Training collapse at about 188 epochs.

[GeoffreyChen777]

Hi, thanks for your great work!

I'm using your code on a custom dataset. I transformed my dataset into your required format and then train the model on it.

At the beginning of the training, everything was fine. But at about 188 epochs, I got this:

Do you have any insights on this such as any parameters changes at this step?

Best wishes.

[MohsenZand]

Thank you! You can start with a smaller initial learning rate (it's in the 'hyp.yaml'). Also, changing the learning rate at epoch ~ 150 would help. Some data augmentations are also possible. Despite being available in the code, they are not utilized in the reported results. So, check out the 'hyp.yaml' and 'flag_sets.py'.

[GeoffreyChen777]

Thanks!