I have some problem about object query.

[Zhong-Zi-Zeng]

The last page of the original paper shows a simple code for DETR. The decoder's input is just a random value of size (100, 256). However, on your GitHub, I can't understand what you did about the object query. Why did you use the embedding layer's weight instead of a random value?

[AjibolaPy]

That's also my confusion. I guess instead of using random values, the embedding weights was used and reshaped. Maybe it's the same. But is it trainable?. I'll appreciate an answer if you've gotten the answer

[Zhong-Zi-Zeng]

I have implemented DETR and found that embedding weights is more convenient than random values when building a model.

[AjibolaPy]

I have implemented DETR and found that embedding weights is more convenient than random values when building a model.

Thanks for helping.
Assuming:

embedding=nn.Embedding(45, 2)


weights=embedding.weight.unsqueeze(1).repeat(1,3)

Did you keep the weights requires_grad=True. This is just my confusion

[Zhong-Zi-Zeng]

Yes, so that the gradient of embedding weight will have the same value.

[AjibolaPy]

Yes, so that the gradient of embedding weight will have the same value.

Thanks for your answer.
I should keep requires_grad set to 'True.' That means it will also be trained during backpropagation, and the value will change. This applies to the embedding weights, specifically the query embedding.

[AjibolaPy]

Yes, so that the gradient of embedding weight will have the same value.

I think it's in (num_queries, batch_size, dim) and not (batch_size, num_queries, dim)

[Zhong-Zi-Zeng]

You are right!

[MLDeS]

I have implemented DETR and found that embedding weights is more convenient than random values when building a model.

@Zhong-Zi-Zeng What do you mean with more convenient? Is it that the results are better? Because, as shown in the DETR colab notebook, if you use, nn.Parameter(torch.rand(100, hidden_dim)) as the queries with 100 being the num_queries and update all parameters of the model, it should still work because the nn.parameter has requires_grad=True by default, so it will still update the parameters, right? It will not be random values thereafter.