def k_diag(self, input_x, return_full=True):
"""Iteratively building the diagonal part (variance) of the NNGP kernel.
Args:
input_x: tensor of input of size [num_data, input_dim].
return_full: boolean for output to be [num_data] sized or a scalar value
for normalized inputs
Sets self.layer_qaa_dict of {layer #: qaa at the layer}
Returns:
qaa: variance at the output.
"""
with tf.name_scope("Kdiag"):
# If normalized input length starts at 1.
if self.use_fixed_point_norm:
current_qaa = self.var_fixed_point
else:
current_qaa = self.weight_var * tf.convert_to_tensor(
[1.], dtype=tf.float64) + self.bias_var
self.layer_qaa_dict = {0: current_qaa}
for l in range(self.depth):
with tf.name_scope("layer_%d" % l):
samp_qaa = interp.interp_lin(self.var_aa_grid,
self.qaa_grid, current_qaa)
samp_qaa = self.weight_var * samp_qaa + self.bias_var
self.layer_qaa_dict[l + 1] = samp_qaa
current_qaa = samp_qaa
if return_full:
qaa = tf.tile(current_qaa[:1], ([input_x.shape[0].value]))
else:
qaa = current_qaa[0]
return qaa
def k_diag(self, input_x, return_full=True):
"""Iteratively building the diagonal part (variance) of the NNGP kernel.
Args:
input_x: tensor of input of size [num_data, input_dim].
return_full: boolean for output to be [num_data] sized or a scalar value
for normalized inputs
Sets self.layer_qaa_dict of {layer #: qaa at the layer}
Returns:
qaa: variance at the output.
"""
current_qaa = self.weight_var * np.array([1.]) + self.bias_var
self.layer_qaa_dict = {0: current_qaa}
for l in range(self.depth):
samp_qaa = interp.interp_lin(self.var_aa_grid, self.qaa_grid,
current_qaa)
samp_qaa = self.weight_var * samp_qaa + self.bias_var
self.layer_qaa_dict[l + 1] = samp_qaa
current_qaa = samp_qaa
if return_full:
qaa = np.repeat(current_qaa[:1], [input_x.shape[0]])
else:
qaa = current_qaa[0]
return qaa
def get_var_fixed_point(self):
with tf.name_scope("get_var_fixed_point"):
# If normalized input length starts at 1.
current_qaa = self.weight_var * tf.constant(
[1.], dtype=tf.float64) + self.bias_var
diff = 1.
prev_qaa_np = 1.
it = 0
while diff > 1e-6 and it < 300:
samp_qaa = interp.interp_lin(self.var_aa_grid, self.qaa_grid,
current_qaa)
samp_qaa = self.weight_var * samp_qaa + self.bias_var
current_qaa = samp_qaa
with tf.Session() as sess:
current_qaa_np = sess.run(current_qaa)
diff = np.abs(current_qaa_np - prev_qaa_np)
it += 1
prev_qaa_np = current_qaa_np
return current_qaa_np, current_qaa
