Fine-tune model code

[dreamcubeblock]

the code in model/point_rae.py:

if self.side is not None:
            alpha_squashed = torch.sigmoid(self.side_alpha)
            side_output = alpha_squashed * x[:, 1:] + (1 - alpha_squashed) * side_output
            concat_f = torch.cat([x[:, 0], side_output.max(1)[0]], dim=-1)
        else:
            # concat_res = torch.cat([x_res[:, 0], x_res[:, 1:].max(1)[0]], dim=-1)
            concat_pred = torch.cat([x_pred.mean(1), x_pred.max(1)[0]], dim=-1)
            # x = torch.cat([x_res, x_pred], dim=1)

            # concat_f = torch.cat([x.mean(1), x.max(1)[0]], dim=-1)

Looks a little different than what you described in your paper.