If you want to take a look, I put on GitHub the train.py that my team and I are using, maybe you will find something you want to include in your train.py:
With the default hps, the train.py use the same configuration of the original paper, some hp is adapted automatically when another change (e.g. lr when batch size change, or l2 reg. when lr change, we took inspiration from here and here ).
This default configuration works well for our use case.
We use AWS SageMaker so some part is not 100% generic but it should be very easy to adapt it to any enviroment.
I take this opportunity to thank you for having made this package available, we are using it successfully to identify assets and anomalies in aerial photos.
Hi @zylo117,
If you want to take a look, I put on GitHub the train.py that my team and I are using, maybe you will find something you want to include in your train.py:
https://github.com/mnslarcher/Yet-Another-EfficientDet-Pytorch/blob/master/train.py
With the default hps, the train.py use the same configuration of the original paper, some hp is adapted automatically when another change (e.g. lr when batch size change, or l2 reg. when lr change, we took inspiration from here and here ).
This default configuration works well for our use case.
We use AWS SageMaker so some part is not 100% generic but it should be very easy to adapt it to any enviroment.
I take this opportunity to thank you for having made this package available, we are using it successfully to identify assets and anomalies in aerial photos.