Very poor results despite high value on validation sets


I encounter something very strange. I cannot see where the problem is since on the validation set, the results seem to be 0.94. However, the IOU and other metrics are extremely low. I suspect I have made a mistake, however, I do not know where the mistake it. The model dimensional seems to be correct, but the results are very poor.

Is there any mentor who would like to have a look at the notebook and point out the problem? Thank you very much.

OK, to answer my own question. The key here is that it allows you to use different optimizers. What I did is to use Adam to a certain point, then use SGD to train very small steps to squeeze in more accuracy. Also, it is suggested to use a learning rate warm-up (although I didn’t do it), and save model checkpoints.

I’m getting that. The reason is it just predicts there to be no numerals anywhere. Most of each example is background.

Hi @R_Wang just to be sure! You solved the problem by yourself :slight_smile:

If you still need a second opinion about your notebook just ask.