P(examination | click) and P(relevance | click).
with tf.compat.v1.name_scope(name="compute_latent_prob"):
exam_prob = tf.math.sigmoid(tf.cast(exam_logits, tf.float32))
rel_prob = tf.math.sigmoid(tf.cast(rel_logits, tf.float32))
is_clicked = tf.greater_equal(tf.cast(clicks, tf.float32), 1.0)
prob_non_clicks = 1 - exam_prob * rel_prob + 1e-6
exam_prob_given_non_clicks = exam_prob * (1 - rel_prob) / prob_non_clicks
rel_prob_given_non_clicks = (1 - exam_prob) * rel_prob / prob_non_clicks
exam_prob_given_non_clicks, rel_prob_given_non_clicks = [
tf.stop_gradient(
tf.compat.v1.where(is_clicked, tf.ones_like(prob), prob))
for prob in [exam_prob_given_non_clicks, rel_prob_given_non_clicks]
]
return exam_prob_given_non_clicks, rel_prob_given_non_clicks
def compute_unreduced_loss(self, labels, logits):
Computes the loss for each element.
After Change
rel_prob_posterior = tf.compat.v1.where(
is_clicked, tf.ones_like(rel_logits_posterior),
tf.sigmoid(rel_logits_posterior))
return tf.stop_gradient(exam_prob_posterior), tf.stop_gradient(
rel_prob_posterior)
def compute_unreduced_loss(self, labels, logits):
Computes the loss for each element.
