28eb8ed1d64fb8d5cfddc58d65fc48aeace6f436,pyro/distributions/transforms/spline.py,ConditionalSpline,condition,#ConditionalSpline#Any#,484

Before Change

        // Rational linear splines have additional lambda parameters
        if self.order == "linear":
            w, h, d, l = self.nn(context)
            l = torch.sigmoid(l.reshape(l.shape[:-1] + (self.input_dim, self.count_bins)))
        elif self.order == "quadratic":
            w, h, d = self.nn(context)
            l = None
        else:

After Change

                    self.order))

        // AutoRegressiveNN and DenseNN return different shapes...
        if w.shape[-1] == self.input_dim:
            w = w.transpose(-1, -2)
            h = h.transpose(-1, -2)
            d = d.transpose(-1, -2)

        else:
            w = w.reshape(w.shape[:-1] + (self.input_dim, self.count_bins))
            h = h.reshape(h.shape[:-1] + (self.input_dim, self.count_bins))
            d = d.reshape(d.shape[:-1] + (self.input_dim, self.count_bins - 1))

        w = F.softmax(w, dim=-1)
        h = F.softmax(h, dim=-1)
        d = F.softplus(d)
        return ConditionedSpline(w, h, d, l, bound=self.bound, order=self.order)

Instances