Method decode

star/models/CNNNet.py:158–264

Decode the input features. Args: x (tensor): layer-0 features. x_1 (tensor): layer-1 features. x_2 (tensor): layer-2 features. x_3 (tensor): layer-3 features. x_4 (tensor): layer-4 featuers. batch (int): The batch size.

(
        self,
        x,
        x_1,
        x_2,
        x_3,
        x_4,
        batch,
        kd_flag=False,
        requires_adaptive_max_pool3d=False,
    )

        return [x, x_1, x_2, x_3, x_4]

    def decode(
        self,
        x,
        x_1,
        x_2,
        x_3,
        x_4,
        batch,
        kd_flag=False,
        requires_adaptive_max_pool3d=False,
    ):
        """Decode the input features.

        Args:
            x (tensor): layer-0 features.
            x_1 (tensor): layer-1 features.
            x_2 (tensor): layer-2 features.
            x_3 (tensor): layer-3 features.
            x_4 (tensor): layer-4 featuers.
            batch (int): The batch size.
            kd_flag (bool, optional): Required to be true for DiscoNet. Defaults to False.
            requires_adaptive_max_pool3d (bool, optional): If set to true, use adaptive max pooling 3d. Defaults to False.

        Returns:
            if kd_flag is true, return a list of output from layer-8 to layer-5
            else return a list of a single element: the output after passing through the decoder
        """
        # -------------------------------- Decoder Path --------------------------------
        x_5 = F.relu(
            self.bn5_1(
                self.conv5_1(
                    torch.cat((F.interpolate(x_4, scale_factor=(2, 2)), x_3), dim=1)
                )
            )
        )
        x_5 = F.relu(self.bn5_2(self.conv5_2(x_5)))

        x_2 = x_2.view(batch, -1, x_2.size(1), x_2.size(2), x_2.size(3))
        x_2 = x_2.permute(0, 2, 1, 3, 4).contiguous()
        x_2 = (
            F.adaptive_max_pool3d(x_2, (1, None, None))
            if requires_adaptive_max_pool3d
            else x_2
        )
        x_2 = x_2.permute(0, 2, 1, 3, 4).contiguous()
        x_2 = x_2.view(-1, x_2.size(2), x_2.size(3), x_2.size(4)).contiguous()

        x_6 = F.relu(
            self.bn6_1(
                self.conv6_1(
                    torch.cat((F.interpolate(x_5, scale_factor=(2, 2)), x_2), dim=1)
                )
            )
        )
        x_6 = F.relu(self.bn6_2(self.conv6_2(x_6)))

        x_1 = x_1.view(batch, -1, x_1.size(1), x_1.size(2), x_1.size(3))
        x_1 = x_1.permute(0, 2, 1, 3, 4).contiguous()