def get_pair_loss(pairwise_label_scores, pairwise_predicted_scores, params):
"""
Paiwise learning-to-rank ranknet loss
Check paper https://www.microsoft.com/en-us/research/publication/
learning-to-rank-using-gradient-descent/
for more information
Args:
pairwise_label_scores: a dense tensor of shape [n_data, n_data]
pairwise_predicted_scores: a dense tensor of shape [n_data, n_data]
n_data is the number of tweet candidates in a BatchPredictionRequest
params: network parameters
mask options: full_mask and diag_mask
Returns:
average loss over pairs defined by the masks
"""
if params.mask == "full_mask":
# full_mask that only covers pairs that have different labels
# (all pairwise_label_scores = 0.5: selfs and same labels are 0s)
mask, pair_count = masks.full_mask(pairwise_label_scores)
elif params.mask == "diag_mask":
# diag_mask that covers all pairs
# (only selfs/diags are 0s)
mask, pair_count = masks.diag_mask(pairwise_label_scores)
else:
mask, pair_count = masks.diag_mask(pairwise_label_scores)
# pairwise sigmoid_cross_entropy_with_logits loss
loss = tf.cond(
tf.equal(pair_count, 0),
lambda: 0., lambda: _get_average_cross_entropy_loss(
pairwise_label_scores, pairwise_label_scores, mask, pair_count))
return loss
def get_hinge_loss(pairwise_label_scores, pairwise_predicted_scores, params):
"""
Paiwise learning-to-rank ranknet loss
Check paper https://www.microsoft.com/en-us/research/publication/
learning-to-rank-using-gradient-descent/
for more information
Args:
pairwise_label_scores: a dense tensor of shape [n_data, n_data]
pairwise_predicted_scores: a dense tensor of shape [n_data, n_data]
n_data is the number of tweet candidates in a BatchPredictionRequest
params: network parameters
mask options: full_mask and diag_mask
Returns:
average loss over pairs defined by the masks
"""
# only full_mask is appropriate in hinge_loss
# hinge_loss needs 0, 1 labels, which later
# converted to -1, 1 labels
mask, pair_count = masks.full_mask(pairwise_label_scores)
# pairwise sigmoid_cross_entropy_with_logits loss
loss = tf.cond(
tf.equal(pair_count, 0), lambda: 0., lambda: _get_average_hinge_loss(
pairwise_label_scores, pairwise_predicted_scores, mask, pair_count)
)
return loss
def get_lambda_pair_loss(pairwise_label_scores, pairwise_predicted_scores,
params, swapped_ndcg):
"""
Paiwise learning-to-rank lambdarank loss
faster than the previous gradient method
Note: this loss depends on ranknet cross-entropy
delta NDCG is applied to ranknet cross-entropy
Hence, it is still a gradient descent method
Check paper http://citeseerx.ist.psu.edu/viewdoc/
download?doi=10.1.1.180.634&rep=rep1&type=pdf for more information
for more information
Args:
pairwise_label_scores: a dense tensor of shape [n_data, n_data]
pairwise_predicted_scores: a dense tensor of shape [n_data, n_data]
n_data is the number of tweet candidates in a BatchPredictionRequest
params: network parameters
swapped_ndcg: swapped ndcg of shape [n_data, n_data]
ndcg values when swapping each pair in the prediction ranking order
mask options: full_mask and diag_mask
Returns:
average loss over pairs defined by the masks
"""
n_data = tf.shape(pairwise_label_scores)[0]
if params.mask == "full_mask":
# full_mask that only covers pairs that have different labels
# (all pairwise_label_scores = 0.5: selfs and same labels are 0s)
mask, pair_count = masks.full_mask(pairwise_label_scores)
else:
# diag_mask that covers all pairs
# (only selfs/diags are 0s)
mask, pair_count = masks.diag_mask(pairwise_label_scores)
# pairwise sigmoid_cross_entropy_with_logits loss
loss = tf.cond(
tf.equal(pair_count, 0),
lambda: 0., lambda: _get_average_cross_entropy_loss(
pairwise_label_scores, pairwise_predicted_scores, mask, pair_count,
swapped_ndcg))
return loss