def split2d_prior(z, hps):
n_z2 = int(z.get_shape()[3])
n_z1 = n_z2
h = conv2d_zeros(z, 2 * n_z1, name="conv")
mean, logsd = tf.split(h, 2, axis=-1)
rescale = tf.get_variable("rescale", [], initializer=tf.constant_initializer(1.))
scale_shift = tf.get_variable("scale_shift", [], initializer=tf.constant_initializer(0.))
logsd = tf.tanh(logsd) * rescale + scale_shift
return gaussian_diag(mean, logsd)
def split2d_prior(z, hps):
n_z2 = int(z.get_shape()[3])
n_z1 = n_z2
h = conv2d_zeros(z, 2 * n_z1, name="conv")
mean, logsd = tf.split(h, 2, axis=-1)
rescale = tf.get_variable("rescale", [],
initializer=tf.constant_initializer(1.))
scale_shift = tf.get_variable("scale_shift", [],
initializer=tf.constant_initializer(0.))
logsd = tf.tanh(logsd) * rescale + scale_shift
return gaussian_diag(mean, logsd)
def prior(y_onehot, hps, name=None):
n_z = hps.top_shape[-1]
h = tf.zeros([tf.shape(y_onehot)[0]]+hps.top_shape[:2]+[2*n_z])
h = conv2d_zeros(h, 2*n_z, name="p")
if hps.ycond:
h += tf.reshape(linear_zeros(y_onehot, 2*n_z, name="y_emb"), [-1, 1, 1, 2 * n_z])
mean, logsd = tf.split(h, 2, axis=-1)
rescale = tf.get_variable("rescale", [], initializer=tf.constant_initializer(1.))
scale_shift = tf.get_variable("scale_shift", [], initializer=tf.constant_initializer(0.))
logsd = tf.tanh(logsd) * rescale + scale_shift
pz = gaussian_diag(mean, logsd)
logp = lambda z1: pz.logp(z1)
eps = lambda z1: pz.get_eps(z1)
sample = lambda eps: pz.sample(eps)
return logp, sample, eps
def prior(name, y_onehot, cfg):
with tf.variable_scope(name):
cfg.top_shape = [1, 1, 768]
n_z = cfg.top_shape[-1]
h = tf.zeros([tf.shape(y_onehot)[0]]+cfg.top_shape[:2]+[2*n_z])
if cfg.learntop:
assert(False)
h = ops._conv2d('p', h, 2*n_z, 3, 1, True)
if cfg.ycond:
assert(False)
h += tf.reshape(ops.dense("y_emb", y_onehot, 2*n_z,
True, init_zero=True), [-1, 1, 1, 2 * n_z])
pz = ops.gaussian_diag(h[:, :, :, :n_z], h[:, :, :, n_z:])
def logp(z1):
objective = pz.logp(z1)
return objective
def sample(eps=None, eps_std=None, temp=1.0):
if eps is not None:
# Already sampled eps. Don't use eps_std
z = pz.sample_eps(eps)
elif eps_std is not None:
# Sample with given eps_std
z = pz.sample_eps(pz.eps * tf.reshape(eps_std, [-1, 1, 1, 1]))
else:
# Sample normally
z = pz.sample(temp)
return z
def eps(z1):
return pz.get_eps(z1)
return logp, sample, eps
def prior(y_onehot, hps, name=None):
n_z = hps.top_shape[-1]
h = tf.zeros([tf.shape(y_onehot)[0]] + hps.top_shape[:2] + [2 * n_z])
h = conv2d_zeros(h, 2 * n_z, name="p")
if hps.ycond:
h += tf.reshape(linear_zeros(y_onehot, 2 * n_z, name="y_emb"),
[-1, 1, 1, 2 * n_z])
mean, logsd = tf.split(h, 2, axis=-1)
rescale = tf.get_variable("rescale", [],
initializer=tf.constant_initializer(1.))
scale_shift = tf.get_variable("scale_shift", [],
initializer=tf.constant_initializer(0.))
logsd = tf.tanh(logsd) * rescale + scale_shift
pz = gaussian_diag(mean, logsd)
logp = lambda z1: pz.logp(z1)
eps = lambda z1: pz.get_eps(z1)
sample = lambda eps: pz.sample(eps)
return logp, sample, eps
def split2d_prior(z, cfg):
n_z = int(z.get_shape()[3])
h = f_('split2d_prior', z, cfg, n_out=n_z * 2)
mean = h[:, :, :, 0::2]
logs = h[:, :, :, 1::2]
return ops.gaussian_diag(mean, logs)