"""Triplet Precision Layer."""
import tensorflow as tf
from deepr.layers import base
from deepr.layers.reduce import WeightedAverage
from deepr.utils.broadcasting import make_same_shape
[docs]class TripletPrecision(base.Layer):
"""Triplet Precision Layer."""
def __init__(self, **kwargs):
super().__init__(n_in=4, n_out=1, **kwargs)
[docs] def forward(self, tensors, mode: str = None):
"""Computes Triplet Precision
Parameters
----------
tensors : Tuple[tf.Tensor]
- positives : shape = (batch, num_events)
- negatives : shape = (batch, num_events, num_negatives)
- mask : shape = (batch, num_events, num_negatives)
- weights : shape = (batch, num_events)
Returns
-------
tf.Tensor
BPR loss
"""
# Retrieve positives and negatives logits
positives, negatives, mask, weights = tensors
positives, negatives = make_same_shape([positives, negatives], broadcast=False)
# One triplet precision per event
event_triplet = WeightedAverage()((tf.cast(positives > negatives, tf.float32), tf.cast(mask, tf.float32)), mode)
# Each event contributes according to its weight
event_weights = weights * tf.to_float(tf.reduce_any(mask, axis=-1))
return tf.div_no_nan(tf.reduce_sum(event_triplet * event_weights), tf.reduce_sum(event_weights))