Welcome to efficientdet-pytorch Discussions!

[rwightman]

👋 Welcome!

We’re using Discussions as a place to connect with other members of our community. We hope that you:

• Ask questions you’re wondering about, especially those not suited for Issues. Issues will be exclusively focused on verifiable bugs and feature requests that benefit all users of the library.
• Share ideas.
• Engage with other community members.
• Respect others. This is a community.

[Ekta246]

I see the Lr scheduler.step_update takes the eval_metric as mAP set to default in the train.py
However, when I use the eval_metric mAP I see the validation loss diverges in a weird manner.

While, when I change the eval_metric to Loss in the LR scheduler update, I see the model has a tremendous improvement learning and the validation loss converges.

efficientdet-pytorch/train.py

Line 416 in 611532d

lr_scheduler.step(epoch + 1, eval_metrics[eval_metric])

However, I believe there is a bug in my mAP evaluation maybe.

Also, I calculate mean Intersection of Union on the validation set and the test set. It is pretty good. It is around 90%
The predictions that the model generates are pretty close to the ground truth. So, it's a good news.