Method sampling

grbm.py:272–305

(self, v_init, num_steps=1, save_gap=1)

    @torch.no_grad()
    def sampling(self, v_init, num_steps=1, save_gap=1):
        v_shape = v_init.shape
        v = v_init.view(v_shape[0], -1)
        var = self.get_var()
        var_mean = var.mean().item()

        if self.inference_method == 'Gibbs':
            samples = self.Gibbs_sampling_vh(v, num_steps=num_steps - 1)
            samples = [xx[0] for xx in samples]  # extract v
        elif self.inference_method == 'Langevin':
            samples = self.Langevin_sampling_v(v,
                                               num_steps=num_steps - 1,
                                               eta=self.Langevin_eta * var_mean,
                                               is_anneal=self.is_anneal_Langevin,
                                               adjust_step=self.Langevin_adjust_step)
        elif self.inference_method == 'Gibbs-Langevin':
            samples = self.Gibbs_Langevin_sampling_vh(
                v,
                num_steps=num_steps - 1,
                num_steps_Langevin=self.Langevin_step,
                eta=self.Langevin_eta * var_mean,
                is_anneal=self.is_anneal_Langevin,
                adjust_step=self.Langevin_adjust_step)
            samples = [xx[0] for xx in samples]  # extract v

        # use conditional mean as the last sample
        h = torch.bernoulli(self.prob_h_given_v(samples[-1], var))
        mu = self.prob_v_given_h(h)
        v_list = [(0, v_init)] + [(ii + 1, samples[ii].view(v_shape).detach())
                                  for ii in range(num_steps - 1)
                                  if (ii + 1) % save_gap == 0
                                  ] + [(num_steps, mu.view(v_shape).detach())]

        return v_list

    @torch.no_grad()
    def positive_grad(self, v):