Instability in training OPANAS from Scratch.

VDIGPKU / OPANAS

[CVPR 2021]OPANAS: One-Shot Path Aggregation Network Architecture Search for Object Detection

Apache License 2.0

44 stars 7 forks source link

Instability in training OPANAS from Scratch. #8

Open cydiachen opened 2 years ago

cydiachen commented 2 years ago

Dear Author: I am now trying to reproduce your results from training from scratch on COCO Dataset. But it occurs to me that the Model cannot covergence with APEXOptimizer hook. Could you provide me with some advices?

tingtingliangvs commented 2 years ago

Hi, thanks for noticing. The convergence of this model is not very stable due to the deep Neck architecture. In my previous work, I tried lowering the initial learning rate and adding batch normalization in Neck to solve the convergence problem. I have no plans for training from scratch in the near future, but issues are welcomed if you run into other problems.