edvardHua/PoseEstimationForMobile

Issues with Hourglass loss during training

jgcbrouns opened this issue · 0 comments

jgcbrouns commented

Hi everyone,

What could be the problem here? I use the .cfg for hourglass with same settings as CPM (from repository). I use the same dataset and all other factors are also the same.

  • CPM model is training correctly and converging to a stable loss after 200.000 steps
  • Hourglass loss converges already after a few steps (while it should'nt) and loss does not further decrease:
    hourglass_loss

The final accuracy of hourglass is also a lot lower than my trained CPM model.

Any ideas? maybe hyperparameters? Notice that I turned off data-augmentation for both CPM and Hourglass, because I am using synthetic data.