About Meta-train label and 3K-base 1K novel

[HuShaohanAI]

Hi, there
Q1:
In faster_rcnn.py, line 121
"proposal_labels = rois_label[b * 128:(b + 1) * 128].data.cpu().numpy()[0]"
why you only use the first roi's label?

Q2:
And in faster_rcnn.py, line 150,
"RCNN_loss_cls = F.cross_entropy(cls_score, rois_label[b * 128:(b + 1) * 128])"
why you compute loss using all class with each attention?
Shouldn't it compute loss only with the class of the attention that it use?

Q3:
In 10-shot task, when shot=10, in the second phase, you say 3K base and 1K novel, so in roibatchLoader, you get 30 base and 10 novel, but in MetaDataset, you only get 10 image and 10 mask for both base and novel class?
Q4:
I didn't fine "voc_2007_train_first_split.txt" file?

[XPZ07]

你是否找到了问题所在，我也很疑惑Q1，为什么只使用第一个roi标签，我认为[0]可以删掉，但这与他的代码不相符合，希望听到你的回复

[XPZ07]

Hello，Have you found the problem? I am also puzzled as to why Q1 only uses the first ROI tag. I think [0] can be deleted, but it does not match his code. I hope to hear your reply

[XPZ07]

我突然觉得你提出的Q2问题非常的合理，我也非常疑惑，期待得到你的回复