def forward()

in criterion.py [0:0]

• 46 lines of code
• 5 McCabe index (conditional complexity)

    def forward(self, outputs, targets):

        batchsize = outputs["sem_cls_prob"].shape[0]
        nqueries = outputs["sem_cls_prob"].shape[1]
        ngt = targets["gt_box_sem_cls_label"].shape[1]
        nactual_gt = targets["nactual_gt"]

        # classification cost: batch x nqueries x ngt matrix
        pred_cls_prob = outputs["sem_cls_prob"]
        gt_box_sem_cls_labels = (
            targets["gt_box_sem_cls_label"]
            .unsqueeze(1)
            .expand(batchsize, nqueries, ngt)
        )
        class_mat = -torch.gather(pred_cls_prob, 2, gt_box_sem_cls_labels)

        # objectness cost: batch x nqueries x 1
        objectness_mat = -outputs["objectness_prob"].unsqueeze(-1)

        # center cost: batch x nqueries x ngt
        center_mat = outputs["center_dist"].detach()

        # giou cost: batch x nqueries x ngt
        giou_mat = -outputs["gious"].detach()

        final_cost = (
            self.cost_class * class_mat
            + self.cost_objectness * objectness_mat
            + self.cost_center * center_mat
            + self.cost_giou * giou_mat
        )

        final_cost = final_cost.detach().cpu().numpy()
        assignments = []

        # auxiliary variables useful for batched loss computation
        batch_size, nprop = final_cost.shape[0], final_cost.shape[1]
        per_prop_gt_inds = torch.zeros(
            [batch_size, nprop], dtype=torch.int64, device=pred_cls_prob.device
        )
        proposal_matched_mask = torch.zeros(
            [batch_size, nprop], dtype=torch.float32, device=pred_cls_prob.device
        )
        for b in range(batchsize):
            assign = []
            if nactual_gt[b] > 0:
                assign = linear_sum_assignment(final_cost[b, :, : nactual_gt[b]])
                assign = [
                    torch.from_numpy(x).long().to(device=pred_cls_prob.device)
                    for x in assign
                ]
                per_prop_gt_inds[b, assign[0]] = assign[1]
                proposal_matched_mask[b, assign[0]] = 1
            assignments.append(assign)

        return {
            "assignments": assignments,
            "per_prop_gt_inds": per_prop_gt_inds,
            "proposal_matched_mask": proposal_matched_mask,
        }