Asking for code component

[trqminh]

Hi @lkeab,
Thank you for your consideration.

I see that the proposed method is mainly in mask_head.py, however, I find the code is hard to keep track. Would you mind showing me where is the code for this component below?

[lkeab]

Hi, thanks for watching our work Transfiner.
The sequence encoder is: class TransformerEncoder, L928 to L957.
The pixel decoder is conv_r1, L935.
The incoherent query sequence sent into the transformer for refinement is here during training, while here during inference.

[trqminh]

Thanks, @lkeab for your answer.
Could I ask about the size of the incoherent query sequence according to those pyramid features maps?
Thank you.

[lkeab]

N x C, N is the number of the detected incoherent query points. The size of N depends on different objects but much smaller than total points in the RoI region.

[trqminh]

Thank you @lkeab .

[128ve900]

May I ask whether different samples have different values ​​of N? Is the input size of the incoherent query sequence indeterminate?