lambda_w = update_lambda_w(lambda_w, lambda_phi[idx, :],
lambda_beta, lambda_m, w_o, beta_o, m_o,
x_batch, K, BATCH_SIZE, D)
lambda_phi = update_lambda_phi(lambda_phi, lambda_pi, lambda_m,
lambda_nu, lambda_w, lambda_beta,
xn, K, D, idx)
# ELBO computation and variational parameter updates
lb = elbo2(x_batch, alpha_o, lambda_pi, lambda_phi[idx, :], m_o,
lambda_m, beta_o, lambda_beta, nu_o, lambda_nu, w_o,
inv(lambda_w), BATCH_SIZE, K)
lb = lb * (N / BATCH_SIZE)
aux_lbs.append(lb)
if len(aux_lbs) == (N / BATCH_SIZE):
lbs.append(np.mean(aux_lbs))
n_iters += 1
aux_lbs = []
if VERBOSE:
print('\n******* ITERATION {} *******'.format(n_iters))
print('lambda_pi: {}'.format(lambda_pi))
print('lambda_beta: {}'.format(lambda_beta))
print('lambda_nu: {}'.format(lambda_nu))
print('lambda_m: {}'.format(lambda_m))
print('lambda_w: {}'.format(lambda_w))
print('lambda_phi: {}'.format(lambda_phi[0:9, :]))
print('ELBO: {}'.format(lb))
if D == 2:
covs = [
lambda_w[k, :, :] / (lambda_nu[k] - D - 1)
for k in range(K)
]
ax_spatial, circs, sctZ = plot_iteration(
ax_spatial, circs, sctZ, lambda_m, covs, xn, i, K)
# Break condition
improve = lb - lbs[n_iters - 1] if n_iters > 0 else lb
if VERBOSE: print('Improve: {}'.format(improve))
if n_iters > 0 and 0 <= improve < THRESHOLD: break
zn = np.array([np.argmax(lambda_phi[n, :]) for n in xrange(N)])
if VERBOSE:
print('\n******* RESULTS *******')
for k in range(K):
print('Mu k{}: {}'.format(k, lambda_m[k, :]))
final_time = time()
exec_time = final_time - init_time
print('Time: {} seconds'.format(exec_time))
print('Iterations: {}'.format(n_iters))
print('ELBOs: {}'.format(lbs[len(lbs) - 10:len(lbs)]))
if D == 2: plt.savefig('generated/scavi_plot.png')
if D == 3:
fig = plt.figure()
ax = fig.add_subplot(111, projection='3d')
ax.scatter(xn[:, 0],
xn[:, 1],
xn[:, 2],
c=zn,
cmap=cm.gist_rainbow,
s=5)
ax.set_xlabel('X')
ax.set_ylabel('Y')
ax.set_zlabel('Z')
plt.show()
plt.gcf().clear()
plt.plot(np.arange(len(lbs)), lbs)
plt.ylabel('ELBO')
plt.xlabel('Iterations')
plt.savefig('generated/scavi_elbos.png')
if args.exportAssignments:
with open('generated/scavi_assignments.csv', 'wb') as output:
writer = csv.writer(output,
delimiter=';',
quotechar='',
escapechar='\\',
quoting=csv.QUOTE_NONE)
writer.writerow(['zn'])
for i in range(len(zn)):
writer.writerow([zn[i]])
if args.exportVariationalParameters:
with open('generated/scavi_variational_parameters.pkl',
'w') as output:
pkl.dump(
{
'lambda_pi': lambda_pi,
'lambda_m': lambda_m,
'lambda_beta': lambda_beta,
'lambda_nu': lambda_nu,
'lambda_w': lambda_w,
'K': K,
'D': D
}, output)
if args.exportELBOs:
with open('generated/scavi_elbos.pkl', 'w') as output:
pkl.dump({
'elbos': lbs,
'iter_time': exec_time / n_iters
}, output)
except IOError:
print('File not found!')
except Exception as e:
if e.args[0] == 'input_format': print('Input must be a pkl file')
elif e.args[0] == 'degrees_of_freedom':
print('Degrees of freedom can not be smaller than D!')
else:
print('Unexpected error: {}'.format(sys.exc_info()[0]))
raise
data = pkl.load(inputfile)
xn = data['xn']
N, D = xn.shape
if VERBOSE: init_time = time()
# Priors
alpha_o = np.array([1.0] * K)
nu_o = np.array([float(D)])
w_o = generate_random_positive_matrix(D)
m_o = np.array([0.0] * D)
beta_o = np.array([0.7])
# Variational parameters intialization
lambda_phi_var = np.random.dirichlet(alpha_o, N) \
if args.randomInit else init_kmeans(xn, N, K)
lambda_pi_var = np.zeros(shape=K)
lambda_beta_var = np.zeros(shape=K)
lambda_nu_var = np.zeros(shape=K) + D
lambda_m_var = np.random.uniform(np.min(xn[:, 0]), np.max(xn[:, 0]), (K, D))
lambda_w_var = np.array([np.copy(w_o) for _ in range(K)])
lambda_phi = tf.Variable(lambda_phi_var, trainable=False, dtype=tf.float64)
lambda_pi_var = tf.Variable(lambda_pi_var, dtype=tf.float64)
lambda_beta_var = tf.Variable(lambda_beta_var, dtype=tf.float64)
lambda_nu_var = tf.Variable(lambda_nu_var, dtype=tf.float64)
lambda_m = tf.Variable(lambda_m_var, dtype=tf.float64)
lambda_w_var = tf.Variable(lambda_w_var, dtype=tf.float64)
# Maintain numerical stability
lambda_pi = tf.nn.softplus(lambda_pi_var)
def main():
# Get data
with open('{}'.format(args.dataset), 'r') as inputfile:
data = pkl.load(inputfile)
xn = data['xn']
N, D = xn.shape
if VERBOSE: init_time = time()
# Priors
alpha_o = [1.0] * K
m_o = np.array([0.0, 0.0])
beta_o = 0.01
delta_o = np.zeros((D, D), long)
np.fill_diagonal(delta_o, 1)
# Variational parameters intialization
lambda_phi = np.random.dirichlet(alpha_o, N) \
if args.randomInit else init_kmeans(xn, N, K)
lambda_beta = beta_o + np.sum(lambda_phi, axis=0)
lambda_m = np.tile(1. / lambda_beta, (2, 1)).T * \
(beta_o * m_o + np.dot(lambda_phi.T, xn))
# Plot configs
if VERBOSE:
plt.ion()
fig = plt.figure(figsize=(10, 10))
ax_spatial = fig.add_subplot(1, 1, 1)
circs = []
sctZ = None
# Inference
n_iters = 0
lbs = []
for _ in range(args.maxIter):
# Variational parameter updates
lambda_pi = update_lambda_pi(lambda_phi, alpha_o)
lambda_phi = update_lambda_phi(lambda_pi, lambda_m, lambda_beta,
lambda_phi, delta_o, xn, N, D)
lambda_beta = update_lambda_beta(lambda_phi, beta_o)
lambda_m = update_lambda_m(lambda_beta, lambda_phi, m_o, beta_o, xn, D)
# ELBO computation
lb = elbo(xn, D, K, alpha_o, m_o, beta_o, delta_o, lambda_pi, lambda_m,
lambda_beta, lambda_phi)
lbs.append(lb)
if VERBOSE:
print('\n******* ITERATION {} *******'.format(n_iters))
print('lambda_pi: {}'.format(lambda_pi))
print('lambda_beta: {}'.format(lambda_beta))
print('lambda_m: {}'.format(lambda_m))
print('lambda_phi: {}'.format(lambda_phi[0:9, :]))
print('ELBO: {}'.format(lb))
ax_spatial, circs, sctZ = plot_iteration(ax_spatial, circs, sctZ,
lambda_m, delta_o, xn,
n_iters, K)
# Break condition
improve = lb - lbs[n_iters - 1]
if VERBOSE: print('Improve: {}'.format(improve))
if (n_iters == (args.maxIter - 1)) \
or (n_iters > 0 and 0 < improve < THRESHOLD):
if VERBOSE and D == 2: plt.savefig('generated/plot.png')
break
n_iters += 1
if VERBOSE:
print('\n******* RESULTS *******')
for k in range(K):
print('Mu k{}: {}'.format(k, lambda_m[k, :]))
final_time = time()
exec_time = final_time - init_time
print('Time: {} seconds'.format(exec_time))
print('Iterations: {}'.format(n_iters))
print('ELBOs: {}'.format(lbs))