def add_least_gaussian2d_ops(self, opts):
""" Add ops searching for the 2d plane in z_dim hidden space
corresponding to the 'least Gaussian' look of the sample
"""
with tf.variable_scope('leastGaussian2d'):
# Projection matrix which we are going to tune
sample_ph = tf.placeholder(
tf.float32, [None, opts['latent_space_dim']],
name='sample_ph')
v = tf.get_variable(
"proj_v", [opts['latent_space_dim'], 1],
tf.float32, tf.random_normal_initializer(stddev=1.))
u = tf.get_variable(
"proj_u", [opts['latent_space_dim'], 1],
tf.float32, tf.random_normal_initializer(stddev=1.))
npoints = tf.cast(tf.shape(sample_ph)[0], tf.int32)
# First we need to make sure projection matrix is orthogonal
v_norm = tf.nn.l2_normalize(v, 0)
dotprod = tf.reduce_sum(tf.multiply(u, v_norm))
u_ort = u - dotprod * v_norm
u_norm = tf.nn.l2_normalize(u_ort, 0)
Mproj = tf.concat([v_norm, u_norm], 1)
sample_proj = tf.matmul(sample_ph, Mproj)
a = tf.eye(npoints) - tf.ones([npoints, npoints]) / tf.cast(npoints, tf.float32)
b = tf.matmul(sample_proj, tf.matmul(a, a), transpose_a=True)
b = tf.matmul(b, sample_proj)
# Sample covariance matrix
covhat = b / (tf.cast(npoints, tf.float32) - 1)
# covhat = tf.Print(covhat, [covhat], 'Cov:')
with tf.variable_scope('leastGaussian2d'):
gcov = opts['pot_pz_std'] * opts['pot_pz_std'] * tf.eye(2)
# l2 distance between sample cov and the Gaussian cov
projloss = tf.reduce_sum(tf.square(covhat - gcov))
# Also account for the first moment, i.e. expected value
projloss += tf.reduce_sum(tf.square(tf.reduce_mean(sample_proj, 0)))
# We are maximizing
projloss = -projloss
optim = tf.train.AdamOptimizer(0.001, 0.9)
optim = optim.minimize(projloss, var_list=[v, u])
self._proj_u = u_norm
self._proj_v = v_norm
self._proj_sample_ph = sample_ph
self._proj_covhat = covhat
self._proj_loss = projloss
self._proj_optim = optim
评论列表
文章目录
