[DIGITS] Retrain custom classes on DetectNet, model doesn't converge.


I’m attempting to train a DetectNet on my own dataset with a single class, currently my dataset is small (about 500 images), so I eventually would expect overfitting, but instead all my metrics stay at 0.
Here below there is the parameters used, in case you see a range (like 0-32) it means that i’ve tried using several values within that range.

  • Image size: 640x352x3
  • Batch size: 8 ( max I can get on a 1080ti)
  • Batch accumulation: 0-32
  • Normalize: pixel-image
  • Learning rate: 0.001-0.0000001
  • Learning rate decay: exp, sig, step
  • Optimizer: SDG, Adam

Is there a reference anywhere on optimal hyperparameters for training a DetectNet to start from ?