Exemplo n.º 1
0
    def REL(self, t_txt, box1, box2):
        #conc = torch.cat((box1.expand_as(box2),box2),1)
        conc = torch.cat((box1[:, -5:].expand_as(box2[:, -5:]), box2[:, -5:]),
                         1)
        if self.use_outer:
            qkern0 = outer(box1[:, -5:].view(1, -1), box2[:, -5:])
            conc = torch.cat((conc, qkern0), 1)

        spat = self.Wrbox(conc)
        if self.fusion == 'mul':
            prod = t_txt.expand_as(spat) * spat
            norm = prod / torch.norm(prod, 2, 1).expand_as(prod)
        elif self.fusion == 'sum':
            norm = t_txt.expand_as(spat) + spat
        elif self.fusion == 'concat':
            norm = torch.cat([t_txt.expand_as(spat), spat], 1)
        else:
            raise NotImplementedError()

        score = self.Wrel1(norm)
        #    score = self.Wrel1(self.RELU(self.Wrel0(norm)))
        if self.debug:
            print
            print "REL>norm:", printVec(norm.t())
            print "REL>score:", printVec(score.t())

        return score
Exemplo n.º 2
0
    def REL(self, t_txt, box1, box2):
        if self.only_spatial:
            conc = torch.cat(
                (box1[:, -5:].expand_as(box2[:, -5:]), box2[:, -5:]), 1)
        else:
            conc = torch.cat((box1.expand_as(box2), box2), 1)

        if self.use_outer:
            qkern0 = outer(box1[:, -5:].view(1, -1), box2[:, -5:])
            conc = torch.cat((conc, qkern0), 1)

        spat = self.Wrbox(conc)
        if self.fusion == 'mul':
            prod = t_txt.expand_as(spat) * spat
            norm = prod / torch.norm(prod, 2, 1).expand_as(prod)
        elif self.fusion == 'sum':
            norm = t_txt.expand_as(spat) + spat
        elif self.fusion == "concat":
            norm = torch.cat([t_txt.expand_as(spat), spat], 1)

        if self.layer == 2:
            norm = self.RELU(self.Wrel0(norm))
        score = self.Wrel1(norm)

        return score
    def LOC(self, t_txt, box):
        if self.use_outer:
            O = []
            for i in range(box.size(0)):
                qkern0 = outer(box[i, self.cnn_feat:].contiguous().view(1, -1),
                               box[i, self.cnn_feat:].view(1, -1))
                O.append(qkern0)

            box_out = torch.cat((box, torch.cat(O, 0)), 1)
            t_box = self.Wbox(box_out)
        else:
            t_box = self.Wbox(box)

        t_sum = t_txt.expand_as(t_box) * t_box
        norm = t_sum / torch.norm(t_sum, 2, 1).expand_as(t_sum)

        score = self.Wout0(norm)
        if self.debug:
            print
            print "LOC>norm:", printVec(norm.t())
            print "LOC>score:", printVec(score.t())
        return score