Method forward

PATH/core/models/backbones/vit.py:510–534

(self, input_var)

        return x.permute(0, 2, 1).reshape(B, -1, Hp, Wp)

    def forward(self, input_var):
        output = {}
        x = input_var['image']

        # pre_input padding for test support
        x = self._normalization(x)

        if not self.proj_padding:
            stride = 32
            output["prepad_input_size"] = [x.shape[-2], x.shape[-1]]  # h, w for sem_seg_postprocess
            target_size = (torch.tensor((x.shape[-1], x.shape[-2])) + (stride - 1)).div(stride, rounding_mode="floor") * stride  # w, h
            padding_size = [  # [l,r,t,b]
                0,
                target_size[0] - x.shape[-1],
                0,
                target_size[1] - x.shape[-2],
            ]
            x = F.pad(x, padding_size, value=0.).contiguous()
            output["image"] = x

        # pre_input padding for test support >>> end

        output['backbone_output'] = self.forward_features(x)
        input_var.update(output)
        return input_var


def vit_aligned_base_patch16(pretrained=False, load_pos_embed=True, **kwargs):