0c81011805c9ab4d6f7f314f674d39e51f5ba8eb,Tars/losses/gan_loss.py,GANLoss,estimate,#GANLoss#Any#Any#,26

Before Change

            sample_data = get_dict_values(sample_data_dict, self.d.var)[0]
            return self.d_criterion(sample_data, sample)  // TODO: detach

        return self.g_criterion(sample)

    @staticmethod
    def d_criterion(sample_data, sample):
        return - torch.log(sample_data) - torch.log(1 - sample)

After Change

        x_dict = self._p2.sample(batch_size=batch_size)

        // set labels
        t_data = torch.ones(batch_size, 1).to(x_data.device)
        t = torch.zeros(batch_size, 1).to(x_data.device)

        if discriminator:
            // sample y from x_data
            y_data_dict = self.d.sample(x_data_dict)
            y_data = get_dict_values(y_data_dict, self.d.var)[0]

            // sample y from x
            y_dict = self.d.sample(x_dict)  // TODO: detach x_dict
            y = get_dict_values(y_dict, self.d.var)[0]

            return self.bce_loss(y_data, t_data) + self.bce_loss(y, t)

        // sample y from x
        y_dict = self.d.sample(x_dict)
        y = get_dict_values(y_dict, self.d.var)[0]

        return self.bce_loss(y, t_data)

    def train(self, train_x, **kwargs):
        self.d.train()

Instances