RuntimeError: selected index k out of range， topK out of range when doing evalutaion

Hi, really appreciate your brilliant work.

I am trying to use DINO to train my own dataset with 2 classes. I modified num_classes=3 and dn_labelbook_size = 3 (I think it is necessary), then start training. After one epoch, I get this error when doing evaluation.

Any advice？

Thanks.

Traceback (most recent call last):
  File "main.py", line 448, in <module>
    main(args)
  File "main.py", line 334, in main
    wo_class_error=wo_class_error, args=args, logger=(logger if args.save_log else None)
  File "/usr/local/lib/python3.7/dist-packages/torch/autograd/grad_mode.py", line 28, in decorate_context
    return func(*args, **kwargs)
  File "/opt/tiger/DINO/engine.py", line 169, in evaluate
    outputs = model(samples, targets)
  File "/usr/local/lib/python3.7/dist-packages/torch/nn/modules/module.py", line 1102, in _call_impl
    return forward_call(*input, **kwargs)
  File "/usr/local/lib/python3.7/dist-packages/torch/nn/parallel/distributed.py", line 886, in forward
    output = self.module(*inputs[0], **kwargs[0])
  File "/usr/local/lib/python3.7/dist-packages/torch/nn/modules/module.py", line 1102, in _call_impl
    return forward_call(*input, **kwargs)
  File "/opt/tiger/DINO/models/dino/dino.py", line 274, in forward
    hs, reference, hs_enc, ref_enc, init_box_proposal = self.transformer(srcs, masks, input_query_bbox, poss,input_query_label,attn_mask)
  File "/usr/local/lib/python3.7/dist-packages/torch/nn/modules/module.py", line 1102, in _call_impl
    return forward_call(*input, **kwargs)
  File "/opt/tiger/DINO/models/dino/deformable_transformer.py", line 356, in forward
    topk_proposals = torch.topk(enc_outputs_class_unselected.max(-1)[0], topk, dim=1)[1] # bs, nq
RuntimeError: selected index k out of range

[HaoZhang534]

Two more question:

1. Have you ever had experiments for DINO under different number of queries?
2. When training with customer dataset, except num_classes and dn_labelbook_size, are there any variables related to the number of classes need to be change?

Q1: Reducing 900 queries to 300 queries will slightly reduce the performance (about 0.5 AP). Q2: No other hyper-parameter to change.

[Pattorio]

Two more question:

1. Have you ever had experiments for DINO under different number of queries?
2. When training with customer dataset, except num_classes and dn_labelbook_size, are there any variables related to the number of classes need to be change?