search

SherlockHolmes221 / DOQ

CVPR2022 Distillation Using Oracle Queries for Transformer-based Human-Object Interaction Detection
Apache License 2.0
25 stars 2 forks source link

training problems #7

Open xz-123-new opened 2 years ago

xz-123-new commented 2 years ago

I got a problem with training this model, looking forward to your generous reply. I train this model on hicodet dataset with resnet-50 as backbone , all the parameters concerning training are set following the paper except batch size, i choose batch size =2 to train because of GPU memory ,after i train 80 epoches and even 150 epoches now, i didn't get the mAP value it should be .For now, i only got ""test_mAP": 0.29916627302565096, "test_mAP rare": 0.23861787027422618, "test_mAP non-rare": 0.3172521595617908 ".Could you please give me some advice about training stradegy to attain the correct number?thanks!

SherlockHolmes221 commented 2 years ago

The batch size is set as 2 for per GPU by default. I do not understand your meaning. advice: the learning rate should be changed according to batch size