def M_step(log_R, log_activation, vote, lambda_val=0.01):
R_shape = tf.shape(log_R)
log_R = log_R + log_activation
R_sum_i = cl.reduce_sum(tf.exp(log_R), axis=-3, keepdims=True)
log_normalized_R = log_R - \
tf.reduce_logsumexp(log_R, axis=-3, keepdims=True)
pose = cl.reduce_sum(vote * tf.exp(log_normalized_R),
axis=-3,
keepdims=True)
log_var = tf.reduce_logsumexp(log_normalized_R +
cl.log(tf.square(vote - pose)),
axis=-3,
keepdims=True)
shape = [1 for i in range(len(pose.shape) - 2)] + [pose.shape[-2], 1]
beta_v = tf.Variable(initial_value=tf.keras.initializers.TruncatedNormal(
mean=15., stddev=3.)(shape),
name="beta_v",
shape=shape)
cost = R_sum_i * (beta_v + 0.5 * log_var)
beta_a = tf.Variable(initial_value=tf.keras.initializers.TruncatedNormal(
mean=100.0, stddev=10)(shape),
name="beta_a",
shape=shape)
cost_sum_h = cl.reduce_sum(cost, axis=-1, keepdims=True)
logit = lambda_val * (beta_a - cost_sum_h)
log_activation = tf.math.log_sigmoid(logit)
return (pose, log_var, log_activation)
def E_step(pose, log_var, log_activation, vote):
normalized_vote = cl.divide(tf.square(vote - pose), 2 * tf.exp(log_var))
log_probs = normalized_vote + cl.log(2 * np.pi) + log_var
log_probs = -0.5 * cl.reduce_sum(log_probs, axis=-1, keepdims=True)
log_activation_logit = log_activation + log_probs
log_activation_logit = log_probs
log_R = log_activation_logit - tf.reduce_logsumexp(log_activation_logit, axis=-2, keepdims=True)
return log_R
def M_step(self, log_R, log_activation, vote, lambda_val=0.01):
log_R = log_R + log_activation
R_sum_i = cl.reduce_sum(tf.exp(log_R), axis=-3, keepdims=True)
log_normalized_R = log_R - \
tf.reduce_logsumexp(log_R, axis=-3, keepdims=True)
pose = cl.reduce_sum(vote * tf.exp(log_normalized_R),
axis=-3,
keepdims=True)
log_var = tf.reduce_logsumexp(log_normalized_R +
cl.log(tf.square(vote - pose)),
axis=-3,
keepdims=True)
cost = R_sum_i * (self.beta_v + 0.5 * log_var)
cost_sum_h = cl.reduce_sum(cost, axis=-1, keepdims=True)
logit = lambda_val * (self.beta_a - cost_sum_h)
log_activation = tf.math.log_sigmoid(logit)
return (pose, log_var, log_activation)
def M_step(log_R, log_activation, vote, lambda_val=0.01):
R_shape = tf.shape(log_R)
log_R = log_R + log_activation
R_sum_i = cl.reduce_sum(tf.exp(log_R), axis=-3, keepdims=True)
log_normalized_R = log_R - tf.reduce_logsumexp(log_R, axis=-3, keepdims=True)
pose = cl.reduce_sum(vote * tf.exp(log_normalized_R), axis=-3, keepdims=True)
log_var = tf.reduce_logsumexp(log_normalized_R + cl.log(tf.square(vote - pose)), axis=-3, keepdims=True)
beta_v = tf.get_variable('beta_v',
shape=[1 for i in range(len(pose.shape) - 2)] + [pose.shape[-2], 1],
initializer=tf.truncated_normal_initializer(mean=15., stddev=3.))
cost = R_sum_i * (beta_v + 0.5 * log_var)
beta_a = tf.get_variable('beta_a',
shape=[1 for i in range(len(pose.shape) - 2)] + [pose.shape[-2], 1],
initializer=tf.truncated_normal_initializer(mean=100.0, stddev=10))
cost_sum_h = cl.reduce_sum(cost, axis=-1, keepdims=True)
logit = lambda_val * (beta_a - cost_sum_h)
log_activation = tf.log_sigmoid(logit)
return(pose, log_var, log_activation)