the details of "object attentive vector channel-wise product rois" during meta training

[Hxx2048]

Hello! Thank you for releasing code, I have some questions about the details in meta training.

In ./model/faster_rcnn/faster_rcnn.py Line 122,

" proposal_labels = rois_label[b * 128:(b + 1) * 128].data.cpu().numpy()[0] "

This code means you just choose first label of 128 rois in each batch as the unique label, and you only use object attentive vectors just for this label's class ? Why do you choose only the first one instead of all the labels of 128 rois?
or from your paper, I thought the code should be " proposal_labels = rois_label[b * 128:(b + 1) * 128].data.cpu().numpy()"

[yanxp]

Hello, proposal_labels = rois_label[b * 128:(b + 1) * 128].data.cpu().numpy()[0] is the first label of 128 rois in each batch, please check it.