def _get_attrank_cross_entropy(labels, logits):
# logits is not safe based on their satement
# do not use this function directly elsewhere
results = labels * math_fns.safe_log(logits) + (
1 - labels) * math_fns.safe_log(1 - logits)
results = (-1) * results
results = tf.reduce_mean(results)
return results
def loop_body(iteration, loss, exp_sum, exps):
temp = tf.gather(exps, [iteration])
temp = tf.reshape(temp, [])
exp_sum = tf.subtract(exp_sum, temp)
# clip exp_sum for safer log
loss = tf.add(loss, math_fns.safe_log(exp_sum))
return tf.add(iteration, 1), loss, exp_sum, exps
def get_listmle_loss(labels, predicted_scores):
"""
listwise learning-to-rank listMLE loss
Note: Simplified MLE formula is used in here (omit the proof in here)
\sum_{s=1}^{n-1} (-predicted_scores + ln(\sum_{i=s}^n exp(predicted_scores)))
n is tf.shape(predicted_scores)[0]
Check paper http://icml2008.cs.helsinki.fi/papers/167.pdf for more information
Args:
labels: a dense tensor of shape [n_data, 1]
n_data is the number of tweet candidates in a BatchPredictionRequest
predicted_scores: a dense tensor of same shape and type as labels
Returns:
average loss
"""
labels = tf.reshape(labels, [-1, 1])
n_data = tf.shape(labels)[0]
predicted_scores = tf.reshape(predicted_scores, [-1, 1])
predicted_scores_ordered_by_labels = _get_ordered_predicted_scores(
labels, predicted_scores, n_data)
loss = (-1) * tf.reduce_sum(predicted_scores)
# sum over 1 to n_data - 1
temp = tf.gather(predicted_scores_ordered_by_labels, [n_data - 1])
temp = tf.reshape(temp, [])
loss = tf.add(loss, temp)
exps = tf.exp(predicted_scores_ordered_by_labels)
exp_sum = tf.reduce_sum(exps)
# clip exp_sum for safer log
loss = tf.add(loss, math_fns.safe_log(exp_sum))
iteration = tf.constant(0)
def _cond(iteration, loss, exp_sum, exps):
return tf.less(iteration, n_data - 2)
def _gen_loop_body():
def loop_body(iteration, loss, exp_sum, exps):
temp = tf.gather(exps, [iteration])
temp = tf.reshape(temp, [])
exp_sum = tf.subtract(exp_sum, temp)
# clip exp_sum for safer log
loss = tf.add(loss, math_fns.safe_log(exp_sum))
return tf.add(iteration, 1), loss, exp_sum, exps
return loop_body
iteration, loss, exp_sum, exps = tf.while_loop(
_cond, _gen_loop_body(), (iteration, loss, exp_sum, exps))
loss = loss / tf.cast(n_data, dtype=tf.float32)
loss += get_listnet_loss(labels, predicted_scores)
return loss
def _get_listnet_cross_entropy(labels, logits):
"""
Used in listnet
cross entropy on top-one probabilities
between ideal/label top-one probabilities
and predicted/logits top-one probabilities
for a query/batch/batchPreidictionRequest
"""
# it is safe to use log on logits
# that come from _get_top_one_probs
# do not use this function directly elsewhere
results = (-1) * labels * math_fns.safe_log(logits)
return results
def loop_body(iteration, loss, exp_sum, G_t):
temp_exp = tf.gather(exps, [iteration])
temp_exp = tf.reshape(temp_exp, [])
exp_sum = tf.add(exp_sum, temp_exp)
temp_reward = tf.gather(dcg_k, [iteration])
temp_reward = tf.reshape(temp_reward, [])
G_t = tf.add(G_t, temp_reward)
# clip exp_sum for safer log
log_exp_sum = math_fns.safe_log(exp_sum)
negative_prediction = (-1) * tf.gather(sorted_predictions,
[iteration])
negative_prediction = tf.reshape(negative_prediction, [])
temp_sum = tf.add(negative_prediction, log_exp_sum)
temp_loss = tf.multiply(G_t, temp_sum)
loss = tf.add(loss, temp_loss)
return tf.subtract(iteration, 1), loss, exp_sum, G_t