def forward()

in mobile_cv/arch/fbnet_v2/fbnet_fpn.py [0:0]

• 20 lines of code
• 15 McCabe index (conditional complexity)

    def forward(self, x: List[torch.Tensor]) -> List[torch.Tensor]:
        """Forward pass is generated programatically to be independent of the
        number of spatial resolutions. The only requirement is that there are
        5 stages per spatial resolution

        Since blocks depend on >=1 other blocks, we have to make sure of the
        ordering of block calculations.
        """
        # store results
        data = [
            [None] * self.num_stages_per_resolution for _ in range(self.num_resolutions)
        ]

        for j in range(self.num_resolutions):
            for i in range(self.num_stages_per_resolution):
                if i == 0 or i == 1:
                    data[j][i] = self.stages[j][i](x[j * 2 + i])
                elif i == 2:
                    a, b, c = data[j][0], data[j][1], data[j - 1][-1]
                    inputs = [t for t in [a, b, c] if t is not None]
                    combined_result = self.stage_combiners[j](inputs)
                    data[j][i] = self.stages[j][i](combined_result)
                elif i == 3 or i == 4:
                    if self.stages[j][i] is not None:
                        data[j][i] = self.stages[j][i](data[j][2])
                # if data[j][i] is not None:
                #     print("res", j, "stage", i, "shape", data[j][i].shape)
        output = [data[j][-2] for j in range(self.num_resolutions)]
        if self.combiner_path == "low_res":
            output = output[::-1]

        # pyre-fixme[7]: Expected `List[torch.Tensor]` but got `List[None]`.
        return output