def forward(self, ps, qs, rs):
constrain(next(self._FC.parameters()))
ps_embedding = self._PsEmbedding(ps)
qs_embedding = self._QsEmbedding(qs)
rs_embedding = self._RsEmbedding(rs)
inferences = self._FC(ps_embedding * qs_embedding * rs_embedding)
regs = self.reg * (torch.norm(ps_embedding) +
torch.norm(qs_embedding) + torch.norm(rs_embedding))
return inferences, regs
def forward(self, ps, qs, rs):
for p in self._FC.parameters():
if len(p.size()) == 1: continue
constrain(p)
ps_embedding = self._PsEmbedding(ps)
qs_embedding = self._QsEmbedding(qs)
rs_embedding = self._RsEmbedding(rs)
inferences = self._FC(
torch.cat([ps_embedding, qs_embedding, rs_embedding], dim=-1))
regs = self.reg * (torch.norm(ps_embedding) +
torch.norm(qs_embedding) + torch.norm(rs_embedding))
return inferences, regs
def forward(self, ps, qs, rs):
constrain(next(self._FC.parameters()))
constrain(next(self._W.parameters()))
ps_embedding = self._PsEmbedding(ps)
qs_embedding = self._QsEmbedding(qs)
rs_embedding = self._RsEmbedding(rs)
gmf_out = ps_embedding * qs_embedding * rs_embedding
mlp_out = self._W(
torch.cat([ps_embedding, qs_embedding, rs_embedding], dim=-1))
inferences = self._FC(F.relu(gmf_out + mlp_out))
regs = self.reg * (torch.norm(ps_embedding) +
torch.norm(qs_embedding) + torch.norm(rs_embedding))
return inferences, regs
def forward(self, ps, qs, rs):
constrain(next(self._FC.parameters()))
ps_embedding = self._PsEmbedding(ps)
qs_embedding = self._QsEmbedding(qs)
rs_embedding = self._RsEmbedding(rs)
ps_embedding_trans = self.mlp_p(ps_embedding.view(-1, 1)).view(
ps_embedding.size())
qs_embedding_trans = self.mlp_q(qs_embedding.view(-1, 1)).view(
qs_embedding.size())
rs_embedding_trans = self.mlp_r(rs_embedding.view(-1, 1)).view(
rs_embedding.size())
inferences = self._FC(
ops_triple(self.arch['triple'], ps_embedding_trans,
qs_embedding_trans, rs_embedding_trans))
regs = self.reg * (torch.norm(ps_embedding) + torch.norm(qs_embedding) + \
torch.norm(rs_embedding))
return inferences, regs
def forward(self, ps, qs, rs):
for i in range(len(PRIMITIVES_TRIPLE)):
constrain(next(self._FC[i].parameters()))
ps_embedding = self._PsEmbedding(ps)
qs_embedding = self._QsEmbedding(qs)
rs_embedding = self._RsEmbedding(rs)
ps_embedding_trans = self._arch_parameters['mlp']['p'](
ps_embedding.view(-1, 1)).view(ps_embedding.size())
qs_embedding_trans = self._arch_parameters['mlp']['q'](
qs_embedding.view(-1, 1)).view(qs_embedding.size())
rs_embedding_trans = self._arch_parameters['mlp']['r'](
rs_embedding.view(-1, 1)).view(rs_embedding.size())
# the weight is already binarized
assert self._arch_parameters['triple'].sum() == 1.
inferences = MixedTriple(ps_embedding_trans, qs_embedding_trans,
rs_embedding_trans,
self._arch_parameters['triple'], self._FC)
regs = self.reg * (torch.norm(ps_embedding) +
torch.norm(qs_embedding) + torch.norm(rs_embedding))
return inferences, regs