crwodpose test, AP (easy), ap (medium), ap (hard) output score are 0 #8
Comments
|
Thanks for your interest in our work. Issues here may be helpful jeffffffli/CrowdPose#3 HRNet/HigherHRNet-Human-Pose-Estimation#26. I will also check our checkpoint next month, sorry I am too busy right now. |
Thank you for your help. I tried to apply the modifications in these two tips, but it didn't work. Did you encounter ap easy=0 when you first designed the code? |
|
Dear author, I have also encountered a similar issue. During training, my loss value cannot decrease quickly. After training for 60 epochs, my loss decreased from 200 to 100 Last round of training: I hope you can help me. |
You trained using a single card, right? Your log shows that the loss has not decreased. This may be caused by the learning rate, so your AP score is 0, but your AP (EASY) is consistent with my problem |
|
Sorry for the delayed reply. Yes, don't worry about AP=0. I have successfully fixed it before, following many issue instructions (including the above two) on the web, but I forgot the details, really a long time ago. I am sure some bugs in the crowdpose evaluation API. How about generating results into JSON file and then running the evaluation code? This should be working. I will check the crowdpose evaluation API in detail later. Also, if you prefer, you can check it too and share in this issue. @da13132 |
|
我尝试使用单卡训练时也出现了LOSS不下降的情况，这可能是学习率以及batchsize的设置导致的，我目前使用单卡，LR=0.0001以及batchsize=8可以正常训练；我最初的设置是单卡batchsize=1以及LR=0.001，这种情况下loss是不下降的，你可以试试调整学习率以及batchsize，或许训练会恢复正常 |
Thanks for your reply, I'll try using a json file for testing. Going further, I will try to use xtcocotools to replace crowdposetools. If I have further findings, I will share them. |
谢谢老哥，我试试 |
|
Thanks for your interest in our work. The model is sensitive to learning rate and batch size. To align with previous methods, we have not tested our models with smaller batch sizes before. So I strongly recommend you follow the official training settings for getting reported performance. But you can try some common rules, such as linear adjustment. Also, welcome to share your experience in this issue. @YSLDTZY |
Thank you for your answer. I previously trained according to official parameters, except for the different number of training cards |
Thank you very much for your excellent work. When I tried to use the swinL checkpoint you released to test with the crowdpose data set, I found that the scores of AP (easy), ap (medium), and ap (hard) were 0 points or about 0.05 points, and the AP score was only 63.4. When I tried to start from scratch Using res50 as the backbone network to train crowdpose, the scores of AP (easy), ap (medium), and ap (hard) are still 0 points. I haven't found the cause of this error yet, and I hope you can help me!
The text was updated successfully, but these errors were encountered: