def _predicted_mean_op(self, activations):
activation, activation_size = activations[-1]
predicted_mean = model_utils.fully_connected(
activation,
activation_size,
self.output_window_size * self.num_features,
name="predicted_mean",
activation=None)
return array_ops.reshape(predicted_mean,
[-1, self.output_window_size, self.num_features])
def _predicted_mean_op(self, activations):
activation, activation_size = activations[-1]
predicted_mean = model_utils.fully_connected(
activation,
activation_size,
self.output_window_size * self.num_features,
name="predicted_mean",
activation=None)
return array_ops.reshape(predicted_mean,
[-1, self.output_window_size, self.num_features])
def _create_hidden_stack(self, activation, activation_size):
activations = []
for layer_number, layer_size in enumerate(self.hidden_layer_sizes):
# TODO(agarwal): Migrate to fully_connected in tf slim
activation = model_utils.fully_connected(
activation, activation_size, layer_size,
name="layer_{}".format(layer_number))
activation_size = layer_size
activations.append((activation, activation_size))
return activations
def _create_hidden_stack(self, activation, activation_size):
activations = []
for layer_number, layer_size in enumerate(self.hidden_layer_sizes):
# TODO(agarwal): Migrate to fully_connected in tf slim
activation = model_utils.fully_connected(
activation,
activation_size,
layer_size,
name="layer_{}".format(layer_number))
activation_size = layer_size
activations.append((activation, activation_size))
return activations
def _predicted_covariance_op(self, activations, num_values):
activation, activation_size = activations[-1]
if self.loss == ARModel.NORMAL_LIKELIHOOD_LOSS:
log_sigma_square = model_utils.fully_connected(
activation,
activation_size,
self.output_window_size * num_values,
name="log_sigma_square",
activation=None)
predicted_covariance = gen_math_ops.exp(log_sigma_square)
predicted_covariance = array_ops.reshape(
predicted_covariance, [-1, self.output_window_size, num_values])
else:
shape = array_ops.stack([
array_ops.shape(activation)[0],
constant_op.constant(self.output_window_size),
constant_op.constant(num_values)
])
predicted_covariance = array_ops.ones(shape=shape, dtype=activation.dtype)
return predicted_covariance
def _predicted_covariance_op(self, activations, num_values):
activation, activation_size = activations[-1]
if self.loss == ARModel.NORMAL_LIKELIHOOD_LOSS:
log_sigma_square = model_utils.fully_connected(
activation,
activation_size,
self.output_window_size * num_values,
name="log_sigma_square",
activation=None)
predicted_covariance = gen_math_ops.exp(log_sigma_square)
predicted_covariance = array_ops.reshape(
predicted_covariance, [-1, self.output_window_size, num_values])
else:
shape = array_ops.stack([
array_ops.shape(activation)[0],
constant_op.constant(self.output_window_size),
constant_op.constant(num_values)
])
predicted_covariance = array_ops.ones(shape=shape, dtype=activation.dtype)
return predicted_covariance
