Can you provide hyperparameter settings that are not mentioned in the paper, especially those that are different in .yaml file provided

Question

Can you provide hyperparameter settings that are not mentioned in the paper, especially those that are different in .yaml file provided

Qiuyu-Li276 opened this issue 2 months ago · 4 comments

Answer 1 · 2024-11-05T03:28:51.000Z

The difference between the learning rate and decay factor in modelEMA and the values mentioned in the paper makes me wonder if there are any other parameters that have also been modified, which are different from the configuration file currently provided

Answer 2 · 2024-11-06T07:27:42.000Z

The decay factor is 0.25, which is the same as in the paper and the code. The learning rate should be set to 0.01 (we have already updated the code), and other hyperparameter settings remain the same as in the code.

Answer 3 · 2024-11-06T11:05:59.000Z

The lr0 in /ultralytics/cfg//default.yaml is still 0.004, are you sure you have already updated the code ?

Answer 4 · 2024-11-07T14:42:21.000Z

@XinhaoLuo666 I also noticed that the lr0 value in default.yaml is set to 0.004. Is there something wrong with your update? Additionally, you mentioned that the 'decay factor is 0.25,' which refers to the weight_decay value in default.yaml.