Jiaming Han

Thanks! The bug is fixed by [this commit](https://github.com/csuhan/opendet2/commit/c060a974a5f9bbc6c26669b9cc89de03f1ec751c).

 We select the same number of background samples to 1) balance foreground and background samples, 2) recall unknown from background.

try this link instead: [https://pan.baidu.com/s/1gg9jLw3](https://pan.baidu.com/s/1gg9jLw3) with verify code: umce

@JulienPeyrelon Thank you for your contribution! BTW, can you rename READMD.txt to README.md?

https://www.kaggle.com/datasets/guofeng/hrsc2016

You can transform a GeometricTensor to standard tensor by changing this line to `outs.append(x.tensor) ` https://github.com/csuhan/ReDet/blob/f4a8e7dadec990f8a92437024416f0fb07a868db/mmdet/models/backbones/re_resnet.py#L529

Hi @gladdduck , I think there are two reasons: (a) One-shot results are more sensitive to random seeds or other training factors. (b) Single-GPU training may be different from 8-GPU...

Do you place the base-training checkpoint here? https://github.com/csuhan/VFA/blob/e35411eb22b4fc48b524debe58dc7c09be2bf9a6/configs/vfa/voc/vfa_split1/vfa_r101_c4_8xb4_voc-split1_10shot-fine-tuning.py#L27 You can re-download it and try again.

Check here~ https://github.com/csuhan/VFA/issues/4#issuecomment-1477499937

检查一下pytorch是否能加载model： `model = torch.load('work_dirs/vfa_r101_c4_8xb4_voc-split1_base-training/iter_18000.pth',map_location='cpu')`