vfm.py 文件源码

def term_slop(self, loc, val, bs, nf, train=True):
        """ Compute the slope for each active feature.
        """
        shape = (bs, nf)

        # Reshape all of our constants
        pr_mu = F.broadcast_to(self.slop_mu.b, shape)
        pr_lv = F.broadcast_to(self.slop_lv.b, shape)
        # This is either zero or a very negative number
        # indicating to sample N(mean, logvar) or just draw
        # the mean preicsely
        if not train:
            pr_lv += self.lv_floor

        # The feature slopes are grouped together so that they
        # all share a common mean. Then individual features slop_delta_lv
        # are shrunk towards zero, which effectively sets features to fall
        # back on the group mean.
        sl_mu = F.reshape(self.slop_delta_mu(loc), shape) + pr_mu
        sl_lv = F.reshape(self.slop_delta_lv(loc), shape) + pr_lv
        coef = F.gaussian(sl_mu, sl_lv)
        slop = F.sum(coef * val, axis=1)

        # Calculate divergence between group mean and N(0, 1)
        kld1 = F.gaussian_kl_divergence(self.slop_mu.b, self.slop_lv.b)
        # Calculate divergence of individual delta means and delta vars
        args = (self.slop_delta_mu.W, self.slop_delta_lv.W)
        kld2 = F.gaussian_kl_divergence(*args)

        return slop, kld1 + kld2