def _pairwise_gradients(self, pxs, nxs, kg_model):
ep, tp = unzip_e_et(pxs)
# indices of negative tuples
en, tn = unzip_e_et(nxs)
pscores = self._scores(ep, tp, kg_model)
nscores = self._scores(en, tn, kg_model)
ind = np.where(nscores + self.margin > pscores)[0]
# all examples in batch satify margin criterion
self.nviolations = len(ind)
if len(ind) == 0:
return
ep = list(ep[ind])
en = list(en[ind])
tp = list(tp[ind])
tn = list(tn[ind])
pg = self.ET[tp] - kg_model.E[ep]
ng = self.ET[tn] - kg_model.E[en]
if self.l1:
pg = np.sign(-pg)
ng = -np.sign(-ng)
else:
raise NotImplementedError()
# entity type gradients
ridx, Sm, n = grad_sum_matrix(tp + tn)
gr = Sm.dot(np.vstack((-pg, -ng))) / n
return {'ET': (gr, ridx)}
def _pairwise_gradients(self, pxs, nxs):
# indices of positive examples
sp, pp, op = unzip_triples(pxs)
# indices of negative examples
sn, pn, on = unzip_triples(nxs)
pscores = self.af.f(self._scores(sp, pp, op))
nscores = self.af.f(self._scores(sn, pn, on))
#print("avg = %f/%f, min = %f/%f, max = %f/%f" % (pscores.mean(), nscores.mean(), pscores.min(), nscores.min(), pscores.max(), nscores.max()))
# find examples that violate margin
ind = np.where(nscores + self.margin > pscores)[0]
self.nviolations = len(ind)
if len(ind) == 0:
return
# aux vars
sp, sn = list(sp[ind]), list(sn[ind])
op, on = list(op[ind]), list(on[ind])
pp, pn = list(pp[ind]), list(pn[ind])
gpscores = -self.af.g_given_f(pscores[ind])[:, np.newaxis]
gnscores = self.af.g_given_f(nscores[ind])[:, np.newaxis]
# object role gradients
ridx, Sm, n = grad_sum_matrix(pp + pn)
grp = gpscores * ccorr(self.E[sp], self.E[op])
grn = gnscores * ccorr(self.E[sn], self.E[on])
#gr = (Sm.dot(np.vstack((grp, grn))) + self.rparam * self.R[ridx]) / n
gr = Sm.dot(np.vstack((grp, grn))) / n
gr += self.rparam * self.R[ridx]
# filler gradients
eidx, Sm, n = grad_sum_matrix(sp + sn + op + on)
geip = gpscores * ccorr(self.R[pp], self.E[op])
gein = gnscores * ccorr(self.R[pn], self.E[on])
gejp = gpscores * cconv(self.E[sp], self.R[pp])
gejn = gnscores * cconv(self.E[sn], self.R[pn])
ge = Sm.dot(np.vstack((geip, gein, gejp, gejn))) / n
#ge += self.rparam * self.E[eidx]
return {'E': (ge, eidx), 'R': (gr, ridx)}
def _gradients(self, xys):
ss, ps, os, ys = unzip_triples(xys, with_ys=True)
yscores = ys * self._scores(ss, ps, os)
self.loss = np.sum(np.logaddexp(0, -yscores))
#preds = af.Sigmoid.f(yscores)
fs = -(ys * af.Sigmoid.f(-yscores))[:, np.newaxis]
#self.loss -= np.sum(np.log(preds))
ridx, Sm, n = grad_sum_matrix(ps)
gr = Sm.dot(fs * ccorr(self.E[ss], self.E[os])) / n
gr += self.rparam * self.R[ridx]
eidx, Sm, n = grad_sum_matrix(list(ss) + list(os))
ge = Sm.dot(
np.vstack((fs * ccorr(self.R[ps], self.E[os]),
fs * cconv(self.E[ss], self.R[ps])))) / n
ge += self.rparam * self.E[eidx]
return {'E': (ge, eidx), 'R': (gr, ridx)}
def _pairwise_gradients(self, pxs, nxs):
# indices of positive triples
sp, pp, op = unzip_triples(pxs)
# indices of negative triples
sn, pn, on = unzip_triples(nxs)
pscores = self._scores(sp, pp, op)
nscores = self._scores(sn, pn, on)
ind = np.where(nscores + self.margin > pscores)[0]
# all examples in batch satify margin criterion
self.nviolations = len(ind)
if len(ind) == 0:
return
sp = list(sp[ind])
sn = list(sn[ind])
pp = list(pp[ind])
pn = list(pn[ind])
op = list(op[ind])
on = list(on[ind])
pg = self.E[op] - self.R[pp] - self.E[sp]
ng = self.E[on] - self.R[pn] - self.E[sn]
if self.l1:
pg = np.sign(-pg)
ng = -np.sign(-ng)
else:
raise NotImplementedError()
# entity gradients
eidx, Sm, n = grad_sum_matrix(sp + op + sn + on)
ge = Sm.dot(np.vstack((pg, -pg, ng, -ng))) / n
# relation gradients
ridx, Sm, n = grad_sum_matrix(pp + pn)
gr = Sm.dot(np.vstack((pg, ng))) / n
return {'E': (ge, eidx), 'R': (gr, ridx)}
def _gradients(self, xys):
ss, ps, os, ys = unzip_triples(xys, with_ys=True)
# define memoized functions to cache expensive dot products
# helps only if xys have repeated (s,p) or (p,o) pairs
# this happens with sampling
@memoized
def _EW(s, o, p):
return dot(self.E[s], self.W[p])
@memoized
def _WE(s, o, p):
return dot(self.W[p], self.E[o])
EW = np.array([_EW(*x) for (x, _) in xys])
WE = np.array([_WE(*x) for (x, _) in xys])
yscores = ys * np.sum(self.E[ss] * WE, axis=1)
self.loss = np.sum(np.logaddexp(0, -yscores))
fs = -(ys * af.Sigmoid.f(-yscores))[:, np.newaxis]
#preds = af.Sigmoid.f(scores)
#fs = -(ys / (1 + np.exp(yscores)))[:, np.newaxis]
#self.loss -= np.sum(ys * np.log(preds))
#fs = (scores - ys)[:, np.newaxis]
#self.loss += np.sum(fs * fs)
pidx = np.unique(ps)
gw = np.zeros((len(pidx), self.ncomp, self.ncomp))
for i in range(len(pidx)):
p = pidx[i]
ind = np.where(ps == p)[0]
if len(ind) == 1:
gw[i] += fs[ind] * np.outer(self.E[ss[ind]], self.E[os[ind]])
else:
gw[i] += dot(self.E[ss[ind]].T,
fs[ind] * self.E[os[ind]]) / len(ind)
gw[i] += self.rparam * self.W[p]
eidx, Sm, n = grad_sum_matrix(list(ss) + list(os))
ge = Sm.dot(np.vstack((fs * WE, fs * EW))) / n
ge += self.rparam * self.E[eidx]
return {'E': (ge, eidx), 'W': (gw, pidx)}
def _pairwise_gradients(self, pxs, nxs):
# indices of positive examples
sp, pp, op = unzip_triples(pxs)
# indices of negative examples
sn, pn, on = unzip_triples(nxs)
pxs, _ = np.array(list(zip(*pxs)))
nxs, _ = np.array(list(zip(*nxs)))
# define memoized functions to cache expensive dot products
# helps only if xys have repeated (s,p) or (p,o) pairs
# this happens with sampling
@memoized
def _EW(s, o, p):
return dot(self.E[s], self.W[p])
@memoized
def _WE(s, o, p):
return dot(self.W[p], self.E[o])
WEp = np.array([_WE(*x) for x in pxs])
WEn = np.array([_WE(*x) for x in nxs])
pscores = self.af.f(np.sum(self.E[sp] * WEp, axis=1))
nscores = self.af.f(np.sum(self.E[sn] * WEn, axis=1))
#print("avg = %f/%f, min = %f/%f, max = %f/%f" % (pscores.mean(), nscores.mean(), pscores.min(), nscores.min(), pscores.max(), nscores.max()))
# find examples that violate margin
ind = np.where(nscores + self.margin > pscores)[0]
self.nviolations = len(ind)
if len(ind) == 0:
return
# aux vars
gpscores = -self.af.g_given_f(pscores)[:, np.newaxis]
gnscores = self.af.g_given_f(nscores)[:, np.newaxis]
pidx = np.unique(list(pp) + list(pn))
gw = np.zeros((len(pidx), self.ncomp, self.ncomp))
for pid in range(len(pidx)):
p = pidx[pid]
ppidx = np.where(pp == p)
npidx = np.where(pn == p)
assert (len(ppidx) == len(npidx))
if len(ppidx) == 0 and len(npidx) == 0:
continue
gw[pid] += dot(self.E[sp[ppidx]].T,
gpscores[ppidx] * self.E[op[ppidx]])
gw[pid] += dot(self.E[sn[npidx]].T,
gnscores[npidx] * self.E[on[npidx]])
gw[pid] += self.rparam * self.W[p]
gw[pid] /= (len(ppidx) + len(npidx))
# entity gradients
sp, sn = list(sp[ind]), list(sn[ind])
op, on = list(op[ind]), list(on[ind])
gpscores, gnscores = gpscores[ind], gnscores[ind]
EWp = np.array([_EW(*x) for x in pxs[ind]])
EWn = np.array([_EW(*x) for x in nxs[ind]])
eidx, Sm, n = grad_sum_matrix(sp + sn + op + on)
ge = (Sm.dot(
np.vstack(
(gpscores * WEp[ind], gnscores * WEn[ind], gpscores * EWp,
gnscores * EWn))) + self.rparam * self.E[eidx]) / n
return {'E': (ge, eidx), 'W': (gw, pidx)}