Closed daihaoma closed 2 days ago
- 确保使用多尺度分割来训练ms_train.json并测试ms_test.json
- 看起来您正在使用单个 GPU,并且 bs=1。您可能需要相应地调整 lr 和 bs。[shceduler] #43 Thanks for your reply, I will re change my lr and bs. i refer to your reply to know that lr = (your_batchsize) / (default_batchsize) * 0.0002. i have only one gpu so is default_batchsize equal to 1 here.
Branch
master branch https://mmrotate.readthedocs.io/en/latest/
📚 The doc issue
Thanks to the authors for their research. Can you please tell me if used any other preprocessing besides splitting the dataset before training. I trained the project with dota dataset and found that the map value is only 0.59. and the best training occurs at epoch10. i checked the labeled boxes during training and found that a part of them is not oriented (Fig.). Is this normal? Here's my log.
20240701_164416.log
Suggest a potential alternative/fix
No response