Ace
thank you. I will change it :+1:
I meet the same situation, are you solve it ?~
Although I rewrite the VOCDataset(but the main code is the same as longcw write). Maybe, the problem you meet is the same as me. The problem is mainly by the...
This is caused by the mmcv, you can refer to the [mmcv FAQ](https://mmcv.readthedocs.io/en/latest/faq.html)
Label Matching can be used in one-stage detector. We will support it soon.
Thank you for pointing this out. We will check it in extreme cases.
you can get this file: https://github.com/hikvision-research/SSOD/blob/main/docs/prepare_data.md
The log file [1-1-40k](https://github.com/hikvision-research/SSOD/blob/13c43fc567ae4c9dfbf74f973e8a5ed20070e984/docs/log/labelmatch_1_1_40k.log) is uploaded. (It's the setting in the ablation study. The final results use longer iterations and stronger resize)
1. We use V100 (GPU=32G) in our setting. You can use checkpoint in the backbone (need to combine label and unlabeled data in training) or use fp16 to save memory....
1. 参数部分我们没有进行具体的设置； 2. ① 主要是有监督部分的数据量比较少，采用8卡的话需要增大epoch数目；这块具体采用8卡或者2卡其实都是可以的，只是8卡情况下具体epoch我们没有试过具体多少更合适 ② 半监督里面有监督预训练模型有影响，但在半监督训练的时候迭代次数一般比较大，这个影响因素不算大