Method forward

models/sparse_ins.py:185–217

(self, features, is_training=True, **kwargs)

        return locations.to(x)

    def forward(self, features, is_training=True, **kwargs):
        output = {}
        coord_features = self.compute_coordinates(features)
        features = torch.cat([coord_features, features], dim=1)
        inst_output = self.inst_branch(
            features, is_training=is_training)
        output.update(inst_output)

        if is_training:
            mask_features = self.mask_branch(features)
            pred_kernel = inst_output['pred_kernel']
            N = pred_kernel.shape[1]
            B, C, H, W = mask_features.shape

            pred_masks = torch.bmm(pred_kernel, mask_features.view(
            B, C, H * W)).view(B, N, H, W)
            pred_masks = F.interpolate(
                pred_masks, scale_factor=self.scale_factor,
                mode='bilinear', align_corners=False)
            output.update(dict(
                pred_masks=pred_masks))
        
        if self.training:
            sparse_inst_losses, matched_indices = self.loss(
                    output,
                    lane_idx_map=kwargs.get('lane_idx_map'),
                    input_shape=kwargs.get('input_shape')
            )
            for k, v in sparse_inst_losses.items():
                sparse_inst_losses[k] = self.sparse_decoder_weight * v
            output.update(sparse_inst_losses)
            output['matched_indices'] = matched_indices
        return output

    def loss(self, output, lane_idx_map, input_shape):
        """