loss_cls: 0.0000, loss_bbox: 0.0000 when training YOLOX on my own dataset

[pumpkin33498]

我用YOLOX训练自己的数据集的时候一直显示loss_cls: 0.0000, loss_bbox: 0.0000，然后测试的时候会有ERROR The testing results of the whole dataset is empty 请问可能是什么原因？

[ZouRuia]

我也遇到了，但是没有提示是error，loss_cls,loss_bbox都有值

[VJoer]

请问这个问题你解决了吗？我也遇到了同样的问题

[VJoer]

我用YOLOX训练自己的数据集的时候一直显示loss_cls: 0.0000, loss_bbox: 0.0000，然后测试的时候会有ERROR The testing results of the whole dataset is empty 请问可能是什么原因？ 找到原因了，在coco.py文件中要设置类别，因为我自己的CoCo数据集也是80类，所以没有改，但是类别的名称和CoCo2017类别名称不一样，你需要将类别名称改一下 . 参考这个链接

[ZouRuia]

不知道为啥，在运行一遍，正常跑通了。没有改任何东西。

[ZouRuia]

之前的是这样的： [>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>] 5000/5000, 17.4 task/s, elapsed: 287s, ETA: 0s writing results to ./result.pkl

Evaluating bbox... Loading and preparing results... The testing results of the whole dataset is empty. 我用的都是coco2017的数据，没有变过，之前跑过faster_rcnn,mask_rcnn,yolo,都可以，跑fcos_50也可以，用同样的学习率，gpu个数去跑fcos_101就出现了上面的问题，但是没有改什么，再跑一遍，就又跑通了。

[VJoer]

之前的是这样的： [>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>] 5000/5000, 17.4 task/s, elapsed: 287s, ETA: 0s writing results to ./result.pkl

Evaluating bbox... Loading and preparing results... The testing results of the whole dataset is empty. 我用的都是coco2017的数据，没有变过，之前跑过faster_rcnn,mask_rcnn,yolo,都可以，跑fcos_50也可以，用同样的学习率，gpu个数去跑fcos_101就出现了上面的问题，但是没有改什么，再跑一遍，就又跑通了。

你是用自己训练的模型还是官方给的预训练模型跑测试的？

[ZouRuia]

没有用预训练模型，自己跑出来的模型测试的。

[ZouRuia]

我的loss_cls: 0.0000, loss_bbox: 0.0000这两个值不是0.0000，是有值的，还正常的值。

[pumpkin33498]

我用YOLOX训练自己的数据集的时候一直显示loss_cls: 0.0000, loss_bbox: 0.0000，然后测试的时候会有ERROR The testing results of the whole dataset is empty 请问可能是什么原因？ 找到原因了，在coco.py文件中要设置类别，因为我自己的CoCo数据集也是80类，所以没有改，但是类别的名称和CoCo2017类别名称不一样，你需要将类别名称改一下 . 参考这个链接

谢谢，我看了一下我的类别应该改了QAQ

[VJoer]

没有用预训练模型，自己跑出来的模型测试的。

我的问题是和楼主一样的训练自己的数据集的时候一直显示loss_cls: 0.0000, loss_bbox: 0.0000，然后测试的时候出现The testing results of the whole dataset is empty.，然后我就在coco.py中增加了类名解决了问题，你可以尝试在coco.py中增加类名，然后在coco.py中添加test=dict( samples_per_gpu=32, classes=classes,

[JarvisKevin]

同！跑coco2017都不行

[hhaAndroid]

@VJoer You can see the training log we released, it is normal.

[Snixells]

Hey there! I had the same problem and maybe the solution helps you too: In your COCO Annotations you have to specify an area (how big the boxes are). If you set that to 0 the MMDet training ignores these boxes. That's why the loss quickly goes to 0 (because the boxes do not exist and it is trained to not find anything at all)

[kaphleamrit2]

@Snixells How to specify an area in COCO annotation. It's already there in my datasets. Could you please explain it again? Thanks.

[fire717]

遇到了同样的问题，coco.py中的类别数量除了要保持一致，名称也要一样才行。

[Kinglee90]

### 可以尝试把lr调低，有可能是lr过大，导致梯度消失

[52THANOS]

我跑yolov7也遇到这个问题了，但是我修改classes也没有解决

[1127244933]

同样遇到这样的问题，第一种文档里说可能是正常现象，第二种属于类别没有改，我用的yolov5，具体是yolov5cocodataset继承了mmdet中cocodataset的类别，在其中修改类别就可以。

[kennyadelaide]

Hey there! I had the same problem and maybe the solution helps you too: In your COCO Annotations you have to specify an area (how big the boxes are). If you set that to 0 the MMDet training ignores these boxes. That's why the loss quickly goes to 0 (because the boxes do not exist and it is trained to not find anything at all)

sir！ how to specify the bbox size ? e.g. the areas field is?

[sjiang95]

Hey there! I had the same problem and maybe the solution helps you too: In your COCO Annotations you have to specify an area (how big the boxes are). If you set that to 0 the MMDet training ignores these boxes. That's why the loss quickly goes to 0 (because the boxes do not exist and it is trained to not find anything at all)

This solution solved my problem. Thanks!

My dataset has no instance mask. So I set the area to the area of the bbox.

See https://github.com/open-mmlab/mmdetection/blob/cfd5d3a985b0249de009b67d04f37263e11cdf3d/mmdet/datasets/coco.py#L146-L147 it seems that mmdetection ignores the corresponding bbox if area <= 0.

But, refer to the definition of the keyword "area" of COCO dataset https://github.com/cocodataset/cocoapi/issues/36#issuecomment-279472475, the area means segmentation area. It makes no sense to ignore bboxes if the custom dataset contains no segmentation mask.

Should this be considered as a bug or feature of mmdetection? @hhaAndroid

[gs-max]

我的yolox跑标准的coco数据集一点事没有，自己的数据集就会显示这两项损失为0.能使的方法都试了，哎。