I have a multi-dimensional parameter,
beta = Normal(loc=tf.zeros([FLAGS.S, FLAGS.S, FLAGS.D], dtype=tf.float32), scale=1.0 * tf.ones(FLAGS.D, dtype=tf.float32))
and its variance approximation
qbeta_loc = tf.get_variable("qbeta_loc", [FLAGS.S, FLAGS.S, FLAGS.D], dtype=tf.float32)
qbeta_scale = tf.nn.softplus(tf.get_variable("qbeta_scale", [FLAGS.S, FLAGS.S, FLAGS.D], dtype=tf.float32))
qbeta = Normal(loc=qbeta_loc, scale=qbeta_scale)
But I know some of its element are going to be always 0. Is there a way to specify this?