Some ambiguities in RTDETRPostProcessor forward()?

[minhhotboy9x]

Hi, I'm trying to understand your eval code. In class RTDETRPostProcessor in RTDETR_PYTORCH\src\zoo\rtdetr\rtdetr_postprocessor.py, I see these lines:

if self.use_focal_loss:
   scores = F.sigmoid(logits)
   scores, index = torch.topk(scores.flatten(1), self.num_top_queries, axis=-1)
   labels = index % self.num_classes
   index = index // self.num_classes
   boxes = bbox_pred.gather(dim=1, index=index.unsqueeze(-1).repeat(1, 1, bbox_pred.shape[-1]))

scores = F.sigmoid(logits) has the shape of (b, queries, cls) but you take the top prediction: torch.topk(scores.flatten(1), self.num_top_queries, axis=-1). As I understand, you change the cls score from different queries into a dim (b, queries*cls) then you just take the topk highest cls scores. However, there would be the case that some cls in the same original queries (the score's shape (b, queries, cls)) would be taken.
Can you explain more about this?