def score_documents_tf(user_obs,
doc_obs,
no_click_mass=1.0,
is_mnl=False,
min_normalizer=-1.0):
"""Computes unnormalized scores given both user and document observations.
This implements both multinomial proportional model and multinormial logit
model given some parameters. We also assume scores are based on inner
products of user_obs and doc_obs.
Args:
user_obs: An instance of AbstractUserState.
doc_obs: A numpy array that represents the observation of all documents in
the candidate set.
no_click_mass: a float indicating the mass given to a no click option
is_mnl: whether to use a multinomial logit model instead of a multinomial
proportional model.
min_normalizer: A float (<= 0) used to offset the scores to be positive when
using multinomial proportional model.
Returns:
A float tensor that stores unnormalzied scores of documents and a float
tensor that represents the score for the action of picking no document.
"""
user_obs = tf.reshape(user_obs, [1, -1])
scores = tf.reduce_sum(input_tensor=tf.multiply(user_obs, doc_obs), axis=1)
all_scores = tf.concat([scores, tf.constant([no_click_mass])], axis=0)
if is_mnl:
all_scores = tf.nn.softmax(all_scores)
else:
all_scores = all_scores - min_normalizer
return all_scores[:-1], all_scores[-1]
def projection_dist(states):
inner = tf.multiply(states - starting_states, goals - starting_states)
upper = tf.reduce_sum(inner, -1)
sign = tf.sign(upper)
result = tf.math.divide(upper, tf.norm(goals - starting_states, ord=2))
term_1 = tf.norm(states - starting_states, 2)
return -1*term_1+result
def normalized_dist(states):
inner = tf.multiply(states - starting_states, goals - starting_states)
upper = tf.reduce_sum(inner, -1)
sign = tf.sign(upper)
result = sign * tf.square(tf.math.divide(upper, tf.norm(goals - starting_states, ord=2)))
term_1 = tf.square(tf.norm(states - starting_states, 2))
term_2 = tf.square(tf.math.divide(upper, tf.norm(goals - starting_states, ord=2)))
return tf.sqrt(epsilon + tf.abs(result - alpha * (term_1 - term_2)))
def kaf(linear, name, kernel='rbf', D=None, gamma=None):
if D is None:
D = tf.linspace(start=-2., stop=2., num=20)
with tf.variable_scope('kaf', reuse=tf.AUTO_REUSE):
if kernel == "rbf":
K = gauss_kernel(linear, D, gamma=gamma)
alpha = tf.get_variable(name, shape=(1, linear.get_shape()[-1], D.get_shape()[0]),
initializer=tf.random_normal_initializer(stddev=0.1))
elif kernel == 'rbf2d':
Dx, Dy = tf.meshgrid(D, D)
K = gauss_kernel2D(linear, Dx, Dy, gamma=gamma)
alpha = tf.get_variable(name,
shape=(1, linear.get_shape()[-1] // 2, D.get_shape()[0] * D.get_shape()[0]),
initializer=tf.random_normal_initializer(stddev=0.1))
else:
raise NotImplementedError()
act = tf.reduce_sum(tf.multiply(K, alpha), axis=-1)
# act = tf.squeeze(act, axis=0)
return act
def __call__(self, x):
K = gauss_kernel(x, self.d, self.gamma)
return tf.reduce_sum(tf.multiply(K, self.alpha), axis=-1)