Cannot reproduce the results for COCO HRNet_w32
asanakoy opened this issue · comments
Thanks for the amazing work and the source code!
I'm trying to train HRNet_w32 on COCO dataset.
But currently I cannot reproduce the results reported in the paper. I barely get AP=12.0 when using the config w32_256x192_adam_lr1e-3.yaml and training on 4 GPUs.
Are the provide hyperparameters correct? Any tips what could go wrong?
Have anybody successfully reproduced the results?
Best,
Artsiom
OK. Looks like I have found the problem.
If you reimplement the training loop yourself, make sure so set weight_decay=0. In the codebase it is just confusing that WD
param from the config is used only with SGD, and for Adam the weigh_decay is hardcoded to 0.0
(the default value).