import tensorflow as tf
def relation_logistic_loss(phs, prs, pts, nhs, nrs, nts):
pos_distance = phs + prs - pts
neg_distance = nhs + nrs - nts
pos_score = -tf.reduce_sum(tf.square(pos_distance), axis=1)
neg_score = -tf.reduce_sum(tf.square(neg_distance), axis=1)
pos_loss = tf.reduce_sum(tf.log(1 + tf.exp(-pos_score)))
neg_loss = tf.reduce_sum(tf.log(1 + tf.exp(neg_score)))
loss = tf.add(pos_loss, neg_loss)
return loss
def attribute_logistic_loss(phs, pas, pvs, pws, nhs, nas, nvs, nws):
pos_distance = phs + pas - pvs
neg_distance = nhs + nas - nvs
pos_score = -tf.reduce_sum(tf.square(pos_distance), axis=1)
neg_score = -tf.reduce_sum(tf.square(neg_distance), axis=1)
pos_score = tf.log(1 + tf.exp(-pos_score))
neg_score = tf.log(1 + tf.exp(neg_score))
pos_score = tf.multiply(pos_score, pws)
neg_score = tf.multiply(neg_score, nws)
pos_loss = tf.reduce_sum(pos_score)
neg_loss = tf.reduce_sum(neg_score)
loss = tf.add(pos_loss, neg_loss)
return loss
def relation_logistic_loss_wo_negs(phs, prs, pts):
pos_distance = phs + prs - pts
pos_score = -tf.reduce_sum(tf.square(pos_distance), axis=1)
pos_loss = tf.reduce_sum(tf.log(1 + tf.exp(-pos_score)))
return pos_loss
def attribute_logistic_loss_wo_negs(phs, pas, pvs):
pos_distance = phs + pas - pvs
pos_score = -tf.reduce_sum(tf.square(pos_distance), axis=1)
pos_loss = tf.reduce_sum(tf.log(1 + tf.exp(-pos_score)))
return pos_loss
def logistic_loss_wo_negs(phs, pas, pvs, pws):
pos_distance = phs + pas - pvs
pos_score = -tf.reduce_sum(tf.square(pos_distance), axis=1)
pos_score = tf.log(1 + tf.exp(-pos_score))
pos_score = tf.multiply(pos_score, pws)
pos_loss = tf.reduce_sum(pos_score)
return pos_loss
def space_mapping_loss(view_embeds, shared_embeds, mapping, eye, orthogonal_weight, norm_w=0.0001):
mapped_ents2 = tf.matmul(view_embeds, mapping)
mapped_ents2 = tf.nn.l2_normalize(mapped_ents2)
map_loss = tf.reduce_sum(tf.reduce_sum(tf.square(shared_embeds - mapped_ents2), 1))
norm_loss = tf.reduce_sum(tf.reduce_sum(tf.square(mapping), 1))
return map_loss + orthogonal_weight * orthogonal_loss(mapping, eye) + norm_w * norm_loss
def orthogonal_loss(mapping, eye):
loss = tf.reduce_sum(tf.reduce_sum(tf.pow(tf.matmul(mapping, mapping, transpose_b=True) - eye, 2), 1))
return loss
def alignment_loss(ents1, ents2):
distance = ents1 - ents2
loss = tf.reduce_sum(tf.reduce_sum(tf.square(distance), axis=1))
return loss
