I ran the pytorch version of the code, and the code about the calculation loss was verified and uncommented, and the test_loss_bbox I got was very large, at 20000, but the train_loss_bbox was very small, and the train_loss kept dropping during training, but the AP would drop and rise, with a difference of nearly 0.02, it should be an overfit, and I wanted to output the verification loss, but testloss bbox is big and has barely changed, I don't know what the reason is, how to fix it?
If I change the number of training rounds, do I need to adjust the learning rate?
Hope to get an answer, thank you!
meirashaf
commented
1 day ago
I ran the pytorch version of the code, and the code about the calculation loss was verified and uncommented, and the test_loss_bbox I got was very large, at 20000, but the train_loss_bbox was very small, and the train_loss kept dropping during training, but the AP would drop and rise, with a difference of nearly 0.02, it should be an overfit, and I wanted to output the verification loss, but testloss bbox is big and has barely changed, I don't know what the reason is, how to fix it? If I change the number of training rounds, do I need to adjust the learning rate? Hope to get an answer, thank you!