def _lookup(h, t, l):
ent = var('ent_embeddings')
mat = var('transfer_matrix')
rel = var('rel_embeddings')
return at(ent, h), at(ent, t), at(mat, l), at(rel, l)
def _lookup(h, t, l):
nor = var('normal_vectors')
ent = var('ent_embeddings')
rel = var('rel_embeddings')
return at(ent, h), at(ent, t), at(nor, l), at(rel, l)
def _lookup(h, t, r):
'''Gets the variables concerning a fact.'''
eRe = var('ent_re_embeddings')
rRe = var('rel_re_embeddings')
eIm = var('ent_im_embeddings')
rIm = var('rel_im_embeddings')
return at(eRe, h), at(eRe, t), at(rRe, r), at(eIm, h), at(eIm,
t), at(rIm, r)
def _score(h, t, l):
'''The term to embed triples.'''
ent = var('ent_embeddings')
rel = var('rel_embeddings')
etr = var('ent_transfer')
rtr = var('rel_transfer')
return at(ent, h), at(etr, h), at(ent, t), at(etr, t), at(rel,
l), at(rtr, l)
def _embedding_def(self):
'''Initializes the variables of the model.'''
e, r, d = self.base[0], self.base[1], self.dimension[0]
ent = var('ent_embeddings', [e, d, 1])
rel = var('rel_matrices', [r, d, d])
yield 'ent_embeddings', ent
yield 'rel_matrices', rel
self._entity = squeeze(at(ent, self.predict_h), [-1])
def _embedding_def(self):
'''Initializes the variables of the model.'''
e, r, d = self.base[0], self.base[1], self.dimension[0]
ent = var('ent_embeddings', [e, d])
rel = var('rel_embeddings', [r, d])
yield 'ent_embeddings', ent
yield 'rel_embeddings', rel
self._entity = at(ent, self.predict_h)
self._relation = at(rel, self.predict_l)
def _embedding_def(self):
'''Initializes the variables of the model.'''
e, r, d = self.base[0], self.base[1], self.dimension[0]
#Defining required parameters of the model, including embeddings of entities and relations, and normal vectors of planes
ent = var("ent_embeddings", [e, d])
rel = var("rel_embeddings", [r, d])
nor = var("normal_vectors", [r, d])
yield 'ent_embeddings', ent
yield 'rel_embeddings', rel
yield 'normal_vectors', nor
self._entity = at(ent, self.predict_h)
self._relation = at(rel, self.predict_l)
def _embedding_def(self):
'''Initializes the variables of the model.'''
e, r = self.base
d, k = self.dimension
ent = var('ent_embeddings', [e, d, 1])
rel = var('rel_embeddings', [r, k, 1])
mat = var('transfer_matrix', [r, k, d])
yield 'ent_embeddings', ent
yield 'rel_embeddings', rel
yield 'transfer_matrix', mat
self._entity = at(ent, self.predict_h)
self._relation = at(rel, self.predict_l)
def _embedding_def(self):
'''Initializes the variables.'''
e, r, d = self.base[0], self.base[1], self.dimension[0]
eRe = var("ent_re_embeddings", [e, d])
eIm = var("ent_im_embeddings", [e, d])
rRe = var("rel_re_embeddings", [r, d])
rIm = var("rel_im_embeddings", [r, d])
yield 'ent_re_embeddings', eRe
yield 'ent_im_embeddings', eIm
yield 'rel_re_embeddings', rRe
yield 'rel_im_embeddings', rIm
self._entity = at(eRe, self.predict_h)
self._relation = at(rRe, self.predict_l)
def _update(self, x):
x = tf.reshape(x, (-1, ) + self.shape)
n1 = self._n
n2 = len(x)
n = n1 + n2
mean = tf.reduce_mean(x, axis=0)
M = (n1 * self._M + n2 * mean) / n
delta = self._M - mean
delta2 = tf.square(delta)
var = tf.var(x, axis=0)
S = (self._S * n1 + var * n2 + n1 * n2 / n * delta2) / n
self._M = M
self._n = n
self._S = S
def _lookup(h, t, l):
ent = var('ent_embeddings')
rel = var('rel_matrices')
return at(ent, h), at(ent, t), at(rel, l)