"""TopOne Max Loss Layer"""
import tensorflow as tf
from deepr.layers import base
from deepr.layers.reduce import WeightedAverage, Average
from deepr.layers.core import Softmax
from deepr.utils.broadcasting import make_same_shape
[docs]class TopOneMax(base.Layer):
"""Vanilla TopOne Max Loss Layer"""
def __init__(self, bpr_max_regularizer=0.0, **kwargs):
self.bpr_max_regularizer = bpr_max_regularizer
super().__init__(n_in=2, n_out=1, **kwargs)
[docs] def forward(self, tensors, mode: str = None):
"""Forward method of the layer
Parameters
----------
tensors : Tuple[tf.Tensor]
- positives : shape = (batch, num_events)
- negatives : shape = (batch, num_events, num_negatives)
Returns
-------
tf.Tensor
TopOne Max loss
"""
positives, negatives = tensors
positives, negatives = make_same_shape([positives, negatives], broadcast=False)
softmax_scores = Softmax()((negatives, tf.ones_like(negatives)))
losses = tf.multiply(softmax_scores, tf.nn.sigmoid(negatives - positives) + tf.nn.sigmoid(tf.square(negatives)))
return Average()(losses, mode)
[docs]class MaskedTopOneMax(base.Layer):
"""Masked TopOne Max Loss Layer"""
def __init__(self, bpr_max_regularizer=0.0, **kwargs):
self.bpr_max_regularizer = bpr_max_regularizer
super().__init__(n_in=4, n_out=1, **kwargs)
[docs] def forward(self, tensors, mode: str = None):
"""Forward method of the layer
(details: https://arxiv.org/pdf/1706.03847.pdf)
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
TopOne Max loss
"""
positives, negatives, mask, weights = tensors
positives, negatives = make_same_shape([positives, negatives], broadcast=False)
softmax_scores = Softmax()((negatives, tf.to_float(mask)))
losses = tf.multiply(softmax_scores, tf.nn.sigmoid(negatives - positives) + tf.nn.sigmoid(tf.square(negatives)))
# One loss per event, average of scores : (batch, num_events)
event_scores = WeightedAverage()((losses, tf.to_float(mask)))
# Each event contributes according to its weight
event_weights = weights * tf.to_float(tf.reduce_any(mask, axis=-1))
event_losses = event_scores * event_weights
return tf.div_no_nan(tf.reduce_sum(event_losses), tf.reduce_sum(event_weights))