Method loss

models/sparse_ins.py:219–259

output : from self.forward lane_idx_map : instance-level segmentation map, [20, H, W] where 20=max_lanes

(self, output, lane_idx_map, input_shape)

        return output

    def loss(self, output, lane_idx_map, input_shape):
        """
        output : from self.forward
        lane_idx_map : instance-level segmentation map, [20, H, W] where 20=max_lanes
        """
        pred_masks = output['pred_masks']
        pred_masks = output['pred_masks'].view(
            pred_masks.shape[0],
            self.inst_branch.num_group,
            self.inst_branch.num_mask,
            *pred_masks.shape[2:])
        pred_logits = output['pred_logits']
        pred_logits = output['pred_logits'].view(
            pred_logits.shape[0],
            self.inst_branch.num_group,
            self.inst_branch.num_mask,
            *pred_logits.shape[2:])
        pred_scores = output['pred_scores']
        pred_scores = output['pred_scores'].view(
            pred_scores.shape[0],
            self.inst_branch.num_group,
            self.inst_branch.num_mask,
            *pred_scores.shape[2:])

        out = {}
        all_matched_indices = []
        for group_idx in range(self.inst_branch.num_group):
            sparse_inst_losses, matched_indices = \
                self.sparse_inst_crit(
                    outputs=dict(
                        pred_masks=pred_masks[:, group_idx, ...].contiguous(),
                        pred_logits=pred_logits[:, group_idx, ...].contiguous(),
                        pred_scores=pred_scores[:, group_idx, ...].contiguous(),
                    ),
                    targets=self.prepare_targets(lane_idx_map),
                    input_shape=input_shape, # seg_bev
                )
            for k, v in sparse_inst_losses.items():
                out['%s_%d' % (k, group_idx)] = v
            all_matched_indices.append(matched_indices)
        return out, all_matched_indices

    def prepare_targets(self, targets):
        new_targets = []