Why class_error always = 100 when training

[hailuu684]

Hi everyone, I am training the dert model using the coco dataset to see how it is going, I saw the class_error=100 in every batch when training in the first epoch. I just do not understand why is that?

[Flyooofly]

Hello, I also encountered this problem. The result of each batch is all 0, how do you solve this problem?

[hailuu684]

The reason could be at this line of code, the accuracy between src_logit and target_class = 0, so it still 100

[Flyooofly]

The reason could be at this line of code, the accuracy between src_logit and target_class = 0, so it still 100

Hello, thanks for the reply. I have also seen similar answers, but I think this is just a display result. Will it affect the final map and various indicators? I trained dozens of epochs but the indicators are all 0

[xuecheng990531]

The reason could be at this line of code, the accuracy between src_logit and target_class = 0, so it still 100

Hello, thanks for the reply. I have also seen similar answers, but I think this is just a display result. Will it affect the final map and various indicators? I trained dozens of epochs but the indicators are all 0

你好，请问您解决这个全0的问题了吗？

[alexHxun]

modify the 71 line of datasets/coco.py : classes = [obj["category_id"] for obj in anno] to classes = [int(obj["category_id"]) for obj in anno]

[1749anonymous]

If your class_error is always = 100, try to increase the batchSize, which is useful to me

[nothing7744]

i met the same question.i guess you can decrease the number of object querry.for me, it's useful.