jomjol / neural-network-autotrain-digital-counter Goto Github PK

Hi,

i had problems with the digit model on a electrical meter. So I looked deeper in the code and saw, the models will be trained often with the history and stored model. This makes a much overfitted model. Because of the shuffling every train, the validation and train data mixed and the overfitting is not visible.

TLDR;

I made various changes and created a notebook with an small effnet-model trained with knowledge distillation and added more data for better results.

Look at Colab-Notebook

If you interested in I can create a pull request, to integrate it.

It's running on my eep-cam very well.