val error:The testing results of the whole dataset is empty.; train：no pos rois

[wwlaoxi]

我用whu数据集训练正常，但换成自己的数据集训练，训练中出现了no pos rois。在测试指标时报错：07/25 08:58:55 - mmengine - ERROR - /RSPrompter-release/tools/../mmdet/evaluation/metrics/coco_metric.py - compute_metrics - 472 - The testing results of the whole dataset is empty.
采用调低学习率、梯度累加方法、取消解码时Prompter中正余弦变换、增大batch_size，都没起作用，一直打印这个 no pos rois。我查看了所有标签，没有空标注。请问如何排查解决呢？谢谢！或者能定位到哪个文件打印的这个错误么？

训练日志
no pos rois
no pos rois
no pos rois
no pos rois
no pos rois
no pos rois
no pos rois
no pos rois
no pos rois
no pos rois
no pos rois
no pos rois
no pos rois
no pos rois
no pos rois
no pos rois
no pos rois
no pos rois
no pos rois
no pos rois
no pos rois
no pos rois
no pos rois
no pos rois
no pos rois
no pos rois
no pos rois
no pos rois
no pos rois
no pos rois
no pos rois
no pos rois
no pos rois
no pos rois
no pos rois
no pos rois
no pos rois
no pos rois
no pos rois
no pos rois
no pos rois
no pos rois
no pos rois
no pos rois
no pos rois
no pos rois
no pos rois
no pos rois
no pos rois
no pos rois
07/25 08:58:39 - mmengine - INFO - Epoch(train) [6][ 50/116] lr: 1.9965e-04 eta: 0:38:10 time: 0.1300 data_time: 0.0045 memory: 3020 loss: 0.0000 loss_rpn_cls: 0.0000 loss_rpn_bbox: 0.0000 loss_cls: 0.0000 acc: 100.0000 loss_bbox: 0.0000 loss_mask: 0.0000
no pos rois
no pos rois
no pos rois
no pos rois
no pos rois
no pos rois
no pos rois
no pos rois
no pos rois
no pos rois
no pos rois
no pos rois
no pos rois
no pos rois
no pos rois
no pos rois
no pos rois
no pos rois
no pos rois
no pos rois
no pos rois
no pos rois
no pos rois
no pos rois
no pos rois
no pos rois
no pos rois
no pos rois
no pos rois
no pos rois
no pos rois
no pos rois
no pos rois
no pos rois
no pos rois
no pos rois
no pos rois
no pos rois
no pos rois
no pos rois
no pos rois
no pos rois
no pos rois
no pos rois
no pos rois
no pos rois
no pos rois
no pos rois
no pos rois
no pos rois
07/25 08:58:45 - mmengine - INFO - Epoch(train) [6][100/116] lr: 1.9965e-04 eta: 0:37:51 time: 0.1266 data_time: 0.0025 memory: 3020 loss: 0.0000 loss_rpn_cls: 0.0000 loss_rpn_bbox: 0.0000 loss_cls: 0.0000 acc: 100.0000 loss_bbox: 0.0000 loss_mask: 0.0000

验证日志
no pos rois
no pos rois
no pos rois
no pos rois
no pos rois
no pos rois
no pos rois
no pos rois
no pos rois
no pos rois
no pos rois
no pos rois
no pos rois
no pos rois
no pos rois
no pos rois
07/25 08:58:47 - mmengine - INFO - Exp name: rsprompter_anchor-nwpu-peft-512_20240725_085623
07/25 08:58:47 - mmengine - INFO - Saving checkpoint at 6 epochs
07/25 08:58:55 - mmengine - INFO - Evaluating bbox...
Loading and preparing results...
07/25 08:58:55 - mmengine - ERROR - /RSPrompter-release/tools/../mmdet/evaluation/metrics/coco_metric.py - compute_metrics - 472 - The testing results of the whole dataset is empty.
07/25 08:58:55 - mmengine - INFO - Epoch(val) [6][21/21] data_time: 0.0066 time: 0.0776

[millet-Ding]

i meet the same problem with you，do you know how to solve it？

[KyanChen]

这表示rpn没有召回出合适的候选框

[millet-Ding]

这表示rpn没有召回出合适的候选框
谢谢，我尝试去解决~ : )

[Kechuan-ln]

我也遇到了类似的问题，请问有解决吗？

[millet-Ding]

我用的seg-mask rcnn 那个可以跑，需要添加一下自己的类别，但效果不够好。但是rsprompter一样会遇到这个问题🙋‍♀️

…

---- Replied Message ---- | From | ***@***.***> | | Date | 10/06/2024 17:13 | | To | KyanChen/RSPrompter ***@***.***> | | Cc | millet-Ding ***@***.***>, Comment ***@***.***> | | Subject | Re: [KyanChen/RSPrompter] val error:The testing results of the whole dataset is empty.; train：no pos rois (Issue #107) | 我也遇到了类似的问题，请问有解决吗？ — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you commented.Message ID: ***@***.***>

[Kechuan-ln]

我用的seg-mask rcnn 那个可以跑，需要添加一下自己的类别，但效果不够好。但是rsprompter一样会遇到这个问题🙋‍♀️
…
---- Replied Message ---- | From | @.> | | Date | 10/06/2024 17:13 | | To | KyanChen/RSPrompter @.> | | Cc | millet-Ding @.>, Comment @.> | | Subject | Re: [KyanChen/RSPrompter] val error:The testing results of the whole dataset is empty.; train：no pos rois (Issue #107) | 我也遇到了类似的问题，请问有解决吗？ — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you commented.Message ID: @.***>

谢谢🙏我会尝试一下的。

[millet-Ding]

期待你能解决🥺

…

---- Replied Message ---- | From | ***@***.***> | | Date | 10/06/2024 18:05 | | To | KyanChen/RSPrompter ***@***.***> | | Cc | millet-Ding ***@***.***>, Comment ***@***.***> | | Subject | Re: [KyanChen/RSPrompter] val error:The testing results of the whole dataset is empty.; train：no pos rois (Issue #107) | 我用的seg-mask rcnn 那个可以跑，需要添加一下自己的类别，但效果不够好。但是rsprompter一样会遇到这个问题‍♀️ … ---- Replied Message ---- | From | @.> | | Date | 10/06/2024 17:13 | | To | KyanChen/RSPrompter @.> | | Cc | millet-Ding @.>, Comment @.> | | Subject | Re: [KyanChen/RSPrompter] val error:The testing results of the whole dataset is empty.; train：no pos rois (Issue #107) | 我也遇到了类似的问题，请问有解决吗？ — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you commented.Message ID: @.***> 谢谢我会尝试一下的。 — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you commented.Message ID: ***@***.***>

[jayjiang-sam]

可以试一下在config里面定义一个metainfo字典
metainfo = {
'classes': ('segs'), ##换成自己数据集标注的类别
'palette': [
(220, 20, 60),
]
}