def _define_maximization_operation(self, num_batches):
"""Maximization operations."""
# TODO(xavigonzalvo): some of these operations could be moved to C++.
# Compute the effective number of data points assigned to component k.
with tf.control_dependencies(self._w):
points_in_k = tf.squeeze(tf.add_n(self._points_in_k), squeeze_dims=[0])
# Update alpha.
if 'w' in self._params:
final_points_in_k = points_in_k / num_batches
num_examples = tf.to_float(tf.reduce_sum(final_points_in_k))
self._alpha_op = self._alpha.assign(
final_points_in_k / (num_examples + MEPS))
else:
self._alpha_op = tf.no_op()
self._train_ops = [self._alpha_op]
# Update means.
points_in_k_expanded = tf.reshape(points_in_k,
[self._num_classes, 1, 1])
if 'm' in self._params:
self._means_op = self._means.assign(
tf.div(tf.add_n(self._w_mul_x), points_in_k_expanded + MEPS))
else:
self._means_op = tf.no_op()
# means are (num_classes x 1 x dims)
# Update covariances.
with tf.control_dependencies([self._means_op]):
b = tf.add_n(self._w_mul_x2) / (points_in_k_expanded + MEPS)
new_covs = []
for k in range(self._num_classes):
mean = self._means.value()[k, :, :]
square_mean = tf.matmul(mean, mean, transpose_a=True)
new_cov = b[k, :, :] - square_mean + self._min_var
if self._covariance_type == FULL_COVARIANCE:
new_covs.append(tf.expand_dims(new_cov, 0))
elif self._covariance_type == DIAG_COVARIANCE:
new_covs.append(tf.expand_dims(tf.diag_part(new_cov), 0))
new_covs = tf.concat(0, new_covs)
if 'c' in self._params:
# Train operations don't need to take care of the means
# because covariances already depend on it.
with tf.control_dependencies([self._means_op, new_covs]):
self._train_ops.append(
tf.assign(self._covs, new_covs, validate_shape=False))
评论列表
文章目录
