def compute_elbo_smm(y, reconstructions, theta, phi_tilde, x_k_samps, log_z_given_y_phi, decoder_type):
# ELBO for latent SMM
with tf.name_scope('elbo'):
# unpack phi_gmm and compute expected theta
mu_theta, sigma_theta = unpack_smm(theta[1:3])
alpha_k = dirichlet.natural_to_standard(theta[0])
expected_log_pi_theta = dirichlet.expected_log_pi(alpha_k)
dof = theta[3] # Student-t degrees of freedom
# make sure that gradient is not propagated through stochastic Dirichlet parameter
with tf.name_scope('block_backprop'):
expected_log_pi_theta = tf.stop_gradient(expected_log_pi_theta)
dof = tf.stop_gradient(dof)
r_nk = tf.exp(log_z_given_y_phi)
# compute negative reconstruction error; sum over minibatch (use VAE function)
means, out_2 = reconstructions # out_2 is either gaussian variances or bernoulli logits.
if decoder_type == 'standard':
neg_reconstruction_error = vae.expected_diagonal_gaussian_loglike(y, means, out_2, weights=r_nk)
elif decoder_type == 'bernoulli':
neg_reconstruction_error = vae.expected_bernoulli_loglike(y, out_2, r_nk=r_nk)
else:
raise NotImplementedError
eta1_phi_tilde, eta2_phi_tilde = phi_tilde
N, K, L, _ = eta2_phi_tilde.get_shape().as_list()
eta1_phi_tilde = tf.reshape(eta1_phi_tilde, (N, K, L))
with tf.name_scope('compute_regularizer'):
# compute E[log q_phi(x,z=k|y)]
with tf.name_scope('log_numerator'):
log_N_x_given_phi = gaussian.log_probability_nat_per_samp(x_k_samps, eta1_phi_tilde, eta2_phi_tilde)
log_numerator = log_N_x_given_phi + tf.expand_dims(log_z_given_y_phi, axis=2)
with tf.name_scope('log_denominator'):
# compute E[log p_theta(x,z=k)]
log_N_x_given_theta = student_t.log_probability_per_samp(x_k_samps, mu_theta, sigma_theta, dof)
log_denominator = log_N_x_given_theta + tf.expand_dims(tf.expand_dims(expected_log_pi_theta, axis=0), axis=2)
regularizer_term = tf.reduce_mean(
tf.reduce_sum(
tf.reduce_sum(
tf.multiply(tf.expand_dims(r_nk, axis=2),
log_numerator - log_denominator),
axis=1), # weighted average over components
axis=0) # sum over data points
) # mean over samples
elbo = tf.subtract(neg_reconstruction_error, regularizer_term, name='elbo')
with tf.name_scope('elbo_summaries'):
details = tf.tuple((neg_reconstruction_error,
tf.reduce_sum(tf.multiply(r_nk, tf.reduce_mean(log_numerator, -1))),
tf.reduce_sum(tf.multiply(r_nk, tf.reduce_mean(log_denominator, -1))),
regularizer_term), name='debug')
return elbo, details
def expected_values(self):
_, m, C, v = niw.natural_to_standard(self.A, self.b, self.beta,
self.v_hat)
exp_log_pi = dirichlet.expected_log_pi(
dirichlet.natural_to_standard(self.alpha))
with tf.name_scope('niw_expectation'):
exp_m = tf.identity(m, 'expected_mean')
C_inv = tf.matrix_inverse(C)
C_inv_sym = tf.divide(tf.add(C_inv, tf.matrix_transpose(C_inv)),
2.,
name='C_inv_symmetrised')
exp_C = tf.matrix_inverse(tf.multiply(C_inv_sym,
tf.expand_dims(
tf.expand_dims(v, 1), 2),
name='expected_precision'),
name='expected_covariance')
return exp_log_pi, exp_m, exp_C
if lbl_tr is not None:
entr_tr, prty_tr = purity(tf.exp(log_z_given_y_phi),
lbl_tr)
tf.summary.scalar('entropy_tr', entr_tr)
tf.summary.scalar('purity_tr', prty_tr)
# useful values for tensorboard and plotting
with tf.name_scope('plotting_prep'):
if 'smm' in config['method']:
mu, sigma = svae.unpack_smm(theta[1:3])
else:
beta_k, m_k, C_k, v_k = niw.natural_to_standard(
theta[1], theta[2], theta[3], theta[4])
mu, sigma = niw.expected_values((beta_k, m_k, C_k, v_k))
alpha_k = dirichlet.natural_to_standard(theta[0])
expected_log_pi = dirichlet.expected_log_pi(alpha_k)
pi_theta = tf.exp(expected_log_pi)
theta_plot = mu, sigma, pi_theta
q_z_given_y_phi = tf.exp(log_z_given_y_phi)
neg_normed_elbo = -tf.divide(tf.reduce_sum(tower_elbo),
size_minibatch)
tf.summary.scalar('elbo/elbo_normed', neg_normed_elbo)
tf.summary.scalar(
'elbo/neg_rec_err',
tf.divide(tf.reduce_sum(tower_neg_rec_err),
size_minibatch))
tf.summary.scalar(
'elbo/regularizer',
tf.divide(tf.reduce_sum(tower_regularizer),
size_minibatch))
def compute_elbo(y, reconstructions, theta, phi_tilde, x_k_samps, log_z_given_y_phi, decoder_type):
# ELBO for latent GMM
with tf.name_scope('elbo'):
# unpack phi_gmm and compute expected theta
with tf.name_scope('expct_theta_to_nat'):
beta_k, m_k, C_k, v_k = niw.natural_to_standard(*theta[1:])
mu, sigma = niw.expected_values((beta_k, m_k, C_k, v_k))
eta1_theta, eta2_theta = gaussian.standard_to_natural(mu, sigma)
alpha_k = dirichlet.natural_to_standard(theta[0])
expected_log_pi_theta = dirichlet.expected_log_pi(alpha_k)
# do not backpropagate through GMM
with tf.name_scope('block_backprop'):
eta1_theta = tf.stop_gradient(eta1_theta)
eta2_theta = tf.stop_gradient(eta2_theta)
expected_log_pi_theta = tf.stop_gradient(expected_log_pi_theta)
r_nk = tf.exp(log_z_given_y_phi)
# compute negative reconstruction error; sum over minibatch (use VAE function)
means, out_2 = reconstructions # out_2 is either gaussian variances or bernoulli logits.
if decoder_type == 'standard':
neg_reconstruction_error = vae.expected_diagonal_gaussian_loglike(y, means, out_2, weights=r_nk)
elif decoder_type == 'bernoulli':
neg_reconstruction_error = vae.expected_bernoulli_loglike(y, out_2, r_nk=r_nk)
else:
raise NotImplementedError
# compute E[log q_phi(x,z=k|y)]
eta1_phi_tilde, eta2_phi_tilde = phi_tilde
N, K, L, _ = eta2_phi_tilde.get_shape().as_list()
eta1_phi_tilde = tf.reshape(eta1_phi_tilde, (N, K, L))
N, K, S, L = x_k_samps.get_shape().as_list()
with tf.name_scope('compute_regularizer'):
with tf.name_scope('log_numerator'):
log_N_x_given_phi = gaussian.log_probability_nat_per_samp(x_k_samps, eta1_phi_tilde, eta2_phi_tilde)
log_numerator = log_N_x_given_phi + tf.expand_dims(log_z_given_y_phi, axis=2)
with tf.name_scope('log_denominator'):
log_N_x_given_theta = gaussian.log_probability_nat_per_samp(x_k_samps,
tf.tile(tf.expand_dims(eta1_theta, axis=0), [N, 1, 1]),
tf.tile(tf.expand_dims(eta2_theta, axis=0), [N, 1, 1, 1]))
log_denominator = log_N_x_given_theta + tf.expand_dims(tf.expand_dims(expected_log_pi_theta, axis=0), axis=2)
regularizer_term = tf.reduce_mean(
tf.reduce_sum(
tf.reduce_sum(
tf.multiply(tf.expand_dims(r_nk, axis=2),
log_numerator - log_denominator),
axis=1), # weighted average over components
axis=0) # sum over minibatch
) # mean over samples
elbo = tf.subtract(neg_reconstruction_error, regularizer_term, name='elbo')
with tf.name_scope('elbo_summaries'):
details = tf.tuple((neg_reconstruction_error,
tf.reduce_sum(tf.multiply(r_nk, tf.reduce_mean(log_numerator, -1))),
tf.reduce_sum(tf.multiply(r_nk, tf.reduce_mean(log_denominator, -1))),
regularizer_term), name='debug')
return elbo, details
def visualize_svae(ax,
config,
log_path,
ratio_tr=0.7,
nb_samples=20,
grid_density=100,
window=((-20, 20), (-20, 20)),
param_device='/cpu:0'):
with tf.device(param_device):
if config['dataset'] in ['mnist', 'fashion']:
binarise = True
size_minibatch = 1024
output_type = 'bernoulli'
else:
binarise = False
size_minibatch = -1
output_type = 'standard'
# First we build the model graph so that we can load the learned parameters from a checkpoint.
# Initialisations don't matter, they'll be overwritten with saver.restore().
data, lbl, _, _ = make_minibatch(config['dataset'],
path_datadir='../datasets',
ratio_tr=ratio_tr,
seed_split=0,
size_minibatch=size_minibatch,
size_testbatch=-1,
binarise=binarise)
# define nn-architecture
encoder_layers = [(config['U'], tf.tanh), (config['U'], tf.tanh),
(config['L'], 'natparam')]
decoder_layers = [(config['U'], tf.tanh), (config['U'], tf.tanh),
(int(data.get_shape()[1]), output_type)]
sample_size = 100
if config['dataset'] in ['mnist', 'fashion']:
data = tf.where(tf.equal(data, -1),
tf.zeros_like(data, dtype=tf.float32),
tf.ones_like(data, dtype=tf.float32))
with tf.name_scope('model'):
gmm_prior, theta = svae.init_mm(config['K'],
config['L'],
seed=config['seed'],
param_device='/gpu:0')
theta_copied = niw.natural_to_standard(tf.identity(gmm_prior[1]),
tf.identity(gmm_prior[2]),
tf.identity(gmm_prior[3]),
tf.identity(gmm_prior[4]))
_, sigma_k = niw.expected_values(theta_copied)
pi_k_init = tf.nn.softmax(
tf.random_normal(shape=(config['K'], ),
mean=0.0,
stddev=1.,
seed=config['seed']))
L_k = tf.cholesky(sigma_k)
mu_k = tf.random_normal(shape=(config['K'], config['L']),
stddev=1,
seed=config['seed'])
with tf.variable_scope('phi_gmm'):
mu_k = variable_on_device('mu_k',
shape=None,
initializer=mu_k,
trainable=True,
device=param_device)
L_k = variable_on_device('L_k',
shape=None,
initializer=L_k,
trainable=True,
device=param_device)
pi_k = variable_on_device('log_pi_k',
shape=None,
initializer=pi_k_init,
trainable=True,
device=param_device)
phi_gmm = mu_k, L_k, pi_k
_ = vae.make_encoder(data,
layerspecs=encoder_layers,
stddev_init=.1,
seed=config['seed'])
with tf.name_scope('random_sampling'):
# compute expected theta_pgm
beta_k, m_k, C_k, v_k = niw.natural_to_standard(*theta[1:])
alpha_k = dirichlet.natural_to_standard(theta[0])
mean, cov = niw.expected_values((beta_k, m_k, C_k, v_k))
expected_log_pi = dirichlet.expected_log_pi(alpha_k)
pi = tf.exp(expected_log_pi)
# sample from prior (first from
x_k_samples = tf.contrib.distributions.MultivariateNormalFullCovariance(
loc=mean, covariance_matrix=cov).sample(sample_size)
z_samples = tf.multinomial(logits=tf.reshape(tf.log(pi), (1, -1)),
num_samples=sample_size,
name='k_samples')
z_samples = tf.squeeze(z_samples)
assert z_samples.get_shape() == (sample_size, )
assert x_k_samples.get_shape() == (sample_size, config['K'],
config['L'])
# compute reconstructions
y_k_samples, _ = vae.make_decoder(x_k_samples,
layerspecs=decoder_layers,
stddev_init=.1,
seed=config['seed'])
assert y_k_samples.get_shape() == (sample_size, config['K'],
data.get_shape()[1])
with tf.name_scope('cluster_sample_data'):
tf.get_variable_scope().reuse_variables()
_, clustering = svae.predict(data,
phi_gmm,
encoder_layers,
decoder_layers,
seed=config['seed'])
# load trained model
saver = tf.train.Saver()
model_path = log_path + '/' + generate_log_id(config)
print(model_path)
latest_ckpnt = tf.train.latest_checkpoint(model_path)
sess_config = tf.ConfigProto(allow_soft_placement=True)
sess = tf.Session(config=sess_config)
saver.restore(sess, latest_ckpnt)
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(sess=sess, coord=coord)
collected_y_samps = []
collected_z_samps = []
for s in range(nb_samples):
y_samps, z_samps = sess.run((y_k_samples, z_samples))
collected_y_samps.append(y_samps)
collected_z_samps.append(z_samps)
collected_y_samps = np.concatenate(collected_y_samps, axis=0)
collected_z_samps = np.concatenate(collected_z_samps, axis=0)
assert collected_y_samps.shape == (nb_samples * sample_size,
config['K'], data.shape[1])
assert collected_z_samps.shape == (nb_samples * sample_size, )
# use 300 sample points from the dataset
data, lbl, clustering = sess.run(
(data[:300], lbl[:300], clustering[:300]))
# compute PCA if necessary
samples_2d = []
if data.shape[1] > 2:
pca = PCA(n_components=2).fit(data)
data2d = pca.transform(data)
for z_samples in range(config['K']):
chosen = collected_z_samps == z_samples
samps_k = collected_y_samps[chosen, z_samples, :]
if samps_k.size > 0:
samples_2d.append(pca.transform(samps_k))
else:
data2d = data
for z_samples in range(config['K']):
chosen = (collected_z_samps == z_samples)
samps_k = collected_y_samps[chosen, z_samples, :]
if samps_k.size > 0:
samples_2d.append(samps_k)
# plot 2d-histogram (one histogram for each of the K components)
from matplotlib.colors import LogNorm
for z_samples, samples in enumerate(samples_2d):
ax.hist2d(samples[:, 0],
samples[:, 1],
bins=grid_density,
range=window,
cmap=make_colormap(dark_colors[z_samples %
len(dark_colors)]),
normed=True,
norm=LogNorm())
# overlay histogram with sample datapoints (coloured according to their most likely cluster allocation)
labels = np.argmax(lbl, axis=1)
for c in np.unique(labels):
in_class_c = (labels == c)
color = bright_colors[int(c % len(bright_colors))]
marker = markers[int(c % len(markers))]
ax.scatter(data2d[in_class_c, 0],
data2d[in_class_c, 1],
c=color,
marker=marker,
s=data_dot_size,
linewidths=0)