kakaobrain / sparse-detr Goto Github PK

What does the structure of the scoring network consist of? The paper only describes that the scoring network is the same as the detection head of the decoder, is it all the same? Thank you for your help!

why didn't you set the real ground truth as ground truth of salient token prediction??

RuntimeError: 0INTERNAL ASSERT FAILED at "/opt/conda/conda-bld/pytorch_1639180594101/work/torch/csrc/jit/ir/alias_analysis.cpp":584, please report a bug to PyTorch. We don't have an op for aten::fill_ but it isn't a special case. Argument types: Tensor, bool,

Hello, nice work!
In your repo, I find that the AP of swin-t detr is reported.
Would you like to kindly offer the checkpoint of 500-epoch swin-t detr with AP of 45.4?
Looking for your reply!
Cordially,
Sean

Is there an implementation of Objectness Score in the code you provide

Hi! Thank you for releasing such a wonderful work. How DAM is generated was a bit unclear to me when reading the paper. Assuming there are N tokens in total from the encoder (considering one feature level, then N = H x W), and M object queries:

1. Regarding "In the case of the dense attention, DAM can be easily obtained by summing up attention maps from every decoder layer", do you mean the cross-attention with shape N x M?
2. Regarding "produces a single map of the same size as the feature map from the backbone", how is this achieved? Could you help walk through the calculation and the shapes of the tensors?
3. Why not directly use the DAM to select the top-k tokens and why have a separate scoring network?

Thanks! I look forward to your reply.

/data/public/rw/team-autolearn/pretrainedmodels/swin/swin_tiny_patch4_window7_224.pth

Could you please share the checkpoint? Thank you in advance!

If my understanding is correct, when you do:

self.class_embed = nn.Linear(hidden_dim, num_classes)

at DeformableDETR() in deformable_detr.py, "num_classes" should instead be "num_classes + 1"

The same thing goes for:

self.class_embed.bias.data = torch.ones(num_classes) * bias_value

in the same DeformableDETR() function, where it should be "num_classes + 1" instead of "num_classes".

Otherwise I'm just getting confused somewhere, but I'm pretty sure there should be an extra background class logit.

That is how it was done in the original DETR code anyways.

Dear authors, could you please provide the code of your model for predicted bounding boxes visualization? Thanks a lot.

When I try to test with the pre-trained model, an error reported matrix contains invalid numeric entries, due to backbone output is nan, is there any solution?